Abstract
The Sibuya distribution arises as the distribution of the waiting time for the first success in Bernoulli trials, where the probabilities of success are inversely proportional to the number of a trial. We study a generalization that can be viewed as the distribution of the excess random variable \(N-k\) given \(N>k\), where N has the Sibuya distribution and k is an integer. We summarize basic facts regarding this distribution and provide several new results and characterizations, shedding more light on its origin and possible applications. In particular, we emphasize the role Sibuya distribution plays in the extreme value theory and point out its invariance property with respect to random thinning operation.
Similar content being viewed by others
References
Aban, I. B., Meerschaert, M. M., Panorska, A. K. (2006). Parameter estimation for the truncated Pareto distribution. Journal of the American Statistical Association, 101, 270–277.
Bondesson, L. (1992). Generalized gamma convolutions and related classes of distributions and densities. Lecture notes in statistics (Vol. 76). Berlin: Springer.
Buddana, A., Kozubowski, T. J. (2014). Discrete Pareto distributions. Economic Quality Control, 29(2), 143–156.
Christoph, G., Schreiber, K. (1998). Discrete stable random variables. Statistics and Probability Letters, 37, 243–247.
Christoph, G., Schreiber, K. (2000). Shifted and scaled Sibuya distribution and discrete self-decomposability. Statistics and Probability Letters, 48(2), 181–187.
Clauset, A., Newman, M. E. J. (2009). Power-law distributions in empirical data. SIAM Review, 2, 661–703.
Devroye, L. (1993). A triptych of discrete distributions related to the stable law. Statistics and Probability Letters, 18, 349–351.
Gabaix, X. (2009). Power laws in economics and finance. Annual Review of Economics, 1, 255–293.
Gradshteyn, I. S., Ryzhik, I. M. (2007). Tables of integrals, series, and products (7th ed.). Amsterdam: Academic Press.
Huillet, T. E. (2012). On Linnik’s continuous-time random walks (new version of Huillet (2000). On Linnik’s continuous-time random walks. Journal of Physics A: Mathematical and General, 33(14), 2631–2652. Available at http://www.reserachgate.net/publication/231129053.
Huillet, T. E. (2016). On Mittag-Leffler distributions and related stochastic processes. Journal of Computational and Applied Mathematics, 296, 181–211.
Johnson, N. L., Kotz, S., Kemp, A. W. (1993). Univariate discrete distributions. New York: John Wiley & Sons.
Johnson, N. L., Kotz, S., Balakrishnan, N. (1994). Continuous univariate distributions (Vol. 1). New York: John Wiley & Sons.
Kozubowski, T.J., Podgórski, K. (2016). Certain bivariate distributions and random processes connected with maxima and minima. Working papers in statistics 2016:9, Department of Statistics, School of Economics and Management, Lund University.
Krishna, H., Singh Pundir, P. (2009). Discrete Burr and discrete Pareto distributions. Statistical Methodology, 6, 177–188.
Lehman, E. L. (1983). Theory of point estimation. New York: John Wiley & Sons.
Newman, M. E. J. (2005). Power laws, Pareto distributions and Zipf’s law. Contemporary Physics, 46, 323–351.
Pakes, A. G. (1995). Characterization of discrete laws via mixed sums and Markov branching processes. Stochastic Processes and their Applications, 55, 285–300.
Pillai, R. N., Jayakumar, K. (1995). Discrete Mittag-Leffler distributions. Statistics and Probability Letters, 23, 271–274.
Rényi, A. (1976). On outstanding values of a sequence of observations. In: Rényi A (ed). Selected papers of A. Rényi (Vol. 3, pp. 50–65). Budapest: Akadémiai Kiadó.
Satheesh, S., Nair, N. U. (2002). Some classes of distributions on the non-negative lattice. Journal of the Indian Statistical Association, 40(1), 41–58.
Sibuya, M., Shimizu, R. (1981). The generalized hypergeometric family of distributions. Annals of the Institute of Statistical Mathematics, 33, 177–190.
Sibuya, M. (1979). Generalized hypergeometric, digamma, and trigamma distributions. Annals of the Institute of Statistical Mathematics, 31, 373–390.
Sornette, D. (2006). Critical phenomena in natural sciences: Chaos, fractals, selforganization and disorder: Concepts and tools (2nd ed.). Berlin: Springer.
Steutel, F. W., van Harn, K. (1979). Discrete analogues of self-decomposability and stability. Annals of Probability, 7, 893–899.
Steutel, F. W., van Harn, K. (2004). Infinitely divisibility of probability distributions on the real line. New York: Marcel Dekker.
Stumpf, M. P. H., Porter, M. A. (2012). Critical truths about power laws. Science, 335, 665–666.
Yule, G. U. (1925). A mathematical theory of evolution based on the conclusions of Dr. J.C. Willis, F.R.S. Philosophical Transactions of the Royal Society of London Series B., 213, 21–87.
Zipf, G. K. (1949). Human behavior and the principle of least effort. Cambridge, MA: Addison-Wesley.
Acknowledgements
The authors thank two anonymous referees for their helpful comments. Research of the first author was partially funded by the European Union’s Seventh Framework Programme for research, technological development and demonstration under Grant Agreement No 318984 - RARE. The second author was partially supported by Riksbankens Jubileumsfond Grant Dnr: P13-1024.
Author information
Authors and Affiliations
Corresponding author
Appendix
Appendix
Proof of Proposition 1
For the variables X and Y in (7), we have
where \(G_{N_\alpha }\) is the PGF given in (11). Thus,
as required. \(\square \)
Proof of Proposition 2
Suppose that, for some \(\alpha \in (0,1)\), N has Sibuya distribution \(GS_1(\alpha ,0)\), given by the PMF (5). Then, for each \(t\in \mathbb R_+\), the value of the process N(t) defined by (8) admits the stochastic representation (45), where \(c(t)=\mathbb P(X_j>t)\). Since for Sibuya distributed N we have (42) with \(c=c(t)\), which, in turn, implies (43), N(t) satisfies (9), as desired.
Next, assume that N(t) satisfies equation (9). Thus, for each \(t\in \mathbb R_+\), we have
Using standard conditioning argument, write
Noting that for \(k<n\) we have \(\mathbb P(N(t)=n|N=k)=0\) while for \(k\ge n\) the variable \(N(t)=n|N=k\) is binomial with parameters k and \(p=1-F(t)\), where F is the common CDF of the \(X_j\)’s, we conclude that
For \(n=0\), we have
We now write \(s=F(t)\in (0,1)\) and \(p_n=\mathbb P(N=n)\) and substitute (58) and (59) into (57), which results in the following equation
Further, by expanding the term \((1-s)^n\) into a power series in s and changing the index of the summation on the left-hand side of (60) to \(j=k-n\), we conclude that
Using standard result for power series, stating that the coefficients \(c_k\) of the product
are given by
following some algebra, we conclude the left-hand side of (61) is of the form (62) with
Thus, in view of the above, coupled with (61), and by the uniqueness of the power series, we conclude that
In particular, for \(k=1\), relation (63) reduces to
leading to
It now follows by induction that the \(\{p_n\}\) coincide with Sibuya probabilities (5), where \(\alpha =p_1=\mathbb P(N=1)\). This concludes the proof. \(\square \)
Proof of Proposition 4
Since, in view of (13), the results of Proposition 4 and Corollary 1 are equivalent, it is enough to establish (16). First, by incorporating the well-known property of the gamma function,
the generalized Sibuya SF (12) can be written as
Next, since for any \(\gamma >0\), we have the asymptotic representation of the Gamma function (see, e.g., Gradshteyn and Ryzhik 2007, formula 8.328.2, p. 895)
the right-hand side of (7) divided by the right-hand side of (16) converges to 1 with \(n\rightarrow \infty \), as desired. \(\square \)
Proof of Proposition 5
By Proposition 7, we have the equality in distribution \(N\mathop {=}\limits ^{d}N(X)\), where \(\{N(t), \,t>0\}\) is a standard Poisson process independent of \(X\mathop {=}\limits ^{d}X_1 X_{\nu -\alpha +1}/X_\alpha \), where all the three variables on the right-hand side are independent and gamma distributed with scale one and the shape parameters indicated by the subindex. The result now follows from (17), the representation (18) for the integer-order moments of N(t), and and the well-known moment formulas for gamma distribution, which produce
\(\square \)
Proof of Proposition 6
By Proposition 7, the PGF of N is given by (33), where \(\phi _X(\cdot )\) is the LT of the variable X defined in (27). To prove the result, it is enough to show that the LT of X is given by (34). To establish the latter, we condition on \(T_{\alpha ,\nu }\) when computing the LT of X, leading to
where f(x) is given in (28) and E is standard exponential with the LT
Thus,
where B(a, b) is the beta function (29). The result now follows by the integration formula 3.227.1 p. 320 of Gradshteyn and Ryzhik (2007). \(\square \)
Proof of Proposition 7
It is known (see, e.g., Devroye 1993) that the generalized hypergeometric distribution of type B3, given in (30) with X as in (32), is of the form
Setting \(a=1\), \(b=1-\alpha +\nu \), and \(c=\alpha \) in (64) produces the \(GS_0(\alpha , \nu )\) distribution. \(\square \)
Proof of Proposition 9
We proceed by showing that the PMF of the variable [W] coincides with that of the \(GS_0(\alpha ,\nu )\) distribution. First, using standard conditioning argument, write
where E has the standard exponential distribution and g is the PDF of \(V_{\alpha ,\nu }\), given by (36). Since
the probability (65) takes on the form
where
Noting that the function \(g(\cdot )\) in (36) is a genuine PDF for each \(\nu \ge 0\) and \(0<\alpha <\nu +1\), we conclude that
A substitution of (66) into (7), followed by some algebra, produces the \(GS_0(\alpha ,\nu )\) distribution. This concludes the proof. \(\square \)
Proof of Proposition 11
To prove the result, we shall use the following sufficient condition for this property to hold (Bondesson 1992, p. 28): A strictly decreasing PMF \(\{p_n\}\), \(n\in \mathbb N_0\), is DSD if
First, we shall show that generalized Sibuya PMF is strictly decreasing in n. To see this, note that the ratio
is strictly increasing in \(n\in \mathbb N_0\). Indeed, the derivative of the function
is positive for all \(x\in \mathbb R_+\), which can be checked by straightforward algebra. Since the ratio (68) converges to 1 as \(n\rightarrow \infty \), we conclude that \(p_{n+1}/p_n<1\) for all \(n\in \mathbb N_0\), showing the monotonicity of the sequence \(\{p_n\}\), \(n\in \mathbb N_0\). This also shows that the maximum on the left-hand side of (67) is attained for \(n=j\), so that the condition (67) becomes
After some algebra, condition (69) can be restated as
Since
and the function
is non-decreasing in \(x\in \mathbb R_+\), we obtain (70). This concludes the proof. \(\square \)
Proof of Proposition 12
According to the remarks following the statement of Proposition 12, condition (42) implies (44), which, in view of (45), is equivalent to (9). The result now follows from Proposition 2. \(\square \)
Proof of Proposition 14
For \(n=1\), the statement is trivial. To prove the result for general \(n\in \mathbb N\), it is enough to show the following fact:
(A) For each \(n\ge 2\), the conditional distribution of \(T_n\) given the \(n-1\) values \(0< t_1< \cdots<t_{n-1}<1\) of the previous jump locations has a uniform distribution on the interval \((t_{n-1}, 1)\).
Indeed, if (A) is true, the PDF of the joint distribution of \((T_1, \ldots ,T_n)\) is easily seen to be given by (48). This, in turn, is the joint PDF of the random vector on the right-hand side of (47). To see this, consider a random vector \((\Gamma _1, \ldots , \Gamma _n)\) of successive arrivals of standard Poisson process, so that \(\Gamma _i=W_1+\cdots + W_i\), \(i=1\ldots n\), where the \(\{W_i\}\) are IID standard exponential variables. Routine calculations show that the PDF of \((\Gamma _1, \ldots , \Gamma _n)\) is of the form
Consider a one-to-one transformation \(T_i=H(\Gamma _i)\), \(i=1,\ldots ,n\), where \(H(x)=1-e^{-x}\) is the common CDF of the \(W_i\)’s, with the inverse of \(H^{-1}(t)=-\log (1-t)\). Since the Jacobian of the inverse transformation is the product
the PDF of \((T_1, \ldots , T_n)\) becomes
which produces (48).
To establish the claim (A) above, we start with \(n=2\), and consider the conditional probability \(\mathbb P(T_2>t|T_1=t_1)\) for \(t_1<t<1\). Using the law of total probability, we obtain
where \((K_i,R_i)\) are the random pairs of record times and their sizes (with \(R_i=1-T_i\)), connected with the sequence \(\{nU_n\}\) (as described in Sect. 5). Note that the probability under the above sum can be written in terms of the \(\{U_n\}\) as
or, equivalently, as
where
When compared with (4), \(p(r_1, k)\) is recognized as the PMF of \(S\sim GS_0(r_1,1)\) and consequently,
Since the quantity on the right-hand side above is the survival function of the uniform distribution on the interval \((t_1,1)\), the result holds for \(n=2\). The proof in the case \(n>2\) is similar.
Under the same notation and using again the law of total probability, we have
where \(A_{n-1}\) denotes the condition \(T_1=t_1, \ldots , T_{n-1}=t_{n-1}\). Similarly as before, the conditional probabilities under the double sum above can be expressed as
where
is recognized as the probability \(\mathbb P(S=m)\) with \(S\sim GS_1(r_1,k)\). Since these probabilities sum up to one across the values of \(m\in \mathbb N_0\), and so do the probabilities \(\mathbb P(K_{n-1}=k)\) across the values of \(k\ge n-1\), we obtain
Since the quantity on the right-hand side above is the survival function of the uniform distribution on the interval \((t_{n-1},1)\), the result follows. \(\square \)
Proof of Proposition 17
Write the estimators as
where
whenever \(y_2-2y_1^2+y_1>0\), while otherwise, \(H_1(y_1,y_2)=0\), \(H_2(y_1,y_2)=1/y_1\) (with \(y_1, y_2\ge 1\)). To prove consistency, apply law of large numbers to the sequence \(Z_i=(X_i, X_i^2)'\) and conclude that the sample mean \(\overline{Z}_n=(M_1,M_2)'\) converges in distribution to the population mean \(m_Z=\mathbb E(Z_i) = (\mu _1,\mu _2)'\), where
are the first two moments of \(GS_1(\beta , \theta )\) distribution (and are well defined when \(\theta >2\beta \)). Since the function H is continuous at \(m_Z\), by continuous mapping theorem, the sequence (71) converges in distribution to \(H(m_Z) = (\beta , \theta )\). The last equality follows straightforward, albeit tedious, algebra. This proves the estimators are consistent.
Next, we establish their asymptotic normality. Assuming the fourth moment of the \(\{X_i\}\) is finite (\(\theta >4\beta \)), by the classical multivariate central limit theorem, we have the convergence in distribution \(\sqrt{n} (\overline{Z}_n - m_Z) \mathop {\rightarrow }\limits ^{d} \text{ N }(0,\Sigma )\), where the right-hand side denotes the bivariate normal distribution with mean vector zero and covariance matrix
A straightforward calculation, facilitated by Propositions 5 and 3, along with basic properties of expectation, shows that
Thus, since the function H is differentiable at \(m_Z\), standard multivariate delta method leads to the conclusion that, as \(n\rightarrow \infty \), the variables
converge in distribution to a bivariate normal vector with mean vector zero and covariance matrix \(\Sigma _{MME} = D\Sigma D'\), where
is the matrix of partial derivatives of the vector-valued function H evaluated at \(m_Z\). A routine, rather lengthy calculation yields
Finally, straightforward matrix multiplication produces the asymptotic covariance matrix \(\Sigma _{MME}\). \(\square \)
About this article
Cite this article
Kozubowski, T.J., Podgórski, K. A generalized Sibuya distribution. Ann Inst Stat Math 70, 855–887 (2018). https://doi.org/10.1007/s10463-017-0611-3
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10463-017-0611-3