The Lukacs–Olkin–Rubin Theorem on Symmetric Cones Without Invariance of the “Quotient”

Journal of Theoretical Probability, Dec 2014

We prove the Lukacs–Olkin–Rubin theorem without invariance of the distribution of the “quotient,” which was the key assumption in the original proof of (Olkin–Rubin in Ann Math Stat 33:1272–1280, 1962). Instead, we assume existence of strictly positive continuous densities of respective random variables. We consider the (cone variate) “quotient” for any division algorithm satisfying some natural conditions. For that purpose, a new proof of the Olkin–Baker functional equation on symmetric cones is given.

A PDF file should load here. If you do not see its contents the file may be temporarily unavailable at the journal website or you do not have a PDF plug-in installed and enabled in your browser.

Alternatively, you can download the file locally and open with any standalone PDF reader:

https://link.springer.com/content/pdf/10.1007%2Fs10959-014-0587-3.pdf

The Lukacs–Olkin–Rubin Theorem on Symmetric Cones Without Invariance of the “Quotient”

Bartosz Koodziejek Mathematics Subject Classification We prove the Lukacs-Olkin-Rubin theorem without invariance of the distribution of the quotient, which was the key assumption in the original proof of (Olkin-Rubin in Ann Math Stat 33:1272-1280, 1962). Instead, we assume existence of strictly positive continuous densities of respective random variables. We consider the (cone variate) quotient for any division algorithm satisfying some natural conditions. For that purpose, a new proof of the Olkin-Baker functional equation on symmetric cones is given. The Lukacs [17] theorem is one of the most celebrated characterizations of probability distributions. It states that if X and Y are independent, positive, non-degenerate random variables such that their sum and quotient are also independent, then X and Y have gamma distributions with the same scale parameter. This theorem has many generalizations. The most important in the multivariate setting was given by Olkin and Rubin [21] and Casalis and Letac [7], where the authors extended characterization to matrix and symmetric cones variate distributions, respectively. There is no unique way of defining the quotient of elements of the cone - of positive definite symmetric matrices +, and in these papers, the authors have considered very general form U = g( X + Y ) X gT ( X + Y ), where g is the so-called division algorithm, that is, g(a) a gT (a) = I for any a +, where I is the identity matrix and g(a) is invertible for any a + (later on, abusing notation, we will write g(x)y = g(x) y gT (x), that is, in this case, g(x) denotes the linear operator acting on +). The drawback of their extension was the additional strong assumption of invariance of the distribution of U under a group of automorphisms. This result was generalized to homogeneous cones in Boutouria et al. [6]. There were successful attempts in replacing the invariance of the quotient assumption with the existence of regular densities of random variables X and Y . Bobecka and Wesoowski [2] assuming existence of strictly positive, twice differentiable densities proved a characterization of Wishart distribution on the cone + for division algorithm g1(a) = a1/2, where a1/2 denotes the unique positive definite symmetric root of a +. These results were generalized to all non-octonion symmetric cones of rank >2 and to the Lorentz cone for strictly positive and continuous densities by Koodziejek [12, 13]. Exploiting the same approach, with the same technical assumptions on densities as in Bobecka and Wesoowski [2], it was proven by Hassairi et al. [11] that the independence of X + Y and the quotient defined through the Cholesky decomposition, i.e., g2(a) = Ta1, where Ta is a lower triangular matrix such that a = Ta TaT +, characterizes a wider family of distributions called Riesz (or sometimes called RieszWishart). This fact shows that the invariance property assumed in Olkin and Rubin [21] and Casalis and Letac [7] is not of technical nature only. Analogous results for homogeneous cones were obtained by Boutouria [4, 5]. In this paper, we deal with the density version of LukacsOlkinRubin theorem on symmetric cones for division algorithm satisfying some natural properties. We assume that the densities of X and Y are strictly positive and continuous. We consider quotient U for an arbitrary, fixed division algorithm g as in the original paper of Olkin and Rubin [21], additionally satisfying some natural conditions. In the known cases (g = g1 and g = g2), this improves the results obtained in Bobecka and Wesoowski, Hassairi et al. and Koodziejek [2, 11, 13]. In general case, the densities of X and Y are given in terms of, so-called, w-multiplicative Cauchy functions, that is functions satisfying f (x) f (w(I )y) = f (w(x)y) , (x, y) where w(x)y = w(x) y wT (x) (i.e., g(x) = w(x)1 is a division algorithm). Consistently, we will call w a multiplication algorithm. Such functions were recently considered in Koodziejek [14]. Unfortunately, we cannot answer the question whether there exists division (or equivalently multiplication) algorithm resulting in characterizing other distribution than Riesz or Wishart. Moreover, the simultaneous removal of the assumptions of the invariance of the quotient and the existence of densities remains a challenge. This paper is organized as follows. We start in the next section with basic definitions and theorems regarding analysis on symmetric cones. The statement and proof of the main result are given in Sect. 4. Section 3 is devoted to consideration of w-logarithmic 2 Preliminaries Cauchy functions and the OlkinBaker functional equation. In that section, we offer much shorter, simpler and covering more general cones proof of the OlkinBaker functional equation than given in Bobecka and Wesoowski, Hassairi et al. and Koodziejek [2,11,13]. In this section, we give a short introduction to the theory of symmetric cones. For further details, we refer to Faraut and Kornyi [8]. A Euclidean Jordan algebra is a Euclidean space E (endowed with scalar product denoted x, y ) equipped with a bilinear mapping (product) (x, y) xy E and a neutral element e in E such that for all x, y, z in E: For x E let L(x) : E E be linear map defined by L(x)y = xy, P(x) = 2L2(x) L x2 . The map P : E E nd(E) is called the quadratic representation of E. An element x is said to be invertible if there exists an element y in E such that L(x)y = e. Then, y is called the inverse of x and is denoted by y = x1. Note that the inverse of x is unique. It can be shown that x is invertible if and only if P(x) is invertible, and in this case, (P(x))1 = P x1 . Euclidean Jordan algebra E is said to be simple if it is not a Cartesian product of two Euclidean Jordan algebras of positive dimensions. Up to linear isomorphism, there are only five kinds of Euclidean simple Jordan algebras. Let K denotes either the real numbers R, the complex ones C, quaternions H or the octonions O and write Sr (K) for the space of r r Hermitian matrices with entries valued in K, endowed with the Euclidean structure x, y = Trace (x y ) and with the Jordan product where x y denotes the ordinary product of matrices and y is the conjugate of y. Then, Sr (R), r 1, Sr (C), r 2, Sr (H), r 2, and the exceptional S3(O) are the first four kinds of Euclidean simple Jordan algebras. Note that in this case xy = 21 (x y + y x), P(y)x = y x y. (x0, x1, . . . , xn) (y0, y1, . . . , yn) = xi yi , x0 y1 + y0x1, . . . , x0 yn + y0xn . (3) To each Euclidean simple Jordan algebra, one can attach the set of Jordan squares i=0 = x2 : x E . The interior is a symmetric cone. Moreover, is irreducible, i.e., it is not the Cartesian product of two convex cones. One can prove that an open convex cone is symmetric and irreducible if and only if it is the cone of some Euclidean simple Jordan algebra. Each simple Jordan algebra corresponds to a symmetric cone; hence, there exist up to linear isomorphism also only five kinds of symmetric cones. The cone corresponding to the Euclidean Jordan algebra Rn+1 equipped with Jordan product (3) is called the Lorentz cone. We denote by G(E) the subgroup of the linear group G L(E) of linear automorphisms which preserves , and we denote by G the connected component of G(E) containing the identity. Recall that if E = Sr (R) and G L(r, R) is the group of invertible r r matrices, elements of G(E) are the maps g : E E such that there exists a G L(r, R) with g(x) = a x aT . K = {k G : ke = e}. A multiplication algorithm is a map G : x w(x) such that w(x)e = x for all x . This concept is consistent with, so-called, division algorithm g, which was introduced by Olkin and Rubin [21] and Casalis and Letac [7], that is a mapping x g(x) G such that g(x)x = e for any x . If w is a multiplication algorithm, then g = w1 (that is, g(x)w(x) = w(x)g(x) = I d for any x ) is a division algorithm and vice versa; if g is a division algorithm, then w = g1 is a multiplication algorithm. One of two important examples of multiplication algorithms is the map w1(x) = P x1/2 . We will now introduce a very useful decomposition in E, called spectral decomposition. An element c E is said to be a idempotent if cc = c = 0. Idempotents a and b are orthogonal if ab = 0. Idempotent c is primitive if c is not a sum of two non-null idempotents. A complete system of primitive orthogonal idempotents is a set (c1, . . . , cr ) such that ci = e and ci c j = i j ci for 1 i j r. The size r of such system is a constant called the rank of E. Any element x of a Euclidean simple Jordan algebra can be written as x = ri=1 i ci for some complete system of primitive orthogonal idempotents (c1, . . . , cr ). The real numbers i , i = 1, . . . , r are the eigenvalues of x. One can then define trace and determinant of x by, respectively, tr x = ri=1 i and det x = ri=1 i . An element x E belongs to if and only if all its eigenvalues are strictly positive. The rank r and dim of irreducible symmetric cone are connected through relation = r + where d is an integer called the Peirce constant. If c is a primitive idempotent of E, the only possible eigenvalues of L(c) are 0, 21 and 1. We denote by E(c, 0), E(c, 21 ) and E(c, 1) the corresponding eigenspaces. The decomposition E = E(c, 0) E(c, 21 ) E(c, 1) is called the Peirce decomposition of E with respect to c. Note that P(c) is the orthogonal projection of E onto E(c, 1). Fix a complete system of orthogonal idempotents (ci )ri=1. Then for any i, j {1, 2, . . . , r }, we write 1 E c j , 2 if i = j. It can be proved (see Faraut and Kornyi [8, Theorem IV.2.1]) that E = i j Ei j Ei j Eii + Ei j , Ei j E jk Eik , if i = k, Ei j Ekl = {0}, if {i, j } {k, l} = . x2 = 21 x 2(ci + c j ), xy 2 = 18 x 2 y 2. The dimension of Ei j is the Peirce constant d for any i = j . When E is Sr (K), if (e1, . . . , er ) is an orthonormal basis of Rr , then Eii = Rei eiT and Ei j = K ei e Tj + e j eiT for i < j and d is equal to di m|RK. For 1 k r , let Pk be the orthogonal projection onto E(k) = E(c1 + . . . + ck , 1), det(k) the determinant in the subalgebra E(k), and, for x , k (x) = det(k)( Pk (x)). Then, k is called the principal minor of order k with respect to the Jordan frame (ck )rk=1. Note that r (x) = det x. For s = (s1, . . . , sr ) Rr and x , we write s (x) = 1(x)s1s2 2(x)s2s3 . . . r (x)sr . s is called a generalized power function. If x = 2s2 . . . rsr . We will now introduce some basic facts about triangular group. For x and y in , let x y denote the endomorphism of E defined by x y = L(xy) + L(x)L(y) L(y)L(x). If c is an idempotent and z E(c, 21 ), we define the Frobenius transformation c(z) in G by x = c1 (z(1))c2 (z(2)) . . . cr1 (z(r1)) k=1 Mapping w2 : T , x w2(x) = tx realizes a multiplication algorithm. For E = Sr (R), we have = +. Let us define for 1 i, j r matrix i j = (kl )1k,lr such that i j = 1 and all other entries are equal 0. Then for Jordan Given a Jordan frame (ci )ri=1, the subgroup of G, i=1 k= j+1 E jk is called the triangular group corresponding to the Jordan frame (ci )ri=1. For any x in , there exists a unique tx in T such that x = txe, that is, there exist (see Faraut and Kornyi [8, Theorem IV.3.5]) elements z( j) rk= j+1 E jk , 1 j r 1 and positive numbers 1, . . . , r such that frame (ci )ri=1, where ck = kk , k = 1, . . . , r , we have z jk = ( jk + k j ) E jk oraz z jk 2 = 2, 1 j, k r , j = k. if z(i) rj=i+1 Ei j , i = 1, . . . , r 1, then there exists (i) = (i+1, . . . , r ) Rri such that z(i) = rj=i+1 j zi j . Then, the Frobenius transformation reads ci (z(i))x = Fi ((i)) x Fi ((i))T , j=i+1 It can be shown (Faraut and Kornyi [8, Proposition VI.3.10]) that for each t T , x and s Rr , and for any z E(ci , 21 ), i = 1, . . . , r , s (t x) = if only s and T are associated with the same Jordan frame (ci )ri=1. We will now introduce some necessary basics regarding certain probability distribution on symmetric cones. Absolutely continuous Riesz distribution Rs,a on is defined for any a and s = (s1, . . . , sr ) Rr such that si > (i 1)d/2, i = 1, . . . , r , though its density Rs,a(dx) = sdim / r (x)e a,x I (x) dx, x , where s is the generalized power function with respect to a Jordan frame (ci )ri=1 and is the Gamma function of the symmetric cone . It can be shown that (s) = (2 )(dim r)/2 rj=1 ( s j ( j 1) d2 ) (see Faraut and Kornyi [8, VII.1.1.]). Riesz distribution was introduced in Hassairi and Lajmi [10]. Absolutely continuous Wishart distribution p,a on is a special case of Riesz distribution for s1 = . . . = sr = p. If a and p > dim / r 1 it has density (det x) pdim / r e a,x I (x) dx, x , where ( p) := ( p, . . . , p). Wishart distribution is a generalization of gamma distribution (case r = 1). In generality, Riesz and Wishart distributions does not always have densities, but due to the assumption of existence of densities in Theorem 4.2, we are not interested in other cases. 3 Functional Equations 3.1 Logarithmic Cauchy Functions As will be seen, the densities of respective random variables will be given in terms of w-logarithmic Cauchy functions, i.e., functions f : R that satisfy the following functional equation f (x) + f (w(e)y) = f (w(x)y), (x, y) where w is a multiplication algorithm. If f is w-logarithmic, then e f is called wmultiplicative. In the following section, we will give the form of w-logarithmic Cauchy functions for two basic multiplication algorithms, one connected with the quadratic representation w1(x) = P(x1/2), w2(x) = tx T . and the other related to a triangular group T , Such functions were recently considered without any regularity assumptions in Koodziejek [14]. It should be stressed that there exists infinite number of multiplication algorithms. If w is a multiplication algorithm, then trivial extensions are given by w(k)(x) = w(x)k, where k K is fixed (Remark 4.3 explains why this extension is trivial when it comes to multiplicative functions). One may consider also multiplication algorithms of the form P(x)tx12 , which interpolate between the two main examples: w1 (which is = 1/2) and w2 (which is = 0). In general, any multiplication algorithm may be written in the form w(x ) = P(x1/2)kx , where x kx K . Functional equation (9) for w1 was already considered by Bobecka and Wesoowski [3] for differentiable functions and by Molnr [19] for continuous functions of real or complex Hermitian positive definite matrices of rank >2. Without any regularity assumptions, it was solved on the Lorentz cone by Wesoowski [23]. Case of w2(x) = tx T for a triangular group T , perhaps a bit surprisingly, leads to a different solution. It was indirectly solved for differentiable functions by Hassairi et al. [11, Proof of Theorem 3.3]. By Faraut and Kornyi [8, Proposition III.4.3], for any g in the group G, det(gx) = (Det g)r/ dim where Det denotes the determinant in the space of endomorphisms on . Inserting a multiplication algorithm g = w(y), y , and x = e, we obtain Det (w(y)) = (det y)dim / r det(w(y)x) = det y det x for any x, y . This means that f (x) = H (det x), where H is generalized logarithmic function, i.e., H (ab) = H (a) + H (b) for a, b > 0, is always a solution to (9), regardless of the choice of multiplication algorithm w. If a w-logarithmic functions f is additionally K -invariant ( f (x) = f (kx) for any k K ), then H (det x) is the only possible solution (Theorem 3.4). In Koodziejek [14], the following theorems have been proved. They will be useful in the proof of the main theorems in this paper. Theorem 3.1 (w1-logarithmic Cauchy functional equation) Let f : function such that f (x) + f (y) = f P x1/2 y , (x, y) Then, there exists a logarithmic function H such that for any x Theorem 3.2 (w2-logarithmic Cauchy functional equation) Let f : function satisfying for any x and y in the cone of rank r , ty T , where T is the triangular group with respect to the Jordan frame (ci )ri=1. Then, there exist generalized logarithmic functions H1, . . . , Hr such that for any x , f (x) = H (det x). f (x) + f (y) = f (tyx) f (x) = k=1 k is the principal minor of order k with respect to (ci )ri=1. If we assume in Theorem 3.2 that f is additionally measurable, then functions Hk are measurable. This implies that there exist constants sk R such that Hk () = sk log and k=1 f (x) = sk log( k (x)) = log k=1 Thus, we obtain the following Remark 3.3 If we impose on f in Theorem 3.2 some mild conditions (e.g., measurability), then there exists s Rr such that for any x , f (x) = log s (x). f (x) = H (det x). Theorem 3.4 Let f : R be a function satisfying (9). Assume additionally that f is K -invariant, i.e., f (kx) = f (x) for any k K and x . Then, there exists a logarithmic function H such that for any x , Lemma 3.5 (w-logarithmic Pexider functional equation) Assume that a, b, c defined on the cone satisfy following functional equation a(x) + b(y) = c(w(x)y), (x, y) Then, there exist w-logarithmic function f and real constants a0, b0 such that for any x , a(x) = f (x) + a0, b(x) = f (w(e)x) + b0, c(x) = f (x) + a0 + b0. 3.2 The OlkinBaker Functional Equation In the following section, we deal with the OlkinBaker functional equation on irreducible symmetric cones, which is related to the Lukacs independence condition (see proof of the Theorem 4.2). Henceforth, we will assume that multiplication algorithm w additionally is homogeneous of degree 1, that is, w(sx) = sw(x) for any s > 0 and x . It is easy to create a multiplication algorithm without this property, for example: w(x) = w1(x), if det x > 1, w2(x), if det x 1. The problem of solving f (x )g(y) = p(x + y)q(x /y), (x , y) (0, )2 for unknown positive functions f , g, p and q was first posed in Olkin [20]. Note that in one-dimensional case, it does not matter whether one considers q(x /y) or q(x /(x + y)) on the right- hand side of (13). Its general solution was given in Baker [1] and later analyzed in Lajk [15] using a different approach. Recently, in Mszros [18] and Lajk and Mszros [16], Eq. (13) was solved assuming that it is satisfied almost everywhere on (0, )2 for measurable functions which are nonnegative on its domain or positive on some sets of positive Lebesgue measure, respectively. Finally, a new derivation of solution to (13), when the equation holds almost everywhere on (0, )2 and no regularity assumptions on unknown positive functions are imposed, was given in Ger et al. [9]. The following theorem is concerned with an adaptation of (13), after taking logarithm, to the symmetric cone case. Theorem 3.6 (OlkinBaker functional equation on symmetric cones) Let a, b, c and d be real continuous functions on an irreducible symmetric cone of rank r . Assume a(x) + b(y) = c(x + y) + d (g (x + y) x) , (x, y) where g1 = w is a homogeneous of degree 1 multiplication algorithm. Then, there exist constants Ci R, i = 1, . . . , 4, E such that for any x and u D = {x : e x }, a(x) = b(x) = c(x) = , x + e(x) + f (x) + C3, d(u) = e(w(e)u) + f (e w(e)u) + C4, where e and f are continuous w-logarithmic Cauchy functions and C1 + C2 = C3 + C4. We will need following simple lemma. For the elementary proof, we refer to Koodziejek [13, Lemma 3.2]. Lemma 3.7 (Additive Pexider functional equation on symmetric cones) Let a, b and c be measurable functions on a symmetric cone satisfying a(x) + b(y) = c(x + y), (x, y) Then, there exist constants , R and E such that for all x Now, we can come back and give a new proof the OlkinBaker functional equation. Proof of Theorem 3.6 In the first part of the proof, we adapt the argument given in Ger et al. [9], where the analogous result on (0, ) was analyzed, to the symmetric cone setting. For any s > 0 and (x, y) 2, we get a(sx) + b(sy) = c(s(x + y)) + d (g(sx + sy)sx) . Since w is homogeneous of degree 1, we have g(sx) = 1s g(x) and so g(sx + sy)sx = g(x + y)x for any s > 0. Subtracting now (14) from (17) for any s > 0, we arrive at the additive Pexider equation on symmetric cone , as (x) + bs (y) = cs (x + y), (x, y) where as , bs and cs are functions defined by as (x) := a(sx) a(x), bs (x) := b(sx) b(x) and cs (x) := c(sx) c(x). Due to continuity of a, b and c and Lemma 3.7, it follows that for any s > 0, there exist constants (s) E, (s) R and (s) R such that for any x , By the definition of as and the above observation, it follows that for any (s, t ) (0, )2 and z ast (z) = at (sz) + as (z). Since (18) holds for any z , we see that (st ) = (s)+(t ) for all (s, t ) (0, )2. That is (s) = k1 log s for s (0, ), where k1 is a real constant. On the other hand (st ), z = (s), z + (t ), sz = (t ), z + (s), t z since one can interchange s and t on the left-hand side. Putting s = 2 and denoting = (2), we obtain , z (t 1) a (x) = a(x) a (sx) = a (x) + k1 log s b(x) = b(x) c(x) = c(x) for s > 0 and x . Analogous considerations for function bs gives existence of constant k2 such that b(sx) = b(x) + k2 log s, where hence c(sx) = c(x) + (k1 + k2) log s and for any s > 0 and x . Functions a , b, c and d satisfy original OlkinBaker functional equation: a (x) + b(y) = c(x + y) + d (g (x + y) x) , (x, y) Taking x = y = v in (22), we arrive at a (v) + b(v) = c(2v) + d( 21 e) = c(v) + (k1 + k2) log 2 + d( 21 e). for t > 0 and z . It then follows that for all s (0, ) and z as (z) = a(sz) a(z) = , z (s 1) + k1 log s. Let us define function a by formula From (20), we get Insert x = w(v)u and y = w(v)(eu) into (22) for 0 < < 1 and (u, v) (D, ) . Using (21), we obtain Let us observe, that due to continuity of b on and lim0 {w(v)(e u)} = w(v)e = v (convergence in the norm generated by scalar product , ), limit as 0 of the left-hand side of the above equality exists. Hence, the limit of the right-hand side also exists and a (w(v)u) + b(v) = c(v) + lim0 {d(u) k1 log } , (u, v) (D, ). Subtracting (24) from (23), we have a (w(v)u) = a (v) + g(u) for u D, v , where g(u) = lim0 {d(u) k1 log }(k1 +k2) log 2d( 21 e). Due to the property (21), equation (25) holds for any u , so we arrive at the w-logarithmic Pexider equation. Lemma 3.5 implies that there exists w-logarithmic function e such that a (x) = e(x) + C1 for any x and a constant C1 R. Function e is continuous, because a is continuous. Coming back to the definition of a , we obtain a(x) = , x + e(x) + C1, x . Analogously for function b, considering equation (22) for x = w(v)(e u) and y = w(v)u after passing to the limit as 0, we show that there exists continuous w-logarithmic function f such that b(x) = , x + f (x) + C2, x for a constant C2 R. The form of c follows from (23). Taking x = w(e)u and y = e w(e)u in (22) for u D, we obtain the form of d. 4 The LukacsOlkinRubin Theorem Without Invariance of The Quotient In the following section, we prove the density version of LukacsOlkinRubin theorem for any multiplication algorithm w satisfying (i) w(sx) = sw(x) for s > 0 and x (ii) differentiability of mapping We assume (ii) to ensure that Jacobian of the considered transformation exists. We start with the direct result, showing that the considered measures have desired property. The converse result is given in Theorem 4.2. For every generalized multiplication w, the family of these w-Wishart measures (as defined in (26)) contains the Wishart laws. For w = w1, there are no other distributions, while the w2-Wishart measures consist of the Riesz distributions. It is an open question whether there is a generalized multiplication w that leads to other probability measures in this family. Theorem 4.1 Let w be a multiplication algorithm satisfying condition (ii) and define g = w1. Suppose that X and Y are independent random variables with densities given by f X (x) = C X e(x) exp , x I (x), fY (x) = CY f (x) exp , x I (x), where e and f are w-multiplicative functions, E and E is the Euclidean Jordan algebra associated with the irreducible symmetric cone . Then, vector (U, V ) = (g(X + Y )X, X + Y ) have independent components. Note that if w(x) = w1(x) = P(x1/2), then there exist positive constants X and Y such that e(x) = (det x)X dim / r and f (x) = (det x)Y dim / r . In this case =: a and (X, Y ) X ,a Y ,a. Similarly, if w(x) = w2(x) = tx, X and Y follow Riesz distributions with the same scale parameter . In general, we do not know whether a = should always belong to . be a mapping defined through Then, (U, V ) = (X, Y ). The inverse mapping 1 : D (x, y) = 1(u, v) = (w(v)u, w(v)(e u)), hence is a bijection. We are looking for the Jacobian of the map 1, that is, the determinant of the linear map dx/du dx/dv dy/du dy/dv J = = Det(w(v)). Det (w (v)) = (det v)dim / r . Now, we can find the joint density of (U, V ). Since (X, Y ) have independent components, we obtain f(U,V )(u, v) = (det v)dim / r f X (w(v)u) fY (w(v)(e u)) E, C X , CY R and w-multiplicative functions f(U,V )(u, v) = (det v)dim / r f X (w(v)u) fY (w(v)(e u)) = C1C2 (det v)dim / r e(w(v)u) f (w(v)(e u)) e , v I (w(v)u)I (w(v)(e u)) = C1C2 (det v)dim / r e(v) f (v)e , v I (v) e(w(e)u) f (w(e)(e u))ID(u), = fU (u) fV (v), what completes the proof. To prove the characterization of given measures, we need to show that the inverse implication is also valid. The following theorem generalizes results obtained in Bobecka and Wesoowski, Hassairi et al. and Koodziejek [2,11,13]. We consider quotient U for any multiplication algorithm w satisfying conditions (i) and (ii) given at the beginning of this section (note that multiplication algorithms w1 and w2 defined in (10) and (11), respectively, satisfy both of these conditions). Respective densities are then expressed in terms of w-multiplicative Cauchy functions. Theorem 4.2 (The LukacsOlkinRubin theorem with densities on symmetric cones) Let X and Y be independent rvs valued in irreducible symmetric cone with strictly positive and continuous densities. Set V = X + Y and U = g (X + Y ) X for any multiplication algorithm w = g1 satisfying conditions (i) and (ii). If U and V are independent, then there exist E and w-multiplicative functions e, f such that (26) holds. In particular, Proof We start from (27). Since (U, V ) is assumed to have independent components, the following identity holds almost everywhere with respect to Lebesgue measure: (det(x + y))dim / r f X (x) fY (y) = fU (g (x + y) x) fV (x + y), where f X , fY , fU and fV denote densities of X , Y , U and V , respectively. Since the respective densities are assumed to be continuous, the above equation holds for every x, y . Taking logarithms of both sides of the above equation (it is permitted since f X , fY > 0 on ), we get a(x) + b(y) = c(x + y) + d (g (x + y) x) , a(x) = log f X (x), b(x) = log fY (x), c(x) = log fV (x) dimr d(u) = log fU (u), for x and u D. The first part of the conclusion follows now directly from Theorem 3.6. Thus, there exist constants E, Ci R, i {1, 2} and w-logarithmic functions e and f such that f X (x) = ea(x) = eC1 e(x)e , x , fY (x) = eb(x) = eC2 f (x)e , x , for any x . Let us observe that if w(x) = w1(x) = P(x1/2), then for Theorem 3.1, there exist constants i , i = 1, 2, such that e(x) = (det x)1 and f (x) = (det x)2 . Since f X and fY are densities, it follows that a = , ki = pi (dim )/ r > 1 and eCi = (det(a)) pi / ( pi ), i = 1, 2. Analogously, if w(x) = w2(x) = tx, then Theorem 3.2 and Remark 3.3 imply that there exist constants si = (si, j )rj=1, si, j > ( j 1)d/2, i = 1, 2, and a = such that X Rs1,a i Y Rs2,a. Remark 4.3 Fix k K and consider w(k)(x) = w(x)k. The w(k)-multiplicative function f satisfies equation Substituting y k1y f (x) f (w(e)ky) = f (w(x)ky). f (x) f (w(e)y) = f (w(x)y), that is w(k)-multiplicative functions are the same as w-multiplicative functions. This leads to the rather unsurprising observation that if we consider Theorem 4.2 with w(x) = P(x1/2)k or w(x) = txk, regardless of k K , we will characterize the same distributions as in points (1) and (2) of Theorem 4.2. With Theorem 4.2, one can easily reprove original LukacsOlkinRubin theorem (version of Olkin and Rubin [22] and Casalis and Letac [7]), when the distribution of U is invariant under a group of automorphisms: fU (u) = C e(w(e)u) f (e w(e)u). The distribution of U is invariant under K , thus density fU is a K -invariant function, that is fU (u) = fU (ku) for any k K . Note that w(e) K , thus e(u) f (e u) = e(ku) f (e ku), (k, u) K D. We will show that both functions e and f are K -invariant. Recall that e(x) e(w(e)y) = e(w(x)y), therefore after taking y = e, we obtain e(x) = e(x)e(e) for any > 0 and x . Inserting u = v into (30), we arrive at Thus, e(v) f (e v) = e(kv) f (e kv) for any (0, 1], v D. Since f (e) = 1 and f is continuous on , by passing to the limit as 0, we get that e is K -invariant and so is f . By Theorem 3.4 and continuity of e and f , we get that there exist constants 1, 2 such that e(x) = (det x)1 and f (x) = (det x)2 ; hence, X and Y have Wishart distributions. Acknowledgments The author thanks J. Wesoowski for helpful comments and discussions. This research was partially supported by NCN Grant No. 2012/05/B/ST1/00554. Open Access This article is distributed under the terms of the Creative Commons Attribution License which permits any use, distribution, and reproduction in any medium, provided the original author(s) and the source are credited.


This is a preview of a remote PDF: https://link.springer.com/content/pdf/10.1007%2Fs10959-014-0587-3.pdf

Bartosz Kołodziejek. The Lukacs–Olkin–Rubin Theorem on Symmetric Cones Without Invariance of the “Quotient”, Journal of Theoretical Probability, 2016, 550-568, DOI: 10.1007/s10959-014-0587-3