Journal of Mathematical Economics 14 (1985) 129-134. North-Holland ON THE SEPARABILITY OF THE QUASI CONCAVE CLOSURE OF AN ADDITIVELY SEPARABLE FUNCTION Uzi SEGAL University of Toronto, Toronto MS8 IA1 Canada Received May 1983, iinal version This paper gives a characterization closure also is additively separable. of the additively accepted separable March 1985 functions whose quasi concave 1. Introduction Let 2 be a binary relation on R’!+. The convexed relation 2 is defined by xky iff Vz[y~conv{w:w~z}]*[x~conv{w:w~z)] [see Hildenbrand (1974)]. Starr (1969), first formally presented this relation, which was used by him and by others in equilibrium analysis of markets with non-convex preference relations [see for example Shaked (1976)]. The question arises which properties of the unconvexed relation hold also for the convexed one. Segal (1983) proved that if the relation 2 on R’!+ can be represented by an additively separable convex function of the form U(x,, . . . , x,) =I ui(xi) where are all convex functions, then the convexed relation 2 can be Ul,. . . > u, represented by an additively separable function if and only if there are U: R+R, d, ,..., d,>Q, and b, ,..., b, E R such that for every i, Ui(xi) = u(d,x,) + bi. Let U be a representation of 2. Denote by U” the quasi concave closure of U, the function which represents 2 and which is equal to U where possible (for an exact definition see section 2 below). This paper gives a full characterization of all the additively separable functions whose quasi concave closure is a transformation of an additively separable function. The above conditions are necessary and sufficient ones if U has first order derivatives and if it is not quasi concave, even if it is not a convex function. In the general case these conditions are not necessary. Necessary and sufficient conditions for this case are presented in section 5. 2. Definitions and propositions The following notations will be used: R”, = {x =(x1, . . . , x,):x1 2 0,. . . , x, 2 0}, x,>O}, A”={x~R!+:~x~=l), and A’;=A”nR”,+. R”++=(x~R”:x~>0,..., 03044068/85/$3.30 0 1985, Elsevier Science Publishers B.V. (North-Holland) 130 U. Segal, Separability of the quasi concave closure A function U:R: +R is quasi concave (quasi convex) if for all x, PER; and 0 5 2 5 1, U(Ax + (1 - 1)~) 2 min {U(x), U(y)} ( smax {U(x), U(y)>). Let U:R’L,+R be a continuous and strictly increasing function [i.e., if for every i, Xi~yi, and for some i, xi>yi, then U(x) > U(y)]. U* denotes the quasi concave closure of U and is defined by U*(x) =inf {sup{ U(y):a(y-xx) =O>: and”). By the properties of U it follows that on the axes U* = U. If then the inf in the definition of U” is obtained, XER:., and it is obtained in a vector EAT (which is unique when n = 2), hence on :CIE A: >. This function is unique, RF,> U*(x)=min(max{U(y):a(y-x)=0} and it is easy to verify that if U represents the relation 2, then U* represents the convexed relation 2. Henceforth, Ti = [ri, $1 denotes a closed segment on the ith axis. 3. The function 17 This section discusses the case of two variables (n=2). Let U(x,, x2) be a strictly increasing function, and assume that U* is a transformation of an additively separable function, that is, U*(x,, x2) = h(u:(x,) +$(x2)). Let A(U) = (x: U*(x) # U(x)) (this set is relatively open in Rt). Theorem 1. Let x E A(U). There exists a rectangle around x (bounded or unbounded) on which both u: and u4 are linear, and such that on its boundary (if exists) U* = U. Proof: Let CIbe the vector in 4: such that U*(x) =max{ U(y):a(y-xx) =0} and let H(x, a) = {y: cc(y-x) = O}. There is a segment L cH(x, a) around x on which U* # U. I first prove that for every ZE L, U*(z)= U*(x). Let ZE L and assume that U*(z) =max{ U(y):fi(y-z) =O>, where P#K Obviously, &x-z)>O, otherwise max{U(y):cc(y-z)=O)=max(U(y):cl(y-x)=O}=U*(x) <max{U(y):B(y-x)=O}~max{U(y):/?(y-z)=O}=U*(z), in contradiction to the definition of U*. It is easy to verify that there is a point w~H(x, a) such that U*(x) = U(w) and bw </?z. U*(z) =max{U(y):P(y-z) =0} >max{U(y):B(y-w)=O)ZU(w)= m U*(x)=max{U(y):cr(y-x)=0) . contradiction to the definition of U*, hence for every z E L, U*(z) = U*(x). There is a (closed) rectangle C= r1 x TZ around x on which the indifference curves of U* are straight lines. I now prove that on C both u: and u; are linear. It follows first that for every y=(y,, y,)~Int C, u: is differentiable at y, iff uz is differentiable at y,. Since an increasing function is almost everywhere differentiable, these two functions are differentiable on C. Let M and N be two indifference curves of U* (see fig. 1). Since U* is a transformation of an additively separable function, points 1 and 6 are on the U. Segal, Separability of the quasi concave closure 131 “2 Fig. 1 5 same indifference curve of U*. The slope of this curve, which is a straight line, equals the value of ur’/ur’ at point 1. On the segment 2-7 this ratio is linear in x1, hence u:(xr) =a& +a,~, +a,. Similarly, one can prove that ut’(x,) = b/(cx, + d), hence on C, u4 is a linear function, or the log of a linear function. Since on C the indifference curves of U* are straight lines, both UT and ut are linear. Assume now that C is maximal in the sense that there is no other rectangle C’ 1C on which u: and u2 are linear. Obviously, on the axes U* = U. Assume that there is a point y E R: + on the boundary of C such that U*(y) # U(y). There is a rectangle PI x Yz around y on which UT and ur are linear. Since (rI+r2) n(Yl x F2)#& one obtains that on (r, u Yr) x (r, u r;) UT and ul are linear, in contradiction to the maximality of C. Q.E.D. 4. The function U: The differentiable case Let x E A( U), and let C = rI x r2 be the maximal rectangle whose existence has been proved in theorem 1. On each existing side of C which does not lie on the axes, the indifference curves of U* and U are tangent to each other. Hence, on each such side u’& is fixed, and the corresponding ui is linear. 132 U. Segal, Separability of the quasi concave closure Since on C UT and UT are linear, and since on the boundary of C, U* = U, it follows that ai is linear on r1 iff Z.Qis linear on TZ. However, if both of them are linear, then on C, U* = U. We assumed that U*(x)# U(x), hence C= R:, that is, U* is a transformation of the function a.x for some a~ A:. It is easy to verify that in this case, for every p, q, r, s 2 0 such that U(p, 0) = U(0, q) = U(r, s), air + CQS1 a,p = a2q (that is, the indifference curve passing through (p, 0) and (0, q) lies above the segment connecting these two points). By the same technique used by Segal (1983), one can prove that there are u:R++R, d,, d,>O, and b,, b,ER, such that Ui(Xi)=u(dixi)+bi, i=l,2. Obviously, this is also a sufficient condition. These results are summarized in the following theorem. Theorem 2. Let ul, u2: R, -rR be strictly increasing, diflerentiable functions. The quasi concave closure of the function U(x,, x2) =u1(x1)+u2(xJ is a transformation of an additively separable function, tff U is quasi concave, or alternatively, if for every p, q, r, s > 0 such that U(p, 0) = U(0, q) = U(r, s), ps>q(p-r), and if there are u:R.+R, d,,d,>O, and b,,b,ER such that ui(xi) = u(dixi) + b,, i = 1,2. Remark. Functions satisfying the second condition of this theorem are not necessarily quasi convex. Counter-examples can easily be found. 5. The function ZJzThe non-differentiable case be the maximal rectangle whose Let XEA(U) and let C(x)=r,xr, existence has been proven in theorem 1. As has been proved already, Ui on Ti is non-linear, i= 1,2. Let ~EA( U) such that y1 E r1 but y, # TZ. There is a maximal rectangle C(y) =Pr x r; around y on which UT and uz are linear and such that on its boundary U* = U. Obviously, r1 = P1 and TXA rl, = 8. Denote G,=u{C(x):x~(r~xR+)nA(U)). If ClG,#r, xR+, then there exists a rectangle rr x Pi on which U* = U, that is, ul(xl) +u,(x,) = h(uT(x,) +$(x2)). h solves this functional equation only if it is a linear function [see Aczel (1966)], and since on rl, UT is linear, u1 too is linear, a contradiction, hence Cl G, =rr x R,. Similarly, let G, = u {C(x):x E (R, x r,) A A(U)}, and it follows that Cl Gz = R, x Tz. It thus follows that u1 and u2 are piecewise linear, and if y, is a kink of UT, or yZ is a kink of u;, then u*(y,, yZ)= U(y,, y2). Moreover, if between two kinks of UT, u1 is linear, or between two kinks of ~2, u2 is linear, then A(U) = 8. Let C = r1 x TZ = [yr, y’J x [y2, y;] be a maximal rectangle as in theorem 1. Assume that on Ti, @(xi) = aixi + bi, i = 1,2, and let c = a1/a2. Assume, without loss of generality, that U(y,, y;) = U(6, yJ, where Sjy;. Let &=6-y,. On U. Segal, Separability of the quasi concave closure 133 [yl, S], u,(x,)=u,(c~~+K)+k, on r2, u~(x,)=u,(c-‘x,+L)+1, and on [y; -8, y’J, ul(xl) =u,(cx, +K) +k’. Moreover, there exists M such that for every x1 e[yl, 7; -81, ul(xl +~‘)=u,(x,)+ M. One can now prove that if ~=(y; -y&i -yl) is an irrational number, then u1 on r1 is a linear function. If E is a rational number, and on rl, u1 is not linear, then there exist a minimal 5>0 and M, such that for every x1 E [yl, y; -51, ul(xl + 5) = ul(xl) + M, and for every x2 E [y2, y)2-c<], u2(x2 + ct) =4(x2) + M. 5 does not depend on x2, ~~(0) exists, and since V* is quasi concave, the ratio c is decreasing with x2. It thus follows that there are sequences 0 = y: < yi < ... , d: 1 d: 2 . . ., Mi, M& . . , and ml, m:, . . . such that dJ;(yj2- y’,- ‘) = 5, and such that on [y’,- ‘, ~$1, u2(x2) = u,(dJlx, + n/l’,) +mi. The same argument holds for ul. These results are summarized in the following theorem. Theorem 3. The quasi concave closure of the function V(x,,x,)=u,(x,) +u2(xJ is a transformation of an additively separable function, iff V satisfies one of the two conditions of Theorem 2, or alternatively, if there are 5 >O, u:[O,t]-R, O=$<y!<..., d!zd:z..., and M!,M?,... such that the segment &[yk 1 , yi] j + M{ is equal to the segment [0,41, and such that on [$‘,$J, Li(xi)=u(d{xi+M$+(j-l)u({), i=l,2, j=l,2,.... 6. The case of n variables Theorem 4. Theorems 2 and 3 can be extended to functions of n variables. ProoJ: Let i#j and let (a, b)‘j=(O, . . . . O,s,O, . . . . O,F,O ,..., 0). For every define V’j:R:+R Obviously, by Vij(xl, x2) = V((x,, x,)‘j). (V*)‘jz(V’j)* (as before, V* denotes the quasi concave closure of V), and if x1x2 = 0, then (V*)‘j=( Vij)*. Let (x1, x2) E R: +, and let (CQ,az) be the vector in A: such that (V’j)*(x,, x,)=sup{V’j(y,, ~,):(a,, a,).((~~, y,)-(x,, xJ)=O}. For every E> 0 define CC’ E A?+ by V:R;-tR c$=(l _&)@I, k= i, =(l -&)cQ, k =j, = e/(n - 2), k$ {Li). It is easy to verify that lim,,,sup{V(y):cr”(y-(x,,x,)ii)=O} s(V’j)*(x,, Hence (V*)‘js (U’j)*, and therefore, ( U*)ij= (U’j)*. Let U(X)=CUi(Xi) such that V*(x) = h(CuT(xJ). On the axes V* hence if for every i, ui is a linear transformation of u:, then h is linear. every i, Ui is the same linear transformation of u*, and V*= V. It follows that if A(V) # 8, i.e., if V is not quasi concave, then there exists i x2). = V, For thus such 134 U. Segal, Separability of the quasi concave closure that ui is not a linear transformation of ur. Since for every j# i, (U*)ij= (w)*, and since (U*)‘j is a transformation of an additively separable function, it follows that (Vi’,* # Vii, otherwise ui is a linear transformation of UT [see Aczel (1966)]. Theorems 2 and 3 apply now for i and every j, hence theorem 4. Q.E.D. References Aczel, J., 1966, Lectures on functional equations and their applications (Academic Press, New York). Hildenbrand, W., 1974, Core and equilibria of a large economy (Princeton University Press, Princeton, NJ). Segal, U., 1983, A theorem on the addivity of the quasi concave closure of an additive convex function, Journal of Mathematical Economics 11, 261-266. Shaked, A., 1976, Absolute approximations to equilibrium in markets with non-convex oreferences. Journal of Mathematical Economics 3, 185-196. Stair, R., 1964, Quasi equilibria in markets with non convex preferences, Econometrica 37, 25-38.