On Generalized Interval Orders
On Generalized Interval Orders
Introduction
A partially ordered set (X, ) is called an interval order if there is a function I assigning to each element x X a closed interval I (x) = [mx, Mx ] of the real line R so that for all x, y X , y x if and only if my > Mx in R. The concept of an interval order was introduced by Peter C. Fishburn in [5] to build models of preferences whose associated indierences may fail to be transitive. Comparing intervals is a frequently encountered problem in preference modelling, decision aid and experimental psychology. This is due to the fact that the comparison judgments in experimental psychology (perception, evaluation of objects, persons) as well as the comparison of alternatives in economic theory (outcomes, objects, candidates, etc) generally are realized through their evaluations on numerical scales, while such evaluations often are imprecise or uncertain. Thus, taking an example from economic theory, an interval order is obtained when one considers that an alternative is preferred to another i its interval is completely to the rightof the other, while any two alternatives, the interval of which have a non-empty intersection are considered indierent. A classical result in interval orders is that of Fishburn in [5]; which examines the issue of representing an interval order by means of two functions u and v (x y if and only if u(x) > v (y )). Numerical representations are used in
1
1296
order to look for maximal elements of binary relations in the set of alternatives. Mathematical optimization tools are applied more easily with numerical representations of binary relations. Existence of maximal elements of binary relations has interested economists for long. This kind of relations appear e.g. in preference theory, in equilibrium theory and in the analysis of dynamic systems. In all of these choice problems, the interpretation of existence of maximal elements means that there exists a choice for which there exist no strictly better choices. In the case of considering the set of alternatives as a topological space, conditions for the existence of maximal elements are given by using topological conditions on the alternatives set as well as continuity assumptions on the relation. One of the more general continuity assumption on nding the maximal choice set of an acyclic binary relation R, is that of upper tc-semicontinuity dened by Alcantud and Rodr iguez-Palmero in [1]. This denition uses the transitive closure of the relation R. As the set of maximal choices (maximal choice set) is empty and, hence, does not help much when R contains a top cycle 2, we turns to the concept of Generalized Optimal Choice Set (GOCS), which is a generalization of the maximal choice set. The origin of GOCS can be found in Schwartz [7]. Schwartz proved that the GOCS is the maximal choice set of the asymmetric part of the transitive closure of R (cf. [14, Corollary 6.2.2]). Van Deemen in [9] proves that if the set of alternatives is nite, then the GOCS is never empty. The existence of maximal elements for interval orders have been mainly investigated by Campbell and Walker in [4] and Alcantud in [2]. More precisely, Campbell and Walker provide sucient conditions to ensure the existence of maximal elements for interval orders and Alcantud characterize the existence of maximal elements of interval orders. In this paper, we extend the notion of interval order in non-transitive binary relations. This new notion which is called generalized interval order it allows us to examine the representation of binary relations by means of two functions, in a general framework, where the axioms of transitivity and irreexivity do not hold. Here, we provide a generalization of the classical result of Fishburn in [5]. Finally, we use the notion of GOCS set (which generalize the notion of maximal choice set), as it is dened by Schwartz in [7] and we extend the results of Campbell and Walker in [4] and Alcantud in [2]. All of the results known in the literature about representations of binary relations and the existence of their maximal elements, make use of conditions (e.g. continuity conditions) based either on the binary relation itself or on its transitive closure. After the above mentioned result of Schwartz, a natural generalization of the
Consider a commodity space X and a binary relation R dened on X . A path from x to y is a nite sequence of distinct points x1 , x2 , ..., xn such that xRx1 , x1Rx2 , ..., xnRy. A path from x to y becomes a cycle when the pair yRx is added. The nite sequence (x1 , x2, ..., xn) is a top cycle in X i (i) (x1 , x2 , ..., xn) is a cycle in X and (ii) there is no x X \ {x1 , x2, ..., xn} and there is no y {x1 , x2 , ..., xn} such that xRy .
2
1297
well known notions (e.g. weakly upper semicontinuity), which will ensure the compatibility of the extended results with the classical ones, is to replace the transitive closure of a binary relation with the asymmetric part of the transitive closure of this relation. This generalization transform the problem of the existence of GOCS into the problem of the existence of maxima of a real-valued function.
We recall some denitions from [1], [5], [8]. Let X be a non-empty universal set of alternatives, and let R X X be a binary relation on X . The asymmetric part P (R) of R is denoted by P (R) = {(x, y ) X X : (x, y ) R and (y, x) / R}. We sometimes abbreviate (x, y ) R as xRy . The binary relation R is a strict partial order if it is irreexive (we never have (x, x) R) and transitive (for all x, y, z X, xRzRy implies xRy ). We often use to denote a strict partial order. The transitive closure of a relation R is denoted by R, that is for all x, y X , (x, y ) R if there exists k N and x0 , ..., xK X such that x = x0 (xk1 , xk ) R for all k {1, ..., K } and xK = y . Clearly, R is transitive and, because the case K = 1 is included, it follows that R R. Acyclicity says that there do not exist K and x0 , x1, ..., xK such that x = x0 , (xk1 , xk ) R for all k {1, ..., K } and xK = x. An element x X is said to be a maximal element of the binary relation R on X , if (y, x) P (R) does not hold for all y X . According to Schwartz, the Generalized Optimal Choice Set (GOCS) of a binary relation R is dened to be the maximal choice set of the asymmetric part of the transitive closure of R (P (R) in symbols). It is clear that for each R X X , P (R) denes a strict partial order on X . An irreexive binary relation R is an interval order if it satises the / R following interval order condition: (x, x ) R, (y , y ) R and (y , x ) imply (x, y ) R. Irreexivity and the interval order condition imply that R is transitive. An asymmetric binary relation R (P (R) = R) is weakly upper semicontinuous if whenever xRy there is a neighborhood U of y such that zRx is false for each z U . Associated with R we consider the binary relation dened as x y if neither xRy nor yRx. The relation is said to be the indierence associated to R. The strong indierence relation for R is dened by x y if for all z X, x z y z .
Denition 1 A binary relation R on X satises the transitive (tc)-interval order condition if and only if: for each x, y, x , y X if (x, x ) R and (y , y ) R then (x, y ) R or (y , x ) R. In this case, R is called generalized interval order.
1298
An equivalent expression for the tc-interval order condition is the following: (x, x ) R, (y , y ) R and (y , x ) / R (x, y ) R
Examples 1 (i) A non-irreexive or non-transitive binary relation satisng the interval order condition is a generalized interval order which is not an interval order (cf. Proposition 6 below). (ii) Let X = {1, 2, 3, 4} and R = {(1, 1), (1, 2), (2, 3), (3, 4)}. Then R is a non-irreexive, non-transitive generalized interval order which does not satises the interval order condition. Denition 2 Let R be a binary relation on X . Then the relation R of Rindierence and the relation R of strong R-indierence are dened as follows: x R y if neither xP (R)y nor yP (R)x and x R y if for all z X, x R z y R z. If R is a strict partial order then the notions R-indierence and strong R-indierence coincide, respectively, with the classical notions of indierence and strong indierence. If is an equivalence relation on a topological space (X, ), then the quo, and its elements tient set by this equivalence relation will be denoted by X which (equivalence classes) by [x] . Let the projection map : X X X carries each point of X to the element of that contains it. In the quotient is open in X if and only if 1(U ) topology induced by , a subset U of X X is open in X . Thus, the typical open set in is a collection of equivalence classes whose union is an open set in X . The quotient topology associated with a topological space (X, ) and an equivalence relation will be denoted by quot (). The equivalent relation on X saturates an A if x A and y x implies y A. In this case we say that A is a -saturated set. If is the projection map associated with the equivalence relation on X , then the condition A = 1 ( (A)) for A is equivalent to saturates A. Denition 3 A binary relation R is -saturated weakly upper semicontinuous (SWUS), if whenever xP (R)y there is a -saturated open neighborhood U of y such that zRx is false for each z U . Finally, we dene a binary relation R on
X R
by :
1299
Results
As we mention in the introduction, we are interested to extend the classical result of Fishburn in [5] to the class of generalized interval orders. First, we give three auxiliary propositions. Proposition 2 Let R be a binary relation on X . Then x P (R) z and z R y imply x P (R) y and x R z and z P (R) y imply x P (R)y . Proof. We prove that x P (R) z and z R y imply x P (R) y . The rest is / P (R). proved similarly. Suppose that x P (R) z and z R y . Then (y, x) Suppose that (x, y ) / P (R). Then x R y . Hence, from z R y we conclude that x R z which contradicts x P (R) z. Thus, x P (R) y. Proposition 3 Let R be a binary relation dened on a nonempty set X . If R satises the interval order condition, then R is a generalized interval order. Proof. Let R be a binary relation and let x, x , y, y X such that (x, x ) R, / R. Then, there exist K N and x0 , ..., xK X such (y , y ) R and (y , x ) that x = x0, (xk1 , xk ) R for all k {1, ..., K } and xK = x . There also exist M N and y0 , ..., yM X such that y = y0 , (ym1 , ym ) R for all m {1, ..., M } and yM = y . Then, from (xK 1 , x ) R, (y , y1 ) R and / R we conclude that (xK 1 , y1 ) R. Hence, (x, y ) R. (y , x ) If R is a partial order, then the notions of generalized interval order and interval order coincide.
Proposition 4 If R is a generalized interval order dened on a nonempty set X , then P (R) is an interval order on X . Proof. It is clear that P (R) is irreexive. Let now x, x , y, y X such / P (R). We must prove that that (x, x ) P (R), (y , y ) P (R) and (y , x ) (x, y ) P (R). We have two cases: / R. Then, by (x, x ) R and (y , y ) R we conclude that Case 1. (y , x ) (x, y ) R. Suppose that (y, x) R. Then (y , x ) R which is impossible. Hence (x, y ) P (R). Case 2. (x , y ) R and (y , x ) R. Then (x, y ) R. If (y, x) R, then (y, y ) R which is impossible. Hence (x, y ) P (R). The proof of the following theorem which generalize the main result of Fishburn in [5], is inspired by the result of Bridges in [3] for interval orders.
Theorem 5 Let R be a binary relation on a set X such that Then the following conditions are equivalent: (i) R is a generalized interval order.
X R
is countable.
1300
(ii) There exist mappings u, v : X R, with v u, such that, for all x, y X xP (R)y if and only if u(x) > v (y ). iii) There exist mappings u, : X , with > 0, such that, for all x, y X xP (R)y if and only if u(x) > u(y ) + (y ). Proof. (iii) (ii) (i) is evident. We prove that (i) (iii). Let R be a X = {[x0]R , [x1]R , ...}. Let also R generalized interval order on X , and let R be the R-quotient relation of R. Then, by Proposition 5, it is easy to verify that xP (R)y if and only if [x]R R [y ]R .
X Hence, R is a strict partial order in . By Proposition 7, R also satisR X X es the interval order condition in R . For each [x]R in , we dene R
nS ([x] )
R
/ R and ([xm] , [xn ] ) R }. Dene also, u([x] ) = N, ([xm ] , [x] ) R R R R R n 2 if S ([x]R ) = and u([x]R ) = 0 if S ([x]R ) = as well as
nT ([x] )
R
v([x]R ) =
For [x] = [xj ] we put ([x] ) = 2j1 + v ([x] ) u([x] ). Clearly, R R R R R S ([x]R ) T ([x]R ); so that v ([x]R ) u([x]R ), and therefore ([x]R ) > 0. Let x, y X such that xP (R)y and let [y ] = [xj ] for some j N. Then, [x] R[xj ] , and therefore T ([y ] ) S ([x] ). R R R R From j S ([x]R )\T ([y ]R ) we conclude that Let : X = . Then,
X R
R R
u([x]R ) (v[y ]R ) + 2j > u([y ]R ) + ([x]R . be the projection map. We put u = u , v = v and
u(x) = u (x) = u([x]R ) > u([y ]R ) + ([y ]R ) = u(y ) + (y ) On the other hand, if (x, y ) / P (R), then S ([x]R ) T ([y ]R ), and so u(y ) + (y ) = u([y ] ) + ([y ] ) u([x] ) = u(x).
R R R
Hence, u(x) u(y ) + (y ) which implies that xP (R)y u(x) > u(y ) + (y ). As a corollary of the above Theorem, we have the main theorems of Fishburn in [5] and Bridges in [3].
1301
We extend the notion of R-maximal element dened by Suzumura in [8], as follows: Denition 4 An element x X is said to be an optimal element of the binary relation R on X , if (y, x) P (R) does not hold for all y X . The set of optimal elements in X constitutes its GOCS. In transitive binary relations the notion of optimal element coincides with the notion of R-maximal element. In order to characterize the existence of optimal elements in GIOs we need the following theorem of Campbell and Walker in [4]. Theorem 6 Let (X, ) be a compact topological space. Then, every weakly upper semicontinuous interval order on X has a maximal element. Lemma 7 Let R be a binary relation dened on a set X and let R be the X X X . A [z ]R in is a maximal element on R -quotient relation of R on R R R if and only if each t [z ]R is an optimal element on X .
X . If we suppose that a Proof. Let [x] be a maximal element on R R t [x ]R is not an optimal element on X , then there is some y X such that
Theorem 8 Let (X, ) be a compact topological space, and let R be a Rsaturated weakly upper semicontinuous GIO. Then, the Generalized Optimal Choice Set of R is non-empty. Proof.
X , quot (R)) satises the supFirstly, we shall show that R in ( R
positions of Theorem 11. Indeed, R is a partial order and by denition of X , quot (R )) is compact. By Proposition 7, we also quot (R)), the space ( R
conclude that R satises the interval order condition. Finally, R is weakly upX . per semicontinuous. To check it, let [x]R R[y ]R for some [x]R , [y ]R R Then xP (R)y . Hence, there is a R -saturated open neighborhood U of y such that zRx is false for each z U . Hence, 1 ( (U )) = U which implies that (U ) is quot (R ))-open neighborhood of [y ] . Thus, R satises the weakly
R
1302
X ( , quot (R)). Then, Lemma 11 implies that the set of optimal elements is R non-empty. We now use Theorem 12 in order to prove the following characterization of the existence of optimal elements in the class of GIOs, which extends the Theorem 7 of Alcantud in [2].
Theorem 9 Let R be a GIO dened on X . The following conditions are equivalent: (a) The Generalized Optimal Choice Set of R is non-empty, (b) there exists a compact topology on X such that R is R -saturated weakly upper semicontinuous. Proof. (b) implies (a) by Theorem 12. We now prove that (a) implies (b). Let M be the set of optimal elements (GOCS) in X . Let be the excluded set topology generated by M [6, p.48] (this has as open sets all those subsets of X which are disjoint from M, together with X itself). Then X is compact under since every open cover of X includes X itself. Hence, {X } is always a nite subcover. It remains to prove that R is R -saturated weakly upper semicontinuous. Indeed, let x, y X such that xP (R)y . Then, for each z M, the set K = {t X : xP (R)t} does not contain z . Hence, M K = . Thus, K is an open neighborhood of y . It is clear that for each t K , (t, x) / R. Indeed, if (t, x) R for some t K , then tP (R)xP (R)t or (x, t) I (R) which implies that (x, t) / P (R), an absurdity. Finally, Proposition 5 implies that K is R -saturated.
References
[1] Alcantud J., Rodriguez-Palmero,C., Orderability of topological spaces by continuous preferences, Int. J. Math. Math. Sci., 27, (2001), 505-512. [2] Alcantud J., Characterization of the existence of maximal elements of acyclic relations, Econ. Theory., 19, (2002), 407-416. [3] Bridges D., Numerical representation of intransitive preference on a countable set, J. Econ. Theory, 30, (1983), 213-217. [4] Campbell, D. E., Walker, M., Maximal elements of weakly continuous relations, J. Econ. Theory, 50, (1990), 459-464. [5] Fishburn P., Intransitive indierence with unequal indierence intervals, J. Math. Psychology, (1970), 144-149. [6] Lynn, A. S., J. Arthur Seebach, Counterexamples in Topology, Campbell, D. E., Springer-Verlag, New York, (1978).
1303
[7] Schwartz, T., The logic of Collective Choice, New York: Columbia University Press. [8] Suzumura K., On constrained dual recoverability theorems, Math. Soc. Sci., 45, (2002), 143-154. [9] Van Deemen M.A., Coalition formation and social choice, Academic Publishers, Dordrecht, 1997. Received: October 21, 2006