Families Of Distributions On The Circle - A Review
1. Theory and Methods of Statistical Inference F.Rotolo
Families of distributions on the circle
A review
Federico Rotolo
federico.rotolo@stat.unipd.it
Department of Statistical Sciences
University of Padua
September 7, 2010
Families of distributions on the circle — A review 1/ 23
2. Theory and Methods of Statistical Inference F.Rotolo
Introduction
Families of Circular Distributions
The Jones & Pewsey distribution
The Generalized von Mises distribution
The Kato & Jones distribution
Two particular submodels
Comparison
Generality
Data modelling
Inferential aspects
Bibliography
Families of distributions on the circle — A review 2/ 23
3. Theory and Methods of Statistical Inference F.Rotolo
Introduction
Data on the circle are present in many applications, whenever
directional data are observed.
(wind direction, earthquake propagation, waves action on moving ships, etc.)
Families of distributions on the circle — A review 3/ 23
4. Theory and Methods of Statistical Inference F.Rotolo
Introduction
Data on the circle are present in many applications, whenever
directional data are observed.
(wind direction, earthquake propagation, waves action on moving ships, etc.)
Distributions on the real line are not suitable for direction, so new
models are needed.
[Jammalamadaka & SenGupta(2001), Mardia & Jupp(1999), Fisher(1993), Mardia(1972)]
Families of distributions on the circle — A review 3/ 23
5. Theory and Methods of Statistical Inference F.Rotolo
Introduction
Data on the circle are present in many applications, whenever
directional data are observed.
(wind direction, earthquake propagation, waves action on moving ships, etc.)
Distributions on the real line are not suitable for direction, so new
models are needed.
[Jammalamadaka & SenGupta(2001), Mardia & Jupp(1999), Fisher(1993), Mardia(1972)]
The most popular circular distributions are:
• von Mises vM(µ, κ)
• wrapped Cauchy wC(µ, ρ)
• Carthwright’s power-of-cosine Cpc(µ, ψ)
• cardioid ca(µ, ρ)
• circular Uniform cU(0; 2π)
Families of distributions on the circle — A review 3/ 23
6. Theory and Methods of Statistical Inference F.Rotolo
Introduction
0.6
0.5 An example
π/2
0.4
Density
π 0
0.3
0.2
3/2π
0.1
0.0
−3 −2 −1 0 1 2 3
Angle
vM(0.48π,1.8) (dash), wC(-0.45π,0.6) (dot),
Cpc(-0.16π,0.6) (long dash), ca(0.89π,0.2) (dot-dash).
Families of distributions on the circle — A review 4/ 23
7. Theory and Methods of Statistical Inference F.Rotolo
Introduction
These simple circular distributions are symmetric and unimodal,
so their flexibility is quite limited.
Families of distributions on the circle — A review 5/ 23
8. Theory and Methods of Statistical Inference F.Rotolo
Introduction
These simple circular distributions are symmetric and unimodal,
so their flexibility is quite limited.
⇓
Recently some more general families of circular distributions
have been proposed:
Families of distributions on the circle — A review 5/ 23
9. Theory and Methods of Statistical Inference F.Rotolo
Introduction
These simple circular distributions are symmetric and unimodal,
so their flexibility is quite limited.
⇓
Recently some more general families of circular distributions
have been proposed:
• Jones & Pewsey [Jones & Pewsey(2005)]
• Generalized von Mises [Gatto & Jammalamadaka(2007)]
• Kato & Jones [Kato & Jones(2010)]
Families of distributions on the circle — A review 5/ 23
10. Theory and Methods of Statistical Inference F.Rotolo
The Jones & Pewsey distribution
The first proposed family of circular distributions [Jones & Pewsey(2005)] is
the three-parameter Jones & Pewsey distribution JP(µ, κ, ψ)
Families of distributions on the circle — A review 6/ 23
11. Theory and Methods of Statistical Inference F.Rotolo
The Jones & Pewsey distribution
The first proposed family of circular distributions [Jones & Pewsey(2005)] is
the three-parameter Jones & Pewsey distribution JP(µ, κ, ψ)
with density
(cosh(κψ) + sinh(κψ) cos(θ − µ))1/ψ
fJP (θ) =
2πP1/ψ (cosh(κψ))
0 ≤ θ < 2π, 0 ≤ µ < 2π, κ ≥ 0, ψ ∈ R, and P1/ψ (·) is the associated Legendre function of the first kind
and order 0.
Families of distributions on the circle — A review 6/ 23
12. Theory and Methods of Statistical Inference F.Rotolo
The Jones & Pewsey distribution
The first proposed family of circular distributions [Jones & Pewsey(2005)] is
the three-parameter Jones & Pewsey distribution JP(µ, κ, ψ)
with density
(cosh(κψ) + sinh(κψ) cos(θ − µ))1/ψ
fJP (θ) =
2πP1/ψ (cosh(κψ))
0 ≤ θ < 2π, 0 ≤ µ < 2π, κ ≥ 0, ψ ∈ R, and P1/ψ (·) is the associated Legendre function of the first kind
and order 0.
All the vM, wC, ca, Cpc and cU distributions can be obtained
as special cases of it.
Families of distributions on the circle — A review 6/ 23
13. Theory and Methods of Statistical Inference F.Rotolo
The Jones & Pewsey distribution
The first proposed family of circular distributions [Jones & Pewsey(2005)] is
the three-parameter Jones & Pewsey distribution JP(µ, κ, ψ)
with density
(cosh(κψ) + sinh(κψ) cos(θ − µ))1/ψ
fJP (θ) =
2πP1/ψ (cosh(κψ))
0 ≤ θ < 2π, 0 ≤ µ < 2π, κ ≥ 0, ψ ∈ R, and P1/ψ (·) is the associated Legendre function of the first kind
and order 0.
All the vM, wC, ca, Cpc and cU distributions can be obtained
as special cases of it.
Two other distributions, the wrapped Normal [Stephens(1963)] and the
wrapped symmetric stable [Mardia(1972)], can be well approximated by
the JP model.
Families of distributions on the circle — A review 6/ 23
14. Theory and Methods of Statistical Inference F.Rotolo
The Jones & Pewsey distribution
Properties
The JP family is symmetric unimodal.
Families of distributions on the circle — A review 7/ 23
15. Theory and Methods of Statistical Inference F.Rotolo
The Jones & Pewsey distribution
Properties
The JP family is symmetric unimodal.
MLE:
ˆ ˆ
µ is asymptotically independent of (ψ, κ),
ˆ
ˆ ˆ
no reparametrization is available to reduce corr(ψ, κ).
Families of distributions on the circle — A review 7/ 23
16. Theory and Methods of Statistical Inference F.Rotolo
The Jones & Pewsey distribution
Properties
The JP family is symmetric unimodal.
MLE:
ˆ ˆ
µ is asymptotically independent of (ψ, κ),
ˆ
ˆ ˆ
no reparametrization is available to reduce corr(ψ, κ).
0.6
π/2
Density
0.4
π 0
0.2
3/2π
0.0
−3 −2 −1 0 1 2 3
Angle µ=4.1, κ=1.8, ψ=−0.6
Families of distributions on the circle — A review 7/ 23
17. Theory and Methods of Statistical Inference F.Rotolo
The Generalized von Mises distribution
A five-parameter class of distributions comprising the vM was
proposed by Maksimov in 1967.
An interesting subclass of it is the four-parameter Generalized
von Mises distribution GvM(µ1 , µ2 , κ1 , κ2 ) [Gatto & Jammalamadaka(2007)]
Families of distributions on the circle — A review 8/ 23
18. Theory and Methods of Statistical Inference F.Rotolo
The Generalized von Mises distribution
A five-parameter class of distributions comprising the vM was
proposed by Maksimov in 1967.
An interesting subclass of it is the four-parameter Generalized
von Mises distribution GvM(µ1 , µ2 , κ1 , κ2 ) [Gatto & Jammalamadaka(2007)]
with density
1
fGvM (θ) = exp{κ1 cos(θ − µ1 ) + κ2 cos 2(θ − µ2 )}
2πG0 (δ, κ1 , κ2 )
0 ≤ θ < 2π, 0 ≤ µ1 < 2π, 0 ≤ µ2 < π, κ1 , κ2 ≥ 0, δ = (µ1 − µ2 )modπ, G0 (δ, κ1 , κ2 ) is the normalizing
constant.
Families of distributions on the circle — A review 8/ 23
19. Theory and Methods of Statistical Inference F.Rotolo
The Generalized von Mises distribution
Properties
The GvM family can be asymmetric and it can have one or two
maxima.
Families of distributions on the circle — A review 9/ 23
20. Theory and Methods of Statistical Inference F.Rotolo
The Generalized von Mises distribution
Properties
The GvM family can be asymmetric and it can have one or two
maxima.
The skewness and the maxima location are mainly controlled by
µ1 and µ2 ,
Families of distributions on the circle — A review 9/ 23
21. Theory and Methods of Statistical Inference F.Rotolo
The Generalized von Mises distribution
Properties
The GvM family can be asymmetric and it can have one or two
maxima.
The skewness and the maxima location are mainly controlled by
µ1 and µ2 , the kurtosis mostly by κ1 and κ2 .
Families of distributions on the circle — A review 9/ 23
22. Theory and Methods of Statistical Inference F.Rotolo
The Generalized von Mises distribution
Properties
The GvM family can be asymmetric and it can have one or two
maxima.
The skewness and the maxima location are mainly controlled by
µ1 and µ2 , the kurtosis mostly by κ1 and κ2 .
1.0
0.8
π/2
0.6
Density
π 0
0.4
3/2π
0.2
0.0
−3 −2 −1 0 1 2 3
Angle µ1=0.8π, µ2=π, κ1=4.8, κ2=4.1
Families of distributions on the circle — A review 9/ 23
23. Theory and Methods of Statistical Inference F.Rotolo
The Generalized von Mises distribution
3.0
3.0
3.0
2.0
2.0
2.0
κ1
κ2
κ2
1.0
1.0
1.0
0.0
0.0
0.0
0 1 2 3 4 5 6 0 1 2 3 4 5 6 0.0 1.0 2.0 3.0
µ2 µ2 κ1
Unimodality(white)/bimodality(green) of the GvM density with µ1 = 0.
For each variable the value chosen for the graph where it is absent is shown by the grey lines.
Families of distributions on the circle — A review 10/ 23
24. Theory and Methods of Statistical Inference F.Rotolo
The Generalized von Mises distribution
3.0
3.0
3.0
2.0
2.0
2.0
κ1
κ2
κ2
1.0
1.0
1.0
0.0
0.0
0.0
0 1 2 3 4 5 6 0 1 2 3 4 5 6 0.0 1.0 2.0 3.0
µ2 µ2 κ1
Unimodality(white)/bimodality(green) of the GvM density with µ1 = 0.
For each variable the value chosen for the graph where it is absent is shown by the grey lines.
! A big portion of the parameter space gives a bimodal distribution.
Families of distributions on the circle — A review 10/ 23
25. Theory and Methods of Statistical Inference F.Rotolo
The Generalized von Mises distribution
3.0
3.0
3.0
2.0
2.0
2.0
κ1
κ2
κ2
1.0
1.0
1.0
0.0
0.0
0.0
0 1 2 3 4 5 6 0 1 2 3 4 5 6 0.0 1.0 2.0 3.0
µ2 µ2 κ1
Unimodality(white)/bimodality(green) of the GvM density with µ1 = 0.
For each variable the value chosen for the graph where it is absent is shown by the grey lines.
! A big portion of the parameter space gives a bimodal distribution.
In general there is no reason to expect bimodality
→ maybe misleading results.
Families of distributions on the circle — A review 10/ 23
26. Theory and Methods of Statistical Inference F.Rotolo
The Generalized von Mises distribution
3.0
3.0
3.0
2.0
2.0
2.0
κ1
κ2
κ2
1.0
1.0
1.0
0.0
0.0
0.0
0 1 2 3 4 5 6 0 1 2 3 4 5 6 0.0 1.0 2.0 3.0
µ2 µ2 κ1
Unimodality(white)/bimodality(green) of the GvM density with µ1 = 0.
For each variable the value chosen for the graph where it is absent is shown by the grey lines.
! A big portion of the parameter space gives a bimodal distribution.
In general there is no reason to expect bimodality
→ maybe misleading results.
When bimodality is expected (e.g. with two groups of data)
→ good model: simpler inference w.r.t. mixture models.
Families of distributions on the circle — A review 10/ 23
27. Theory and Methods of Statistical Inference F.Rotolo
The Generalized von Mises distribution
Membership of the Exponential Family
The most interesting property of the GvM model is that the
reparametrization
λ = (κ1 cos µ1 , κ1 sin µ1 , κ2 cos 2µ2 , κ2 sin 2µ2 )T
Families of distributions on the circle — A review 11/ 23
28. Theory and Methods of Statistical Inference F.Rotolo
The Generalized von Mises distribution
Membership of the Exponential Family
The most interesting property of the GvM model is that the
reparametrization
λ = (κ1 cos µ1 , κ1 sin µ1 , κ2 cos 2µ2 , κ2 sin 2µ2 )T
makes it possible to express the density as
fGvM (θ | λ) = exp{λT t(θ) − k(θ)},
a four-parameter exponential family, indexed by λ ∈ [−1, 1]4 .
t(θ) = (cos θ, sin θ, cos 2θ, sin 2θ)T , k(λ) = log(2π) + log G0 (δ, λ(1) , λ(2) ), λ(1) = (λ1 , λ2 )T ,
(2) T (1) (2)
λ = (λ3 , λ4 ) and δ = (arg λ − arg λ /2)modπ.
Families of distributions on the circle — A review 11/ 23
29. Theory and Methods of Statistical Inference F.Rotolo
The Generalized von Mises distribution
Membership of the Exponential Family
The most interesting property of the GvM model is that the
reparametrization
λ = (κ1 cos µ1 , κ1 sin µ1 , κ2 cos 2µ2 , κ2 sin 2µ2 )T
makes it possible to express the density as
fGvM (θ | λ) = exp{λT t(θ) − k(θ)},
a four-parameter exponential family, indexed by λ ∈ [−1, 1]4 .
t(θ) = (cos θ, sin θ, cos 2θ, sin 2θ)T , k(λ) = log(2π) + log G0 (δ, λ(1) , λ(2) ), λ(1) = (λ1 , λ2 )T ,
(2) T (1) (2)
λ = (λ3 , λ4 ) and δ = (arg λ − arg λ /2)modπ.
Thus it has many good inferential properties, like the uniqueness
of the MLEs, when they exist, and the asymptotic normality of
the estimator.
Families of distributions on the circle — A review 11/ 23
30. Theory and Methods of Statistical Inference F.Rotolo
The Kato & Jones distribution
The four-parameter Kato & jones distribution KJ is obtained by
applying a M¨bius transformation to a vM-distributed random
o
variable [Kato & Jones(2010)] .
Families of distributions on the circle — A review 12/ 23
31. Theory and Methods of Statistical Inference F.Rotolo
The Kato & Jones distribution
The four-parameter Kato & jones distribution KJ is obtained by
applying a M¨bius transformation to a vM-distributed random
o
variable [Kato & Jones(2010)] .
The M¨bius transformation is a (closed under composition)
o
circle-to-circle function M¨µ,ν,r : Ξ → Θ given by
o
e iΞ + re iν
e iΘ = e iµ ,
re i(Ξ−ν) + 1
with 0 ≤ µ, ν < 2π and 0 ≤ r < 1.
Families of distributions on the circle — A review 12/ 23
32. Theory and Methods of Statistical Inference F.Rotolo
The Kato & Jones distribution
The four-parameter Kato & jones distribution KJ is obtained by
applying a M¨bius transformation to a vM-distributed random
o
variable [Kato & Jones(2010)] .
The M¨bius transformation is a (closed under composition)
o
circle-to-circle function M¨µ,ν,r : Ξ → Θ given by
o
e iΞ + re iν
e iΘ = e iµ ,
re i(Ξ−ν) + 1
with 0 ≤ µ, ν < 2π and 0 ≤ r < 1.
If Ξ ∼ vM(0, κ), then Θ = M¨µ,ν,r (Ξ) ∼ KJ(µ, ν, r , κ) has density
o
κ{ξ cos(θ−η)−2r cos ν}
1 − r 2 exp 1+r 2 −2r cos(θ−γ)
fKJ (θ) = 2 − 2r cos(θ − γ)
,
2πI0 (κ) 1 + r
r 4 + 2r 2 cos(2ν) + 1, η = µ + arg[r 2 {cos(2ν) + i sin(2ν)} + 1], γ = µ + ν.
p
0 ≤ θ < 2π, ξ =
Families of distributions on the circle — A review 12/ 23
33. Theory and Methods of Statistical Inference F.Rotolo
The Kato & Jones distribution
Properties
The KJ distribution can be symmetric or asymmetric.
Families of distributions on the circle — A review 13/ 23
34. Theory and Methods of Statistical Inference F.Rotolo
The Kato & Jones distribution
Properties
The KJ distribution can be symmetric or asymmetric.
It includes the vM, the wC and the cU models as special cases.
Families of distributions on the circle — A review 13/ 23
35. Theory and Methods of Statistical Inference F.Rotolo
The Kato & Jones distribution
Properties
The KJ distribution can be symmetric or asymmetric.
It includes the vM, the wC and the cU models as special cases.
It can also be either unimodal or bimodal, but conditions for
unimodality are not straigthforward.
Families of distributions on the circle — A review 13/ 23
36. Theory and Methods of Statistical Inference F.Rotolo
The Kato & Jones distribution
Properties
The KJ distribution can be symmetric or asymmetric.
It includes the vM, the wC and the cU models as special cases.
It can also be either unimodal or bimodal, but conditions for
unimodality are not straigthforward.
π/2
0.3
Density
0.2
π 0
0.1
3/2π
0.0
−3 −2 −1 0 1 2 3
Angle µ=0.3π, ν=0.95π, r=0.7, κ=2.3
Families of distributions on the circle — A review 13/ 23
37. Theory and Methods of Statistical Inference F.Rotolo
The Kato & Jones distribution
3.0
3.0
0.8
2.0
2.0
κ
κ
r
0.4
1.0
1.0
0.0
0.0
0.0
0 1 2 3 4 5 6 0 1 2 3 4 5 6 0 1 2 3 4 5 6
ν ν r
Unimodality(white)/bimodality(yellow) of the KJ density.
For each variable the value chosen for the graph where it is absent is shown by the grey lines.
Families of distributions on the circle — A review 14/ 23
38. Theory and Methods of Statistical Inference F.Rotolo
The Kato & Jones distribution
3.0
3.0
0.8
2.0
2.0
κ
κ
r
0.4
1.0
1.0
0.0
0.0
0.0
0 1 2 3 4 5 6 0 1 2 3 4 5 6 0 1 2 3 4 5 6
ν ν r
Unimodality(white)/bimodality(yellow) of the KJ density.
For each variable the value chosen for the graph where it is absent is shown by the grey lines.
! The portion of the parameter space originating a bimodal
distribution is appreciably smaller than in the GvM case
Families of distributions on the circle — A review 14/ 23
39. Theory and Methods of Statistical Inference F.Rotolo
The Kato & Jones distribution
3.0
3.0
0.8
2.0
2.0
κ
κ
r
0.4
1.0
1.0
0.0
0.0
0.0
0 1 2 3 4 5 6 0 1 2 3 4 5 6 0 1 2 3 4 5 6
ν ν r
Unimodality(white)/bimodality(yellow) of the KJ density.
For each variable the value chosen for the graph where it is absent is shown by the grey lines.
! The portion of the parameter space originating a bimodal
distribution is appreciably smaller than in the GvM case
→ better for general applications.
Families of distributions on the circle — A review 14/ 23
40. Theory and Methods of Statistical Inference F.Rotolo
The Kato & Jones distribution
Circle-circle regression
The most interesting property of the KJ distribution is its role in
circular regression.
Families of distributions on the circle — A review 15/ 23
41. Theory and Methods of Statistical Inference F.Rotolo
The Kato & Jones distribution
Circle-circle regression
The most interesting property of the KJ distribution is its role in
circular regression.
The considered regression model [Downs & Mardia(2002)] is
xj + β1
Yj = β0 ¯ εj , xj ∈ Ω,
β1 xj + 1
with β0 ∈ Ω, β1 ∈ C and {arg(εj )} independent angular errors.
Families of distributions on the circle — A review 15/ 23
42. Theory and Methods of Statistical Inference F.Rotolo
The Kato & Jones distribution
Circle-circle regression
The most interesting property of the KJ distribution is its role in
circular regression.
The considered regression model [Downs & Mardia(2002)] is
xj + β1
Yj = β0 ¯ εj , xj ∈ Ω,
β1 xj + 1
with β0 ∈ Ω, β1 ∈ C and {arg(εj )} independent angular errors.
The use of the KJ distribution for circular errors is a general
extention of the model with vM and the wC distributions, in use
untill now [Downs & Mardia(2002), Kato et al.(2008)].
Families of distributions on the circle — A review 15/ 23
43. Theory and Methods of Statistical Inference F.Rotolo
The Kato & Jones distribution
Circle-circle regression
The most interesting property of the KJ distribution is its role in
circular regression.
The considered regression model [Downs & Mardia(2002)] is
xj + β1
Yj = β0 ¯ εj , xj ∈ Ω,
β1 xj + 1
with β0 ∈ Ω, β1 ∈ C and {arg(εj )} independent angular errors.
The use of the KJ distribution for circular errors is a general
extention of the model with vM and the wC distributions, in use
untill now [Downs & Mardia(2002), Kato et al.(2008)].
Since both the the regression curve and the KJ distribution are
expressed in terms of M¨bius transformations this framework
o
seems very promising.
Families of distributions on the circle — A review 15/ 23
44. Theory and Methods of Statistical Inference F.Rotolo
The Kato & Jones distribution
Two particular submodels
ν = 0 Symmetric and unimodal
π
ν=± Asymmetric and uni/bi-modal
2
Families of distributions on the circle — A review 16/ 23
45. Theory and Methods of Statistical Inference F.Rotolo
The Kato & Jones distribution
Two particular submodels
ν = 0 Symmetric and unimodal
The kurtosis varies, the skewness being fixed
π
ν=± Asymmetric and uni/bi-modal
2
Families of distributions on the circle — A review 16/ 23
46. Theory and Methods of Statistical Inference F.Rotolo
The Kato & Jones distribution
Two particular submodels
ν = 0 Symmetric and unimodal
The kurtosis varies, the skewness being fixed
It includes the vM, wC and cU distributions
π
ν=± Asymmetric and uni/bi-modal
2
Families of distributions on the circle — A review 16/ 23
47. Theory and Methods of Statistical Inference F.Rotolo
The Kato & Jones distribution
Two particular submodels
ν = 0 Symmetric and unimodal
The kurtosis varies, the skewness being fixed
It includes the vM, wC and cU distributions
As for the JP distribution, µ is asymptotically independent of
ˆ
ˆ and κ
r ˆ
π
ν=± Asymmetric and uni/bi-modal
2
Families of distributions on the circle — A review 16/ 23
48. Theory and Methods of Statistical Inference F.Rotolo
The Kato & Jones distribution
Two particular submodels
ν = 0 Symmetric and unimodal
The kurtosis varies, the skewness being fixed
It includes the vM, wC and cU distributions
As for the JP distribution, µ is asymptotically independent of
ˆ
ˆ and κ
r ˆ
A reparametrization (r , κ) → (s(r , κ), κ) is proposed which
reduces both the asymptotic correlation between ˆ and κ and
s ˆ
the asymptotic variance of κ.
ˆ
π
ν=± Asymmetric and uni/bi-modal
2
Families of distributions on the circle — A review 16/ 23
49. Theory and Methods of Statistical Inference F.Rotolo
The Kato & Jones distribution
Two particular submodels
ν = 0 Symmetric and unimodal
The kurtosis varies, the skewness being fixed
It includes the vM, wC and cU distributions
As for the JP distribution, µ is asymptotically independent of
ˆ
ˆ and κ
r ˆ
A reparametrization (r , κ) → (s(r , κ), κ) is proposed which
reduces both the asymptotic correlation between ˆ and κ and
s ˆ
the asymptotic variance of κ.
ˆ
π
ν=± Asymmetric and uni/bi-modal
2
The skewness varies, the kurtosis being fixed
Families of distributions on the circle — A review 16/ 23
50. Theory and Methods of Statistical Inference F.Rotolo
The Kato & Jones distribution
Two particular submodels
ν = 0 Symmetric and unimodal
The kurtosis varies, the skewness being fixed
It includes the vM, wC and cU distributions
As for the JP distribution, µ is asymptotically independent of
ˆ
ˆ and κ
r ˆ
A reparametrization (r , κ) → (s(r , κ), κ) is proposed which
reduces both the asymptotic correlation between ˆ and κ and
s ˆ
the asymptotic variance of κ.
ˆ
π
ν=± Asymmetric and uni/bi-modal
2
The skewness varies, the kurtosis being fixed
Good performances in modelling real data
Families of distributions on the circle — A review 16/ 23
51. Theory and Methods of Statistical Inference F.Rotolo
Comparison
Generality
Generality: in terms of known densities comprised as special cases.
Families of distributions on the circle — A review 17/ 23
52. Theory and Methods of Statistical Inference F.Rotolo
Comparison
Generality
Generality: in terms of known densities comprised as special cases.
Special cases JP GvM KJ
von Mises • • •
circular Uniform • • •
wrapped Cauchy • ◦ •
cardioid • ◦ ◦
Cartwright’s power-of-cosine • ◦ ◦
•=Yes; ◦=No
Families of distributions on the circle — A review 17/ 23
53. Theory and Methods of Statistical Inference F.Rotolo
Comparison
Generality
Generality: in terms of known densities comprised as special cases.
Special cases JP GvM KJ
von Mises • • •
circular Uniform • • •
wrapped Cauchy • ◦ •
cardioid • ◦ ◦
Cartwright’s power-of-cosine • ◦ ◦
•=Yes; ◦=No
• The JP model is the most general
Families of distributions on the circle — A review 17/ 23
54. Theory and Methods of Statistical Inference F.Rotolo
Comparison
Generality
Generality: in terms of known densities comprised as special cases.
Special cases JP GvM KJ
von Mises • • •
circular Uniform • • •
wrapped Cauchy • ◦ •
cardioid • ◦ ◦
Cartwright’s power-of-cosine • ◦ ◦
•=Yes; ◦=No
• The JP model is the most general
• The vM distribution, which is the most important and widely
used one, belongs to all of the three models
Families of distributions on the circle — A review 17/ 23
55. Theory and Methods of Statistical Inference F.Rotolo
Comparison
Generality
Generality: in terms of known densities comprised as special cases.
Special cases JP GvM KJ
von Mises • • •
circular Uniform • • •
wrapped Cauchy • ◦ •
cardioid • ◦ ◦
Cartwright’s power-of-cosine • ◦ ◦
•=Yes; ◦=No
• The JP model is the most general
• The vM distribution, which is the most important and widely
used one, belongs to all of the three models
• The poorest family, in this sense, is the GvM model
Families of distributions on the circle — A review 17/ 23
56. Theory and Methods of Statistical Inference F.Rotolo
Comparison
Data modelling
Very few empirical examples are available.
[Further work would be useful in future in this sense...]
JP [Jones & Pewsey(2005)]
Families of distributions on the circle — A review 18/ 23
57. Theory and Methods of Statistical Inference F.Rotolo
Comparison
Data modelling
Very few empirical examples are available.
[Further work would be useful in future in this sense...]
JP [Jones & Pewsey(2005)]
with symmetric data, JP distribution performs significantly
better than the vM, ca and wC models
Families of distributions on the circle — A review 18/ 23
58. Theory and Methods of Statistical Inference F.Rotolo
Comparison
Data modelling
Very few empirical examples are available.
[Further work would be useful in future in this sense...]
JP [Jones & Pewsey(2005)]
with symmetric data, JP distribution performs significantly
better than the vM, ca and wC models
its advantage is no more significant in presence of heavy
tails, requiring a mixture model with a cU distribution
Families of distributions on the circle — A review 18/ 23
59. Theory and Methods of Statistical Inference F.Rotolo
Comparison
Data modelling
Very few empirical examples are available.
[Further work would be useful in future in this sense...]
JP [Jones & Pewsey(2005)]
with symmetric data, JP distribution performs significantly
better than the vM, ca and wC models
its advantage is no more significant in presence of heavy
tails, requiring a mixture model with a cU distribution
KJ [Kato & Jones(2010)]
Families of distributions on the circle — A review 18/ 23
60. Theory and Methods of Statistical Inference F.Rotolo
Comparison
Data modelling
Very few empirical examples are available.
[Further work would be useful in future in this sense...]
JP [Jones & Pewsey(2005)]
with symmetric data, JP distribution performs significantly
better than the vM, ca and wC models
its advantage is no more significant in presence of heavy
tails, requiring a mixture model with a cU distribution
KJ [Kato & Jones(2010)]
with asymmetric data, the GvM model fits better than
simpler distributions and the KJ model and its asymmetric
submodel are even better.
Families of distributions on the circle — A review 18/ 23
61. Theory and Methods of Statistical Inference F.Rotolo
Comparison
Data modelling
Very few empirical examples are available.
[Further work would be useful in future in this sense...]
JP [Jones & Pewsey(2005)]
with symmetric data, JP distribution performs significantly
better than the vM, ca and wC models
its advantage is no more significant in presence of heavy
tails, requiring a mixture model with a cU distribution
KJ [Kato & Jones(2010)]
with asymmetric data, the GvM model fits better than
simpler distributions and the KJ model and its asymmetric
submodel are even better.
circular-circular regression: improvement in performances for
the KJ model w.r.t. its submodels, but no comparison with
other distributions
Families of distributions on the circle — A review 18/ 23
62. Theory and Methods of Statistical Inference F.Rotolo
Comparison
Inferential aspects
• The GvM distribution belongs to the exponential family
⇒ if the MLE exists, then it is unique
⇒ even numerical solutions are very reliable
Families of distributions on the circle — A review 19/ 23
63. Theory and Methods of Statistical Inference F.Rotolo
Comparison
Inferential aspects
• The GvM distribution belongs to the exponential family
⇒ if the MLE exists, then it is unique
⇒ even numerical solutions are very reliable
Explicit estimates exists for some parameters in some cases
Families of distributions on the circle — A review 19/ 23
64. Theory and Methods of Statistical Inference F.Rotolo
Comparison
Inferential aspects
• The GvM distribution belongs to the exponential family
⇒ if the MLE exists, then it is unique
⇒ even numerical solutions are very reliable
Explicit estimates exists for some parameters in some cases
• The two other models have no particularly good properties
in general
Families of distributions on the circle — A review 19/ 23
65. Theory and Methods of Statistical Inference F.Rotolo
Comparison
Inferential aspects
• The GvM distribution belongs to the exponential family
⇒ if the MLE exists, then it is unique
⇒ even numerical solutions are very reliable
Explicit estimates exists for some parameters in some cases
• The two other models have no particularly good properties
in general
• The KJ distribution has a slight advantage in the
reparametrization (r , κ) → (s(r , κ), κ) useful in general to
reduce both the asymptotic correlation with and the
asymptotic variance of κ ˆ
Families of distributions on the circle — A review 19/ 23
66. Theory and Methods of Statistical Inference F.Rotolo
Bibliography I
Downs, T. D. & Mardia, K. V. (2002).
Circular regression.
Biometrika 89, 683–697.
Fisher, N. I. (1993).
Statistical Analysis of Circular Data.
Cambridge: Cambridge University Press.
Gatto, R. & Jammalamadaka, S. R. (2007).
The generalized von Mises distribution.
Statistical Methodology 4, 341–353.
Jammalamadaka, S. R. & SenGupta, A. (2001).
Topics in circular statistics.
Singapore: World Scientific.
Families of distributions on the circle — A review 20/ 23
67. Theory and Methods of Statistical Inference F.Rotolo
Bibliography II
Jones, M. C. & Pewsey, A. (2005).
A family of simmetric distributions on the circle.
J. Am. Statist. Assoc. 100, 1422–1428.
Kato, S. & Jones, M. C. (2010).
A family of distributions on the circle with links to, and
applications arising from, M¨bius transformation.
o
J. Am. Statist. Assoc. 105, 249–262.
Kato, S., Shimizu, K. & Shieh, G. S. (2008).
A circular-circular regression model.
Statistica Sinica 18, 633–645.
Families of distributions on the circle — A review 21/ 23
68. Theory and Methods of Statistical Inference F.Rotolo
Bibliography III
Maksimov, V. M. (1967).
Necessary and sufficient conditions for the family of shifts of
probability distributions on the continuous bicompact groups.
Theoria Verojatna 12, 307–321.
Mardia, K. V. (1972).
Statistics of directional data.
London: Academic Press.
Mardia, K. V. & Jupp, P. E. (1999).
Directional statistics.
Chichester: Wiley.
Stephens, M. A. (1963).
Random walk on a circle.
Biometrika 50, 385–390.
Families of distributions on the circle — A review 22/ 23