Journal of Applied Statistics

ISSN: 0266-4763 (Print) 1360-0532 (Online) Journal homepage: http://www.tandfonline.com/loi/cjas20

A generalized normal distribution

Saralees Nadarajah

To cite this article: Saralees Nadarajah (2005) A generalized normal distribution, Journal of
Applied Statistics, 32:7, 685-694, DOI: 10.1080/02664760500079464

To link to this article: https://doi.org/10.1080/02664760500079464

Published online: 12 Apr 2011.

Submit your article to this journal

Article views: 1823

View related articles

Citing articles: 129 View citing articles

Full Terms & Conditions of access and use can be found at

http://www.tandfonline.com/action/journalInformation?journalCode=cjas20
Journal of Applied Statistics
Vol. 32, No. 7, 685 –694, September 2005

A Generalized Normal Distribution

SARALEES NADARAJAH
Department of Statistics, University of Nebraska, Lincoln, USA

ABSTRACT Undoubtedly, the normal distribution is the most popular distribution in statistics. In
this paper, we introduce a natural generalization of the normal distribution and provide a
comprehensive treatment of its mathematical properties. We derive expressions for the nth
moment, the nth central moment, variance, skewness, kurtosis, mean deviation about the mean,
mean deviation about the median, Rényi entropy, Shannon entropy, and the asymptotic
distribution of the extreme order statistics. We also discuss estimation by the methods of moments
and maximum likelihood and provide an expression for the Fisher information matrix.

KEY WORDS : Estimation, entropy, generalized normal distribution, moments, normal distribution,
order statistics

Introduction
A random variable X is said to have the normal distribution with parameters m and s if its
p.d.f. is given by
 
1 (x
m)2
f (x) ¼ pffiffiffiffiffiffi exp
(1)
2ps 2 s2

for
1 , x , 1,
1 , m , 1 and s . 0. A natural generalization of equation (1) is to

replace the power 2 with power s . 0 of any order. Thus, the p.d.f. will take the form
n x
ms o
 
f (x) ¼ K exp
  (2)
s

Simple integration shows that the normalizing constant K is given by

s
K¼
2sG(1=s)

In addition to containing equation (1) as the particular case for s ¼ 2, Equation (2) also
contains the Laplace distribution as the particular case for s ¼ 1. Like the normal distri-
bution, equation (2) is bell-shaped and unimodal with mode at x ¼ m. Furthermore,
f (m) ¼ K. Figure 1 illustrates some possible shapes of equation (2).

Correspondence Address: Saralees Nadarajah, Department of Statistics, University of Nebraska, Lincoln,

Nebraska 68583, USA.

0266-4763 Print=1360-0532 Online=05=070685–10 # 2005 Taylor & Francis

DOI: 10.1080=02664760500079464
686 S. Nadarajah

Figure 1. The generalized normal p.d.f. (equation (2)) for m ¼ 0 and s ¼ 1

If x  m then the c.d.f. of X can be written as

Ðx
s
1 exp {
(m
y)s =ss }dy
F(x) ¼ (3)
2sG(1=s)

On setting z ¼ ((m
y)=s)s , equation (3) reduces to

Ð1
((m
x)=s)s z1=s
1 exp (
z)dz
F(x) ¼ (4)
2G(1=s)

By using the complementary incomplete gamma function defined by

ð1
G(a; x) ¼ ta
1 exp (
t)dt
x

equation (4) can be rewritten as

G(1=s; (ðm
xÞ=s)s )

F(x) ¼ (5)
2G(1=s)

Similarly, if x . m then one can show that

G(1=s; (ðx
mÞ=s)s )

F(x) ¼ 1
(6)
2G(1=s)

Note that as in the normal distribution, the median of X is m.

The hazard rate function defined by h(x) ¼ f (x)={1
F(x)} is an important quantity
characterizing life phenomena. It is immediate from equations (2), (5) and (6) that the
 A Generalized Normal Distribution 687

Figure 2. The generalized normal hazard rate function (equation (7)) for m ¼ 0 and s ¼ 1

hazard rate function is given by

8
> s exp {
jðx
mÞ=sjs }
>
< ; if x  m
s{2G(1=s)
G(1=s; (ðm
xÞ=s)s )}
h(x) ¼ (7)
>
> s exp {
jðx
mÞ=sjs }
: ; if x . m
sG(1=s; (ðx
mÞ=s)s )
Note that h(x)  (s=s)xs
1 and h(
x)  s exp½
{(x þ m)=s}s ={2sG(1=s)} as x ! 1.
Furthermore, h(m) ¼ s={sG(1=s)}. Some possible shapes of equation (7) for m ¼ 0
and s ¼ 1 are shown in Figure 2. It appears that h(x) is an increasing function of x
for s  1.
In the rest of this paper, we provide a comprehensive treatment of the mathematical
properties of equation (2) and also discuss estimation issues. In particular, expressions
are derived for the nth moment and the nth central moment (including the first four
moments, variance, skewness and kurtosis), the mean deviation about the mean, the
mean deviation about the median, Rényi and Shannon entropies, and the asymptotic
distribution of the extreme order statistics. Finally, we discuss estimation by the methods
of moments and maximum likelihood and derive an expression for the Fisher information
matrix.
In addition to the complementary incomplete gamma function mentioned above, the
calculations of this paper use the Euler’s psi function defined by C(x) ¼ d log G(x)=dx.
The properties of this special function can be found in Prudnikov et al. (1990) and
Gradshteyn & Ryzhik (2000).

Moments
If X has the p.d.f. (equation (2)) then the standardized variable Z ¼ (X
m)=s will have
the p.d.f.
s exp {
jzjs }
f (z) ¼
2G(1=s)
688 S. Nadarajah

It is easily seen that the kth moment of Z is given by

 
k 1 þ (
1)k kþ1
E(Z ) ¼ G
2G(1=s) s

Thus, the nth moment of X can be obtained as

E(X n ) ¼ E½(m þ sZ)n 

Xn  
n
¼ mn
k sk E(Z k )
k¼0 k
 
P n
mn nk¼0 (s=m)k {1 þ (
1)k }G(ðk þ 1Þ=s)
k
¼
2G(1=s)

In particular, the first four moments can be worked out as

E(X) ¼ m (8)
s2 G(3=s)
E(X 2 ) ¼ m2 þ
G(1=s)
3ms2 G(3=s)
E(X 3 ) ¼ m3 þ
G(1=s)

and

6m2 s2 G(3=s) s4 G(5=s)

E(X 4 ) ¼ m4 þ þ
G(1=s) G(1=s)

The nth central moment can be obtained as

ð1
n n n x
mn s exp {
(ðx
mÞ=s)s }
E½(X
m)  ¼ s {1 þ (
1) } dx
m s 2sG(1=s)
ð
ssn {1 þ (
1)n } 1 n
¼ z exp (
zs )dz
2G(1=s) 0
ð
sn {1 þ (
1)n } 1 ðnþ1Þ=ðs
1Þ
¼ y exp (
y)dy
2G(1=s) 0

sn {1 þ (
1)n }G(ðn þ 1Þ=s)

¼
2G(1=s)

In particular, the first three central moments, skewness and the kurtosis of X are given by

s2 G(3=s)
Var(X) ¼ (9)
G(1=s)
E½{X
E(X)}3 ¼ 0
s4 G(5=s)
E½{X
E(X)}4 ¼
G(1=s)
Skewness(X) ¼ 0
 A Generalized Normal Distribution 689

and
G(1=s) G(5=s)
Kurtosis(X) ¼ (10)
G2 (3=s)
respectively. Note that the expectation is still m and the skewness is still zero. But both the
variance and the kurtosis measures depend on the shape parameter s. Figure 3 shows that
these two measures decrease with respect to s.

Mean Deviations
The amount of scatter in a population is evidently measured to some extent by the totality
of deviations from the mean and median. These are known as the mean deviation about the
mean and the mean deviation about the median – defined by
ð1
d1 (X) ¼ jx
E(X)j f (x)dx

1

and
ð1
d2 (X) ¼ jx
Median(X)j f (x)dx

1

Figure 3. Variation of variance and kurtosis (given by equations (9) and (10), respectively) versus
s ¼ 0:1; 0:2; . . . ; 10 and s ¼ 1. The y-axis is in log scale
690 S. Nadarajah

respectively. For the generalized normal p.d.f. (equation (2)), both E(X) and Median(X)
are equal to m. Thus, the required measure is EjX
mj. Note that one can write
sI
EjX
mj ¼ (11)
sG(1=s)
where I denotes the integral
ð1 n x
ms o
I¼ (x
m) exp
dx: (12)
m s

Setting y ¼ {(x
m)=s}s , equation (12) can be rewritten as

ð  
s2 1 2=ðs
1Þ s2 2
y exp (
y) dy ¼ G (13)
s 0 s s
Substituting equation (13) into equation (11), one obtains
sG(2=s)
EjX
mj ¼
G(1=s)

Rényi and Shannon Entropies

An entropy of a random variable X is a measure of variation of the uncertainty. Rényi
entropy is defined by
ð 
1
J R (g) ¼ log f g (x)dx (14)
1
g
where g . 0 and g = 1 (Rényi, 1961). For the generalized normal pdf given by
equation (2),
ð1 ð1
sg exp {
gjðx
mÞ=sjs }
f g (x)dx ¼ dx

1
1 (2s)g Gg (1=s)
ð
sg 2sg
ð1=sÞ G(1=s) 1 s exp {
jðx
mÞ=sg
ð1=sÞ js }
¼ dx
(2s)g Gg (1=s) s
1 2sg
ð1=sÞ G(1=s)
sg 2sg
ð1=sÞ G(1=s)
¼
(2s)g Gg (1=s) s

Thus, equation (14) takes the expression

 
log g s
JR (g) ¼
log (15)
s(g
1) 2sG(1=s)

Shannon entropy defined by E½
log f (X) is the particular case of equation (14) for
g " 1. Limiting g " 1 in equation (15) and using L’Hospital’s rule, one obtains
 
1 s
E½
log f (X) ¼
log
s 2sG(1=s)
Song (2001) observed that the gradient of the Rényi entropy J R0 (g) ¼ (d=dg)JR (g)
is related to the loglikelihood by JR0 (1) ¼
(1=2)Var½log (f (X)). This equality and
the fact that the quantity
JR0 (1) remains invariant under location and scale transform-
 A Generalized Normal Distribution 691

ations motivated Song to propose
2JR0 (1) as a measure of the shape of a distribution.
From equation (15), the first derivative is
 
1 1 log g
JR0 (g) ¼

s g(g
1) (g
1)2

Using L’Hospital’s rule again, one gets the expression

1

2JR0 (1) ¼
s
for the measure proposed by Song (2001). This measure plays a similar role as the
kurtosis measure in comparing the shapes of various densities and measuring heaviness
of tails.

Asymptotics
If X1 ; . . . ; Xn is a random sample from equation (2) and if X
¼ (Xp1 ffiffiþ
ffi    þ Xn )=npdenotes
ffiffiffiffiffiffiffiffiffiffiffiffiffiffi
the sample mean then by the usual central limit theorem n(X

E(X))= Var(X)
approaches the standard normal distribution as n ! 1. Sometimes one would be inter-
ested in the asymptotics of the extreme values Mn ¼ max (X1 ; . . . ; Xn ) and
mn ¼ min (X1 ; . . . ; Xn ). Note from equations (5) and (6) that

1 t
m1
s n t
ms o
1
F(t)  exp

2G(1=s) s s

and
  
1 t þ m1
s t þ ms
F(
t)  exp

2G(1=s) s s

as t ! 1. Thus, it can be seen that

    
1
F(t þ xg(t)) t þ xg(t)
m 1
s t
ms t þ xg(t)
m s
lim ¼ lim exp

t!1 1
F(t) t!1 t
m s s
(16)

and
    
F(
t
xg(t)) t þ xg(t) þ m 1
s t þ ms t þ xg(t) þ m s
lim ¼ lim exp

t!1 F(
t) t!1 tþm s s
(17)

as t ! 1. If one chooses g(t)  (ss =s)tð1
sÞ as t ! 1 then both the limits in equations
(16) and (17) would reduce to exp (
x) as t ! 1. Hence, it follows from Theorem 1.6.2
in Leadbetter et al. (1983) that there must be norming constants an . 0, bn , cn . 0 and
dn such that
Pr {an (Mn
bn )  x} ! exp {
exp (
x)}
692 S. Nadarajah

and

Pr {cn (mn
dn )  x} ! 1
exp {
exp (x)}

as n ! 1. The form of the norming constants can also be determined. For instance, using
Corollary 1.6.3 in Leadbetter et al. (1983), one can determine that an ¼ (s=s)( log n)1
ð1=sÞ
and bn ¼ s( log n)1=s .

Estimation Issues
We consider estimation by two methods: the method of moments and the method of
maximum likelihood. Let x1 ; . . . ; xn be a random sample from equation (2). Under the
method of moments, equating E(X), Var(X) and Kurtosis(X) in equations (8), (9), (10),
respectively, with the corresponding sample estimates
1X n
s1 ¼ xi ;
n i¼1

1X n
s2 ¼ (xi
s1 )2
n i¼1

and
1 X n
s3 ¼ (xi
s1 )4
ns22 i¼1

respectively, one obtains the system of equations

m ¼ s1 (18)
s2 G(3=s)
¼ s2 (19)
G(1=s)
and
G(1=s)G(5=s)
¼ s3 (20)
G2 (3=s)
The estimates for m and s can be obtained directly from equations (18) and (20), respect-
ively (the latter equation p
has to be solved numerically).
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi From equation (19), the estimate
for s can be obtained as s2 G(1=^s)=G(3=^s), where s^ denotes the estimate for s.
The log-likelihood for a random sample x1 ; . . . ; xn from equation (2) is:
  X n  
s xi
ms
log L(m; s; s) ¼ n log
  (21)
2sG(1=s) i¼1
s

The derivatives of this log-likelihood with respect to m, s and s are:

( )
@ log L s X s
1
X
s
1
¼ s (xi
m)
(m
xi )
@m s xi m x i ,m

@ log L n s X n
¼
þ sþ1 jxi
mjs
@s s s i¼1
 A Generalized Normal Distribution 693

and

    X n    
@ log L n 1 1 xi
m xi
m
¼ C þ1
  log 
@s s s s i¼1
s s

which can be solved simultaneously for m, s and s.

For interval estimation of (m; s; s) and tests of hypothesis, one requires the Fisher infor-
mation matrix. For a single observation (i.e. n ¼ 1), the second-order derivatives of
equation (21) are:

 
@2 log L s(s
1) ðx
mÞs
2
¼

@m2 s2  s 
 
@2 log L  s 2 ðx
mÞs
1
 
¼ sign(m
x)
@m@s s s 
   
@2 log L 1 ðx
mÞs ðx
mÞ
 
¼ 1 þ s log
@m@s ðx
mÞ  s  s 
  
@2 log L 1 ðx
mÞs
¼ 2 1
s(s þ 1)  
@s 2 s s 
   
@2 log L 1 ðx
mÞs ðx
mÞ
 
¼  1 þ s log
@s@s s s   s 

and

        2
@2 log L ðx
mÞs  
¼

1
1 þ
2
C
1
þ
1 0 1
C
  logðx
mÞ
@s 2 s 2 s s s 2 s  s   s 

Now, using the facts

 
ð1
ðx
mÞm s

E   ¼ zm exp (
zs )dz
s  G(1=s) 0
ð1
1
¼ yðmþ1Þ=ðs
1Þ exp (
y)dy
G(1=s) 0
G(m þ ð1=sÞ)
¼
G(1=s)
 m  
ð1
ðx
mÞ  
E   logðx
mÞ ¼ s zm log z exp (
zs )dz
s   s  G(1=s) 0
ð1
1
¼ yðmþ1Þ=ðs
1Þ log y exp (
y)dy
sG(1=s) 0
G(ðm þ 1Þ=s)C(ðm þ 1Þ=s)
¼
sG(1=s)
694 S. Nadarajah

and
"     # ð1
ðx
mÞm ðx
mÞ 2 s

E   
log  ¼ zm ( log z)2 exp (
zs )dz
s  s  G(1=s) 0
ð1
1
¼ 2 yðmþ1Þ=ðs
1Þ ( log y)2 exp (
y)dy
s G(1=s) 0
G(ðm þ 1Þ=s){C2 (ðm þ 1Þ=s)
C0 (ðm þ 1Þ=s)}
¼
s2 G(1=s)

one can show that the elements of the Fisher information matrix are
 2 
@ log L s(s
1)G(1
ð1=sÞ)
E
¼
@m2 s2 G(1=s)
 2 
@ log L
E
¼0
@m@s
 2 
@ log L
E
¼0
@m@s
 2 
@ log L s
E
2
¼ 2
@s s
 2    
@ log L 1 1
E
¼
1þC 1þ
@s@s ss s

and
 2     
@ log L 1 2 1 1 0 1
E
¼ 1 þ C þ C
@s2 s2 s s s2 s
     

1 1 1 1
þ G 1þ C2 1 þ
C0 1 þ
s s s s

Note that like in the normal distribution, the estimates of m and s are independent.

References
Gradshteyn, I. S. & Ryzhik, I. M. (2000) Table of Integrals, Series, and Products 6th edn (San Diego: Academic
Press).
Leadbetter, M. R., Lindgren, G. & Rootzén, H. (1987) Extremes and Related Properties of Random Sequences
and Processes (New York: Springer-Verlag).
Prudnikov, A. P., Brychkov, Y. A. & Marichev, O. I. (1990) Integrals and Series Vols 1, 2 and 3 (Amsterdam:
Gordon and Breach Science Publishers).
Rényi, A. (1961) On measures of entropy and information, in: Proceedings of the 4th Berkeley Symposium on
Mathematical Statistics and Probability, Vol. I, pp. 547 –561 (Berkeley: University of California Press).
Song, K.-S. (2001) Rényi information, loglikelihood and an intrinsic distribution measure. Journal of Statistical
Planning and Inference, 93, pp. 51–69.