Mathematical Social Sciences 11 (1986) 19% 199 195 North- Holland NOTE ON LEXICOGRAPHIC PROBABILITY RELATIONS Uzi SEGAL Department of Econornrcs, Unrversrty of Toronto, Toronto MSS IAI, Canada Commumcated by F.W. Roush Received 10 December 1984 Revrsed 6 May 1985 In thus note 1 propose a defmnron for lextcographic probabihty relatrons and discuss the question, under what conditrons 1s a probabihty relatron lexicographic. Key words: Lexrcographrc probabrhty relatron; trght probabtlity relatton. 1. Introduction In this note I examine conditions under which a probability relation on a set of events is lexicographic. Chipman (1971) discussed the question of the minimal ordinal CTsuch that there exists an order-preserving function from an ordered set A (in his; paper A is fR with any order on it) to lRa with the lexicographic order on it. Since there always exists an ordinal /? such that there is an order-preserving function from A to the lexicographically ordered lRP (see Chipman), one can define an order as lexicographic if there is no order-preserving function from A to R. According to this definition, a countable set cannot be ordered lexicographically. Consider, however, the following example. Let Q be the set of all the finite unions of ra tional intervals [a, p) in [0,2). Q is closed under complementation and finite intersections. Define on Q a relation 2 as follows: for every A, BE Q, A Z B iff 11(~1n[O,l))>~(Bft[O,l)) or p(AfI[O,l))=p(BfI[O,l)) and p(An[l,2))> ~/t/In[l, 2)) (p denotes the Lebesgue measure). Q is countable, but seems to be lexicographic. Indeed, there exists no probability function P on Q such that A 2 B iff I’( 4) ZIP(B) (see Section 2). A natural definition of lexicographic orders seems therefore as follows. An order I? on a set X is lexicographic if there exist two orders R, and R2 on X such that for every x,y~ X, xRy iff xRly, but not yR,x; or xR,y, yR,x, and xR,y. (R, and R, may themselves be lexicographic orders.) This definition may encompass too much. For example, the decimal writing of real numbers induces on them an apparent lexicographic order. It seems, therefore, that if one wishes to avoid defining too simple orders as lexicographic, yet does not OIOS-4896/86/$3.50 CC)1986, Elsevter Scrence Publishers B.V. (North-Holland) U. Segai / Lexicographic 196 probablhty relations want to restrict the set of lexicographic orders too narrowly, one must use some properties of the space itself, as, for example, its being a vector space (see Hausner, 1954), or a space of sets, as below. In Section 2 I define lexicographic probability relations. In Section 3 I present conditions implying that a tight probability relation is lexicographic. The examples of Section 4 prove that these are sufficient, but not necessary conditions. 2. Definitions Let S be a set of states of the world and let Q 5 2’ be closed under complementation and finite intersections. Elements of Q are called events, denoted by A, B, C etc. Let Z be a binary relation on Q. A > B iff A 2 B but not B 2 A, and A-B iff A2B and B2A. A denotes S\A. Definition. A relation 2 on Q c 2’ is called a probability relation if (a) 2 is complete and transitive; (b) for every AEQ, A20 and S>0; (c) for every A,B,CEQ such that A(IC=BflC=O, A2B iff AUCkBUC. Proposition. Let 2 and AflC=BflD=O, be a probability relation on Q and let A, B, C, DE Q. If A -B then CLD iff AUC2BUD. Proof. A-B#(A\B)U(C\(BUD))-(BUC)\(AUD); C2De (BUC)\(AUD) eAUCZBUD. hence, L (D\C)U(B\(AUC)) 0 Definition. A probability relation 2 on Q is lexicographic if there exists an event A >0 such that: (1) ,?i includes an infinite number of non-equivalent events; (2) for every B, CE Q, BLC iff Bela > Cola, or Bfl&--Cola and BiTA CnA. Lemma 1. A probability relation 2 on Q is lexicographic iff there exists A > 0 satisfying (l), such that (3) for every B;C c A, B > Ce B > CUA. Proof.(2)~(3):LetB,C5AsuchthatB>C.BnA=B>C=(CUA)nA,henceby (2), B > CUA. (3)*(2): Let B2C. (Bfl&UA2B1CkCflA; hence, BnAkCnA. Assume I/, Segal / Lexrcographlc probability relatrons that BnA-CfiA. (BnA)U(BnA)=B2C=(CfM)U(CflA). tion, BnA z cnA. lfBnA>cnA, thenBLBnA>(cnA)uA2C. Ifsnii-cnA Cfl A, then by the proposition, BZ C. 0 111the next section I will use the following definitions 197 By the proposiandBnAZ and theorem, taken from Savage (1972): Definition. Let Q c 2’ be a space it. P: Q-+ [0, l] is a probability P(A) + P(B) - P(A fll?). P strictly agrees with 2 iff A Z B * P(A) 2 of events and let 2 be a probability relation on function if P(O)=O, P(S)= 1, and P(AUB)= agrees with 2 iff A 2 B H P(A) 1 P(B). P almost P(B). Definition. A probability relation Z on Q c 2’ is fine if for every A > 0 there exists a partition of S, with all its elements less probable than A. A probability relation is tight if for every A >B there exists C> 0 such that BflC=O and A > BUC. Theorem. If 2 on Q is fine, then there is one and only one probability function almost agreeing with 2. I.> 3. Lexicographic If 2 is fine and tight, then this function strictly agrees with probability relations Obviously, if L is lexicographic, then it cannot be represented by a strictly agreeing probability function. In this section I demonstrate conditions under which a probability relation which cannot be represented by a strictly agreeing probability function is lexicographic. Lemma 2. Let 2 be a tight probability relation on Q and let P be an almost agreeing probability function on Q. If there exists A > 0 satisfying (I), such that (4) P(A) = 0, and for every DE Q, P(D) = 0 implies A 2 D, then 2 is lexicographic. Proof. Let A > 0 satisfying (1) and (4), and let B, CC A, B > C. By the definition, HUA > CUA. Since 2 is tight, there exists D>0 such that (CUA)flD=O and P(B)=P(BUA)>P(CUAUD)= HUA > CUAUD. By (4), P(AUD)>O. P(C) + P(A UD)> P(C) = P(CUA); hence, B > CUA. By Lemma 1, 2 is lexicographic. 0 This lemma note. cannot be reversed. A counter-example appears at the end of this U. &gal / Lexicographic probability reMions 198 Definition. A set of events Z= {Ai : ie I} is a chain if for every i, jc I, i# jl it follows that A,5Ai or Aj($Ai. 2 is a maximal chain of events which have a certain property if for every event B$Z which has this property, ZU {B} is not such a chain. Lemma 3. Let 2 be a probability relation on Q. IfP almost agrees with 2, then there exists a maximal chain of events with probability 0. Proof. Denote by r the set of all the chains of events with probability 0. Elements of f are partially ordered by the inclusion relation. Let {Zj :jE J} be a chain of elements out of r. Obviously, U jezZj is an upper boundary for this chain. By the Lemma of Zorn (see Halmos, 1960), there exist in rmaximal elements; hence, there exists a maximal chain of events with probability 0. Obviously, more than one such chain may exist. Cl Theorem. Let 2 be a tight probability relation on Q. If in one of the maximal chains of events with probability 0 there exists a maximal event (containing all the events of this chain), then 2 is lexicographic. Proof. Let A be a maximal event of a maximal chain of events with probability 0. 2 is tight; hence, there exists no B > A such that P(B) = 0. The theorem now follows from Lemma 2. 0 4. Examples The question arises whether there must always exist a maximal event with prob ability 0. If so, then by the theorem a tight probability relation which can be represented by an almost agreeing probability function (but not by a strictly agreeing probability function) must be lexicographic. The answer to this question is not necessarily affirmative. Example. S= [0, a~), Q is the set of all the finite unions of intervals [a, /3) in S (8~ 00). Define on Q a probability relation Z as follows. A >B iff &A)>p(B), or p(A) =p(B) = 00 and &I)>&). 2 is tight. Obviously, the only one probability function P on Q is given by P(A) = 0, dA)<=+ 1, ~(A)=oc Of course, there is no maximal event with probability 0, and 2 is not lexicographic. Indeed, let A EQ such that p(A <- and let BCa such that 0<&3)1 p(A). B > 0, but 0UA 2 B. By Lemma 1, 2 is not lexicographic. I/. Segal / Lexwographrc probabllrty relatrons 199 Denote this probability relation on [0,00) by 2 *. The following example presents a tight and lexicographic probability relation which can be represented by an almost agreeing probability function, but does not include any maximal events with probability 0. In other words, Lemma 2 cannot be reversed. Example. S= [- 1, oo), Q is the set of all the finite unions of intervals [Q, p) in S. A 2 B iff A fl [0,00) > *Bn [0, a), or A n [0,m) -*Bn[O,m) and &lfl[-1,O))r p(Bn I - 1,w. Acknowlegements I am grateful to R. Holzman, for their comments. J. Mirrlees, J. Sobel, and especially M.E. Yaari, References J.C Chipman On the lexicographic representation of preference ordering, in: J.C. Chlpman, L. Horowitz, M.K. Richter and H.F. Sonnenschem, eds., Preference, Utlhty and Demand (Academic Press, New York, 1971). P.RHaImos, Naive Set Theory (Van Nostrand, Princeton, 1960). M Hausner, Multidimensional utihtles, m: R.M. Thrall, C.H. Coombs, and R.L. Davis, eds., Decision Processes (Wiley, New York, 1954). L.1 Savage The Foundations of Statistics, second printing (Dover, New York, 1972).