228 SOME SPECIAL DISTRIBUTIONS

Rayleigh: f (t) = (t/α2 ) exp{−t2 /(2α2 )}, t > 0.

(a)
√
Lognormal: f (t) = 1/(tσ 2π) exp{−(n t − μ)2 /2σ 2 }.
(b)
Pareto: f (t) = αθα /tα+1 , t > θ, and = 0 otherwise.
(c)
Weibull: f (t) = (α/β)tα−1 exp(−tα /β), t > 0.
(d)
Logistic: f (t) = (1/β) exp{−(t − μ)/β}[1 + exp{−(t − μ)/β}]−2 , −∞ <
(e)
t < ∞.
28. Consider the PDF
1/2  !
λ λ(x − μ)2
f (x) = exp − , x>0
2πx3 2μ2 x

and = 0 otherwise. An RV X with PDF f is said to have an inverse Gaussian

distribution with parameters μ and λ, both positive. Show that

EX = μ, var(X) = μ3 /λ and
  1/2

λ 2tμ2
M(t) = E exp(tX) = exp 1− 1− .
μ λ

29. Let f be the PDF of a N(μ, σ 2 ) RV:

(a) For what value of c is the function cf n , n > 0, a pdf?
(b) Let Φ be the DF of Z ∼ N(0, 1). Find E{ZΦ(Z)} and E{Z 2 Φ(Z)}.

5.4 BIVARIATE AND MULTIVARIATE NORMAL DISTRIBUTIONS

In this section we introduce the bivariate and multivariate normal distributions and inves-
tigate some of their important properties. We note that bivariate analogs of other PDFs are
known but they are not always uniquely identified. For example, there are several versions
of bivariate exponential PDFs so-called because each has exponential marginals. We will
not encounter any of these bivariate PDFs in this book.

Definition 1. A two-dimensional RV (X, Y) is said to have a bivariate normal distribution

if the joint PDF is of the form
1
f (x, y) = * e−Q(x,y)/2 , (1)
2πσ1 σ2 1 − ρ2
− ∞ < x < ∞, −∞ < y < ∞,

where σ1 > 0, σ2 > 0, |ρ| < 1, and Q is the positive definite quadratic form
 
2 2
1 x − μ1 x − μ1 y − μ2 y − μ2
Q(x, y) = − 2ρ + . (2)
1 − ρ2 σ1 σ1 σ2 σ2

Figure 1 gives graphs of bivariate normal PDF for selected values of ρ.

BIVARIATE AND MULTIVARIATE NORMAL DISTRIBUTIONS 229

(a)
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
−3
−2
−1
0 −3
1 −2
−1
2 0
1 ρ = −0.9
3 2
3
(b)
0.2
0.18
0.16
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
−3
−2
−1
0
1 –3
2 −1 −2
1 0
3 3 2 ρ = −0.5

Fig. 1 Bivariate normal with μ1 = μ2 = 0, σ1 = σ2 = 1, and ρ = −0.9, −0.5, 0.5, 0.9.

230 SOME SPECIAL DISTRIBUTIONS

(c)

0.2

0.15

0.1
−3
0.05 −2
−1
0 0
−3
−2 1 ρ = 0.5
−1
0 2
1
2 3
3
(d)

ρ = 0.9

Fig. 1 (continued).

We first show that (1) indeed defines a joint PDF. In fact, we prove the following result.

Theorem 1. The function defined by (1) and (2) with σ1 > 0, σ2 > 0, |ρ| < 1 is a joint
PDF. The marginal PDFs of X and Y are, respectively, N(μ1 , σ12 ) and N(μ2 , σ22 ), and ρ is
the correlation coefficient between X and Y.
BIVARIATE AND MULTIVARIATE NORMAL DISTRIBUTIONS 231
 ∞
Proof. Let f1 (x) = f (x, y) dy. Note that
−∞

2 2
y − μ2 x − μ1 x − μ1
(1 − ρ2 )Q(x, y) = −ρ + (1 − ρ2 )
σ2 σ1 σ1
!2 2
y − [μ2 + ρ(σ2 /σ1 )(x − μ1 )] x − μ1
= + (1 − ρ2 ) .
σ2 σ1

It follows that
! ∞
+ ,
1 −(x − μ1 )2 exp −(y − βx )2 /[2σ22 (1 − ρ2 )]
f1 (x) = √ exp * √ dy, (3)
σ1 2π 2σ12 −∞ σ2 1 − ρ2 2π

where we have written

σ2
βx = μ2 + ρ (x − μ1 ). (4)
σ1

The integrand is the PDF of an N(βx , σ22 (1 − ρ2 )) RV, so that

 
2
1 1 x − μ1
f1 (x) = √ exp − , −∞ < x < ∞.
σ1 2π 2 σ1

Thus
 ∞  ∞ !  ∞
f (x, y) dy dx = f1 (x) dx = 1,
−∞ −∞ −∞

and f (x, y) is a joint PDF of two RVs of the continuous type. It also follows that f1 is the
marginal PDF of X, so that X is N(μ1 , σ12 ). In a similar manner we can show that Y is
N(μ2 , σ22 ).
Furthermore, we have
!
f (x, y) 1 −(y − βx )2
= * √ exp , (5)
f1 (x) σ2 1 − ρ2 2π 2σ22 (1 − ρ2 )

where βx is given by (4). It is clear, then, that the conditional PDF fY|X (y | x) given by (5)
is also normal, with parameters βx and σ22 (1 − ρ2 ). We have

σ2
E{Y | x} = βx = μ2 + ρ (x − μ1 ) (6)
σ1

and

var{Y|x} = σ22 (1 − ρ2 ). (7)

232 SOME SPECIAL DISTRIBUTIONS

In order to show that ρ is the correlation coefficient between X and Y, it suffices to

show that cov(X, Y) = ρσ1 σ2 . We have from (6)

E(XY) = E{E{XY|X}}
 !
σ2
= E X μ2 + ρ (X − μ1 )
σ1
ρσ2 2
= μ1 μ2 + σ .
σ1 1

It follows that

cov(X, Y) = E(XY) − μ1 μ2 = ρσ1 σ2 .

Remark 1. If ρ2 = 1, then (1) becomes meaningless. But in that case we know

(Theorem 4.5.1) that there exist constants a and b such that P{Y = aX + b} = 1. We
thus have a univariate distribution, which is called the bivariate degenerate (or singular)
normal distribution. The bivariate degenerate normal distribution does not have a PDF
but corresponds to an RV (X, Y) whose marginal distributions are normal or degenerate
and are such that (X, Y) falls on a fixed line with probability 1. It is for this reason that
degenerate distributions are considered as normal distributions with variance 0.

Next we compute the MGF M(t1 , t2 ) of a bivariate normal RV (X, Y). We have, if f (x, y)
is the PDF given in (1) and f1 is the marginal PDF of X,
 ∞  ∞
M(t1 , t2 ) = et1 x+t2 y f (x, y) dx dy,
−∞
∞
−∞
 ∞ !
= fY|X (y | x)et2 y dy et1 x f1 (x) dx
−∞
∞
−∞
 !
1 22 σ2
= e f1 (x) exp σ2 t2 (1 − ρ ) + t2 μ2 + ρ (x − μ1 )
t1 x 2
dx
−∞ 2 σ1
  ∞
1 σ2
= exp σ22 t22 (1 − ρ2 ) + t2 μ2 − ρt2 μ1 et1 x e(ρσ2 /σ1 )xt2 f1 (x) dx.
2 σ1 −∞

Now
  
∞ 2
σ2 1 σ2
e(t1 +ρt2 σ2 /σ1 )x f1 (x) dx = exp μ1 t1 + ρ t2 + σ12 t1 + ρt2 .
−∞ σ1 2 σ1

Therefore,

σ12 t12 + σ22 t22 + 2ρσ1 σ2 t1 t2

M(t1 , t2 ) = exp μ1 t1 + μ2 t2 + . (8)
2

The following result is an immediate consequence of (8).

BIVARIATE AND MULTIVARIATE NORMAL DISTRIBUTIONS 233

Theorem 2. If (X, Y) has a bivariate normal distribution, X and Y are independent if and
only if ρ = 0.

Remark 2. It is quite possible for an RV (X, Y) to have a bivariate density such that the
marginal densities of X and Y are normal and the correlation coefficient is 0, yet X and Y
are not independent. Indeed, if the marginal densities of X and Y are normal, it does not
follow that the joint density of (X, Y) is a bivariate normal. Let
 
1 1 −1
f (x, y) = exp (x 2
− 2ρxy + y 2
) (9)
2 2π(1 − ρ2 )1/2 2(1 − ρ2 )
 !
1 −1 2 2
+ exp (x + 2ρxy + y ) .
2π(1 − ρ2 )1/2 2(1 − ρ2 )

Here f (x, y) is a joint PDF such that both marginal densities are normal, f (x, y) is not
bivariate normal, and X and Y have zero correlation. But X and Y are not independent. We
have
1
f1 (x) = √ e−x /2 ,
2
−∞ < x < ∞,
2π
1 −y2 /2
f2 (y) = √ e , −∞ < y < ∞,
2π
EXY = 0.

Example 1. (Rosenberg [93]). Let f and g be PDFs with corresponding DFs F and G.
Also, let

h(x, y) = f (x)g(y)[1 + α(2F(x) − 1)(2G(y) − 1)], (10)

where |α| ≤ 1 is a constant. It was shown in Example 4.3.1 that h is a bivariate density
function with given marginal densities f and g.
In particular, take f and g to be the PDF of N(0, 1), that is,

1
f (x) = g(x) = √ e−x /2 ,
2
−∞ < x < ∞, (11)
2π

and let (X, Y) have the joint PDF h(x, y). We will show that X + Y is not normal except in
the trivial case α = 0, when X and Y are independent.
Let Z = X + Y. Then

EZ = 0, var(Z) = var(X) + var(Y) + 2 cov(X, Y).

It is easy to show (Problem 2) that cov(X, Y) = α/π, so that var(Z) = 2[1 + (α/π)]. If Z
is normal, its MGF must be
2
[1+(α/π)]
Mz (t) = et . (12)
234 SOME SPECIAL DISTRIBUTIONS

Next we compute the MGF of Z directly from the joint PDF (10). We have

M1 (t) = E{etX+tY }
 ∞ ∞
2
= et + α etx+ty [2F(x) − 1][2F(y) − 1]f (x)f (y) dx dy
−∞ −∞
 ∞ 2
2
= et + α e [2F(x) − 1]f (x) dx
tx
.
−∞

Now
 ∞  ∞
2
etx [2F(x) − 1]f (x) dx = −2 etx [1 − F(x)]f (x) dx + et /2
−∞ −∞
 ∞ ∞ !
2 1 1 2
= e −2
t /2
exp − (x + u − 2tx) du dx
2
−∞ x 2π 2
!
1 2
 ∞  ∞ exp − [x + (v + x) − 2tx] 2
2 2
= et /2 − dv dx
−∞ 0 π
 ∞ 
2 exp{−v2 /2 + (v − t)2 /4} ∞ exp{−[x + (v − t)/2]2 }
= et /2 − √ √ dx dv
π −∞ π
0
!
1
 ∞ exp − [(v + t)2 /2]
2 2 2
= et /2 − 2et /2 √ dv
2 π
0
!
2 2 t
= et /2 − 2et /2 P Z1 > √ , (13)
2
where Z1 is an N(0, 1) RV.
It follows that
 !2
2 2 2 1
M1 (t) = et + α et /2 − 2et /2 P Z1 > √
 2
! 2
2 t
= et 1 + α 1 − 2P Z1 > √ . (14)
2

If Z were normally distributed, we must have Mz (t) = M1 (t) for all t and all |α| ≤ 1,
that is,
 ! 2
t2 (α/π)t2 t2 t
e e = e 1 + α 1 − 2P Z1 > √ . (15)
2

For α = 0, the equality clearly holds. The expression within the brackets on the right side
2
of (15) is bounded by 1 + α, whereas the expression e(α/π)t is unbounded, so the equality
cannot hold for all t and α.

Next we investigate the multivariate normal distribution of dimension n, n ≥ 2. Let M

be an n × n real, symmetric, and positive definite matrix. Let x denote the n × 1 column
BIVARIATE AND MULTIVARIATE NORMAL DISTRIBUTIONS 235

vector of real numbers (x1 , x2 , . . . , xn ) and let μ denote the column vector (μ1 , μ2 , . . . , μn ) ,
where μi (i = 1, 2, . . . , n) are real constants.

Theorem 3. The nonnegative function

!
(x − μ) M(x − μ)
f (x) = c exp − − ∞ < xi < ∞, (16)
2
i = 1, 2, . . . , n,

defines the joint PDF of some random vector X = (X1 , X2 , . . . , Xn ) , provided that the
constant c is chosen appropriately. The MGF of X exists and is given by
!
t M−1 t

M(t1 , t2 , . . . , tn ) = exp t μ + , (17)
2

where t = (t1 , t2 , . . . , tn ) and t1 , t2 , . . . , tn are arbitrary real numbers.

Proof. Let
  ! n
∞ ∞
(x − μ) M(x − μ) 

I=c ··· exp t x − dxi . (18)
−∞ −∞ 2
i=1

Changing the variables of integration to y1 , y2 , . . . , yn by writing xi − μi = yi , i = 1, 2, . . . , n,

and y = (y1 , y2 , . . . , yn ) , we have x − μ = y and
 
∞ ∞
y My 
n
I = c exp(t μ) ··· exp t y − dyi . (19)
−∞ −∞ 2
i=1

Since M is positive definite, it follows that all the n characteristic roots of M, say
m1 , m2 , . . . , mn , are positive. Moreover, since M is symmetric there exists an n × n orthog-
onal matrix L such that L ML is a diagonal matrix with diagonal elements m1 , m2 , . . . , mn .
Let us change the variables to z1 , z2 , . . . , zn by writing y = Lz, where z = (z1 , z2 , . . . , zn ),
and note that the Jacobian of this orthogonal transformation is |L|. Since L L = In , where
In is an n × n unit matrix, |L| = 1 and we have
 
∞ ∞
z L MLz 
n

I = c exp(t μ) ··· exp t Lz − dzi . (20)
−∞ −∞ 2
i=1

If we write t L = u = (u1 , u2 , . . . , un ) then t Lz = n

i=1 ui zi . Also L ML =
  n 2
diag(m1 , m2 , . . . , mn ) so that z L MLz = i=1 mi zi . The integral in (20) can therefore
be written as
n  ∞   - 
 m1 z2i
n
2π u2i
exp ui zi − dzi = exp .
−∞ 2 mi 2mi
i=1 i=1
236 SOME SPECIAL DISTRIBUTIONS

If follows that
" n #
(2π) n/2  u2
I = c exp(t u) exp i
. (21)
(m1 m2 · · · mn )1/2 i=1
2mi

Setting t1 = t2 = · · · = tn = 0, we see from (18) and (21) that

 ∞  ∞
c(2π)n/2
··· f (x1 , x2 , . . . , xn )dx1 dx2 · · · dxn = .
−∞ −∞ (m1 m2 · · · mn )1/2

By choosing

(m1 m2 · · · mn )1/2
c= (22)
(2π)n/2

we see that f is a joint PDF of some random vector X, as asserted.

Finally, since

(L ML)−1 = diag(m−1 −1 −1

1 , m2 , . . . , mn ),

we have

n
u2 i
= u (L M−1 L)u = t M−1 t.
mi
i=1

Also

|M−1 | = |L M−1 L| = (m1 m2 · · · mn )−1 .

It follows from (21) and (22) that the MGF of X is given by (17), and we may write

1
c= . (23)
{(2π)n |m−1 |}1/2

This completes the proof of Theorem 3.

Let us write M−1 = (σij )i,j=1,2,...,n . Then

ti2
M(0, 0, . . . , 0, ti , 0, . . . , 0) = exp ti μi + σii
2

is the MGF of Xi , i = 1, 2, . . . , n. Thus each Xi is N(μi , σii ), i = 1, 2, . . . , n. For i = j, we

have for the MGF of Xi and Xj

M(0, 0, . . . , 0, ti , 0, . . . , 0, tj , 0, . . . , 0)
" #
σii ti2 + 2σij ti tj + tj2 σjj
= exp ti μi + tj μj + .
2
BIVARIATE AND MULTIVARIATE NORMAL DISTRIBUTIONS 237

This is the MGF of a bivariate normal distribution with means μi , μj , variances σii , σjj ,
and covariance σij . Thus we see that

μ = (μ1 , μ2 , . . . , μn ) (24)

is the mean vector of X = (X1 , . . . , Xn ),

σii = σi2 = var(Xi ), i = 1, 2, . . . , n, (25)

and

σij = ρij σi σj , i = j; i, j = 1, 2, . . . , n. (26)

The matrix M−1 is called the dispersion (variance-covariance) matrix of the multivariate
normal distribution.
If σij = 0 for i = j, the matrix M−1 is a diagonal matrix, and it follows that the RVs
X1 , X2 , . . . , Xn are independent. Thus we have the following analog of Theorem 2.

Theorem 4. The components X1 , X2 , . . . , Xn of a jointly normally distributed RV X are

independent if and only if the covariances σij = 0 for all i = j (i, j = 1, 2, . . . , n).

The following result is stated without proof. The proof is similar to the two-variate case
n
except that now we consider the quadratic form in n variables: E{ i=1 ti (Xi − μi )}2 ≥ 0.

Theorem 5. The probability that the RVs X1 , X2 , . . . , Xn with finite variances satisfy at
least one linear relationship is 1 if and only if |M| = 0.
Accordingly, if |M| = 0 all the probability mass is concentrated on a hyperplane of
dimension < n.

Theorem 6. Let (X1 , X2 , . . . , Xn ) be an n-dimensional RV with a normal distribution. Let

Y1 , Y2 , . . . , Yk , k ≤ n, be linear functions of Xj (j = 1, 2, . . . , n). Then (Y1 , Y2 , . . . , Yk ) also
has a multivariate normal distribution.

Proof. Without loss of generality let us assume that EXi = 0, i = 1, 2, . . . , n. Let


n
Yp = Apj Xj , p = 1, 2, . . . , k; k ≤ n. (27)
j=1

Then EYp = 0, p = 1, 2, . . . , k, and


n
cov(Yp , Yq ) = Api Aqj σij , (28)
i,j=1

where E(Xi Xj ) = σij , i, j = 1, 2, . . . , n.

238 SOME SPECIAL DISTRIBUTIONS

The MGF of (Y1 , Y2 , . . . , Yk ) is given by

⎧ ⎛ ⎞⎫
⎨ 
n 
n ⎬
M ∗ (t1 , t2 , . . . , tk ) = E exp ⎝t1 A1j Xj + · · · + tk Akj Xj ⎠ .
⎩ ⎭
j=1 j=1

k
Writing uj = p=1 tp Apj , j = 1, 2, . . . , n, we have
 " n #

∗
M (t1 , t2 , . . . , tk ) = E exp ui Xi
⎛ i=1
⎞
1 
n
= exp ⎝ σij ui uj ⎠ by (17)
2
i,j=1
⎛ ⎞
1  n  k
= exp ⎝ σij tl tm Ali Amj ⎠
2
i,j=1 l,m=1
⎛ ⎞
1  k  n
= exp ⎝ tl tm Ali Amj σij ⎠
2
l,m=1 i,j=1
⎧ ⎫
⎨1  k ⎬
= exp tl tm cov(Yl , Ym ) . (29)
⎩2 ⎭
l,m=1

When (17) and (29) are compared, the result follows.

Corollary 1. Every marginal distribution of an n-dimensional normal distribution is

univariate normal. Moreover, any linear function of X1 , X2 , . . . , Xn is univariate normal.

Corollary 2. If X1 , X2 , . . . , Xn are iid N(μ, σ 2 ) and A is an n × n orthogonal transforma-

tion matrix, the components Y1 , Y2 , . . . , Yn of Y = AX , where X = (X1 , . . . , Xn ) , are
independent RVs, each normally distributed with the same variance σ 2 .

We have from (27) and (28)


n 
cov(Yp , Yq ) = Api Aqi σii + Api Aqj σij
i=1 i=j

0 if p = q,
=
σ2 if p = q,
n n 2
since i=1 Api Aqi = 0 and j=1 Apj = 1. It follows that
" n #
∗ 1 2 2
M (t1 , t2 , . . . , tn ) = exp tl σ
2
l=1

and Corollary 2 follows.

BIVARIATE AND MULTIVARIATE NORMAL DISTRIBUTIONS 239

Theorem 7. Let X = (X1 , X2 , . . . , Xn ) . Then X has an n-dimensional normal distribution

if and only if every linear function of X

X t = t1 X1 + t2 X2 + · · · + tn Xn

has a univariate normal distribution.

Proof. Suppose that X t is normal for any t. Then the MGF of X t is given by

1
M(s) = exp bs + σ 2 s2 . (30)
2

Here b = E{X t} = 1 tij μi = t μ, where μ = (μ1 , . . . , μn ), and σ 2 = var(X t) =

n

var( ti Xi ) = t M−1 t, where M−1 is the dispersion matrix of X. Thus

1
M(s) = exp t μs + t M−1 ts2 . (31)
2

Let s = 1 then
1
M(1) = exp t μ + t M−1 t , (32)
2

and since the MGF is unique, it follows that X has a multivariate normal distribution. The
converse follows from Corollary 1 to Theorem 6.

Many characterization results for the multivariate normal distribution are now available.
We refer the reader to Lukacs and Laha [70, p. 79].

PROBLEMS 5.4

1. Let (X, Y) have joint PDF

!
1 8 x2 31 xy y2 4 71
f (x, y) = √ exp − − x+ + − y+ ,
6π 7 7 16 32 8 9 3 16

for −∞ < x < ∞, −∞ < y < ∞.

(a) Find the means and variances of X and Y. Also find ρ.
(b) Find the conditional PDF of Y given X = x and E{Y|x}, var{Y|x}.
(c) Find P{4 ≤ Y ≤ 6|X = 4}.
2. In Example 1 show that cov(X, Y) = α/π.
3. Let (X, Y) be a bivariate normal RV with parameters μ1 , μ2 , σ12 , σ22 , and ρ. What is
the distribution of X + Y? Compare your result with that of Example 1.
4. Let (X, Y) be a bivariate normal RV with parameters μ1 , μ2 , σ12 , σ22 , and ρ, and let
U = aX + b, a = 0, and V = cY + d, c = 0. Find the joint distribution of (U, V).
240 SOME SPECIAL DISTRIBUTIONS

5. Let (X, Y) be a bivariate normal RV with parameters μ1 = 5, μ2 = 8, σ12 = 16, σ22 = 9,

and ρ = 0.6. Find P{5 < Y < 11 | X = 2}.
6. Let X and Y be jointly normal with means 0. Also, let

W = X cos θ + Y sin θ, Z = X cos θ − Y sin θ.

Find θ such that W and Z are independent.

7. Let (X, Y) be a normal RV with parameters μ1 , μ2 , σ12 , σ22 , and ρ. Find a necessary
and sufficient condition for X + Y and X − Y to be independent.
8. For a bivariate normal RV with parameters μ1 , μ2 , σ1 , σ2 , and ρ show that
1 1 ρ
P(X > μ1 , Y > μ2 ) = + tan−1 * .
4 2π 1 − ρ2
 
[Hint: The required probability is P (X − μ1 )/σ1 > 0, (Y − μ2 )/σ2 > 0 . Change
to polar coordinates and integrate.]
9. Show that every variance–covariance matrix is symmetric positive semidefinite and
conversely. If the variance–covariance matrix is not positive definite, then with prob-
ability 1 the random (column) vector X lies in some hyperplane c X = a with
c = 0.
10. Let (X, Y) be a bivariate normal RV with EX = EY = 0, var(X) = var(Y) = 1, and
cov(X, Y) = ρ. Show that the RV Z = Y/X has a Cauchy distribution.
11. (a) Show that
! n  

1 xi2
xi e−xi /2
2
f (x) = exp − 1+
(2π)n/2 2
1

is a joint PDF on Rn .
(b) Let (X1 , X2 , . . . , Xn ) have PDF f given in (a). Show that the RVs in any proper
subset of {X1 , X2 , . . . , Xn } containing two or more elements are independent
standard normal RVs.

5.5 EXPONENTIAL FAMILY OF DISTRIBUTIONS

Most of the distributions that we have so far encountered belong to a general family of
distributions that we now study. Let Θ be an interval on the real line, and let {fθ : θ ∈ Θ}
be a family of PDFs (PMFs). Here and in what follows we write x = (x1 , x2 , . . . , xn ) unless
otherwise specified.

Definition 1. If there exist real-valued functions Q(θ) and D(θ) on Θ and Borel-
measurable functions T(x1 , x2 , . . . , xn ) and S(x1 , x2 , . . . , xn ) on Rn such that

fθ (x1 , x2 , . . . , xn ) = exp{Q(θ)T(x) + D(θ) + S(x)}, (1)

we say that the family {fθ , θ ∈ Θ} is a one-parameter exponential family.