ta one ene THE DISTRIBUTION OF LINEAR COMBINATIONS OF ORDER STATISTICS FROM THE UNIFORM DISTRIBUTION By HERBERT WEISBERG New York University 1, Introduction, In this paper we derive an algorithm for computing the distribu tion function of an arbitrary linear combination of order statistics from a uniform distribution, Suppose U,, is the ith smallest observation from a sample of size n from the uniform distribution on (0, 1}, with the convention Up = 0, Uys, = 1. Consider a set of 5 integers {k;} such that (a) ky =O < ky < ky <0 0 and any x, we seek PALS dU uy) Sx}. Our approach is to generalize a formula derived by Dempster and Kleyle (1968). 2, Derivation of the algorithm, Let X,= Uj Cat = 0. G=1,2,--n, Let Jamays Define ¢1, €25 + ¢ by Ch = 41 di for imi," S y= Cian for j¢(ky,-* Then we have (2a) ou For the special case S = n, Dempster and Kleyle (1968) have shown that e g __(¢-x)" 2.2) Py eXiSxe= @) 1% . } Bi 4 fias(ej— 41) for 0 S x S cy. where r is the largest positive integer such that x < ¢,. In the general case S Sn, we wish to allow 1d U ayy = Lhe Xe Chyerth = Chyart2 = =k, = Cay for s = 1,2, Let k,—k,-y = Py $=, Ssand n—ks = ry 4. Then we wish to let the first rr, ¢'s take the value ¢), the next r2 take the value ¢3), ete, Let ce41) = 0. In this situation (2.2) is not applicable unless r, = | for alls. Received March 6, 1970; revised October 19, 1970, 704LINEAR COMBINATIONS OF ORDER STATISTICS 705 Suppose, however, that we define (2.3) ba, rent) = cay t(—Dh, for h> OG =A ss = S. Sinsiy A) = rsa. 1h Pe ertsee Then we have LEMMA 1. img P{Df=1 b(h)X, S x} = PTS. dU yu, Sx). Proor. Let A, = {Y?-1 b(Iir)X; S x} Am{Yhar dU Sx} = (Li 1 Xi Liao GayXet 7 Sx}. nf) =o} -| a(x ssfnm A, © Ay, and A = (J A,. Therefore Mi P(A) = P(\a1 A,) = lim, P{A,}- Suppressing for convenience the dependence of 6, on A we have from (2.2) that . yes) bea Sams a BT ie j= B) sakcvs iT vst, =x) ~ ja eB Thes-B) where m is the largest integer such that x S Cia). Let ke (by x)" Te 2 bles) -h) so that (24) PSI bX, $x} =1-TE 7 Lema 2. Fhe) TD! where (on Thnk and {\ is the forward difference operator defined by AYf@) = ah 'f(x+h)-ah 4). Ic) =706 HERBERT WEISBERG (oj= 3)" +1 bi Tie j= 0) pinay (j= 8) fo), ~1 0 for al i, the general problem can be handied by reordering and shifting variables, making use of the symmetry in the situation. For example 2U,3,—Uy) = X;+2X,+2X, has the same distribution as 2X, +2¥2+%3 = Up)+Upy. 3. Application. Following the notation of Wilks (1962) we define the (k—1)— variate Dirichlet distribution D(v,, ¥2,-* 4-15 Ys) by the density Cnty Pn 5.0) TG Hh (1 x) ok EF for x20, =0 otherwise. It is easily shown that the joint distribution of Ua, Ua,+y~ Uays Unseen Ularortty yy» £00 K 7S as in (1.1), is Dek yy kay. kys MDG Rt D). Let pi, P25 ° Pats Pa Fepresent the cell probabilities for a multinomial popu- lation with k categories. For a Bayesian analysis it is common to assume a con- jugate prior of the form D(m:, m2. ---5 m) for py, + py-1- Suppose my, 12, -- Me are integers. Let m, i = 1,---k, be the observed frequency for the ith category and n = Y4., m,. Then the posterior distribution of py, «py. is Don +m mtn)708 HERBERT WEISBERG Suppose we wish to make posterior probability statements about events of the form {Y% ap, S x} for real numbers a,,-+- a, and x. Let vj = Yio, (+n). j= 1, k. Then we have GA) PUL ap. Sx} = PIL aU Ups») S%} = Plat EP (ay aie Woy SX} where U,) is the jth smallest observation from a sample of size (yy—1) from the uniform distribution on (0, 1}. Thus the algorithm of Section 2 can be applied. For example, suppose we have k = 5, and we assume the improper prior P(O, 0, 0, 0; 0) suggested by Lindley (1964) for (P1, P2, Pa, Ps). Suppose also that =5 ny = 10 -2 my = 1S 0 ny = 10 a, 2 ng = 10 as = +5 ns = 6. From (3.1) PALI apy S x} = P{S—3U 10) as) Wes)— Was) $9) = P(3U 10) + Was) + Wyss) +3U las) Z SX}, where the order statistics are from a sample of size 50. A computer program to implement the algorithm of Section 2 has been success- fully run and used to obtain the following results for this example. x Pap, S x} 1.0 9998 8 9992 6 9967 4 9885 2 9660 0 9150 -2 8196 -4 6738 ~ 6 4929 ~ 8 3119 10 1669 -12 0741 14 0269 -16 0079 18 0018 -20 0003.