Open Journal of Statistics, 2012, 2, 106-114 http://dx.doi.org/10.4236/ojs.2012.21011 Published Online January 2012 (http://www.SciRP.org/journal/ojs)
Wrapped Skew Laplace Distribution on Integers: A New Probability Model for Circular Data
K. Jayakumar1, Sophy Jacob2 1University of Calicut, Kerala, India 2MES Asmabi College, Kerala, India Email: {jkumar19, sophyjacob}@rediffmail.com
Received September 19, 2011; revised October 21, 2011; accepted October 31, 2011
ABSTRACT In this paper we propose a new family of circular distributions, obtained by wrapping discrete skew Laplace distribution on Z = 0, ±1, ±2, around a unit circle. In contrast with many wrapped distributions, here closed form expressions exist for the probability density function, the distribution function and the characteristic function. The properties of this new family of distribution are studied.
Keywords: Circular Distributions; Trigonometric Moments; Wrapped Laplace Distribution
1. Introduction son, wrapped Cauchy are discussed in [9]. We give a brief description of circular distribution in Section 2. In Circular data arise in various ways. Two of the most com- Section 3 we introduce and study Wrapped Discrete mon correspond to circular measuring instruments, the Skew Laplace Distribution. Section 4 deals with the es- compass and the clock. Data measured by compass usu- timation of the parameters using the method of moments. ally include wind directions, the direction and orienta- tions of birds and animals, ocean current directions, and 2. Circular Distributions orientation of geological phenomena like rock cores and fractures. Data measured by clock includes times of arri- A circular distribution is a probability distribution whose val of patients at a hospital emergency room, incidences total probability is concentrated on the circumference of of a disease throughout the year, where the calendar is a circle of unit radius. Since each point on the circum- regarded as a one-year clock. Circular or directional data ference represents a direction, it is a way of assigning also arise in many scientific fields, such as Biology, Ge- probabilities to different directions or defining a direc- ology, Meteorology, Physics, Psychology, Medicine and tional distribution. The range of a circular random vari- Astronomy [1]. able Θ measured in radians, may be taken to be 0, 2π Study on directional data can be dated back to the 18th orπ, π . century. In 1734 Daniel Bernoulli proposed to use the Circular distributions are of two types: they may be resultant length of normal vectors to test for uniformity discrete - assigning probability masses only to a count- of unit vectors on the sphere [2]. In 1918 von Mises in- able number of directions, or may be absolutely continuous. troduced a distribution on the circle by using characteri- In the latter case, the probability density function f θ zation analogous to the Gauss characterization of the exists and has the following basic properties. normal distribution on a line [2]. Later, interest was re- 1) f θ 0 newed in spherical and circular data by [3-5]. 2π Circular distributions play an important role in model- 2) f θθd1 ing directional data which arise in various fields. In re- 0 3) f θ f θ 2πk , for any integer k. That is cent years, several new unimodal circular distributions f θ is periodic with period 2π . capable of modeling symmetry as well as asymmetry have been proposed. These include, the wrapped versions Wrapped Distributions of skew normal [6], exponential [7] and Laplace [8]. Wrapped distributions provide a rich and useful class One of the common methods to analyze circular data is of models for circular data. known as wrapping approach [10]. In this approach, The special cases of the wrapped normal, wrapped Pois- given a known distribution on the real line, we wrap it
Copyright © 2012 SciRes. OJS K. JAYAKUMAR ET AL. 107 around the circumference of a circle with unit radius. ||x Technically this means that if X is a random variable on 1 e , &x 0 fx (6) 2 ||x the real line with distribution function F x , the ran- 1 e , &x 0 dom variable X w of the wrapped distribution is defined by where, 0 , are inserted into Equation (5), the prob-
XXw mod 2π (1) ability mass function of the resulting discrete distribution takes an explicit form in terms of the parameters p* = and the distribution function of X w is given by 1 e and q* e . Fw θ F θ 2πkF 2πk , k (2) Definition 3.1 A random variable Y has a discrete * k 0, 1, 2, skew Laplace distribution with parameters p (0,1) and q* (0,1) denoted by DSL pq**, , if By this approach, we are accumulating probability over all the overlapping points x θθ,2 π,4θπ, fk PYk So if g θ represents a circular density and f x the *** 11pq pk,0,1,2, (7) density of the real valued random variable, we have ** 1 pq qk* , 0,1,2, g θ f θ 2πk , (3) k The characteristic function of Y is given by 02θ π 1(1)pq** By this technique, both discrete and continuous ttR, (8) wrapped distributions may be constructed. In particular, 1(1)pe**it qeit if X has a distribution concentrated on the points k In this paper, we study the probability distribution ob- x , k 0, 1, 2,and m is an integer, the prob- 2πm tained by wrapping discrete skew Laplace distribution on Z 0,,1, 2 around a unit circle. ability function of X w is given by As we know, reduction modulo 2π wraps the line onto 2πr Pr X pr km the circle, reduction modulo 2πm (if m is a positive in- w th m k (4) teger) wraps the integers onto the group of m root of 1, r 0,1, , m 1 regarded as a subgroup of the circle. That is, if X is a random variable on the integers, then Θ, defined by where “p” is the probability function of the random vari- able X. 2πX (mod2πm) 2πr 3. Wrapped Discrete Skew Laplace is a random variable on the lattice , rm0,1, , 1 Distribution m on the circle. The probability function of Θ is given by 3.1. Discrete Skew Laplace Distribution Equation (4). In particular, if X has a discrete skew Laplace distribu- Discrete Laplace distribution was introduced by [11] tion with parameters p* and q* , then the probability following [12], who defined a discrete analogue of the 2πr normal distribution. The probability mass function of a distribution of the wrapped random variable is general Discrete Normal random variable Y can be writ- m ten in the form given by fk 2πr PYk , PprkmrmΘ , 0,1, , 1 m k f j . (5) rkm j 11pqq k 0, 1, 2, k 1 pq
11pq where “f” is the probability density function of a normal pr distribution with mean µ and variance 2 [13]. 1 pq Using Equation (5), for any continuous random vari- 11)pqp( rkm able X on R, we can define a discrete random variable Y 1 pq that has integer support on Z. When the skew Laplace k density where pp * mod 2π m and
Copyright © 2012 SciRes. OJS 108 K. JAYAKUMAR ET AL.
qq * mod 2π m rmpq0,1, , 1 and , 0,1 and we denote it by 1 1(1)pqrkm rr()km WDSL pqm, , . qpp 1 pq kk 1 Following are the graphs of wrapped discrete skew 11pq Laplace distribution for various values of κ, σ and m. In 11ppmrpm1 q m (9) Figure 1, the graph of the pdf of wrapped discrete skew 1 pq Laplace distribution for κ = 0.25, σ = 1 and for m = 5, 10, pqmr 1 m ,r0,1, , m 1 20, 30, 40, 50 and 100 are given. In Figure 2, the graph of the pdf of wrapped discrete mr m r m 1(1)pqqp11pqskew Laplace distribution for κ = 0.5, σ = 1 and for m = 5, 1 pq 11pqmm10, 20, 30, 40, 50 and 100 are given. The graph of the pdf of wrapped discrete skew Laplace distribution for κ = for rm0,1 , , 1 and pq,0,1 0.25, σ = 1 and for m = 5, 10, 20, 30, 40, 50 and 100 are Again, we have given in Figure 3.
m1 pw θ 1 3.2. Special Cases r 0 When either “p” or “q” converges to zero, we obtain the Hence P represents a probability distribution. w following two special cases: ~WDSL (,0,) p m with Definition 3.2 An angular random variable “Θ” is p 0,1 is a wrapped geometric distribution with said to follow wrapped skew Laplace distribution on in- probability mass function tegers with parameters p, q and m if its probability mass 1 ppr function is given by 2πr (11) PrmΘ m , 0,1, , 1. mr m r m m 1 p 11pqqp11pq pw θ (10) while ~(0,,)WDSL q m with q 0,1 is a wrapped 1 pq 11pqmm geometric distribution with probability mass function
Figure 1. Wrapped discrete skewed Laplace distribution for σ = 1 , κ = 0.25 and for different values of “m”.
Copyright © 2012 SciRes. OJS K. JAYAKUMAR ET AL. 109
Figure 2. Wrapped discrete skewed Laplace distribution for σ = 1 , κ = 0.25 and for different values of “m”.
Figure 3. Wrapped discrete skewed Laplace distribution for σ = 1 , κ = 0.25 and for different values of “m”.
Copyright © 2012 SciRes. OJS 110 K. JAYAKUMAR ET AL.
r 2πr 1 qq mr m r m k qppq11 PΘ m ,0r ,1,, m 1. (12) 1(1)pq mq1 F θ mm r0 1 pq 11pq when p = q, we have (14) mr r mk m m k1 2πr 1(pp p ) qppqqp1 1 (1 )(1 )(1 ) PΘ m , mp1(1)p (13) 11(1)pq pmm q rm0,1, , 1 rm 0,1, , 1 . which is the probability mass function of wrapped dis- crete Laplace distribution. 3.4. Probability Generating Function and Characteristic Function of WDSL pqm, , 3.3. Distribution Function of WDSL pqm, , The probability generating function of WDSL pqm,, The distribution function, F θ is given by is given by 11pq m1 s Psqmmr1( p )11p m q mmmr p (1)() q sp mm 11pq p 1q r0 q (15) mmm m m 11pq 1pqsq11sp q 11pq pmm1q qs1 sp Also, we have P11 . where nm 0 (mod ) when s einm2π we have If t is the characteristic function of a linear ran- dom variable X, then the characteristic function of the in2π m 1(1pq) wrapped random variable, X , is n , for Pe w 1 pe1in2π mn qei2π m n 0, 1, 2, [2]. Thus for the wrapped discrete skew Laplace distribution, we have
in nEe , for n 0, 1, 2, in2πr Er em , for 0,1, ,m 1 (16) m1 11pq in2π r qpppqmr1 m r 1 m 1 m pqmr(1 m ) e m mm r0 111pq p q On simplification it reduces to 11pq n , n 0, 1, 2, ,n 0 (mod m ) (17) 1epqinm2π 1einm2π
Again, we have nn2π 2π 111pq pqpq cos ipq sin 11pq mm n in2π in2π 22 nni mm nn2π 2π 1epq e pq 1cossinpq p q p q mm where and n2π n2π 11pq1 pqpq cos 1(1)pqpq sin m m n 22 n 2 2 nn2π 2π nn2π 2π 1()cos()sinpq p q p q 1pq (p q )cos (p q )sin( ) mm mm (18) (19)
Copyright © 2012 SciRes. OJS K. JAYAKUMAR ET AL. 111
Proposition 3.1 If Θ ~WDSL(,,)pqm then Θd ΘΘ1 2 11 pq n where Θ and Θ are two independently distributed n j inm2π inm2π 1 2 1epq 1e wrapped g eometric random variables with probability 11 mass functions inm2π inm2π r r 1ep 1eq 2πr pp1 2πr qq1 , PΘ1 = m PΘ2 = m 11pp11qq m 11p m 11q 11 Proof. We have, ABeeinm2π CDinm2π 11pq n 1 1epqinm2π 1einm2π where A, B, C, D > 0, A – B = C – D = 1 and A , 1 p
nm0, 1, 2, , n 0 (mod ) p 1 q B ,,C D and 1, pqq 11 1p1qTherefore, nn , 121epqinm2π 1einm2π n nn12 nk Therefore, k 11 1(1)pq inm2π inm2π nn() j1 (eAB )(e CD ) 12 inm2π i2πnm 1epq 1e k (eABinm2π )(e CD inm2π ) 3.5. Infinite Divisibility 3.6. Stability with Respect to Geometric We know that the geometric distribution with probability Summation mass function, PX k(1 u )k uu , 0,1, is infi- nitely divisible, so wrapped geometric distribution is Proposition 3.2. Let 12,, be identically and in- infinitely divisible. Hence Θ ~(,,)WDSLpqm is infi- dependently distributed WDSL (,,p q m ) angular ran- nitely divisible. By the well known factorisation proper- dom variables, and let Nu be a geometric random va- ties of geometric law [14], Θ ~WDSL (,,) p q m admits a riable with probability mass function (1 uu )k1 , representation involving wrapped negative binomial dis- k 1, 2, , independent of j s . Then the angular tribution, d, k 1,2, where Nu nn12 nk n j random variable d (mod2π) has the WDSL d WW . j n1n2 j1 Consider s, rm, distribution with 2 p sq srand 2 p pq11 pqu pq 11 p qu 4 pq
Proof. Now we show that the above function coincides with
Let 1,2, be identically and independently distrib- the characteristic function of WDSL s, r, m distri- uted an gular ran dom variables following WDSL (,,pqm ) bution with s and r as given above. Setting Equation (20) and Nu is a geometric random variable with mean 1 u . equal to ns, r , which is the characteristic function Conditio ning on the distribution of Nu we can write the of WDSL s, r, m distribution, produces the equation characteristic function of the right hand side of the above 1(1)sr equation as (Equation (20)) n (21) um, in2π in2π i mod 2π mm 1 Nu i1 kk1 EEee1uu1esr (1e ) k 1 upqm,, (20) up11q 11u pqm,, in2π in2π 1e pmm1eq 1 upq 1 1 up11 q inm2π in2π m 1epq 1e 1 upq 1 1 (22)
Copyright © 2012 SciRes. OJS 112 K. JAYAKUMAR ET AL.
That is, Dividing Equation (23) by Equation (24), we get in2π in2π sq mm r . Substituting the value of r in Equation (23) we get 1es 1erpqu1 1 p in22in 11pqusqps 11 sqpq (25) mm 11sr 1e p 1eq On simplification, Equation (25) reduces to qs2 s q p11 p q u p 0 (26) 1 upq1 1 2 That is, f sqssqp 11 p qup. which shou ld hold for each p, q. This will happen when the following two equations hold simultaneously. Since pp, q 0, f 00 , and fupq1110 . 11pqur11srq (23) Therefore, f s admits a unique solution in the in- 11pqus11srp (24) terval 0, 1 and is given by
2 qp11 p qu qp 11 p qu 4 qp 2q (27) 2 p 2 qp11 p qu qp 11 p qu 4 qp
Remark 3.1. Wrapped discrete skew Laplace distribu- is given by tion is infinitely divisible since a circular random vari- 11pq able obtained by wrapping an infinitely divisible random ()n in2π in2π variable is infinitely divisible by [1,11] . mm 1epq1e 3.7. Trigonometric Moments The above expression can also be expressed in the th The n trigonometric moment of the WDSL pqm, , form
n2π ()sinpq m 1 i tan1 222 n2π 1cpq p q os n2π n2π m np 11 q 1 pqp qcospqsin e mm
in ne where n 0,1 is the p th mean resultant length and and n 0, 2π is the p th mean direction, for n 1,2, , n2π pq sin 2 11pq 1 pqpqnm cos2π 1 m n n tan (29) n2π 1 (28) 1cpq p q os 2 2 m pq sinn 2 m The length of the mean resultant vector, 11 pq 22 111 22 nn2π 2π 1cossinpqp q p q mm and the mean direction, 2π pq sin 111 m 1 tan tan 1 2π 1cpq p q os m
Copyright © 2012 SciRes. OJS K. JAYAKUMAR ET AL. 113
The circular variance, 11pq V 11 0 22 nn2π 2π 1cpq p q os p q sin mm The circular standard deviation,
22 nn2π 2π 1cpq p q os p q sin mm 2ln ln 0 2 11pq The measure of skewness, Using Equations (33)-(34) and for a fixed value of “m” we can find estimates for “p” and “q”. 032 where E sin sin 2 12 V0 2 We have, The measure of kurtosis, 2π 1cospq p q 4 a1 m 0 2 (1V0 ) where 2 E cos 2 2 2 2π V b1 0 pq sin m 4. Estimation That is Method of Moments 2π 2π apq11sin b 1 pqpq cos mm Let 1,, k be a random sample of size n taken from the WDSL (,,pqm ) distribution with parameters p, q, which gives and m. Then the nth sample trigonometric moment about the zero direction, 2π 2π bq111cos aq sin b mm mann ib n (30) p (35) 2π 2π where bq11 bcos a 1 sin 1 k mm anj cosn (31) k j1 Substituting the value of “p” in terms of “q” in Equa- and tion (33) or in Equation (34) we will get an equation in 1 k “q” and solving that we can find the estimate of “q” and bnnj sin (32) thus “p”. k j1 Corresponding population moment is ni nn REFERENCES Equating the sample moments to the corresponding popula- [1] K. V. Mardia and P. E. Jupp, “Directional Statistics,” 2nd tion moments, we get nn a and nn b for n 1, 2, . Thus, we have Edition, Wiley, New York, 2001. [2] K. V. Mardia, “Statistics of Directional Data,” Academic 2π Press, London, 1972. 11pq1 pqpq cos m [3] R. A. Fisher, “Dispersion on a Sphere,” Proceedings of a1 22(33) 2π 2π the Royal Society of London A, Vol. 217, No. 1130, 1953, 1cpq p q os p q si n pp. 295-305. doi:10.1098/rspa.1953.0064 m m [4] J. A. Greenwood and D. Durand, “The Distribution of and Length and Components of the Sum of n Random Unit Vectors,” Annals of Mathematical Statistics, Vol. 26, No. 2π 2, 1955, pp. 233-246. doi:10.1214/aoms/1177728540 11pqpq sin m [5] G. S. Watson and E. J. William, “On the Construction of b1 22 nn2π 2π Significance Tests on the Circle and the Sphere,” Biometrika, 1cpq p q os p q sin Vol. 43, 1956, pp. 344-352. mm [6] A. Pewsey, “A Wrapped Skew—Normal Distribution on (34) the Circle,” Communications in Statistics: Theory and
Copyright © 2012 SciRes. OJS 114 K. JAYAKUMAR ET AL.
Methods, Vol. 29, No. 11, 2000, pp. 2459-2472. 294. doi:10.2307/1403799 doi:10.1080/03610920008832616 [11] S. Inusah and T. J. Kozubowski, “A Discrete Analogue of [7] S. R. Jammalamadaka and T. J. Kozubowski, “A Wr- the Laplace Distribution,” Journal of Statistical Planning apped Exponential Circular Model,” Proceedings of the and Inference, Vol. 136, No. 3, 2006, pp. 1090-1102. Andhra Pradesh Academy of Sciences, Vol. 5, 2001, pp. doi:10.1016/j.jspi.2004.08.014 43-56. [12] A. W. Kemp, “Characterization of a Discrete Normal [8] S. R. Jammalamadaka and T. J. Kozubowski, “A New Distribution,” Journal of Statistical Planning and Infer- Family of Circular Models: The Wrapped Laplace Dis- ence, Vol. 63, No. 2, 1997, pp. 223-229. tributions,” Advances and Applications in Statistics, Vol. doi:10.1016/S0378-3758(97)00020-7 3, No. 1, 2003, pp. 77-103. [13] P. J. Szablowski, “Discrete Normal Distribution and Its [9] S. R. Jammalamadaka and A. Sengupta, “Topics in Cir- Relationship with Jacobi Theta Functions,” Statistics and cular Statistics,” World Scientific, Singapore, 2001. Probability Letters, Vol. 52, No. 3, 2001, pp. 289-299. doi:10.1142/9789812779267 doi:10.1016/S0167-7152(00)00223-6 [10] P. E. Jupp and K. V. Mardia, “A Unified View of the [14] W. Feller, “An Introduction to Probability Theory and Its Theory of Directional Statistics, 1975-1988,” Interna- Applications,” John Wiley & Sons, New York, 1957. tional Statistical Review, Vol. 57, No. 3, 1989, pp. 261-
Copyright © 2012 SciRes. OJS