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

net/publication/237670753

Symmetric Unimodal Models for Directional Data Motivated by Inverse

Stereographic Projection

Article  in  JOURNAL OF THE JAPAN STATISTICAL SOCIETY · January 2010

DOI: 10.14490/jjss.40.045

CITATIONS READS

14 104

3 authors, including:

Toshihiro Abe Arthur Pewsey

Nanzan University Universidad de Extremadura
18 PUBLICATIONS   149 CITATIONS    62 PUBLICATIONS   1,556 CITATIONS   

SEE PROFILE SEE PROFILE

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

Grupo EME View project

All content following this page was uploaded by Arthur Pewsey on 25 February 2015.

The user has requested enhancement of the downloaded file.

J. Japan Statist. Soc.
Vol. 40 No. 1 2010 45–61

SYMMETRIC UNIMODAL MODELS FOR DIRECTIONAL

DATA MOTIVATED BY INVERSE STEREOGRAPHIC
PROJECTION

Toshihiro Abe*, Kunio Shimizu** and Arthur Pewsey***

In this paper, a modified inverse stereographic projection, from the real line to
the circle, is used as the motivation for a means of resolving a discontinuity in the
Minh–Farnum family of circular distributions. A four-parameter family of symmetric
unimodal distributions which extends both the Minh–Farnum and Jones–Pewsey fam-
ilies is proposed. The normalizing constant of the density can be expressed in terms
of Appell’s function or, equivalently, the Gauss hypergeometric function. Important
special cases of the family are identified, expressions for its trigonometric moments
are obtained, and methods for simulating random variates from it are described.
Parameter estimation based on method of moments and maximum likelihood tech-
niques is discussed, and the latter approach is used to fit the family of distributions
to an illustrative data set. A further extension to a family of rotationally symmetric
distributions on the sphere is briefly made.

Key words and phrases: Appell’s function, Gauss hypergeometric function, Jones–
Pewsey family, maximum likelihood estimation, Möbius transformation, von Mises
distribution.

1. Introduction
As the papers of Shimizu and Iida (2002), Minh and Farnum (2003) and
Jones and Pewsey (2005) attest, recent years have seen renewed interest in the
development of flexible models for directional data distributed on the unit circle
or sphere. Historically, four general approaches have been used to generate circu-
lar distributions. These are projection (or offsetting), conditioning, wrapping and
(inverse) stereographic projection. The reader is directed to the books by Mardia
and Jupp (1999, Section 3) and Jammalamadaka and SenGupta (2001, Section
2.1.1) for the details of these various constructions. In this paper, we consider
families of distributions that are either directly generated, or are motivated, by
the latter of these general techniques, inverse stereographic projection.
Minh and Farnum (2003) used inverse stereographic projection, or equiva-
lently bilinear (Möbius) transformation, of distributions defined on the real line
to induce distributions on the circle. Their construction based on this approach
proceeds as follows. Let f (x) denote the probability density function (pdf) of
the t-distribution on the real line with m (a positive integer) degrees of freedom,
Received April 7, 2009. Revised February 26, 2010. Accepted April 27, 2010.
*School of Fundamental Science and Technology, Keio University, Yokohama, Japan.
**Department of Mathematics, Faculty of Science and Technology, Keio University, Yokohama, Japan.
***Mathematics Department, Escuela Politécnica, University of Extremadura, 10003 Cáceres, Spain.
46 TOSHIHIRO ABE ET AL.

i.e.
Γ((m + 1)/2)
(1.1) f (x) = √ , −∞ < x < ∞.
πmΓ(m/2)(1 + x2 /m)(m+1)/2
Then, applying inverse stereographic projection, defined by the (generally) one-
to-one mapping,
sin θ
(1.2) x=u+v = u + v tan(θ/2), −π ≤ θ < π,
1 + cos θ
√
with u = 0 and v = m, and writing m = 2n + 1, for n = 0, 1, . . ., leads to the
pdf on the circle of radius v,
Γ(n + 1)
(1.3) f (θ) = √ (1 + cos θ)n .
2n+1 πΓ(n + 1/2)
Given its construction, the distribution with this density might be referred to
as a type of t-distribution on the circle, or the circular t-distribution induced
by inverse stereographic projection if one were being more precise. However,
the distribution with density (1.3) already has a name in the literature, namely
Cartwright’s power-of-cosine distribution. When n = 0 (i.e. m = 1), the cir-
cular uniform distribution is obtained, clearly induced by inverse stereographic
projection of the Cauchy distribution on the real line.
The three-parameter family of symmetric circular distributions proposed by
Jones and Pewsey (2005) has density

(cosh(κψ))1/ψ (1 + tanh(κψ) cos(θ − µ))1/ψ

(1.4) f (θ) = ,
2πP1/ψ (cosh(κψ))
where −π ≤ µ < π is a location parameter (equal, in general, to the mean/mo-
dal/median direction), κ ≥ 0 is a concentration parameter, ψ ∈ R is a shape
index, and P1/ψ is the associated Legendre function of the first kind of degree
1/ψ and order 0. The family with density (1.4) is an important one because
it includes most of the standard symmetric circular distributions. Specifically,
it contains the circular uniform (κ = 0, or ψ → ±∞ and κ finite), von Mises
(ψ = 0), wrapped Cauchy (ψ = −1) and cardioid (ψ = 1) distributions, as well
as Minh and Farnum’s (2003) (or, equivalently, Cartwright’s power-of-cosine)
distribution, with the range of n now extended to include the whole of the positive
real line, (ψ > 0, κ → ∞). Due to the relation that exists between the associated
Legendre function and the Gauss hypergeometric function (a combination of
equations 8.772.3 and 9.134.2 of Gradshteyn and Ryzhik (2007, pp. 970 and
1009)), namely

(1.5) P1/ψ (z) = z 1/ψ 2 F1 (−1/(2ψ), (1 − 1/ψ)/2; 1; (z 2 − 1)/z 2 ),

(1.4) can be re-expressed as

(1 + tanh(κψ) cos(θ − µ))1/ψ

f (θ) = .
2π 2 F1 (−1/(2ψ), (1 − 1/ψ)/2; 1, tanh2 (κψ))
 INVERSE STEREOGRAPHIC PROJECTION 47

Here the Gauss hypergeometric function is defined as

1 b−1
1 t (1 − t)c−b−1
2 F1 (a, b; c; z) = dt
B(b, c − b) 0 (1 − tz)a
or, alternatively, as
∞
 (a)r (b)r z r
2 F1 (a, b; c; z) = , |z| < 1,
r=0
(c)r r!

where 
a(a + 1) · · · (a + r − 1), r ≥ 1,
(a)r =
1, r = 0,
is referred to as Pochhammer’s symbol. Note that in (1.5) the value of 2 F1 is
positive if ψ < 0 (from the definition of the Gauss hypergeometric function), and
is also positive if ψ > 0 due to the transformation formula (Abramowitz and
Stegun (1972), 15.3.3, p. 559) 2 F1 (a, b; c; z) = (1 − z)c−a−b 2 F1 (c − a, c − b; c; z).
Making use of the substitutions 1/ψ = −(1+n/2) and tanh(κψ) = −2κn (ρ | r)/n,
where n is a positive integer, ρ > 0, r > 0 and κn (ρ | r) = ρr/{1 + (ρ2 + r2 )/n},
this last representation of the density of the Jones–Pewsey distribution can be
seen to be closely related to the pdf of the t-distribution on the circle proposed
by Shimizu and Iida (2002), which is given by

{1 − (2/n)κn (ρ | r) cos(θ − µ)}−n/2−1

f (θ) = .
2π 2 F1 (n/4 + 1/2, n/4 + 1; 1; (2κn (ρ | r)/n)2 )
Thus, the Jones–Pewsey family also includes the t-distribution on the circle of
Shimizu and Iida (2002), covered as it is by ψ-values in the range (−1, 0).
The distribution with density (1.3) lacks continuity, in the sense that (1.1)
tends to the standard normal pdf as m → ∞ but a degenerate distribution
concentrated at the point θ = 0 is obtained when n → ∞ in (1.3). In order to
resolve this problem we employ the transformation (1.2) with u = 0 and v > 0
independent of n. A geometric representation of this transformation is portrayed
in Figure 1.
Now inverse stereographic projection of the t-distribution on the real line
yields a different circular t-distribution from that given in (1.3), with density

v (1 + tan2 (θ/2))
(1.6) f (θ) = √ × ,
2 mB(m/2, 1/2) (1 + v 2 tan2 (θ/2)/m)(m+1)/2

where m = 2n + 1 and B(p, q) denotes the beta function, defined by B(p, q) =

 1 p−1
0 t (1 − t)q−1 dt. In this paper, we shall refer to the distribution with density
(1.6) as the modified Minh–Farnum distribution. As m → ∞, the pdf in (1.6)
tends to
     
v θ v2 θ
(1.7) f (θ) = √ 1 + tan2 exp − tan2 .
2 2π 2 2 2
48 TOSHIHIRO ABE ET AL.

(a) (b)

Figure 1. Graphical representation of the stereographic projection x = v tan(θ/2) for a)

0 < θ < π/2; b) π/2 < θ < π.

This is the density of the distribution obtained when the modified inverse stereo-
graphic projection described above is applied to the standard normal distribution
on the real line. Thus, the transformation (1.2) with v > 0 independent of n has
continuity from the t to the normal distributions
√ on the circle. √
The density (1.7) is unimodal if v ≥ 2, and is bimodal if v < 2. An
alternative representation of (1.6) is

v0 (1 + cos θ)n
(1.8) f (θ) = × ,
B(n + 1/2, 1/2)(1 + v02 )n+1 (1 − z cos θ)n+1
√
where v0 = v/ 2n + 1 and z = (v02 −1)/(v02 +1). Finally, we note that the Jones–
Pewsey density tends to that of a Minh–Farnum distribution when κ → ∞.
The remainder of the paper is structured as follows. In Section 2 we propose
a new family of circular distributions which extends both the Jones–Pewsey and
unimodal modified Minh–Farnum distributions. The basic properties of this new
family, such as its unimodality, special cases and trigonometric moments, are
considered in Section 3. We also describe how random variates can be simulated
from it using the inversion and acceptance-rejection methods. Section 4 addresses
the estimation of the distribution’s parameters. There we discuss the use of the
method of moments and maximum likelihood approaches. Maximum likelihood
techniques are then employed to fit the family to an illustrative data set in
Section 5. A form of stereographic projection which can be used to generate
distributions on the sphere of arbitrary dimension is introduced in Section 6.
The paper concludes, in Section 7, with a p-dimensional extension of the new
family of circular distributions proposed in Section 2.

2. A new extended family of unimodal symmetric circular distribu-

tions
In this section we introduce the new family of unimodal symmetric distribu-
tions on the circle which extends both the Jones–Pewsey family of distributions as
well as the unimodal distributions contained within the modified Minh–Farnum
family. Its density results on combining the defining structure of the pdf in (1.8),
 INVERSE STEREOGRAPHIC PROJECTION 49

(a) (b)

(c) (d)

Figure 2. Density plots of (2.1) for µ = 0 and: (a) ψ = 0.5, κ2 = 0; (b) ψ = 0.5, κ2 = 1;
(c) ψ = 1, κ2 = 0; (d) ψ = 1, κ2 = 1. The six densities plotted in each panel are those
for κ1 = ∞, 2, 1.5, 1, 0.5, 0 in order of decreasing height at 0. The densities for κ1 = 0, ∞
in each panel are the densities for the bounding Jones–Pewsey and modified Minh–Farnum
distributions, respectively. The horizontal lines in panels (a) and (c) delimit the density of the
circular uniform distribution. Note that the vertical scales vary between panels.

derived by inverse stereographic projection, with the transformation technique

based on the hyperbolic tangent function employed in the construction of Jones
and Pewsey (2005). The resulting density is given by

(1 + tanh(κ1 ψ) cos(θ − µ))1/ψ

(2.1) f (θ; µ, κ1 , κ2 , ψ) = Cψ (κ1 , κ2 ) ,
(1 − tanh(κ2 ψ) cos(θ − µ))1/ψ+1

where −π ≤ µ < π is a location parameter (which, in general, corresponds to

the distribution’s mean/modal/median direction), κ1 , κ2 ≥ 0 are concentration
parameters and ψ > 0 is an index which also effects the peakedness and tails of the
distribution. Note that f (θ; µ, −κ1 , −κ2 , ψ) = f (θ; µ+π, κ1 , κ2 , ψ) if −π ≤ µ < 0,
= f (θ; µ − π, κ1 , κ2 , ψ) if 0 ≤ µ < π. The overall shape of the distribution is
determined by the last three of these parameters. The normalizing constant
Cψ (κ1 , κ2 ) is the reciprocal of the integral

π
(1 + tanh(κ1 ψ) cos θ)1/ψ
(2.2) dθ
−π (1 − tanh(κ2 ψ) cos θ)1/ψ+1

1
2(1 + tanh(κ1 ψ))1/ψ (1 − z1 t)1/ψ
= t−1/2 (1 − t)−1/2 dt,
(1 − tanh(κ2 ψ))1/ψ+1 0 (1 − z2 t)1/ψ+1

where z1 = 2 tanh(κ1 ψ)/(1+tanh(κ1 ψ)) and z2 = −2 tanh(κ2 ψ)/(1−tanh(κ2 ψ)).

Here, the transformation cos θ = 1 − 2t has been used to obtain the right-hand
50 TOSHIHIRO ABE ET AL.

side of (2.2). The last integral is expressible in terms of Appell’s function, F1 ,

and the Gauss hypergeometric function, 2 F1 , as

1  
(1 − z1 t)1/ψ 1 1 1
t−1/2 (1 − t)−1/2 dt = πF1 ; − , 1 + ; 1; z1 , z2
0 (1 − z2 t)1/ψ+1 2 ψ ψ
 
1 1 z 1 − z2
= π(1 − z2 )−1/2 2 F1 , − ; 1; ,
2 ψ 1 − z2
due to the relation (Appell and Kampé de Fériet (1926, Eq. (28), p. 24))
F1 (a; b1 , b2 ; b1 + b2 ; z1 , z2 ) = (1 − z2 )−a 2 F1 (a, b1 ; b1 + b2 ; (z1 − z2 )/(1 − z2 )), where

1
1 ta−1 (1 − t)c−a−1
F1 (a; b1 , b2 ; c; z1 , z2 ) = dt.
B(a, c − a) 0 (1 − z1 t)b1 (1 − z2 t)b2
The flexibility of density (2.1) is illustrated in Figure 2 for µ = 0 and various
combinations of the other three parameters. Note how, as ψ and κ2 increase, κ1
has increasingly less effect on the shape of the density.

3. Basic properties
3.1. Unimodality
Given that the two concentration parameters are constrained so that κ1 , κ2 ≥
0, it follows that distributions with density (2.1) are always unimodal with f (µ) >
f (µ ± π), apart from the special case of the circular uniform distribution which,
formally, has no mode. The circular uniform distribution is obtained when κ1 =
κ2 = 0 or ψ → ∞ and κ1 and κ2 are finite.
To show that, for all other cases, distributions from the proposed family are
unimodal with mode at µ, we set, without loss of generality, µ equal to 0 and
consider the derivative of (2.1) with respect to θ, i.e.
 
d (1 + tanh(κ1 ψ) cos θ)1/ψ
dθ (1 − tanh(κ2 ψ) cos θ)1/ψ+1
sin θ(1 + tanh(κ1 ψ) cos θ)1/ψ−1
=−
(1 − tanh(κ2 ψ) cos θ)1/ψ+2
 
(1 + ψ) tanh(κ2 ψ) + tanh(κ1 ψ)
× tanh(κ1 ψ) tanh(κ2 ψ) cos θ + .
ψ tanh(κ1 ψ) tanh(κ2 ψ)
It follows from the constraints on the parameters that the density is unimodal
if |[(1 + ψ) tanh(κ2 ψ) + tanh(κ1 ψ)]/[ψ tanh(κ1 ψ) tanh(κ2 ψ)]| > 1. Now, (1 +
tanh(κ1 ψ))1/ψ > (1 − tanh(κ1 ψ))1/ψ and (1 − tanh(κ2 ψ))1/ψ+1 <
(1 + tanh(κ2 ψ))1/ψ+1 . Thus, (1 + tanh(κ1 ψ))1/ψ /(1 − tanh(κ2 ψ))1/ψ+1 > (1 −
tanh(κ1 ψ))1/ψ /(1 + tanh(κ2 ψ))1/ψ+1 . This implies that f (0) > f (±π). In sum-
mary, then, distributions with density (2.1) are either circular uniform or, more
generally, unimodal with mode at µ. The fact that they are also symmetric about
µ follows directly from (2.1).
√
We note that when κ1 → ∞, 1/ψ = (m − 1)/2, κ2 = −1/ψ log( m/v),
m ≥ 1 and v > 0, the unimodality condition reduces to that of the modified
 INVERSE STEREOGRAPHIC PROJECTION 51

Minh–Farnum distribution, namely |[m(1 − v 2 )]/[m − v 2 ]| ≥ 1. In addition, as

m → ∞, this becomes the unimodality condition of the circular standard normal
distribution induced by modified inverse stereographic projection.

3.2. Special cases

In this subsection we consider four special cases of the new family. The first
three cover the Jones–Pewsey family and the unimodal distributions contained
within the modified Minh–Farnum family, whilst the last one corresponds to
a distribution with a pole at 0. We will assume throughout, without loss of
generality, that µ = 0 in (2.1).

Case 1. κ2 = 0, the Jones–Pewsey distribution with positive power 1/ψ.

For this value of κ2 , the density becomes f (θ; 0, κ1 , 0, ψ) = C(κ1 , 0)(1 +
tanh(κ1 ψ) cos θ)1/ψ , where C(κ1 , 0) = (cosh(κ1 ψ))1/ψ /[2πP1/ψ (cosh(κ1 ψ))].

Case 2. κ1 = 0, the Jones–Pewsey distribution with negative power −(1 +

1/ψ). For this choice of κ1 , the density can be written as
(1 − tanh(κ2 ψ))1/ψ+1 (1 − tanh(κ2 ψ) cos θ)−1/ψ−1
f (θ; 0, 0, κ2 , ψ) =
2π 2 F1 (1/ψ + 1, 1/2; 1; z2 )
(1 − tanh(κ2 ψ) cos θ)−1/ψ−1
= ,
2π 2 F1 (1/(2ψ) + 1/2, 1/(2ψ) + 1; 1; tanh2 (κ2 ψ))
where the quadratic transformation formula (Abramowitz and Stegun (1972,
15.3.16, p. 560)) 2 F1 (a, b; 2b; z) = (1 − z/2)−a 2 F1 (a/2, (a + 1)/2; b + 1/2; z 2 (2 −
z)−2 ) has been employed to obtain the second equality.

Case 3. κ1 → ∞, the unimodal distributions contained in the modified

Minh–Farnum family with density (1.6). When κ1 → ∞, f (θ; 0, ∞, κ2 , ψ) =
Cψ (∞, κ2 )(1 + cos θ)1/ψ /(1 − tanh(κ2 ψ) cos θ)1/ψ+1 , where
(1 − tanh(κ2 ψ))1/ψ+1
Cψ (∞, κ2 ) =
21/ψ+1 B(1/2, 1/ψ + 1/2)2 F1 (1/ψ + 1, 1/2; 1/ψ + 1; z2 )
(1 − tanh(κ2 ψ))1/ψ+1 (1 + tanh(κ2 ψ/2))
= .
21/ψ+1 B(1/2, 1/ψ + 1/2) (1 − tanh(κ2 ψ/2))
For this case note that f (±π) = 0. Substituting, 1/ψ = (m − 1)/2 and κ2 =
√
−1/ψ log( m/v) for m ≥ 1 and v > 0, gives
√ v
Cψ (∞, log( m/v)(m−1)/2 ) = √ .
mB(m/2, 1/2)(1 + v 2 /m)(m+1)/2
The density (2.1) can then be expressed as
 √  
m−1 m 2
f θ; 0, ∞, log ,
2 v m−1
v(1 + tan2 (θ/2))
= √ ,
2 mB(m/2, 1/2)(1 + v 2 tan2 (θ/2)/m)(m+1)/2
52 TOSHIHIRO ABE ET AL.

which is identical to that given previously in (1.6).

Case 4. κ2 → +∞, a new distribution with a pole at 0. When κ2 → +∞,

f (θ; 0, κ1 , +∞, ψ) = Cψ (κ1 , +∞)(1 + tanh(κ1 ψ) cos θ)1/ψ /(1 − cos θ)1/ψ+1 , where

Cψ (κ1 , +∞) = (1 + tanh(κ1 ψ)/2)1/ψ 2 F1 (−1/ψ, −1/ψ − 1/2; −1/ψ; z1 )

× B(−1/ψ − 1/2, 1/2)
√
= eκ1 ψ [2/(1 − tanh(κ1 ψ))]1/ψ Γ(−1/ψ)/[ πΓ(−1/ψ − 1/2)].

Note that the ratio of gamma functions in this last expression is non-negative
since Γ(−1/ψ) = (−1)n+1 Γ(n + 1 − 1/ψ)/[(1/ψ)(1/ψ − 1) · · · (1/ψ − n)] and
Γ(−1/ψ − 1/2) = (−1)n+1 Γ(n + 1/2 − 1/ψ)/[(1/ψ + 1/2)(1/ψ − 1/2) · · · (1/ψ +
1/2 − n)], for n < 1/ψ < n + 1 (n = 2, 3, 4, . . .), and hence that the normalizing
constant is indeed non-negative.

3.3. Trigonometric moments

As in the previous subsection, we also assume here that µ = 0 in (2.1). The
trigonometric moments of the distribution are given by {φm : m = 1, 2, 3, . . .},
where φm = αm + iβm , with αm = E(cos mΘ) and βm = E(sin mΘ) being the
mth order cosine and sine moments of the random angle Θ, respectively. Because
the distribution is symmetric about µ = 0, it follows that the sine moments are
all zero. Thus, φm = αm . To obtain the cosine moments, we will make use of
the relationships that exist between them and the (standard) moments of cos Θ.
The latter can be expressed as integral expressions involving Appell’s function,
F1 , or the Gauss hypergeometric function, 2 F1 . Specifically, in terms of Appell’s
function,

π
m (cos θ)m (1 + tanh(κ1 ψ) cos θ)1/ψ
E(cos Θ) = 2Cψ (κ1 , κ2 ) dθ
0 (1 − tanh(κ2 ψ) cos θ)1/ψ+1
1 m
k
= m Ck (−2)
πF1 (1/2, −1/ψ, 1 + 1/ψ; 1, z1 , z2 ) k=0
   
1 1 1 1 1
× F1 k + , − , 1 + ; k + 1, z1 , z2 B k + , ,
2 ψ ψ 2 2
for all m = 0, 1, 2, . . ., and, in terms of the Gauss hypergeometric function,
∞  m  
2Cψ (κ1 , κ2 )(1 + tanh(κ1 ψ))1/ψ  1 (−2)k z1n
m
E(cos Θ) = − m Ck
(1 − tanh(κ2 ψ))1/ψ+1 n=0 k=0
ψ n n!
   
1 1 1 1
× 2 F1 1 + , n + k + ; n + k + 1; z2 B n + k + , .
ψ 2 2 2
Using these expressions, the first three cosine moments of Θ are given by

(1 − z2 )1/2 1
k
α1 = 1 Ck (−2)
π 2 F1 (1/2, −1/ψ; 1; (z1 − z2 )/(1 − z2 )) k=0
 INVERSE STEREOGRAPHIC PROJECTION 53
   
1 1 1 1 1
× F1 k + , − , 1 + ; k + 1, z1 , z2 B k + ,
2 ψ ψ 2 2
(1 − z2 ) F1 (3/2, −1/ψ, 1 + 1/ψ; 2, z1 , z2 )
1/2
= 1− ,
2 F1 (1/2, −1/ψ; 1; (z1 − z2 )/(1 − z2 ))
α2 = 2E(cos2 Θ) − 1
2(1 − z2 )1/2 2
k
= 2 Ck (−2)
π 2 F1 (1/2, −1/ψ; 1; (z1 − z2 )/(1 − z2 )) k=0
   
1 1 1 1 1
× F1 k + , − , 1 + ; k + 1, z1 , z2 B k + , −1
2 ψ ψ 2 2
4(1 − z2 ) F1 (3/2, −1/ψ, 1 + 1/ψ; 2, z1 , z2 )
1/2
= 1−
2 F1 (1/2, −1/ψ; 1; (z1 − z2 )/(1 − z2 ))
3(1 − z2 )1/2 F1 (5/2, −1/ψ, 1 + 1/ψ; 3, z1 , z2 )
+
2 F1 (1/2, −1/ψ; 1; (z1 − z2 )/(1 − z2 ))
and
α3 = 4E(cos3 Θ) − 3E(cos Θ)
4(1 − z2 )1/2 3
k
= 3 Ck (−2)
π 2 F1 (1/2, −1/ψ; 1; (z1 − z2 )/(1 − z2 )) k=0
   
1 1 1 1 1
× F1 k + , − , 1 + ; k + 1, z1 , z2 B k + ,
2 ψ ψ 2 2
 
(1 − z2 )1/2 F1 (3/2, −1/ψ, 1 + 1/ψ; 2, z1 , z2 )
−3 1−
2 F1 (1/2, −1/ψ; 1; (z1 − z2 )/(1 − z2 ))

9(1 − z2 )1/2 F1 (3/2, −1/ψ, 1 + 1/ψ; 2, z1 , z2 )

= 1−
2 F1 (1/2, −1/ψ; 1; (z1 − z2 )/(1 − z2 ))

3
− z2 )1/2 F1 (k + 1/2, −1/ψ, 1 + 1/ψ; k + 1, z1 , z2 )
k (1
+4 3 Ck (−2)
k=2 2 F1 (1/2, −1/ψ; 1; (z1 − z2 )/(1 − z2 ))
 
1 1
×B k+ , .
2 2
Higher-order moments can be obtained similarly.

3.4. Simulation
Lacking any obvious direct construction which leads to the distribution with
density (2.1), here we present two approaches to simulating pseudo-random vari-
ates from it based on the more generally applicable inversion and acceptance-
rejection methods.
Applying the inversion method, if U is a uniform random number on (0,1)
then, in order to generate a variate on (−π, π) from the distribution with pa-
rameters µ = 0, κ1 , κ2 and ψ, one is required to solve numerically for θ in the
θ
equation F (θ) = U , where F (θ) = −π f (φ)dφ is the value of the distribution
function at θ. Trivially, the equation required to be solved can be re-expressed
as F (θ) − U = 0 and thus numerical methods of root finding can be used to solve
54 TOSHIHIRO ABE ET AL.

for θ. For instance, the approximation to the root at the (n + 1)st iteration of
the Newton–Raphson method is given by θn+1 = θn − (F (θn ) − U )/f (θn ). As the
distribution is symmetric about µ = 0, an initial value for the root, θ0 , can be
chosen on the basis of whether U < 1/2 (θ0 ∈ (−π, 0)) or U > 1/2 (θ0 ∈ (0, π)).
Of course, if U = 1/2 then θ = 0. From our experience, the Newton–Raphson
method works well, except for when the density is very peaked. Other standard
numerical methods for obtaining the roots of non-linear equations which do not
make use of derivatives are more reliable in such circumstances.
For cases of the distribution which are not overly peaked, an approach based
on the acceptance-rejection method will generally be more efficient than the one
described above. In its crudest form, such an approach requires the genera-
tion of two uniform random numbers: U1 on (−π, π), and U2 on (0, fm ), where
fm = Cψ (κ1 , κ2 )(1+tanh(κ1 ψ))1/ψ /(1−tanh(κ2 ψ))1/ψ+1 is the modal value of the
density. If U2 > f (U1 ) then U1 is rejected. Otherwise it is accepted as a random
variate from the distribution. For fixed values of κ1 , κ2 and ψ, this approach only
requires the initial computation of fm and, for each acceptance-rejection step,
the two uniform numbers U1 and U2 and the value of f (U1 ). Whilst the computa-
tional effort required at each acceptance-rejection step is obviously less than that
for the inversion method, the overall efficiency of this approach will depend on
the acceptance rate, 1/(2πfm ). At the expense of increasing the computational
effort, the acceptance rate can be increased by using a non-rectangular envelope
which more closely bounds the target density.
For the simulation of variates from cases of the distribution with non-zero
values of µ, one simply follows either of the recipes described above to obtain θ
and then takes θ + µ (mod 2π).

4. Parameter estimation
Our extended family of distributions with density (2.1) has four parameters
which, in general, will all require estimation. Let θ1 , . . . , θn be an independent
and identically distributed sample drawn from the distribution with density (2.1).
In this section we will consider the estimation of the four parameters based on
the method of moments and maximum likelihood approaches.

4.1. Method of moments

In principle, method of moments (MM) estimation would proceed by equat-
ing the sample mean direction θ̄ with µ, the first three sample cosine moments

about θ̄, i.e. ā1 , ā2 and ā3 , where ām = n−1 ni=1 cos(m(θi − θ̄)), with the expres-
sions for α1 , α2 and α3 given in Subsection 3.3, and solving for µ, κ1 , κ2 and ψ.
Whilst, from the first of these equations, the MM estimate of µ is clearly µ̃ = θ̄,
the MM estimates of the other three parameters would involve the solution of a
system of three highly complicated non-linear simultaneous equations in κ1 , κ2
and ψ. For this reason, we do not feel encouraged to promote this particular
approach to point estimation.
 INVERSE STEREOGRAPHIC PROJECTION 55

4.2. Maximum likelihood

Using the results presented in Section 2, the log-likelihood function can be
expressed as

(4.1) l(µ, κ1 , κ2 , ψ)
n n
= −n log(2π) − log(1 + tanh(κ1 ψ)) + log(1 + tanh(κ2 ψ))
ψ 2
  
1 1 2(tanh(κ1 ψ) + tanh(κ2 ψ))
− n log 2 F1 , − ; 1;
2 ψ (1 + tanh(κ1 ψ))(1 + tanh(κ2 ψ))
 
1 1
+n + log(1 − tanh(κ2 ψ))
ψ 2
1 n
+ log(1 + tanh(κ1 ψ) cos(θi − µ))
ψ i=1
  n
1 
− 1+ log(1 − tanh(κ2 ψ) cos(θi − µ)).
ψ i=1

Maximum likelihood (ML) point estimation then reduces to the constrained max-
imization of (4.1) over the points in the parameter space.
The elements of the score vector are just the first-order partial derivatives
of (4.1) with respect to each of the parameters. The expressions for some of
these partial derivatives are rather involved, and so, with space restrictions in
mind, we do not present them here. For those interested in deriving them, we
would recommend the use of a symbolic mathematical computing package, such
as Mathematica.
In addition to its symbolic capabilities, we used Mathematica’s NMaximize
command to carry out the actual numerical optimization of (4.1). This command
has four options which determine the method used to perform the maximization.
These are: i) the Nelder–Mead simplex, ii) simulated annealing, iii) differential
evolution and iv) random search. The default option is the Nelder–Mead simplex,
which we found to be fairly robust. From our experience, NMaximize generally has
no difficulty in homing in on the ML solution. However, for some of the samples
that we analyzed, the log-likelihood surface was found to have multiple maxima.
For others, the algorithm sometimes converged to a point on the boundary of the
parameter space. Thus it is advisable to explore the parameter space carefully
so as to try and ensure that the maximum likelihood solution really is identified.
This can be achieved by using a wide range of different initial values from which
to start the iterative search off from.
Standard asymptotic theory applies for interior points of the parameter
space, but not on its boundaries. The calculations involved in obtaining the
second-order partial derivatives of the log-likelihood function required for the
observed and expected information matrices are yet more involved than those
for the first-order partial derivatives and so, again, we would recommend the use
of a symbolic mathematical computing package to obtain them. Denoting the
expected values of minus the second derivatives of the log-likelihood by iµµ , iµκ1 ,
56 TOSHIHIRO ABE ET AL.

etc., it transpires that iµκ1 = iµκ2 = iµψ = 0. Thus, the maximum likelihood
estimate of the location parameter µ is asymptotically independent of those for
the other three parameters. However, the other elements of the expected infor-
mation matrix are generally non-zero. Estimates of the asymptotic covariance
matrix for the ML estimates can be obtained by inverting the information ma-
trices evaluated at the maximum likelihood estimates. These matrices can then
be used together with standard asymptotic normal theory for ML estimators to
calculate confidence intervals and regions for the parameters. Alternatively, pro-
file likelihood methods together with standard asymptotic chi-squared theory for
the likelihood ratio test can be employed. The latter requires only minor modi-
fications to the numerical routines used to obtain the ML estimates and so will
generally be less computationally demanding. However, the decision as to which
approach to use will generally depend on the software available to the user.

5. An example
As our illustrative example, we present an analysis of the same grouped
data set of n = 714 observations as was considered in Jones and Pewsey (2005).
The data consist of the ‘vanishing angles’ of non-migratory British mallard ducks
taken from Table 1 of Mardia and Jupp (1999). As reported in Jones and Pewsey
(2005), the test of Pewsey (2002) provides no evidence against circular reflective
symmetry for these data (p-value = 0.124).
The results obtained from fitting the new family of distributions and the
Jones–Pewsey and modified Minh–Farnum submodels are presented in Table 1.
A histogram of the data together with the densities for the three fits are presented
in Figure 3. From the latter it can be seen that the density of the fit for the
full family is bounded above and below by the densities of the fits for the two
limiting submodels. Moreover, there is evidence, particularly around the mode,
that the fit for the full family is more similar to the fit for the modified Minh–
Farnum distribution than that for the Jones–Pewsey. This visual impression is
confirmed by the results for the maximum values of the log-likelihood presented in
Table 1. The maximum likelihood solution for the full family is an interior point
of the parameter space, and so applying standard asymptotic distribution theory
for the likelihood ratio test, there is some evidence that it offers a significant
improvement in fit over the Jones–Pewsey family (p-value = 0.070), but not over
the modified Minh–Farnum family (p-value = 0.123).
Focusing on the results for the two model selection criteria included in Ta-
ble 1, the AIC identifies the full family as providing a slightly better fit than the
modified Minh–Farnum submodel. The BIC penalizes parameter-heavy models
more and identifies both the modified Minh-Farnum and Jones-Pewsey three-
parameter submodels as providing better fits than the full four-parameter family.
Chi-squared goodness-of-fit tests at the 5% significance level marginally re-
jected the fit of the Jones–Pewsey submodel (p-value = 0.047), but did not reject
the fit of the modified Minh–Farnum submodel (p-value = 0.120) nor that of the
full family (p-value = 0.082). Again, these p-values provide more evidence of the
 INVERSE STEREOGRAPHIC PROJECTION 57

Figure 3. Histogram of the vanishing angles of 714 British mallard ducks together with the max-
imum likelihood fits for the proposed model (solid) and the modified Minh–Farnum (dashed)
and Jones–Pewsey (dot-dashed) distributions. The data and densities are plotted on approxi-
mately (µ̂ − π, µ̂ + π).

Table 1. Maximum likelihood estimates for three fits to the duck data, and the corresponding
maximized log-likelihood (MLL), AIC and BIC values.

Distribution κ̂1 κ̂2 µ̂ ψ̂ MLL AIC BIC

Jones–Pewsey 0 1.19 −0.82 0.54 −869.36 1744.72 1758.43
Proposed Model 0.74 0.42 −0.82 2.35 −867.72 1743.44 1761.72
Modified Minh–Farnum ∞ 0.35 −0.80 3.05 −868.91 1743.82 1757.53

superior fit of the modified Minh–Farnum submodel for these data.

Finally, approximate 95% confidence intervals for µ, κ1 , κ2 and ψ, obtained
from their respective profile log-likelihood functions, are (−0.87, −0.77), (0, ∞),
(−1.35, −0.28) and (2.35, 2.40). Of course, the interpretation of these confidence
intervals is hampered by the fact that they only provide information about each
individual parameter, not their relations with the others. Nevertheless, the inter-
vals for µ, κ2 and ψ give fairly tight information regarding the probable values
for these three parameters. The interval for κ1 simply indicates that models
across the whole range from the Jones–Pewsey to the modified Minh–Farnum
are potential generators of the data. Obviously, confidence regions for different
combinations of the parameters would provide insight as to the probable joint
values of the parameters.

6. A p-dimensional version of stereographic projection

A three-dimensional version of stereographic projection which preserves an-
gles is mentioned by Watson (1983). However, when applying it, the Jacobian
does not take a simple form. Thus, here we propose a projection which maps the
exterior point x ∈ Rp−1 into the southern hemisphere of the p-dimensional sphere
of radius v, Spv , and the interior point x ∈ Rp−1 into the northern hemisphere of
58 TOSHIHIRO ABE ET AL.

(a) (b)

Figure 4. The p-dimensional generalization of stereographic projection from the point x in Rp−1
to the point ξ on the surface of the sphere in Rp for a) 0 < θp−2 < π/2; b) π/2 < θp−2 < π.

Spv . As such, this projection is somewhat different from that of Watson. Figure 4
provides a graphical representation of the proposed projection.
Let p ≥ 3, v > 0, and Spv := {ξ ∈ Rp ; |ξ| = v}. For −π ≤ φ ≤ π and
0 ≤ θi ≤ π (i = 1, . . . , p − 2), we consider the polar coordinates
⎛ ⎞ ⎛ ⎞
ξ1 v cos φ sin θ1 · · · sin θp−2
⎜ ⎟ ⎜ ⎟
⎜ ξ2 ⎟ ⎜ v sin φ sin θ1 · · · sin θp−2 ⎟
⎜ ⎟ ⎜ ⎟
⎜ .. ⎟ ⎜ .. ⎟
(6.1) ⎜ ⎟=⎜ ⎟.
⎜ . ⎟ ⎜ . ⎟
⎜ ⎟ ⎜ ⎟
⎜ξ ⎟ ⎜ v cos θp−3 sin θp−2 ⎟
⎝ p−1 ⎠ ⎝ ⎠
ξp v cos θp−2

Then, for any x = (x1 , . . . , xp−1 ) ∈ Rp−1 , there exists a unique point ξ =
(ξ1 , . . . , ξp ) ∈ Spv such that
 
vξ1 vξp−1
(x1 , . . . , xp−1 ) = ,...,
v + ξp v + ξp
 
v cos φ sin θ1 . . . sin θp−2 v cos θp−3 sin θp−2
= ,..., ,
1 + cos θp−2 1 + cos θp−2

where ξp = −v. The Jacobian is

   
 ∂(x , x , . . . , x  v sinp−2 θp−2  ∂(x1 , x2 , . . . , xp−1 ) 
 1 2 p−1 ) 
  =  
 ∂(φ, θ1 , . . . , θp−2 )  (1 + cos θp−2 )p−1  ∂(v, φ, θ1 , . . . , θp−3 ) 
 p−1
v
= sin θ1 · · · sinp−2 θp−2 .
1 + cos θp−2

This leads to the following proposition.

Proposition 6.1. For i = 1, . . . , p − 2, suppose φ and θi are randomly cho-

sen angles in the intervals (−π, π) and (0, π), respectively. Let g(φ, θ1 , . . . , θp−2 )
 INVERSE STEREOGRAPHIC PROJECTION 59

denote the pdf of (φ, θ1 , . . . , θp−2 ). Then (6.1) is a random point on the sphere
of radius v. Let x = (x1 , . . . , xp−1 ) ∈ Rp−1 and f (x) denote the pdf of x. Then,
g can be written in terms of f as
 
 ∂(x , x , . . . , x 
 1 2 p−1 ) 
g(φ, θ1 , . . . , θp−2 ) = f (x)  
 ∂(φ, θ1 , . . . , θp−2 ) 
 
v cos φ sin θ1 · · · sin θp−2 v cos θp−3 sin θp−2
=f ,..., J,
1 + cos θp−2 1 + cos θp−2

where  p−1
v
J= sin θ1 · · · sinp−2 θp−2 .
1 + cos θp−2

We note that, when p = 2, the projection proposed above is equivalent to

that in Theorem 1 of Minh and Farnum (2003).

7. The p-dimensional generalization of density (2.1)

In this section, we derive the p-dimensional generalization of the pdf (2.1),
namely

(7.1) g(φ, θ1 , . . . , θp−2 )

(1 + tanh(κ1 ψ) cos(θp−2 − µ))1/ψ+1−p/2 
p−2
= Cψ,p (κ1 , κ2 ) sini θi ,
(1 − tanh(κ2 ψ) cos(θp−2 − µ))1/ψ+p/2
i=1

where ψ > 0, κ1 , κ2 ≥ 0, p ≥ 3, and the normalizing constant Cψ,p (κ1 , κ2 ) is the

reciprocal of

π
π
π
(1 + tanh(κ1 ψ) cos θp−2 )1+1/ψ−p/2 
p−2
··· sini θi dθ1 · · · dθp−2 dφ
−π 0 0 (1 − tanh(κ2 ψ) cos θp−2 )1/ψ+p/2 i=1
 
(1 + tanh(κ1 ψ))1+1/ψ−p/2 p−1 p−1
= 2p−2 ωp−1 B ,
(1 − tanh(κ2 ψ)) 1/ψ+p/2 2 2
   
p−1 1 p 1 p
× F1 ;− 1 + − , + , p − 1; z1 ; z2
2 ψ 2 ψ 2
 
(1 + tanh(κ1 ψ))1+1/ψ−p/2 −(p−1)/2 p−1 p−1
= 2p−2 ωp−1 (1 − z 2 ) B ,
(1 − tanh(κ2 ψ))1/ψ+p/2 2 2
   
p−1 1 p z1 − z2
× 2 F1 ,− 1 + − ; p − 1; ,
2 ψ 2 1 − z2

where ωp−1 is the surface area of the unit sphere in Rp−1 , given by ωp−1 =
2π (p−1)/2 /Γ((p − 1)/2). This reduces to a simpler form when p = 3, namely

π
π
(1 + tanh(κ1 ψ) cos θ)−1/2+1/ψ
sin θdθdφ
−π 0 (1 − tanh(κ2 ψ) cos θ)3/2+1/ψ
60 TOSHIHIRO ABE ET AL.
 
(1 + tanh(κ1 ψ))−1/2+1/ψ 1 1 3 1
= 8π F1 1; − , + ; 2; z1 , z2
(1 − tanh(κ2 ψ))3/2+1/ψ 2 ψ 2 ψ
(1 + tanh(κ1 ψ))−1/2+1/ψ 1
= 8π
(1 − tanh(κ2 ψ))3/2+1/ψ (1/2 + 1/ψ)(z1 − z2 )
  1/2+1/ψ 
1 − z1
× 1− .
1 − z2

In (p − 1)-dimensional Euclidean space, Rp−1 , a multivariate Pearson type

VII distribution has pdf
 −(p−1)/2−ν
Γ(ν + (p − 1)/2) x x
fp−1 (x) = (πA2 )−(p−1)/2 1 + 2 ,
Γ(ν) A

where x ∈ Rp−1 , A = 0, ν > 0. By an application of Proposition 6.1, the

corresponding pdf g(φ, θ1 , . . . , θp−2 ) on Spv is

g(φ, θ1 , . . . , θp−2 )
⎛  2 ⎞−(p−1)/2−ν
Γ(ν + (p − 1)/2) v sin θp−2
= (πA2 )−(p−1)/2 ⎝1 + ⎠ J.
Γ(ν) A(1 + cos θp−2 )

A special case of the Pearson type VII distribution, with ν = n/2 and A2 = n,
is the multivariate t-distribution with n degrees of freedom. When transformed
onto the sphere using inverse stereographic projection, the resulting distribution
has density

(7.2) g(φ, θ1 , . . . , θp−2 )

Γ((n + p − 1)/2)
= (πn)−(p−1)/2
Γ(n/2)
⎛  2 ⎞−(n+p−1)/2
1 v sin θp−2
× ⎝1 + ⎠ J.
n 1 + cos θp−2

Note that as n → ∞, this pdf converges to a density on Spv induced by the

(p − 1)-dimensional standard normal distribution, N (0, I), namely
 
1 v 2 sin2 θp−2
g(φ, θ1 , . . . , θp−2 ) = exp − J.
(2π)(p−1)/2 2(1 + cos θp−2 )2

Formally, p ≥ 3. However, if we set p = 2, we obtain the normal distribution on

the circle with density
 
v v 2 sin2 θ0
g(θ0 ) = exp − .
(2π)1/2 (1 + cos θ0 ) 2(1 + cos θ0 )2

Replacing θ0 by θ, this density is equivalent to that in (1.7), which is the density

of the circular normal induced by the modified inverse stereographic projection.
 INVERSE STEREOGRAPHIC PROJECTION 61

An alternative form for (7.2) is

g(φ, θ1 , . . . , θp−2 )
(1 + cos θp−2 )(n−(p−1))/2 
p−2
Γ((n + p − 1)/2)v p−1
= sini θi ,
Γ(n/2)(πn)(p−1)/2 (1 + u20 )(n+p−1)/2 (1 − w cos θp−2 )(n+p−1)/2 i=1
√
where u0 = v/ n and w = (u20 − 1)/(u20 + 1). Transforming the two cosine
components appearing in this last density using tanh(κ1 ψ) and tanh(κ2 ψ), re-
spectively, and adjusting the normalizing constant appropriately, we obtain the
density in (7.1). Thus, (7.1) can indeed be considered to be the p-dimensional
generalization of the density in (2.1).

Acknowledgements
The work of Dr. Shimizu was supported in part by The Ministry of Edu-
cation, Culture, Sports, Science and Technology of Japan under a Grant-in-Aid
of the 21st Century Center of Excellence (COE) for Integrative Mathematical
Sciences: Progress in Mathematics Motivated by Social and Natural Sciences.
Dr. Pewsey would like to express his most sincere gratitude to the Department
of Mathematics at Keio University for its warm hospitality and COE funding
during the research visit which led to the production of this paper.

References

Abramowitz, M. and Stegun, I. A. (Eds.) (1972). Handbook of Mathematical Functions, Dover,

New York.
Appell, P. and Kampé de Fériet, J. (1926). Fonctions Hypergéométriques et Hypersphériques:
Polynomes d’Hermite, Gauthier–Villars, Paris.
Gradshteyn, I. S. and Ryzhik, I. M. (2007). Table of Integrals, Series, and Products, seventh
ed , Academic Press, London.
Jammalamadaka, S. R. and SenGupta, A. (2001). Topics in Circular Statistics, World Scientific,
Singapore.
Jones, M. C. and Pewsey, A. (2005). A family of symmetric distributions on the circle, J. Amer.
Stat. Assoc., 100, 1422–1428.
Mardia, K. V. and Jupp, P. E. (1999). Directional Statistics, Wiley, Chichester.
Minh, D. L. P. and Farnum, N. R. (2003). Using bilinear transformations to induce probability
distributions, Commun. Stat.–Theory Meth., 32, 1–9.
Pewsey, A. (2002). Testing circular symmetry, Canadian J. Stat., 30, 591–600.
Shimizu, K. and Iida, K. (2002). Pearson type VII distributions on spheres, Commun. Stat.–
Theory Meth., 31, 513–526.
Watson, G. S. (1983). Statistics on Spheres, Wiley, New York.

View publication stats