COMMUNICATIONS IN STATISTICS Theory and Methods Vol. 33, No. 5, pp. 1007–1029, 2004 On the Power-Variance Family of Probability Distributions Vladimir Vinogradov* Department of Mathematics, Ohio University, Athens, Ohio, USA ABSTRACT We derive new properties for the power-variance family of probability distributions, which are also referred to as Tweedie laws. They emerge in the theory of generalized linear models and have a wide range of applications. We present decomposition criteria for Tweedie distributions which generalize Cochran’s and Raikov’s theorems. These criteria are interpreted in terms of the additivity of a shape parameter. The intersection of the power-variance family with the class of distributions which have no indecomposable components is determined. We reveal continuity of this family and obtain convenient common formulas for the cumulants. Key Words: Cumulant; Decomposition criterion; Indecomposable component; Power-variance family; Tweedie law; Weak convergence. *Correspondence: Vladimir Vinogradov, Department of Mathematics, Ohio University, 321 Morton Hall, Athens, OH 45701, USA; Fax: (740) 593-9805; E-mail: [email protected]. 1007 DOI: 10.1081/STA-120029821 Copyright # 2004 by Marcel Dekker, Inc. 0361-0926 (Print); 1532-415X (Online) www.dekker.com ORDER REPRINTS 1008 Vinogradov 1. DESCRIPTION AND BASIC PROPERTIES OF TWEEDIE DISTRIBUTIONS This paper pertains to the three-parametric power-variance family of distributions, which are also referred to as Tweedie probability laws. This family is very heterogeneous (see below). We establish new algebraic and topological properties for members of this class, which incorporate famous results derived previously for certain individual representatives of the Tweedie family. In addition, we obtain common formulas for the cumulants and standardized cumulants of Tweedie distributions. The power-variance family is characterized in this section. We formulate and clarify our main results in Sec. 2. The same section contains the simple proofs. For convenience of the reader, the major proofs are deferred to the concluding Sec. 3. Note that although the properties of the family considered are interesting and elegant in their own right, but this family is currently attracting a growing attention in view of various applications (see Hougaard, 2000; Vinogradov, 2002, Remarks 1.1.i and 2.5.ii). Also, each member of this family is obtained as a weak limit for a wide class of r.v.’s. This stipulates an applicability of the family of Tweedie distributions for approximation. See Jørgensen (1997, Ch. 4) or Jørgensen and Vinogradov (2002) for more detail. Next, let us introduce the power-variance family of probability distributions. Hereinafter, we denote both such distributions and the r.v.’s possessing such distributions by Twp ðm; lÞ. Each particular Tweedie law is uniquely characterized by specific values of power parameter p 2 D :¼ R1 nð0; 1Þ, scaling parameter l 2 L :¼ ð0; þ1Þ, and location (or mean) parameter m. The latter parameter is such that m 2 ½0; þ1Þ if p 2 ð 1; 0Þ; m 2 R1 if p ¼ 0; m 2 ð0; þ1Þ if p 2 ½1; 2, and m 2 ð0; þ1 if p 2 ð2; þ1Þ. Hereinafter, we refer to the values of parameters specified above as their admissible values. For simplicity of notation, we denote the above domains of the location parameter by Op . The probabilistic meaning of parameters p, m and l is clarified by (1.6). Also, the boundary @D of D is f0; 1g. It is known that Tweedie laws with the same value of p satisfy the following scaling relationship: d c Twp ðm; lÞ ¼ Twp ðc m; cp 2 lÞ; ð1:1Þ where c > 0 is an arbitrary fixed real (see Jørgensen, 1997, formula (4.7)). d Hereinafter, sign ‘¼’ is understood in the sense that the distributions (of r.v.’s) coincide. Also, each Tweedie distribution Twp ðm; lÞ can be ORDER REPRINTS Power-Variance Family 1009 characterized by its cumulant-generating function zp;m;l ðsÞ :¼ log E expfs Twp ðm; lÞg, where s belongs to a certain subset of R1 . Now, set 1 l m1 p ; yð p; m; lÞ :¼ ð1:2Þ j1 pj where p 2 Dnf1g. For each fixed p 2 ð2; þ1Þ and l 2 L, set yð p; 1; lÞ :¼ lim yð p; m; lÞ ¼ 0: ð1:3Þ m!þ1 Also, for p 2 Dnf1; 2g, define Bp;l :¼ j1 Proposition 1.1. pjð2 pÞ=ð1 j2 pj pÞ l1=ð p 1Þ Assume that p 2 D, m 2 Op and l 2 L. Then (i) For p 2 ð 1; 0, n zp;m;l ðsÞ ¼ Bp;l ðyð p; m; lÞ þ sÞð2 pÞ=ð1 pÞ yð p; m; lÞð2 pÞ=ð1 pÞ o ; pÞ=ð1 pÞ o ; yð p; m; lÞ if p 2 ð 1; 0Þ and s 2 R1 if p ¼ 0. where s (ii) ð1:4Þ : z1;m;l ðsÞ ¼ m l ðes=l For p 2 ð1; 2Þ, n zp;m;l ðsÞ ¼ Bp;l ðyð p; m; lÞ 1Þ, where s 2 R1 . (iii) sÞð2 pÞ=ð1 pÞ yð p; m; lÞð2 where s < yð p; m; lÞ. (iv) z2;m;l ðsÞ ¼ l logð1 For p 2 ð2; þ1Þ, n zp;m;l ðsÞ ¼ Bp;l yð p; m; lÞð2 s=yð2; m; lÞÞ, where s < yð2; m; lÞ. (v) pÞ=ð1 pÞ ðyð p; m; lÞ sÞð2 where s yð p; m; lÞ. z3 pÞ=ð1 pÞ o ; (vi) zp;m;l ðsÞ ¼ þ1 for all the remaining values of s. (vii) For arbitrary fixed admissible values of p, m and l, zp;m;l ðsÞ and sÞ are the inverse functions. p;1=m;l ð Proof of Proposition 1:1. (i)–(vi) All these representations for zp;m;l ðÞ can be derived from Jørgensen (1997, formula (4.16)). (vii) The proof is straightforward. & ORDER REPRINTS 1010 Vinogradov Remark 1.1. (i) Part (vii) of Proposition 1.1 is new, although it is well known for a few special cases (see Tweedie, 1984, p. 580). It yields new results on the first passage times for a class of Lévy processes constructed starting from the Tweedie laws with p 0 (see Vinogradov, 2002, Sec. 4). (ii) For an arbitrary fixed p 2 D, Tweedie laws Twp ðm; lÞ; m 2 Op ; l 2 Lg described in Proposition 1.1 comprise a two-parametric family of distributions (with respect to parameters m 2 Op and l 2 L). Each such family constitutes a reproductive exponential dispersion model in the same sense as in Jørgensen (1997, Ch. 3). Moreover, the family of Tweedie laws characterized in Proposition 1.1 is a unique class of the probability distributions that for each fixed p 2 D, simultaneously satisfy (1.1) and constitute a reproductive exponential dispersion model (see Jørgensen, 1997, Th. 4.1 and Prop. 4.2). Also, the just quoted theorem yields that each Tweedie law Twp ðm; lÞ 2 ID, where ID denotes the set of all univariate infinitely divisible distributions. d Next, an application of (1.1) with c ¼ 1=m yields that Twp ðm; lÞ ¼ 2 p m Twp ð1; l m Þ. This motivates the consideration of the following shape parameter, which is hereinafter denoted by fp : fp :¼ l m2 p : ð1:5Þ Clearly, fp 0. Also, it is obvious that fp ¼ 0 if and only if either f p 2 ð 1; 0; m ¼ 0g or f p 2 ð2; þ1Þ; m ¼ þ1g. These two cases correspond to stable distributions with index a 2 ð0; 1Þ [ ð1; 2 (see below). However, it is not customary to refer to fp as a ‘‘shape parameter’’ in these two cases. Parameter fp is of importance for the other members of the family of Tweedie laws (see (2.10)). It follows from (1.5)–(1.6) that fp equals the reciprocal of the squared coefficient of variation of r.v. Twp ðm; lÞ. In the case when p ¼ 3 (the inverse Gaussian family), f3 was employed by Chhikara and Folks (1989). Now, Tweedie laws satisfy the variance-to-mean relationship of a d power type. Namely, the variance of r.v. Y ¼ Twp ðm; lÞ is as follows: VarðY Þ ¼ mp =l ð1:6Þ (see Jørgensen, 1997, formula (4.2)). Property (1.6) justifies referring to the class of Tweedie distributions as the power-variance family (compare to Hougaard, 2000, Subsecs. 7.5.1 and 8.6.2). Also, set a ¼ að pÞ :¼ ð2 pÞ=ð1 pÞ: ð1:7Þ ORDER REPRINTS Power-Variance Family 1011 It is clear that parameters p 2 D and a 2 ð 1; 1Þ [ ð1; 2 are in one-toone correspondence. The three-parametric family described above was introduced independently by Tweedie (1984), Hougaard (1986), and Bar-Lev and Enis (1986). Recall that the family of Tweedie laws is quite heterogeneous. It includes continuous normal distributions ( p ¼ 0), gamma distributions ( p ¼ 2) and inverse Gaussian distributions ( p ¼ 3), discrete scaled Poisson distributions ( p ¼ 1) and mixed non-central gamma distributions with zero degrees of freedom ( p ¼ 3=2). In addition, Tweedie distributions with p 2 ð2; þ1Þ are derived by an exponential tilting of positive stable laws with index a ¼ að pÞ 2 ð0; 1Þ. The corresponding value of the exponential tilting parameter equals yð p; m; lÞ (see (1.2)). Namely, it is true that for each fixed p 2 ð2; þ1Þ, the probability densities fp;m;l ðÞ and fp;1;l ðÞ of r.v. Twp ðm; lÞ and positive stable r.v. Twp ð1; lÞ, respectively, are related as follows: fp;m;l ðxÞ e ðyð p;m;lÞxþzp;1;l ð yð p;m;lÞÞÞ fp;1;l ðxÞ: ð1:8Þ Here, x 2 R1þ , and function zp;1;l ðÞ is given in Proposition 1.1.v. Note that both these densities are equal to zero for negative values of x. Evidently, positive stable laws with index a 2 ð0; 1Þ are included in the power-variance family corresponding to f p 2 ð2; þ1Þ; m ¼ þ1g. Various properties of their densities fp;1;l ðÞ can be found in Zolotarev (1986, Sec. 2.5), where a different parametrization was used. Next, Tweedie distributions with p 2 ð 1; 0Þ are derived by an exponential tilting of extreme stable laws with index a ¼ að pÞ ¼ ð2 pÞ=ð1 pÞ 2 ð1; 2Þ and skewness parameter b ¼ 1. The corresponding value of the exponential tilting parameter equals yð p; m; lÞ (see (1.2)). By analogy to (1.8), one derives that for each p 2 ð 1; 0Þ, the probability densities fp;m;l ðÞ and fp;0;l ðÞ of r.v. Twp ðm; lÞ and extreme stable r.v. Twp ð0; lÞ, respectively, are related as follows: fp;m;l ðxÞ eyð p;m;lÞx zp;0;l ðyð p;m;lÞÞ fp;0;l ðxÞ: ð1:9Þ Here, x 2 R1 , and function zp;0;l ðÞ is given in Proposition 1.1.i. It is clear that extreme stable laws with index a 2 ð1; 2Þ and skewness parameter b ¼ 1 are included in the power-variance family corresponding to f p 2 ð 1; 0Þ; m ¼ 0g. Various properties of their densities fp;0;l ðÞ can be found in Zolotarev (1986, Sec. 2.5). In particular, the lower tail of each such density is of a power type, whereas its upper tail is lighter than that of a normal density. ORDER REPRINTS 1012 Vinogradov We refer to Tweedie distributions Twp ðm; Þ for which either f p 2 ð 1; 0Þ; m ¼ 0g or f p 2 ð2; þ1Þ; m ¼ þ1g as non-normal stable Tweedie laws. The union of this subfamily with the family of normal distributions fTw0 ðm; lÞ; m 2 O0 ; l 2 Lg is termed the class of stable Tweedie laws. All the other representatives of the power-variance family are referred to as non-stable Tweedie laws. They pertain to f p 2 Dnf0g; m 2 ð0; þ1Þg. Now, each Tweedie r.v. Twp ðm; lÞ with p 2 ð1; 2Þ has a mixed, compound Poisson-gamma distribution (see Hougaard, 2000). In particular, this distribution has an absolutely continuous component, while PfTwp ðm; lÞ ¼ 0g ¼ expf fp =ð2 pÞg > 0. Recall that each member of the power-variance family is infinitely divisible. Proposition 1.2, Theorem 1.1 and Remark 1.2 provide Lévy representations for the logarithm of the characteristic function of Tweedie distributions. However, we first should introduce some auxiliary notation. Let us denote the analytic continuation of the gamma function onto Cnf0; 1; 2; . . .g by GðÞ. This continuation relies on the next formula: Gðz 1Þ :¼ GðzÞ=ðz 1Þ. In addition, for arbitrary fixed real g and x > 0, we consider the complement of the incomplete gamma function, which is hereinafter denoted by Gðg; xÞ. Set Z 1 Gðg; xÞ :¼ yg 1 e y dy ð1:10Þ x (compare to Johnson et al., 1992, formula (1.85)). The following result provides the density pp;m;l ðÞ of Lévy measure of r.v. Twp ðm; lÞ with respect to Lebesgue measure in the case when p 2 Dn@D. It will be employed in the proof of Theorems 1.1 and 2.2.i. Proposition 1.2 (Compare to Küchler and Sørensen, 1997, Formula (2.1.17)). Assume that r.v. Twp ðm; lÞ is such that p 2 Dn@D, m 2 Op and l 2 L. Then ðiÞ In the case when p 2 ð1; þ1Þ, Twp ðm; lÞ is spectrally positive, there exists density pp;m;l ðÞ of its Lévy measure with respect to Lebesgue measure, and for an arbitrary x > 0, pp;m;l ðxÞ ¼ j p 1j1=ð1 pÞ 1=ð p l jGð1=ð p 1Þj 1Þ j xj ð1þð2 pÞ=ð1 pÞÞ e yð p;m;lÞjxj : ð1:11Þ Here, yð p; m; lÞ is given by (1.2)–(1.3). (ii) In the case when p 2 ð 1; 0Þ, Twp ðm; lÞ is spectrally negative, and there exists density pp;m;l ðÞ of its Lévy measure with respect to Lebesgue measure that admits representation (1.11) for an arbitrary x < 0. ORDER REPRINTS Power-Variance Family 1013 Proof of Proposition 1:2. Following Küchler and Sørensen (1997, Sec. 2.1), we will use cumulant-generating functions of probability distributions instead of the logarithm of their characteristic functions. This enables one to employ Lévy representation of the cumulantgenerating function of an infinitely divisible distribution (see Küchler and Sørensen, 1997, formula (2.1.8)). It is convenient to consider the next four cases separately. (i) If p 2 ð1; 2Þ then Proposition 1.2 follows from the formulas given in Küchler and Sørensen (1997, p. 12). (ii) Assume that p 2 ð2; þ1Þ. Then an application of a formula given in Bertoin (1996, p. 73) and some scaling arguments yield that Z 1 zp;1;l ðsÞ ¼ ðesx 1Þ pp;1;l ðxÞ dx; 0þ where s 0. The rest follows from Küchler and Sørensen (1997, formula (2.1.13)). (iii) Assume that p 2 ð 1; 0Þ. By (1.7), a ¼ að pÞ 2 ð1; 2Þ. Then it is straightforward to demonstrate that for such a, Z 1 Gð1 aÞ ð sÞa ¼ ðesx 1 s xÞ dð x a Þ: 0 Here, s 0 and the analytic continuation of the gamma-function is used. A combination of this representation with simple scaling arguments implies that Z 0 zp;0;l ðsÞ ¼ ðesx 1 s xÞ pp;0;l ðxÞ dx; 1 where s 0. The rest follows from Küchler and Sørensen (1997, formula (2.1.13)). (iv) If p ¼ 2 then Proposition 1.2 follows from a representation of z2;m;l ðÞ in terms of the Frullani integral (see Bertoin, 1996, p. 73). & Evidently, Proposition 1.2 implies that for each p 2 Dn@D, Lévy measure n p;m;l ðÞ of r.v. Twp ðm; lÞ is as follows: Z pp;m;l ðxÞ dx: ð1:12Þ n p;m;l ðAÞ :¼ A Here, A R1 nf0g is an arbitrary Lebesgue-measurable set. ORDER REPRINTS 1014 Vinogradov The next result provides the classical form of Lévy representations for the logarithm of characteristic function cp;m;l ðuÞ of r.v. Twp ðm; lÞ. In addition, it contains the explicit formulas for the corresponding Lévy measure. It is obvious that log cp;m;l ðuÞ ¼ zp;m;l ði uÞ; ð1:13Þ where zp;m;l ði uÞ denotes the analytic continuation of the corresponding cumulant-generating function of r.v. Twp ðm; lÞ, which is defined in Proposition 1.1. Theorem 1.1. (i) Assume that p 2 Dn@D, m 2 Op and l 2 L. Then For each p 2 ð 1; 0Þ, log cp;m;l ðuÞ ¼ Z 0 ðeiux 1 i u xÞ n p;m;l ðdxÞ: ð1:14Þ 1 Here, Le´vy measure n p;m;l ðÞ is such that for each y < 0, j p 1j1=ð1 pÞ 1=ð p n p;m;l fð 1; yg ¼ l jGð1=ð p 1Þj (ii) 1Þ 2 G 1 p ; yð p; m; lÞ j yj : p ð1:15Þ For each p 2 ð1; þ1Þ, log cp;m;l ðuÞ ¼ Z 1 ðeiux 1Þ n p;m;l ðdxÞ: ð1:16Þ 0þ Here, Le´vy measure n p;m;l ðÞ is such that for each y > 0, j p 1j1=ð1 pÞ 1=ð p n p;m;l f½ y; 1Þg ¼ l jGð1=ð p 1Þj 1Þ G 2 1 p ; yð p; m; lÞ y : p ð1:17Þ Proof of Theorem 1:1. (i) In order to derive (1.14), we apply the arguments used in the proof of Proposition 1.2.ii as well as representations (1.12)–(1.13). In turn, a combination of (1.10)–(1.12) implies (1.15). (ii) The validity of (1.16) is easily obtained by combining the arguments used in points (i), (ii) and (iv) of the proof of Proposition 1.2 with (1.12)–(1.13). In addition, representation (1.17) easily follows from a combination of (1.10)–(1.12). & ORDER REPRINTS Power-Variance Family 1015 Remark 1.2. (i) Lévy representation for the logarithm of the characteristic function of normal r.v. Tw0 ðm; lÞ is well known. Namely, log c0;m;l ðuÞ ¼ m i u u2 =ð2 lÞ. (ii) It follows from Proposition 1.1.ii that in the case when p ¼ 1, Lévy representation for the logarithm of the characteristic function of scaled Poisson r.v. Tw1 ðm; lÞ is as follows: Z 1 log c1;m;l ðuÞ ¼ ðeiux 1Þ n 1;m;l ðdxÞ: 1 Here, n 1;m;l ðÞ ¼ m l d1=l . Hence, in this case Lévy measure is degenerate being a multiple of the Dirac point mass d1=l concentrated at 1=l. (iii) It follows from (1.16)–(1.17) that all Tweedie distributions with p 2 ð1; þ1Þ belong to the class of infinitely divisible distributions on R1þ . In contrast, (1.14)–(1.15) imply that all Tweedie laws with p 2 ð 1; 0Þ are spectrally negative. 2. RESULTS We start with a technical result that is essential for the formulation and proof of our main Theorem 2.1. Lemma 2.1. Fix p 2 Dnf0; 1g. Let all li ’s, 1 i n, belong to L. Suppose that all mi ’s, 1 i n, belong to Op nf0g. Fix arbitrary real ci > 0, where 1 i n. Assume that c1 1 l1 m11 p ¼ ¼ cn 1 ln mn1 p ; ð2:1Þ and set m :¼ n X ci mi : ð2:2Þ i¼1 Let l 2 L. Then l m1 p ¼ c1 1 l1 m11 p ð2:3Þ if and only if n X ð p 2Þ=ð p 1Þ ci !p 1 : ð2:4Þ Proof of Lemma 2:1 is straightforward. & l¼ i¼1 1=ð p 1Þ li ORDER REPRINTS 1016 Vinogradov Our main theorem can be regarded as a decomposition criterion for Tweedie laws with p 2 Dn@D. Theorem 2.1. 1 i n. Let p 2 Dn@D, and fix arbitrary real ci > 0, where d (i) Assume that r.v. W ¼ Twp ðm; lÞ, where m 2 Op and l 2 L. Consider d independent r.v.’s fWi ; 1 i ng such that Wi ¼ Twp ðmi ; li Þ. Here, mi ’s and li ’s, 1 i n, are certain constants which belong to Op and L, respectively. Then the validity of decomposition d W¼ n X ci Wi ð2:5Þ i¼1 implies the fulfillment of conditions (2.1)–(2.3). (ii) Let independent r.v.’s Wi ’s have the same distributions as in part (i). Suppose that condition (2.1) is fulfilled, and define r.v. W by formula d (2.5). Then W ¼ Twp ðm; lÞ, where the values of parameters m and l are given by formulas (2.2) and (2.4), respectively. Proof of Theorem 2.1 is deferred to Sec. 3. d Remark 2.1. (i) In the case when r.v. W ¼ Twp ðm; lÞ is not non-normal stable, (1.6) yields that a combination of relationships (2.1) and (2.3) can be interpreted as the property that r.v.’s W and ci Wi possess common ratios of mean to variance. Here, 1 i n. (ii) The assumption of independence of r.v.’s fWi ; 1 i ng imposed in Theorem 2.1 is not necessary. This follows from two counter-examples to Cochran’s theorem, which is a special case of Theorem 2.1 and is given below as Theorem 2.5. The reader is referred to James (1952) and Dykstra and Hewett (1972) for these counterexamples. (iii) In the case when p ¼ 2, Theorem 2.1.ii is stated in Johnson et al. (1994, p. 340). In addition, in the case when p ¼ 3, the validity of Theorem 2.1 is mentioned in Chhikara and Folks (1989, p. 13, Property 2). The problem of existence and construction of decompositions of r.v. d W ¼ Twp ð; Þ into a finite sum of independent random components such that at least one of these components is not distributed according to Tweedie law Twp ð; Þ with the same value of power parameter p is of interest for the arithmetic of probability distributions. In this respect, let us give a few relevant definitions. ORDER Power-Variance Family REPRINTS 1017 Definition 2.1 (cf., e.g., Linnik, 1964, p. 79). R.v. Y and its distribution function F ðÞ are said to be indecomposable if the relationship F ¼ F1 F2 ð2:6Þ implies that F1 or F2 is a degenerate distribution function. Here, F1 and F2 are certain distribution functions which are commonly referred to as the factors of F , and sign ‘ ’ denotes the operation of convolution. It follows from (2.6) that if F1 is a factor of F then for each a 2 R1 , F1 Ea is also a factor of F . Hereinafter, Ea denotes the degenerate distribution function which is concentrated at point a. Factor F1 Ea is commonly referred to as that which is equivalent to F1 , whereas each component of the type Ea will be hereinafter termed a trivial factor of decomposition (2.6). Any other component that can emerge on the right-hand side of (2.6) will be called a proper factor. In the sequel, we will occasionally limit our consideration to decompositions into proper factors only (see Theorems 2.2.ii and 2.3.i) or into the components which have no trivial factors (see Theorem 2.4.i). Definition 2.2 (see Linnik and Ostrovskii, 1977, p. 55). The set of all distribution functions on R1 which have no indecomposable components is hereinafter denoted by I0 . It is known that I0 is a proper subclass of ID (cf., e.g., Linnik and Ostrovskii, 1977, Th. 3.5.1). Definition 2.3 (see Lukacs, 1970, p. 245). A family of probability distributions is called factor-closed if the factors of every element of the family belong necessarily to the family. Hereinafter, each such family is referred to as an FC-family. Remark 2.2. (i) An FC-family does not necessarily belong to class ID. Thus, it is known that the class of binomial distributions is factor-closed (cf., e.g., Teicher, 1954). However, each binomial distribution is not infinitely divisible. (ii) Assume that each member of a certain family of univariate probability distributions belongs to IDnI0 . That is, each member of the family is infinitely divisible, but it has an indecomposable component. Then such family cannot be factor-closed, since this indecomposable component does not belong to ID. Hence, it is not included in the family. ORDER REPRINTS 1018 Vinogradov Let us consider another family of infinitely divisible distributions, which is slightly more general than the class of scaled Poisson distributions. Definition 2.4. Consider a family of univariate infinitely divisible distributions, which are described by the following characteristic function: Rm;g;h ðtÞ ¼ exp i t m þ g ðeith 1Þ : ð2:7Þ Here, m 2 R1 , whereas g and h are arbitrary non-negative real numbers. Hereinafter, we will term the three-parametric family of distributions defined by (2.7) the family of Poisson-type distributions. The intersection of the power-variance family with I0 is described by the following result. Theorem 2.2. (i) Consider a particular Tweedie distribution Twp ðm; lÞ, where p 2 D, l 2 L, and m 2 Op . Then Twp ðm; lÞ 2 I0 if and only if p 2 @D. (ii) The family of normal distributions Tw0 ðm; lÞ; m 2 R1 ; l 2 L is factor-closed provided that the trivial factors are excluded. (iii) The family (2.7) of Poisson-type distributions is factor-closed. Proof of Theorem 2:2. (i) The proof is deferred to Sec. 3. (ii) The proof can be found in Lukacs (1970, Sec. 8.2). (iii) See Lukacs (1970, Corollary to Theorem 8.2.2). & The next two statements can be regarded as decomposition criteria for Tweedie laws with p 2 @D. Theorem 2.3 is a version of the famous Cramér’s theorem (see Cramér, 1970, Th. 19 and also Johnson et al., 1994, pp. 102–103). In turn, Theorem 2.4 slightly generalizes Raikov’s theorem (see Raikov, 1938 and also Johnson et al., 1992, p. 173). Theorem 2.3 (Cramér’s Theorem). 1 i n. Fix arbitrary real ci > 0, where d (i) Assume that r.v. W ¼ Tw0 ðm; lÞ, where m 2 O0 and l 2 L. Consider independent r.v.’s fWi ; 1 i ng such that decomposition (2.5) holds, and neither of r.v.’s Wi is a trivial factor. Then each d Wi ¼ Tw0 ðmi ; li Þ, where mi ’s and li ’s, 1 i n, are certain constants which belong to O0 and L, respectively, and conditions (2.2) and (2.4) (with p ¼ 0) are fulfilled. (ii) Let independent r.v.’s Wi ’s have the same distributions as in part d (i). Define r.v. W by formula (2.5). Then W ¼ Tw0 ðm; lÞ, where the values ORDER REPRINTS Power-Variance Family 1019 of parameters m and l are given by formulas (2.2) and (2.4) (with p ¼ 0), respectively. d Proof of Theorem 2:3. (i) Fix r.v. W ¼ Tw0 ðm; lÞ, where m 2 O0 and l 2 L. By Theorem 2.2.ii, all its proper components Wi ’s are normally distributed. Next, Remark 1.2.i yields that the values mi ’s and li ’s which characterize the distributions of normal r.v.’s fWi ; 1 P i ng may be chosen arbitrarily, provided that (2.2) is fulfilled, and ni¼1 c2i =li ¼ 1=l. It remains to note that in the case when p ¼ 0, this relationship coincides with (2.4). (ii) The proof is obtained by a combination of (1.1) and Proposition 1.1.i. & Theorem 2.4 (Raikov’s Theorem). 1 i n. Fix arbitrary real ci > 0, where d (i) Assume that r.v. W ¼ Tw1 ðm; lÞ, where m 2 O1 and l 2 L. Consider independent r.v.’s fWi ; 1 i ng such that decomposition (2.5) holds, and the distributions of all Wi ’s do not contain trivial factors. Then d each Wi ¼ Tw1 ðmi ; li Þ, where mi ’s and li ’s, 1 i n, are certain constants which belong to O1 and L, respectively, and condition (2.2) along with conditions (2.1) and (2.3) (with p ¼ 1) are fulfilled. (ii) Let independent r.v.’s Wi ’s have the same distributions as in part (i). Suppose that condition (2.1) is fulfilled (with p ¼ 1), and define r.v. W d by formula (2.5). Then W ¼ Tw1 ðm; lÞ, where the value of parameter m is given by formula (2.2), and that of l is given by condition (2.3) (with p ¼ 1). Proof of Theorem 2:4. (i) By (1.1), it suffices to prove this statement in the case when c1 ¼ ¼ cn ¼ 1. Note that the class fTw1 ðm; lÞ; m 2 O1 ; l 2 Lg of scaled Poisson laws is a proper subclass of the family (2.7) of Poisson-type distributions. This is easily derived by a combination of Proposition 1.1.ii with (2.7). In particular, given m 2 O1 and l 2 L, one should set m ¼ 0, g ¼ m l and h ¼ 1=l in formula (2.7) for ch.f. Rm;g;h ðtÞ. d Next, fix r.v. W ¼ Tw1 ðm; lÞ that has ch.f. R0;ml;1=l ðtÞ. Consider its decomposition into independent components Wi ’s, 1 i n, such that neither of these components has a trivial factor. It then follows from Theorem 2.2.iii that for each 1 i n, ch.f. RðiÞ ðtÞ of r.v. Wi satisfies (2.7). Subsequently, one concludes that parameter m in representation (2.7) for RðiÞ ðtÞ equals zero, since the distribution of each Wi is assumed to not contain a trivial factor. Therefore, we obtain that for ORDER REPRINTS 1020 Vinogradov each 1 i n, RðiÞ ðtÞ ¼ exp gi ðeithi 1Þ ; ð2:8Þ where all gi ’s and hi ’s are certain positive real numbers. Now, a combination of Proposition 1.1.ii with (1.1) yields that d Tw1 ðm; lÞ ¼ l 1 d Tw1 ðm l; 1Þ ¼ l 1 Poissðm lÞ: ð2:9Þ Here, PoissðnÞ denotes a Poission r.v. with mean n. A subsequent combination of (2.8)–(2.9) with Proposition 1.1.ii implies that each d Wi ¼ Tw1 ðgi =hi ; hi 1 Þ. Hence, it has a scaled Poisson distribution. Finally, an application of Linnik (1964, Th. 5.1.1) along with the induction arguments yields that all hi ’s are the same and equal to l. The rest is trivial. (ii) The proof is obtained by a combination of (1.1) with Proposition 1.1.ii. & Remark 2.3. (i) Apparently, the results of Theorem 2.2 and Remark 2.2.ii reveal a reason behind the difference in the sets of conditions imposed in the decomposition criteria given by Theorems 2.1 and 2.3–2.4. Namely, each subfamily of Tweedie distributions Twp ðm; lÞ; m 2 Op ; l 2 Lg with a fixed value of p from the interior of D is not factor-closed. In contrast, each subfamily of Tweedie distributions Twp ðm; lÞ; m 2 Op ; l 2 L with a fixed value of p belonging to the boundary of D is factor-closed. (ii) In the case when p ¼ 0, a combination of (2.1)–(2.3) is sufficient but not necessary for decomposition (2.5) to be held. Sufficiency follows by combining (1.1) with Theorem 2.3. The refutation of necessity can be done by constructing trivial counter-examples. The details are left to the reader. Now, we proceed with a series of statements pertaining to Theorems 2.1–2.3. The following result can be interpreted as the additivity property of shape parameter fp , which is defined by (1.5). Also, in the sequel we will employ the nth partial sum of the sequence of independent r.v.’s P fWi ; 1 i ng. On this reason, it is convenient to set Sn :¼ ni¼1 Wi : Proposition 2.1. Let p 2 D, and fix arbitrary real ci > 0, where 1 i n. d Assume that r.v. W ¼ Twp ðm; lÞ, where m 2 Op , and l 2 L. Consider d independent r.v.’s fWi ; 1 i ng such that Wi ¼ Twp ðmi ; li Þ. Here, mi ’s and ORDER REPRINTS Power-Variance Family 1021 li ’s, 1 i n, are certain constants which belong to Op and L, respectively. Suppose that (2.2) is fulfilled, and set fpðiÞ :¼ li mi2 p . Then (i) Under these assumptions, the validity of conditions (2.1) and (2.3) implies the fulfillment of the next property: fp ¼ n X fðiÞ p : ð2:10Þ i¼1 (ii) Suppose that (2.10) is satisfied, and either min1in fðiÞ p > 0 and VarðW Þ ¼ VarðSn Þ: Pn i¼1 fðiÞ p ¼ 0, or ð2:11Þ Then conditions (2.1) and (2.3) are fulfilled. Proof of Proposition 2.1 is deferred to Sec. 3. Remark 2.4. (i) Let us clarify the reason of imposing an assumption on the properties of shape parameter in Proposition 2.1.ii. Namely, Pthe n ðiÞ f in the case when i¼1 p > 0, one of the terms that emerge on the right-hand side of (2.10), say fð1Þ p , can be zero. This would imply the violation of (2.1). Hence, we should exclude the case when min1in fðiÞ p ¼ 0. (ii) It is clear that in the case when min1in fðiÞ p > 0, each Wi has a finite variance, and both terms of Eq. (2.11) are well defined. Moreover, condition (2.11) that emerges in Proposition 2.1.ii is essential. Thus, one d can easily construct three gamma variables Wi ¼ Tw2 ðmi ; li Þ, 1 i n, such that m1 ¼ m2 þ m3 , l1 ¼ l2 þ l3 , but l2 =m2 6¼ l3 =m3 . Suppose that r.v.’s W2 and W3 are independent. Then the probability density of the sum W2 þ W3 , which can be derived from Johnson et al. (1994, formula (17.110)) is not that of a gamma variable. Hence, it differs from that of r.v. W . The original proof of the next celebrated result was given by Cochran (1934, Th. II). Below we will present a simpler derivation of his statement, which relies on a combination of Theorem 2.1 with Proposition 2.1. Recall that r.v. Sn is defined above Proposition 2.1. d Theorem 2.5 (Cochran’s Theorem). Let r.v. W ¼ Tw2 ðm; m=2Þ, where m > 0 is an integer. Assume that independent r.v.’s fWi ; 1 i ng are such ORDER REPRINTS 1022 Vinogradov d that Wi ¼ Tw2 ðmi ; li Þ, where all li ’s belong to L, and all mi ’s are positive integers. Then the distributions of r.v.’s W and Sn coincide if and only if d P m ¼ ni¼1 mi and each li ¼ mi =2. Proof of Theorem 2:5. First, let us not impose a constraint that mi ’s and m are integers. Then the statement follows from Theorem 2.1 and Proposition 2.1. Thus, a combination of (2.1) with (2.3) implies that l1 =m1 ¼ ¼ ln =mn ¼ l=m ¼ 1=2. Subsequently, a combination of this P P equation with (2.10) yields that m=2 ¼ l ¼ ni¼1 li ¼ ni¼1 mi =2. Now, assume that all mi ’s and m can only take on positive integer values. Then a combination of Proposition 1.1.iv with the above d d arguments yields that r.v.’s W ¼ Tw2 ðm; m=2Þ and Wi ¼ Tw2 ðmi ; mi =2Þ are chi-square P distributed with m and mi degrees of freedom, respectively, n such that i¼1 mi ¼ m. The latter relationship coincides with Cochran (1934, cond. (2)). & Next, we present a continuity property of the Tweedie family. d Hereinafter, sign ‘!’ is understood as weak convergence. Theorem 2.6. Fix p0 2 D, m0 2 Op0 , and l0 2 L. Consider a certain subfamily of Tweedie laws fTwp ðm; lÞg, where p 2 D, m 2 O, and l 2 L d are such that p ! p0 , m ! m0 and l ! l0 . Then fTwp ðm;lÞg! Twp0 ðm0 ;l0 Þ. Proof of Theorem 2:6. First, let us demonstrate that d Twp ðm0 ; l0 Þ ! Twp0 ðm0 ; l0 Þ ð2:12Þ as p ! p0 . It is easily seen that unless p0 2 f1; 2g, (2.12) follows from parts (i), (iii) and (v) of Proposition 1.1 and an obvious continuity of function zp;m;l ðÞ with respect to p. Hence, it remains to consider the following cases: (i) p ! 2. Then a combination of parts (iii) and (v) of Proposition 1.1 implies that for arbitrary fixed m 2 O2 and l 2 L, and either for an arbitrary fixed s yð p; m0 ; l0 Þ if p > 2 or for an arbitrary fixed s < yð p; m0 ; l0 Þ if p < 2, zp;m0 ;l0 ðsÞ ¼ fp 2 2 exp 1 p p logð1 p s=yð p; m0 ; l0 ÞÞ 1 : ð2:13Þ ORDER REPRINTS Power-Variance Family 1023 Subsequently, the expression 2 p log ð1 s=yð p; m0 ; l0 ÞÞ ¼ expfð2 exp 1 p pÞ Constg ð1 þ oð1ÞÞ as p ! 2. Hence, it is easy to show that the expression on the right-hand side of (2.13) is equivalent to ð1 pÞ 1 fp log ð1 s=yð p; m0 ; l0 ÞÞ ! l logð1 s=yð2;m0 ;l0 ÞÞ as p ! 2. By Proposition 1.1.iv, this limit coincides with z2;m0 ;l0 ðsÞ. (ii) p # 1. The proof can be derived by employing certain analytical techniques similar to those used in part (i) (compare to Hougaard, 2000, p. 506). Next, consider the general case when p ! p0 , m! m0 and l ! l0 . Evidently, zp;m;l ðsÞ zp;m0 ;l0 ðsÞ þ zp;m;l ðsÞ zp;m0 ;l0 ðsÞ . A combination of Proposition 1.1.i–vi with (1.2)–(1.4) yields that for an arbitrary fixed p 2 D, zp;m;l ðsÞ ! zp;m0 ;l0 ðsÞ as m ! m0 and l ! l0 . The rest follows from (2.12). & Remark 2.5. (i) Theorem 2.6 is of the same spirit as the continuity property of the family of stable laws (see Zolotarev, 1986, Property 2.4). For p0 2 @D, the validity of Theorem 2.6 is mentioned in Küchler and Sørensen (1997, p. 13). See also Hougaard (1986, Lm. 2.a) for the case when p0 2 ½2; 1Þ. (ii) Theorem 2.6 provides a basis for the derivation of new results on a near-equilibrium behavior of both random movements of equities and rational prices of stock options in the presence of downward jumps in the logarithmic returns on equities (see Vinogradov, 2002, Sec. 3)). In order to formulate the next statement, we need to introduce some additional notation. Let kp;m;l ðlÞ denote the lth cumulant of Twp ðm; lÞ, where l 1 is an integer. For l 2, denote rp;m;l ðlÞ :¼ kp;m;l ðlÞ=kp;m;l ð2Þl=2 . These coefficients are frequently referred to as the standardized cumulants (see McCullagh and Nelder, 1989, p. 351). In particular, rp;m;l ð3Þ and rp;m;l ð4Þ are skewness g1 and kurtosis g2 of r.v. Twp ðm; lÞ, respectively. Also, consider a parametric family of special functions of an integer argument (compare to Tweedie, 1984, formula (23)): Pp ðlÞ :¼ l 2 Y ðm ð p 1Þ þ 1Þ: m¼0 Here, l 2 is an integer, and p 2 D. In particular, P2 ðlÞ ¼ GðlÞ. ð2:14Þ ORDER REPRINTS 1024 Vinogradov The following theorem stipulates that the cumulants of each Tweedie law can be expressed by a common, easy-to-use computational formula. Also, it emphasizes the importance of shape parameter fp defined by (1.5). d Theorem 2.7. Fix an arbitrary Tweedie r.v. W ¼ Twp ðm; lÞ, where p 2 D, l 2 L, and m 2 Op . Assume that l 2 is an arbitrary fixed integer. Then (i) kp;m;l ðlÞ ¼ Pp ðlÞ ml fp ðl 1Þ ð2:15Þ ; where Pp ðlÞ is defined by (2.14). (ii) d In the case when r.v. W ¼ Twp ðm; lÞ is not non-normal stable, rp;m;l ðlÞ ¼ Pp ðlÞ fp ðl 2Þ=2 : ð2:16Þ Proof of Theorem 2:7. (i) (2.15) is obtained by induction. Its derivation is based on a successive differentiation of function zp;m;l ðsÞ, which is given in Proposition 1.1. The induction base corresponds to l ¼ 2 and coincides with (1.6). d (ii) Assume that r.v. W ¼ Twp ðm; lÞ is not non-normal stable. Then a combination of (1.1) and (1.5) implies that for an arbitrary real c > 0, fp ðc W Þ ¼ fp ðW Þ ¼ fp : ð2:17Þ Subsequently, a combination of (1.5)–(1.6) implies that the standard deviation sW ¼ m fp 1=2 < 1. Set Y :¼ W =sW . Then (1.1) yields that d p=2 Y ¼ Twp ðf1=2 p ; fp Þ. By (2.17), fp ðY Þ ¼ fp ðW Þ. In order to obtain (2.16), it remains to note that lth cumulant of r.v. Y equals lth standardized cumulant of r.v. W . & Remark 2.6. (i) By (2.16), each Tweedie law Twp ðm; lÞ which is not non-normal stable, has skewness g1 ¼ p=f1=2 and kurtosis p g2 ¼ p ð2p 1Þ=fp . Hence, Tweedie distributions with positive (respectively, negative) p are skewed to the right (respectively, to the left). They are leptokurtic, with the exception of the normal Tweedie laws Tw0 ðm; lÞ. (ii) In the case when p ¼ 0, (2.15) is consistent with the formulas for cumulants of normal distribution with mean m and variance 1=l. Also, in the case when p ¼ 1, (2.16) can be derived by combining (2.9) with Johnson et al. (1992, formula (4.9)). ORDER REPRINTS Power-Variance Family 1025 In the case when p ¼ 2, (2.15) can be obtained by applying Johnson et al. (1994, formula (17.9)). Also, in the case when p ¼ 3, (2.15) is consistent with Chhikara and Folks (1989, formula (2.8)). (iii) In the case when p 2 Dnð 1; 0Þ, (2.16) can be derived by a combination of Tweedie (1984, formulas (22)–(23)). (iv) In the cases when either f p 2 ð2; þ1Þ; m ¼ þ1g or f p 2 ð 1; 0Þ; m ¼ 0g, which correspond to non-normal stable Tweedie distributions, (2.15) should be understood in a generalized sense, whereas (2.16) does not hold. In these two cases, (2.15) implies that for an arbitrary integer l 2, kp;m;l ðlÞ ¼ 1. In the former case, this is also true for l ¼ 1. 3. MAJOR PROOFS Proof of Theorem 2.2.i. First, it is known that both normal and Poisson families Tw0 ðm; lÞ and Tw1 ðm; 1Þ belong to I0 (see Cramér, 1970, Theorem 19; Raikov, 1938, Sec. 3; or Johnson et al., 1992, p. 172). Next, we derive that for an arbitrary fixed l 2 L, subfamily fTw1 ðm; lÞ; m 2 O1 g 2 I0 . This follows from the validity of this property in the special case when l ¼ 1 by combining (2.9) with Linnik (1964, Th. 5.1.1) and applying the induction arguments. The rest of proof relies on the use of Proposition 1.2. In the case when p 2 ð 1; 0Þ [ ½3=2; þ1Þ, Linnik (1964, Th. 12.1.1) yields that the distribution, whose density of Lévy measure with respect to Lebesgue measure is given by Proposition 1.2, does not belong to I0 . Finally, in the case when p 2 ð1; 3=2Þ, Linnik and Ostrovskii (1977, Th. 6.7.2) implies that the distribution whose density of Lévy measure with respect to Lebesgue measure is given by Proposition 1.2 does not belong to I0 . & Proof of Theorem 2:1. By (1.1), it suffices to prove the theorem in the case when c1 ¼ ¼ cn ¼ 1. Proof of (ii). First, assume that n ¼ 2. Subsequently, if at least one of r.v.’s fW1 ; W2 g is extreme stable then the other one should also be extreme stable. This is easily derived from a combination of (2.1) with (1.2)–(1.3). The rest follows from Samorodnitsky and Taqqu (1994, Property 1.2.1). Observe that in this case, their conditions coincide with (2.2) and (2.4). Now, assume that both independent Tweedie r.v.’s fW1 ; W2 g are not stable. Then a combination of (2.1) with (1.2) yields that yð p; m1 ; l1 Þ ¼ yð p; m2 ; l2 Þ. Subsequently, let us employ Proposition 1.1 and keep in ORDER REPRINTS 1026 Vinogradov mind (1.4). One obtains that cumulant-generating function of r.v. W1 þ W2 admits the representation given in Proposition 1.1 with the d same p, m ¼ m1 þ m2 , and l defined by (2.4). Hence, W ¼ Twp ðm; lÞ. In the case when n > 2, the proof is easily obtained by induction in n. Proof of (i). The proof relies on equating cumulant-generating functions of r.v.’s W and Sn , which is defined above Proposition 2.1. It is based on specific representations for zp;m;l ðÞ given in Proposition 1.1 and hence, should be carried out separately in four distinct cases, which correspond to p 2 ð 1; 0Þ, p 2 ð1; 2Þ, p ¼ 2 and p 2 ð2; þ1Þ. Here, we only consider the case when p 2 ð2; þ1Þ, since the method of proof employed for each of these cases is essentially the same. The proof is carried out by induction in n. The induction base corresponds to n ¼ 1 and is trivial. Next, suppose that n 2 and prove the induction step. To this end, we equate cumulant-generating functions of r.v.’s W and Sn , where the latter r.v. is defined above Proposition 2.1. Thus, n zp;m;l ðsÞ ¼ Bp;l yðp;m;lÞð2 ¼ n X i¼1 pÞ=ð1 pÞ n Bp;li yð p; mi ;li Þð2 ðyð p;m;lÞ pÞ=ð1 pÞ sÞð2 ðyðp; mi ; li Þ pÞ=ð1 pÞ sÞð2 o pÞ=ð1 pÞ o ; ð3:1Þ where s yðp; m;lÞ. In addition, the cumulant-generating functions of both r.v.’s should be equal to þ1 for s > yðp; m; lÞ. It is easily seen that log E exp fs Sn g ¼ þ1 for s > min1in yð p; mi ; li Þ and is given by the expression that emerges on the right-hand side of (3.1) for s min1in yð p; mi ; li Þ. Hence, min yð p; mi ; li Þ ¼ yð p; m; lÞ: ð3:2Þ 1in It is clear that we can assume without loss of generality that min1in yð p; mi ; li Þ ¼ yð p; mn ; ln Þ. Now, if m ¼ þ1 then mn ¼ þ1. This follows from the fact that 0 ¼ yð p; m; lÞ ¼ yð p; mn ; ln Þ. Note that zp;1;l ðsÞ and all zp;mi ;li ðsÞ are finite and negative for s 2 ð 1; 0Þ. Hence, Bp;l > Bp;ln , where Bp;l is defined by (1.4). Therefore, ð^ lÞ1=ð p 1Þ :¼ l1=ð p 1Þ ln1=ð p 1Þ > 0: ð3:3Þ ORDER REPRINTS Power-Variance Family 1027 Subsequently, an application of (3.1) yields that zp;1;^l ðsÞ ¼ zp;1;l ðsÞ zp;1;ln ðsÞ n 1 X ¼ i¼1 n Bp;li yðp;mi ;li Þð2 pÞ=ð1 pÞ ðyðp;mi ;li Þ sÞð2 pÞ=ð1 pÞ o ; where ^ l is defined by (3.3) and s yðp;1; ^ lÞ ¼ 0. The rest follows from the induction hypothesis. Next, suppose that m 2 ð0; 1Þ, which implies that 0 < mn < m. Otherwise, one gets a contradiction, since n 2 and min1in 1 mi > 0. Then a combination of (1.2) and (3.2) yields the validity of (3.3). Set ^ :¼ m mn 2 ð0; þ1Þ. It is easily checked that yð p; m ^; ^ m lÞ ¼ yð p; m; lÞ ¼ yð p; mn ; ln Þ: Hence, one may ascertain that zp;^m;^l ðsÞ ¼ zp;m;l ðsÞ ¼ n 1 X i¼1 zp;mn ;ln ðsÞ n Bp;li yðp;mi ;li Þð2 pÞ=ð1 pÞ ðyðp;mi ;li Þ sÞð2 pÞ=ð1 pÞ o ; ^; ^ where s yðp; m lÞ. Also, the cumulant-generating functions of r.v.’s ^ ^; ^ m; lÞ and Sn 1 are equal to þ1 for s > yðp; m lÞ. The rest follows Twp ð^ from the induction hypothesis. & Proof of Proposition 2:1. By analogy to (2.17), it suffices to prove Proposition 2.1 in the case when c1 ¼ ¼ cn ¼ 1. (i) The proof of the fact that a combination ofP(2.1) and (2.3) yields the validity of (2.10) is straightforward, since m ¼ ni¼1 mi . (ii) In the case when fp ¼ 0, the derivation of (2.1) and (2.3) is trivial. Evidently, each fðiÞ p 0 and hence, it is equal to zero. Now, assume that fp ¼ l m2 p > 0. Subsequently, a combination of (1.6) and (2.11) implies that mp =l ¼ n X mpi =li : ð3:4Þ i¼1 Next, let us multiply the left- and right-hand sides of Eqs. (2.10) and (3.5), and keep in mind the validity of (2.2). One gets that n X i¼1 m2i !2 ¼ m2 ¼ n X i¼1 mip =li n X j¼1 lj m2j p : ORDER REPRINTS 1028 Vinogradov After simple algebra, one derives that the latter formula yields the following equation: X X lj p 2 p li p 2 p X X mi mj ¼ m mj þ mj mi 2 : li i lj 1i<jn 1i<jn Note that under the conditions imposed, all mi ’s are positive and finite. Hence, the latter formula can be rewritten as follows: sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi !2 X X lj p 2 p li p 2 p mi mj m mi ¼ 0: li lj j 1i<jn Evidently, this implies that l1 m11 p ¼ ¼ ln m1n p . The rest is trivial. & ACKNOWLEDGMENTS I thank G.M. Feldman, A. Feuerverger and Y.-H. Wang for their advice and an anonymous referee for very useful suggestions. REFERENCES Bar-Lev, S. K., Enis, P. (1986). Reproducibility and natural exponential families with power variance functions. Ann. Statist. 14:1507–1522. Bertoin, J. (1996). Lévy Processes. Cambridge University Press. Chhikara, R. S., Folks, J. L. (1989). The Inverse Gaussian Distribution. Theory, Methodology, and Applications. New York: Marcel Dekker. Cochran, W. G. (1934). The distribution of quadratic forms in a normal system, with applications to the analysis of covariance. Proc. Camb. Phil. Soc. 30:178–191. Cramér, H. (1970). Random Variables and Probability Distributions. 3rd ed. Cambridge University Press. Dykstra, R. L., Hewett, J. E. (1972). Examples of decompositions of chi-squared variables. Amer. Statist. 26(4):42–43. Hougaard, P. (1986). Survival models for heterogeneous populations derived from stable distributions. Biometrika 73:387–396. Hougaard, P. (2000). Analysis of Multivariate Survival Data. New York: Springer. ORDER Power-Variance Family REPRINTS 1029 James, G. S. (1952). Notes on a theorem of Cochran. Proc. Camb. Phil. Soc. 48:443–446. Johnson, N. L., Kotz, S., Kemp, A. W. (1992). Discrete Univariate Distributions. 2nd ed. New York: Wiley. Johnson, N. L., Kotz, S., Balakrishnan, N. (1994). Continuous Univariate Distributions. Vol. 1. 2nd ed. New York: Wiley. Jørgensen, B. (1997). The Theory of Dispersion Models. London: Chapman & Hall. Jørgensen, B., Vinogradov, V. (2002). Convergence to Tweedie models and related topics. In: Balakrishnan, N., ed. Advances on Theoretical and Methodological Aspects of Probability and Statistics. London: Taylor & Francis, pp. 473–489. Küchler, U., Sørensen, M. (1997). Exponential Families of Stochastic Processes. New York: Springer. Linnik, Yu. V. (1964). Decomposition of Probability Distributions. New York: Dover. Linnik, Yu. V., Ostrovskii, I. V. (1977). Decomposition of Random Variables and Vectors. Providence: AMS. Lukacs, E. (1970). Characteristic Functions. 2nd ed. London: Griffin. McCullagh, P., Nelder, J. A. (1989). Generalized Liner Models. 2nd ed. Boca Raton: Chapman & Hall=CRC. Raikov, D. A. (1938). On the decomposition of Gauss’ and Poisson’s laws. Izvestia Akademii Nauk SSSR. Series A, 91–124 (in Russian). Samorodnitsky, G., Taqqu, M. S. (1994). Stable Non-Gaussian Random Processes: Stochastic Models with Infinite Variance. New York: Chapman & Hall. Teicher, H. (1954). On the factorization of distributions. Ann. Math. Statist. 25:769–774. Tweedie, M. C. K. (1984). An index which distinguishes between some important exponential families. In: Ghosh, J. K., Roy, J., eds. Statistics: Applications and New Directions. Proceedings of the Indian Statistical Institute Golden Jubilee International Conference, Calcutta, Indian Statistical Institute, pp. 579–604. Vinogradov, V. (2002). On a class of Lévy processes used to model stock price movements with possible downward jumps. C.R. Math. Rep. Acad. Sci. Canada 24(4):152–159. Zolotarev, V. M. (1986). One-dimensional Stable Distributions. Providence: AMS. Copyright of Communications in Statistics: Theory & Methods is the property of Taylor & Francis Ltd and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use.
© Copyright 2026 Paperzz