Double-Exponential Utility Functions David E. Bell Mathematics of Operations Research, Vol. 11, No. 2. (May, 1986), pp. 351-361. Stable URL: http://links.jstor.org/sici?sici=0364-765X%28198605%2911%3A2%3C351%3ADUF%3E2.0.CO%3B2-G Mathematics of Operations Research is currently published by INFORMS. Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at http://www.jstor.org/about/terms.html. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in the JSTOR archive only for your personal, non-commercial use. Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at http://www.jstor.org/journals/informs.html. Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed page of such transmission. The JSTOR Archive is a trusted digital repository providing for long-term preservation and access to leading academic journals and scholarly literature from around the world. The Archive is supported by libraries, scholarly societies, publishers, and foundations. It is an initiative of JSTOR, a not-for-profit organization with a mission to help the scholarly community take advantage of advances in technology. For more information regarding JSTOR, please contact support@jstor.org. http://www.jstor.org Mon Feb 11 16:52:16 2008 MATHEMATICS OF OPERATIONS RESEARCH Vol. l I, No. 2, May 1986 Printed in U.S.A. DOUBLE-EXPONENTIAL UTILITY FUNCTIONS* DAVID E. BELL Harvard University Ordinal additivity between two attributes X and Y is a property that permits the utility function u(x, y) to be represented as cp(o(x) + w(y)) for some functions cp, u and w where cp may be thought of as a single attribute utility function over o + w . In applications cp is usually taken to be either linear (the additive decomposition) or exponential (the multiplicative decomposition). The utility function can be shown to be exactly one of these forms whenever the two attributes are mutually utility independent. In this paper we introduce a relaxation of utility independence and show that it too implies ordinal additivity. We identify the class of utility functions cp that are consistent with the relaxed assumptions. One member of this class, cp(z) = -exp(bexp(-cz)), which we call double-exponential, seems particularly appealing, for when b and c are positive it is increasing, risk averse and decreasingly risk averse for all z . The multiattribute decomposition corresponding to it -exp(b exp(- c(a + w))) may be represented as u(x, y) = c exp(cu(x)w(y)) which we call the exponential product. 1. Introduction. Let u(x, y) be a two-attribute utility function on a continuous y*], then the conditional utility function u(x Iy) interval space X x Y = [xO,x*]x is defined by In this paper we will restrict attention to the case where x and y are both scalar, for generalizations to the vector case will normally be obvious. For simplicity, we will also assume that preferences for X increase from x0 to x* and, for Y, increase from to y*. Without loss we will presume u(xO, = 0 and u(x*, y*) = 1 and denote u(x*, = a and u(xO,y*) = b. Though the letter u will be used for three functions u(x, y), u(x ( y ) and u(y I x), no confusion should arise. The equation above may be rearranged to illustrate the assessment problem: U(X,y ) = bu(y 1 xO)+ [ a + (1 - a)u(y I x*) - bu(y I xO)]u(xly). (1) Without simplifying assumptions we need to assess two constants a and b, two one-attribute utility functions u(y ( xO)and u(y I x*) as well as, in principle, an infinite number of one-attribute utility functions, u(x 1y). Keeney and Raiffa (1976) exploited the utility independence assumption that u(x 1 y) = u(x IyO) for ally. If, in addition, u(y I x) = u(y 1x4, Keeney (1968) showed that known as the multiplicative form. The special case in which b = 1 - a is known as the additive form. Kirkwood (1976) considered the case in which u(x y ) could be assumed always to *Received June 4, 1984; revised April 11, 1985. A MS 1980 subject classification. Primary: 90A 10. IAOR 1973 subject classification. Main: Decision theory. Cross references: Decision. O R / MS Index 1978 subject classification. Primary: 857 Utility/preference/theory. Key wordr. Utility independence, ordinal additivity, utility theory. 0364-765X/86/ 1102/0351$01.25 Copyright 0 1986. The Institute of Management Sciences/Operations Research Society of America 352 DAVID E. BELL be a member of the same parametric family, for example u(x 1 y) = - exp(- c(y)x). In this case the assessment of u(x ly) is reduced to determining the one-dimensional function c(y). Kirkwood described this kind of condition as Xparametrically dependent on Y. Bell (1979) approached the assessment problem by assuming an interpolation condition: If X is interpolation independent of Y and vice versa, Bell showed that which requires the assessment of 3 constants and 4 one-attribute utility functions. While interpolation independence has many advantages for the assessment of complex multiattribute utility functions, it does have a disadvantage not shared by utility independence or parametric dependence. If u(x 1 and u(x 1 y*) are both exponential, one might wish that so too should be u(x 1 y) for other values of y. More generally we might ask that local risk aversion properties shared by u(x lyO) and u(x 1 y*) would be common in all u(x 1 y). Pratt (1964) defined (local) risk aversion for a utility function u(x) by the function r(x) = - u"(x)/u'(x), a quantity that is proportional to the risk premium for small (local) gambles. Define the conditional risk aversion of X given Y, by the function r(x 1 y) = - U"(XI y)/u'(x 1 y). We note the following properties: (i) r(x I y) is independent of x and y if and only if X is utility independent of Y and for all x and y, for some constant c, u(x 1 y) = cx or - ce -'" (ii) r(x 1 y) is independent of y if and only if X is utility independent of Y, and (iii) r(x ( y ) is independent of x if and only if X is parametrically dependent on Y with the exponential form. These properties suggest that insight about multiattribute assessment procedures may be gained by considering the problem as one of assessing the conditional risk aversion functions. In this paper we exploit the interpolation idea but apply it to r(x I y) rather than u(x 1 y). 2. Interpolating risk aversion. Suppose that, for some function a(y), For the sake of a name, we will call this condition risk aversion interpolation independence with the abbreviation X RAII Y. This property includes utility independence as a special case and, in some cases, it includes parametric dependence as a special case and also. For if u(x 1 y4 and u(x 1 y*) are exponential, so too is u(x 1 y); if u(x 1 u(x 1 y*) have constant proportional risk aversion, so too has u(x 1 y). However, (4) does not preserve the logarithmic form. For example, if u(x 1 = log(x + a,) and U(XIy *) = log(x + a,), then (4) does not produce the result u(x 1 y) = log(x a(y)). Nor does it preserve the more general power function family of (x + alb. [These forms do satisfy the interpolation equation + but we leave a study of that relationship for another time.] DOUBLE-EXPONENTIAL UTILITY FUNCTIONS 353 An assumption of X RAII Y together with Y RAII X we will denote by X MRAII Y, or, equivalently Y MRAII X ( M standing for "mutually"). We note that X MRAII Y is a strictly stronger condition than X RAII Y, for example u ( x , y ) = [ x + l / c ( y ) ] exp(- x c ( y ) ) satisfies X RAII Y but not Y RAII X. The use of r ( x 1 y ) implicitly requires that u ( x , y ) be differentiable at least twice. The proofs that accompany the following results require u ( x , y ) to be differentiable at least three times. Since the proofs are neither elegant nor insightful they have been gathered in a final section of the paper. LEMMA1. X is utility independent of Y and Y RA 11 X if and only if u ( x , y ) has an additive or multiplicative decomposition. Lemma 1 provides slightly weaker assumptions than those required by Keeney for the multiplicative form. 2. If X MRAII Y then X and Y are ordinally additive. LEMMA Our main result will utilize a family of utility functions characterized by a property of the risk aversion function as described by the following lemma. LEMMA3. The risk aversion function r ( x ) = - u " ( x ) / u f ( x ) satisfes the condition r'(x) = cr(x) + d for some constants c # 0 and d , if and only if where b > 0, k is any constant, and 0 is f 1. The indirect representation of u in this result may mask some of its more useful special cases. For example if k is zero then u ( x ) = - exp(b exp(- ex)) a function which for all x is increasing, concave and exhibits decreasing risk aversion so long as b and c are each positive. We will refer to this form as the double exponential. If k is a positive integer and z = becx then either a family that might be useful for small values of k . The case r'(x) = c2 leads to a u ( x )which is the cumulative of a normal distribution, the case r ' ( x ) = - c2 to an integral of the form j:oexp(t2)dt. For expositional ease it is useful to have a name for the family of utility functions characterized by the equation r'(x) = cr(x) + d. DEFINITION.A utility function u ( x ) with risk aversion function r ( x ) is one of the extended double exponentials if and only if r f ( x )= c r ( x ) + d for some constants c and d (that might be zero). Note that another family of utility functions can be characterized by their relation) ~ the linear and ship between r f ( x )and r ( x ) .The equation r f ( x )= - k r ( ~characterizes exponential families ( k = O), the logarithmic utility function (k = l), and the power family (0 # k f 1). THEOREM.The following conditions are equivalent: (i) X MRAII Y, (ii) X and Y are ordinally additive and X RAZZ Y . (iii) u ( x , y ) = +(u(x)+ w ( y ) )for some functions +, u and w, where the extended double exponential family. + is a member of 354 DAVID E. BELL 3. Discussion. The assessment of utility functions is a task requiring great care. Decision makers are often neither consistent with the axioms of von Neumann and Morgenstern, nor do they always satisfy certain established desiderata such as, in the case of the attribute money, being risk averse and decreasingly so. Although procedures have been devised to construct a single attribute utility function to be consistent with directly assessed points and with declared local properties [Meyer and Pratt 1968, Schlaifer 19711, a common route taken is to find a family of utility functions with the desired local properties and then to select the available parameters in a way that best approximates the direct assessments. Multiattribute utility assessment is also best thought of as an exercise in the approximation of functions [Fishburn 19771. The task is made more delicate by the desire to preserve local properties of the true function in the approximation. Even though the interpolation scheme (2) is quite sound as an approximation to the absolute value of the conditional utility function it does not preserve desirable qualitative properties of the generating functions. The interpolation scheme appears to permit a great deal of flexibility in approximating the utility function, yet, as Lemma 2 showed, it and the equivalent relation for r(y I x) require X and Y to be ordinally additive. The ordinally additive representation, u(x, y) = +(u(x) + w(y)) when associated, through +, with families of single attribute utility functions, provides a range of multiattribute decompositions. If is linear we have the additive form, if is exponential we have the multiplicative. If + is the double exponential, u(x) = - exp(b exp(- cx)), then we have u(x, y) = - exp(b exp(- c ( u + w))). This may be written equivalently as u(x, y) = exp(bv,(x)v,(y)) a representation for which the term exponentialproduct seems appropriate. + + 4. Proofs. + PROOFOF LEMMA 1. If X UI Y then u(x, y) = f(y) g(y)h(x) for some functions In this case r(y I x) = - (f "(y) g"(y)h (x)) f, g and h, where h(x) = u(x 1 /( f'(y) + g'(y)h(x)) so that Y RAII X only if + where I have used the scaling convention u(xOlyO)= 0, u(x* Iy0) = 1 to deduce h(xO)= 0, h(x*) = I . Clearing the denominators in the equation we have = afl'f'(f' + g') + (1 - a)fl'f'f'+ (1 - a)gl'flf' - f r ' f ' ( f ' + g') or Now if f"g' - gUf' = 0 then Y UI X immediately since then r(y 1 xO)= r(y 1 x*). If (ag' + f')h = f'(l - a ) then f' = kg' for some constant k so that f(y) = kg(y) + c. Hence in this case, Y UI X also. Now we use Keeney (1968) to show that u is multiplicative. 355 DOUBLE-EXPONENTIAL UTILITY FUNCTIONS PROOFOF LEMMA 2. The proof is rather lengthy. The reader who wishes merely to be satisfied of the plausibility of the result may be able to do so by consideration of equation (7) in light of Lemma 4 (later in this section). Integrating (4) we have U t (1 y~) = eS(~')u'(xy*) 1 Since u(x, y) = f ( y ) + g(y)u(xly) for some functions f, g we also have au/ax = g(y) u'(x 1 y). Therefore we can say that, for arbitrary functions a , , a,, b, and b, and a constant k , , we have where b,(xO)= b2(y0)= 0 and b,(x*) = b2(y*) = 1. By symmetry, where d,(xO)= d,(y") = 0 and d,(x*) = d,(y*) = 1. (In future we will abbreviate d,(xO) as dp and d,(x*) as d: and so on.) To show X OA Y we must show that log(au/dx) - log(au/ay) = v(x) + w(y) for some functions u and w (Luce and Tukey 1964), that is, we must show a , + a, + k,b,b, - c, - c, - k2dld2= V ( X ) ~ ( y for ) some functions o and w. Equivalently, we must establish that k,b,b2 - k2d,d2 = o,(x) + w2(.y) for some functions u, and w,. Since by = dp = 0 we have w2(y) = 0 and similarly o,(x) = 0. Hence we must show that k,b,b, = k,d,d,. We will do this by demonstrating, somewhat tediously, that this condition is implied by (5) and (6). Differentiating (5) with respect t o y we have a2u/axay = (a; k,b,b;)(au/ax) and, similarly, a2u/ayax = (c; + k2d;d2)(au/ay). If these second order derivatives are continuous they are identical so that + + (a; + k,b,b;)exp(a, + a, + k, b, b,) = (c; + k,d;d,)exp(c, + c2 + k2dld2) is a necessary and sufficient condition for X MRAII Y. Substitutingy = y o in (7) we have (a;' k,b,b;")exp(a, a,") = c;exp(c, larly, a; exp(ap + a,) = (ciO+ k2d;Od2)exp(c~ cZ). + + (7) + c a . Simi- + Substituting y = y * in (7) we have Similarly, (a; + k,b;)exp(a: + a, + k,b,) = (c;* + k2d;*d2)exp(c: + c, + k,d,). From these four equations we may identify a;, c;, d; and b;: a; + k2d;Od2)exp(c2- a2 + = (ciO C: - a:), k , d ~ = ( a ~ * + k , b , b ~ * ) e x p ( a , + a , * + k , b , - c , - c : - k , d , ) - c a; nd (9) (10) 356 DAVID E. BELL Also, c;' = a;Oexp(af- cf c;* = (a;' + k,b;)exp(a: + a: - c:), - c: (12) + a: - c:), + a,* - cf - c:) - ciO, = (a;* + k,b;*)exp(a: + a: + k , - k , - c: - c:) - ci*. k2di0= a;* exp(a; k,d;* (I3) ( 14) (I5) Substituting in (7) using (8), (9), ( l o ) , ( 1 1 ) : exp(al + a, + k,b,b,)[(l - b,)(ciO+ k2d;Od2)exp(c2- a, + e 0l - a 0, ) + b,(c;* + k2d;*d2)exp(c:+ c, + k2d2- a: - a, - k l b 2 ) ] =[(I - d2)(ai0+ k,b,b;o)exp(a, - c , + 0: - c:) +d,(a;* + k,b,b;*)exp(a, + a: + k l b l - c , - c: - k2d,)]exp(c,+ c 2 + k2d,d2). Now using (14) and (15): e x p ( k , b l b 2 ) [ (l b,)(c;O(l - d2)+ d2a;* exp(ap + a: - cp - c:))exp(cp - a:) + b,((l - d2)c;*+ d2(a;* + k,b;*)exp(a: + a: + k , - k , - cy - c:)) + k2d2- a: - k,b,)] = exp(k,dld2)[(l- d,)(a; + k,b,bf)exp(a: - c:) + d2(a;* + k,b,b;*)exp(a; - c: + k , b , - k 2 d l ) ] . x eip(c: Now use (12) and (13): e x p ( k ,b,b,) [ ( 1 - b , ) ( ( l - d2)a;0exp(a:- c:) + d2a;*exp(a: - c:)) + b , ( ( l - d2)(ap+ k ,bp)exp(a; - c: + k2d2- k,b,) + d2(a;* + k,b;*)exp(a: = exp(k2d,d2) [(1 - - c: + k , - k , + k2d2- k,b,))] d2)(ai0+ k,bl bp)exp(a: - c;) + d2(a;* + k,b,b;*)exp(a: - c: + k l b , - k i d l ) ] . (16) Let A' = aiOexp(a; - c:), BO A* = a;* exp(a: - c?) and = k,biOexp(a; - c:), B* = k , b;* exp(a; - c:). W e may assume that not all o f these four constants are zero for otherwise we would have a2u/axay = 0 which would imply that u ( x , y ) is additive. Temporarily, let us use 357 DOUBLE-EXPONENTIAL UTILITY FUNCTIONS the notation + b, [(A' + B0)(1 - d,) + (A* + ~ * ) d , e * l - ~ ~ ] e x ~ ( k-, dk,b,) , and so that (16) may be written as exp(k,b,b,)f(x, Y) = exp(k2d1d2) g(x, Y). (17) Differentiate this with respect to x: and use (17) to deduce klb;b2 - k2d;d2 = g f / g - f'/ f. We have + ( ( A 0 + BO)(l- d 2 ) + ( A * + B*)d2eki-*2)(exp(k2d2- klb2))] and x (B* + (A* + blB*)(kl - k2d;(x)/b;(x)))]. Note that f(x*, y ) = [(A' g(x*, y ) = [(A' + B ')(I - d,) + (A* + ~ * ) d , e ~ ~ - ~ ~ ] -e xk1b2) ~ ( k ~and d~ + BO)(l- d2)+ (A* + B*)d2ekl-*2] Note that f(x*, y ) = g(x*, y)exp(k,d, - klb2). In what follows we wish to assume that b;(x) i 0 Z d;(y). Note that b,(x) = loguf(x lyO)- loguf(x ly*) so that b;(x) = r(xly*) - r(x lyO). It' b;(x) = 0 in a subinterval I of [xO,x*]then X RAII Y implies X UI Y on this subinterval. By Lemma y*]. Hence it remains to 1 we know that b,(x) = d,(x) and b,(y) = d2(y) on I x prove that b, = dl and b, = d, on intervals in which b; # 0. The idea is to partition [xO,x*]into intervals I , , . . . , I,, and y*] into J l , . . . , Jkand prove the result on each I, x 4 separately, then rely on the continuity of b,, d l , b, and d, to prove the result for the entire space. Substituting x = x0 in (18), and letting so = dio/b;', we obtain 358 DAVID E. B E L L Similarly, substitution of x* in (18), with s* = d ; * / b , gives (klb2 - k,s*d2) -~ + ~ - d,) ( 1 d 2 e k l - k 2 ( ~+* ( A * + B * ) ( k , - k2s*))+ g ( x O ,y ) e x p ( k l b 2- k2d2)- g ( x * , y ) g(x*3 Y ) + Let z = d 2 ( y )and w = k , b , - k2d2so that k , b , = w k2z. The function b, may be regarded as a function of z because d ; ( y ) .z. 0. Also let l l ( z )= g ( x O y, ) , 12(z)= g ( x * , y), then (19) becomes and (20) becomes + ( A * + B * ) ( k l - k2s*))+ l l e w- 1,. 12(w + k,z(l - s*)) = B O ( l - z ) + Finally, we may write (21) and (22) as I l w + 1,e-"'= q , ( z ) and It will be important later to remember that 1, and 1, are linear in z and that q l and 9, are quadratic in z. The proof will continue by consideration of two cases, the first is when differentiating (23) and (24) and solving for w' produces new information about w, the second is when it doesn't. So, differentiating both (23) and (24) with respect to z : + Iiw' + I;e-"' - w'I,e-" = q', , I;W + l2w' - l i e w - I l w J e w= q ; . I;w (25) (26) Eliminating w' we have (1, - l l e w ) ( q ', 1;w - l i e w )= ( q ; - 1;w + l ; e w ) ( l 1- 12e-"'). Now use (23) and (24) to clear this equation of e" and e - " : = (1, Collecting terms in w: i 1; 11 1 + 1,w - q , ) q; - 1;w + - (12w - q2) . 359 DOUBLE-EXPONENTIAL UTILITY FUNCTIONS Since the term in w2 cancels to zero we conclude that either this equation is an identity or w is representable as the ratio of two polynomials in z. But now use (24) and Lemma 4 (see below) to see that in this latter case w must be constant. If (23) and (24) produce equivalent differential equations in w' and since w(0) = 0 is known, we may conclude that (23) and (24) are equivalent. That is, any solution w to (23) is a solution to (24) and vice versa. We can show that there cannot be two distinct solutions to (23) and (24) for suppose that w = a and w = b were both solutions for some particular choice of z. We would have I,a+12e-"=q,, 1 2 a + I , e a = q 2 , I , b + 1 2 e - b = q , and 1 2 b + l , e b = q 2 from which we may deduce that l;(e-" - eCb) = l:(ea - eb), a condition only possible if a = b or if 1, = I, = 0. This latter case is ruled out since d;(y) # 0. Equation (23) has exactly one solution only if either /,I2 < 0 for all z or if q, = t21, for some constant t,. For consider the equation ax + be-" = c where a, b and c are constants and ab > 0. We may as well assume b > 0. This equation has two solutions whenever c > a(1ogb - loga + I), one solution when c = a(1ogb - loga 1) and no solution otherwise. By contrast, if ab < 0 this equation has at most one solution since ax be-" is a strictly monotone function of x . So the only way we could have exactly one solution to (23) and yet have Ill2> 0 is if + + If Ill2> 0 holds for some z it must hold for an open interval in which case (27) must hold for all z (by analytic continuation). By Lemma 4 1, = t,l, and q, = t211for some constants t, and t,. If /,I2 < 0 for all z then 1, = - t,l, for some constant t, > 0. Substitute for I, in (23) and (24) and note that a root of l,(z) = 0 must also be a root of q, and q2. Hence, for some linear functions 1, and 1, we may write (23) and (24) as ~ - t , e - ~ = land ~ - t3w - e w= I,. P I (29) Recall that (23) and (24) yield the same equation in w'. From (28) w'[l + t3e-"1 = 1; and from (29) w'[t3 + e W ]= - 1; where 1; and 1; are constants. If 1; = 0 then let S be the subset of the real line on which w' # 0. On S we must have 1 + t,eCw = 0 so that w is constant on S. Since w is continuous, it is constant everywhere. Assuming, therefore, that 1; # 0 we have Differentiating this twice with respect to z shows that w must be constant. The case in which q2 = t211is equivalent to that in which 1; = 0. We have shown that (5) and ( 6 ) require that k,b2 - k2d2be constant, and therefore zero. Since b: = d: = 1 we have also shown that k, = k, so that b, = d,. By symmetry, b, = d l . Hence k,blb2= k2dld2,hence u is additive. This ends the proof of Lemma 2. z PROOFOF LEMMA3. Differentiating the integral gives u'(x) = becx. Differentiating once more gives + czk+leg' where = Hence r(x) = - c(k + 1 + 8z), so that rt(x) = - c28z = c(r(x) + c(k + I)), from which we conclude that d, the constant in the lemma statement, is equal to c2(k + 1). 360 DAVID E. BELL + + If rt(x) = cr(x) + d (c # 0), then log(cr d)2 = 2cx logai where a, is a nonzero constant. Hence cr + d = a, exp ex. Integrating once more we have - c log ~ ' ( x +) ~ 2dx + k, = 2(ao/c)expcx which is equivalent to u'(x) = ~ e " " e"PCX e ~ for constants a, b and A so that Substitute t = @becY where 0 = + 1 is chosen so that @b> 0. We have dt = c@be"dy = ct dy, and eaY = (t/@b)"icso that Let k = a / c - 1 and recognize that A/c can be any sign and we have *Lo b exp cx k 0, U(X)= t e dt where b > 0. The following result, used repeatedly in the proof of Lemma 2, is undoubtedly proved in some textbook but I have been unable to locate a reference. LEMMA4. If p l , p2, p3 and p, are polynomials such that p , /p2 = exp(p3/p4) then p I / p 2 and p,/p, are constants. PROOF. We may assume that p , and p, have no factors in common otherwise these could be cancelled. Similarly with p3 and p,. If p2 has a root, say a where p2(a) = 0 then p4(a) must be zero also. Let p2(x) = (X - a)kp,(x) for a polynomial p,, where p5(a) # 0 and k is an integer. We know that Now p,/p5 behaves "sensibly" at x = a but (x - a)kexp(p,/p,) does not. More precisely,p,/p, is analytic at x = a but (x - a)k exp(p,/p,) has an essential singularity at x = a. (See Apostol 1957.) Hence p2 can have no real roots. If p, has a root, say p4(b) = 0, then p3(b) < 0 is implied since p, is not zero and, therefore, we must have pl(b) = 0. Let p,(x) = (x - b)'p,(x) where p,(b) # 0. Then Now p2/p6 is nicely behaved at x = b but (x - b)'exp(-p3/p4) goes to infinity as x -t b. Hence p4 has no real roots. Now consider the equation p2/p1 = exp(-p3/p4). We already know that p2 and p4 have no real roots and by a similar sequence of arguments neither does p , . But, in fact, there is no need to restrict our consideration to real roots. By analytic continuation, if pl/p2 = exp(p3/p,) on the real line it must hold in the complex plane also (Apostol, p. 519). The above arguments now apply for complex roots. (Think of complex roots as factors of the form x2 k where k +O. Now consider behavior as x2+ - k.) Hence p , and p, have no real or complex roots and so must be constants. + PROOFOF THEOREM:(i) -+(ii) by Lemma 2. To show (ii) + (iii) note that X RAII Y is equivalent to r(x 1 y) = a(x) + b(x)a(y) so that r'(x I y) = c(x) + d(x)a(y) for some functions a, b, c and d. Hence 36 1 DOUBLE-EXPONENTIAL UTILITY FUNCTIONS +W) say. Now a u / a x = u'+'(u - v " / v ' - v'+"/+'. Denoting + and a2u/ax2= on+' ulu'+" so that r ( x 1 y ) = by r ( v w ) we have + -+"/+' so that Substituting for r ( x I y ) and rl(x ( y ) in (30) and rearranging we get an expression of the form r f ( v+ W ) = h ( x ) + j ( x ) r ( v + w ) (32) for some functions h and j . If w ( y ) is not constant, so w' # 0 then differentiating (32) with respect to y we have r"(v + w ) = j(x)r'(v w). Either r'(v + w ) = 0 in which case r'(v w ) is trivially linear in r ( v + w ) or j ( x ) is constant. If j ( x ) is constant then so too is h ( x ) . Hence r'(v + w ) is linear in r ( v w ) which shows that is one of the extended double-exponential family. Now to show (iii) +(i). In this case we know r'(v w ) = cr(v w ) + d . If c = 0 then r ( v + w ) = d ( u + w). Substituting in (31) we see that r ( x 1 y ) = - v " / v ' dvl(v w ) which is of the form a ( x ) b ( x ) a ( y ) as required. Hence X RAII Y and, by symmetry Y RAII X. If c # 0 then log(cr d)2 = 2c(v w ) + logai or cr d = aoe c ( c + w ) . Substituting once more in (31) gives + + + + + + + + + + + + which is also of the form a ( x ) + b ( x ) a ( y ) . Acknowledgments. John Pratt drew my attention to the undesirable risk aversion properties of interpolated conditional utility functions. Richard Meyer and Richard Stone provided me a refresher course in complex analysis. References Apostol, Tom M. (1957). Mathematical Analysis, Addison-Wesley, Reading, Mass. Bell, David E. (1979). Consistent Assessment Procedures for Conditional Utility Functions. Oper. Res. 27 1054- 1066. Fishburn, Peter C. (1977). Approximations of Two-Attribute Utility Functions. Math. Oper. Res. 2 30-44. Keeney, Ralph L. (1968). Quasi-Separable Utility Functions. Naval Res. Logisf. Quart. 15 551-565. Kirkwood, Craig W. (1976). Parametrically Dependent Preferences for Multiattributed Consequences. Oper. Res. 24 92-103. Luce, R. Duncan and Tukey, John W. (1964). Simultaneous Conjoint Measurement: A New Type of Fundamental Measurement. J. Math. Psychology 1 1-27. Meyer, Richard F. and Pratt, John W. (1968). The Consistent Assessment and Fairing of Preference Functions. IEEE Trans. Systems Sci. Cybernet. SSC-4 270-278. Pratt, John W. (1964). Risk Aversion in the Small and in the Large. Econornetrica 32 122-136. Schlaifer, Robert 0 . (1971). Computer Programs for Elementary Decision Analysis. Division of Research, Graduate School of Business Administration, Harvard University, Boston. Thompson, Silvanius P. (1914). Calculus Made Easy. MacMillan, New York, Second Edition. GRADUATE SCHOOL O F BUSINESS ADMINISTRATION, HARVARD UNIVERSITY, SOLDIERS FIELD, BOSTON, MASSACHUSETTS 02163