Journal of Mathematical Psychology 47 (2003) 244–258 Probability weights in rank-dependent utility with binary even-chance independence David E. Bella, and Peter C. Fishburnb a Harvard Business School, Boston, MA 02163, USA b AT&T Shannon Laboratory, USA Received 28 August 2001; revised 8 July 2002 Abstract We examine the effects of a weak version of expected utility’s independence axiom on the probability weighting function in rankdependent utility. Our weak independence axiom says that a 50–50 lottery between a two-outcome gamble and its certainty equivalent is indifferent to the certainty equivalent. A variety of nonlinear probability weighting functions satisfy this axiom, but most weighting functions proposed by others do not. Nevertheless, the axiom accommodates weighting functions that are quite similar to the inverse S-shaped concave–convex functions of others that overvalue small probabilities and undervalue large probabilities. r 2003 Elsevier Science (USA). All rights reserved. 1. Introduction It is widely acknowledged that expected utility (EU) theory (von Neumann & Morgenstern, 1944; Marschak, 1950; Herstein & Milnor, 1953; Luce & Raiffa, 1957; Raiffa, 1968; Fishburn, 1982) is descriptively inadequate as a theory of individual choice for decision under risk despite its appeal as a normative guide for rational decision making. A prominent feature of some nonexpected utility theories that accommodate prevalent behavioral violations of EU, including rank-dependent utility (RDU), cumulative prospect theory, and close relatives (Quiggin, 1982, 1993; Luce, 1991; Luce & Fishburn, 1991, 1995; Tversky & Kahneman, 1992; Wakker & Tversky, 1993; Schmidt & Zank, 2001; Safra & Segal, 2001), is that outcome probabilities enter into assessments of risky prospects or gambles through a transformation that subjectively weights objective probabilities. The most common weighting effects are inflation of small probabilities and devaluation of large probabilities (Edwards, 1961; Tversky & Fox, 1995; Tversky & Wakker, 1995; Wu & Gonzalez, 1996; Prelec, 1998; Gonzalez & Wu, 1999; Abdellaoui, 2000; Bleichrodt & Pinto, 2000; Luce, 2001) that give rise to the familiar inverse S-shape of the probability weighting Corresponding author. Fax: +1-617-812-6515. E-mail address: [email protected] (D.E. Bell). function. Several theories propose different weighting functions for outcomes interpreted as gains and losses, but we consider only a single function in this paper. We often refer to it as a pwf, short for probability weighting function. Fig. 1 illustrates pwfs p that inflate small probabilities ½pðpÞ4p and deflate large probabilities ½pðpÞop: Figs. 1a and b have pðpÞ ¼ p near p ¼ 0:375; the others have pð12Þ ¼ 12: All but Fig. 1b satisfy an independence axiom that is inconsistent with prevailing views on pwfs. This axiom, referred to below as binary even-chance independence, is the central topic of the paper. Despite the fact that it departs substantially from prevailing views, we will see that it allows weighting functions which are similar to those proposed by others. The EU assumption most responsible for behavioral violations of EU is the independence axiom or substitution principle (Samuelson, 1952; Herstein & Milnor, 1953; Jensen, 1967). Our purpose here is to examine the effects of a severely restricted independence axiom on non-EU and the pwf. We do this in the context of RDU (Quiggin, 1993), which is perhaps the most widely analyzed descriptive alternative to EU. It will become apparent that our restricted independence axiom is antithetical to some views about RDU and is incompatible with conditions proposed by others for probability weights. There is, however, no fundamental conflict between the axiom and RDU, and, as suggested above, 0022-2496/03/$ - see front matter r 2003 Elsevier Science (USA). All rights reserved. doi:10.1016/S0022-2496(02)00023-8 D.E. Bell, P.C. Fishburn / Journal of Mathematical Psychology 47 (2003) 244–258 1 1 0.9 0.9 0.8 0.8 0.7 0.7 0.6 0.6 0.5 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0 (a) 0 0.2 0.4 0.6 0.8 1 1 0 (b) 0 0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1 1 0.9 0.9 0.8 0.8 0.7 0.7 0.6 0.6 0.5 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0 (c) 245 0 0.2 0.4 0.6 1 0.8 0 (d) 0 0:547 Fig. 1. (a) A pwf that satisfies BECI. (b) pðxÞ ¼ expððlnðxÞÞ Þ for all 0oxo1: (c) p based on a w defined by a quintic on [0,0.5]. (d) p based on a w defined by a linear plus exponential on [0,0.5] with c ¼ 8 (refer to Example 6). we will argue that the axiom’s effects on RDU can be practically indistinguishable from the effects of other conditions that constrain probability transformations. In this vein, the paper can be viewed as an attempt to see how much of EU’s independence axiom can be retained while preserving the integrity of the RDU viewpoint. At the same time, our axiom accommodates but does not require a variety of behavioral possibilities that are not associated with the RDU model and are not usually allowed by constraints proposed by others. The restricted independence axiom we focus on says that if x is a two-outcome gamble with certainty equivalent c; then the even-chance lottery that yields x or c; each with probability 12; is also indifferent to c: We refer to this as binary even-chance independence, or BECI. Our choice of the mixing probability 12 is motivated by symmetry, both intuitional and analytical. A fixed mixing probability other than 12 could be considered, but we will not do so here. We shall see later (Theorem 1) that if every mixing probability were allowed for the indifference conclusion, then RDU would reduce to EU. The following example conveys the essence of BECI and issues it confronts. Example 1. Paul has decided that he is indifferent between $3000 with certainty and a binary gamble x that pays $10,000 with probability 38 and nothing otherwise. He is then asked to compare options A and B that involve a fair coin toss. A: Get $3000 regardless of whether heads or tails occurs; D.E. Bell, P.C. Fishburn / Journal of Mathematical Psychology 47 (2003) 244–258 246 5 4 0.8 3 0.7 0.6 2 0.5 1 0.4 0 0.3 0.75 -1 0.8 0.85 0.9 0.2 -2 0.1 0 (a) 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 -3 (b) Fig. 2. (a) w defined by straight line segments on [0,0.5] and extrapolated by Eq. (1). (b) The curve in (a) extrapolated to [0.75, 0.904]. B: Get $3000 if heads occurs and x if tails occurs. Should Paul be indifferent between A and B? Will he in fact be indifferent between them? EU says that Paul should be indifferent between A and B because this is implied by independence, even in the restricted BECI version. But if Paul is unfamiliar with EU, he may still be indifferent between A and B: After all, if heads occurs, he gets $3000 in either case, and if tails occurs, he is back to the original comparison between $3000 and x: A different impression could arise when outcome probabilities are specified holistically: A has pr. 1 for $3000; 5 B has pr. 16 for $0, pr. 12 for $3000, pr. 3 16 for $10,000. Without the fair-coin dependency, it is far from obvious that Paul will be indifferent between gambles A and B: He may decide in fact that he definitely prefers one to the other. This would violate BECI and EU, but is not inconsistent with RDU. Nevertheless, we will show that indifference between A and B and general adherence to BECI can also be consistent with RDU and with violations of EU that are well known from Allais’s (1953, 1979) analyses and later works. We will show that BECI’s modest infusion of the classical independence axiom into RDU can at least partially preserve the integrity of RDU and allow pwfs that are similar to—but not identical with—forms proposed by others. The next section outlines common versions of RDU and EU and comments on independence axioms. We assume for both models that the outcome set is a real interval and that the utility function u for outcomes is continuous and increasing over the interval. We denote the pwf for RDU by p and assume that it is an increasing and continuous map from ½0; 1 into ½0; 1 with pð0Þ ¼ 0 and pð1Þ ¼ 1: After the RDU and EU models are introduced, we note that the generalization of BECI to binary independence with biased coins reduces RDU to EU. Section 3 begins our analysis of BECI by noting in the RDU context that it is equivalent to the functional equation p p ½1 pðpÞ 2 1þp ¼ pðpÞ 1 p for 0ppp1: ð1Þ 2 The combination of RDU and BECI places no new constraints on u and allows any p that satisfies (1) and is an increasing onto map. We will see that (1) has substantial mathematical implications. It constrains p in interesting ways and can give rise to unrealistic pwfs (see Fig. 2). On the other hand, it allows forms for p that emulate effects observed by others that justify its serious consideration in the RDU context. Two consequences of (1) noted in Section 3 are called conditions near 0 and near 1. The condition near 0 says that pðpÞ=pð2pÞ approaches 1 pð12Þ as p-0: This happens to be violated by most nonlinear forms for p proposed in the literature (see Section 5). Another consequence of (1) is that p cannot be concave near 0 and convex near 1 when pð12Þa12: Although this also goes against prevailing wisdom about p; we argue that it should not rule out BECI as a potential descriptive condition. The final part of Section 3 is based on the observation that specification of p on an interval of length 12 determines it throughout ½0; 1 when (1) is presumed. However, with p continuous and increasing on the D.E. Bell, P.C. Fishburn / Journal of Mathematical Psychology 47 (2003) 244–258 length-12 interval, its extension by (1) might fail to increase throughout ½0; 1 and can even have singularities. Examples illustrate these phenomena. Because we presume throughout that p is increasing and continuous on ½0; 1; we use w in place of p when considering effects of (1). For example, if w is increasing and continuous on a length-12 interval in ½0; 1; and if (1) is assumed for w in place of p throughout ½0; 1; then w might not be increasing over ½0; 1: Section 4 addresses the increasingness issues when w is assumed to be continuous and increasing on ½0; 12 with wð0Þ ¼ 0owð12Þo1: With an additional assumption of differentiability on ½0; 12; we establish a sufficient condition for increasingness over ½0; 1 based solely on the behavior of w on ½0; 12: The condition is then used to prove increasingness throughout ½0; 1 of some basic analytical forms for w on ½0; 12 when w is extended to ½0; 1 by (1). Section 5 demonstrates that p for (1) can have approximately the inverse S-shape often attributed to pwfs. We then consider six parametric forms for p proposed by others. Three of the six violate (1), and the other three are consistent with (1) if and only if pðpÞ ¼ p for all p: Hence, if C is a set of conditions which implies one of the six parametric forms then, in the RDU context, either C and BECI are incompatible or the combination of the two reduces RDU to EU. The pwfs allowed by BECI that most closely resemble forms proposed by others are continuous and increasing but not fully differentiable. This is true of Fig. 1a, where the left and right derivatives of p are not always equal: see Section 5. Fig. 1b pictures a differentiable pwf that does not satisfy BECI. The reason that BECI approximations of such traditional transformations are not fully differentiable is that differentiable pwfs that satisfy BECI usually require pð12Þ ¼ 12: Examples are shown in Figs. 1c and d. We elaborate on these and on differentiability in Section 6, where it is shown (under stated conditions) that if w satisfies (1) and is differentiable at 12 then it is differentiable everywhere. A simple example (not discussed later) is p defined as 12p þ p2 on ½0; 12 and by (1) for ð12; 1: This function is S-shaped, continuous and differentiable throughout ½0; 1: 2. Preference between gambles Let X be the set of finite-support probability distributions or gambles x; y; y on a nondegenerate real interval I of outcomes a; b; y: The convex combination lx þ ð1 lÞy of x; yAX with 0plp1 is the gamble with probability lxðaÞ þ ð1 lÞyðaÞ for aAI: We intend that the convex combination lx þ ð1 lÞy is a gamble entirely defined by its probability distribution. The related compound gamble, x with 247 probability l and y with probability 1 l; which we do not consider, may or may not be evaluated equivalently. Let h be a preference-or-indifference relation on X with indifference B and strict preference g defined by xBy if xhy and yhx; xgy if xhy and not ðyhxÞ: Let x!y mean that ygx: We assume that preference increases over I so that a!b when a; bAI and aob: As mentioned above, we denote a pwf by p and assume that p is strictly increasing and continuous from ½0; 1 onto ½0; 1; so pð0Þ ¼ 0 and pð1Þ ¼ 1: If pðpÞ ¼ p for all pA½0; 1 then p is the identity. We assume also for the RDU model and its EU specialization that the utility function u on outcomes is continuous and strictly increasing from I into R: Let Xn denote the subset of gambles that have positive probabilities for exactly n points in I; so Xn ¼ fxAX : jfaAI : xðaÞ40gj ¼ ng: We identify X1 with I and refer to members of X2 as binary gambles. Suppose xAXn has positive probability pi ¼ xðai Þ for outcome ai with a1 oa2 o?oan : The RDU utility for x with respect to p and u is " ! !# n1 n n X X X p pi p pi uðaj Þ Epu ðxÞ ¼ j¼1 i¼j i¼jþ1 þ pðpn Þuðan Þ: When p is the identity, this simplifies to the EU utility n X Eu ðxÞ ¼ pj uðaj Þ: j¼1 The RDU representation for h with respect to p and u is xhy3Epu ðxÞXEpu ðyÞ; for all x; yAX : The corresponding EU representation is xhy3Eu ðxÞXEu ðyÞ; for all x; yAX : More general treatments of these representations, which allow arbitrary outcome sets and do not presume continuity, are available in Fishburn (1982) for EU and Quiggin (1993) and Abdellaoui (2002) for RDU. Quiggin (1993) provides extensive analyses of axioms and economic applications for RDU. As background for BECI, we note four independence axioms that hold for the EU representation but are violated by most RDU proposals. IA1. For all x; y; zAX and all 0olo1 : xgy ) lx þ ð1 lÞzgly þ ð1 lÞz: IA2. For all x; y; zAX : xBy ) 12x þ 12zB12y þ 12z: IA3. For all xAX2 ; all aAI and all 0olo1: xBa ) lx þ ð1 lÞaBa: BECI. For all xAX2 and all aAI: xBa ) 12x þ 12aBa: 248 D.E. Bell, P.C. Fishburn / Journal of Mathematical Psychology 47 (2003) 244–258 Despite arguments in favor of the axioms, which include two-stage options (see Example 1) with fair ðl ¼ 1 1 2Þ or biased ðla2Þ coins and avoidance of money pumps (Raiffa, 1968), extensive empirical analyses have shown that they are often violated in practice. IA1 is Jensen’s (1967) axiom, and IA2 is due to Herstein and Milnor (1953). In conjunction with ordering and Archimedean axioms, each implies the existence of u on I that satisfies the EU representation. BECI is a weak version of the Herstein–Milnor axiom, and IA3 is a biased-coin generalization of BECI. Although we suggest that BECI could have descriptive validity and will argue for its compatibility with RDU, we will not do this for IA3. There are two reasons. The first is that Allais-type examples severely challenge the descriptive robustness of IA3. viors of p near 0 and 1. By setting p ¼ 2t and then p ¼ 1 2t in (1), we get 1 pðtÞ 1 for 0oto ; ð2Þ p þ t ¼ 1 þ pðtÞ 2 pð2tÞ 2 1 1 pð1 tÞ 1 p t ¼ pð1 2tÞ for 0oto : ð3Þ 2 1 pð1 2tÞ 2 Let a¼p 1 : 2 Then, under continuity, (2) and (3) give pðtÞ ¼ 1 a; the condition near 0: lim t-0 pð2tÞ the condition near 1: Example 2. Paula is indifferent between $2,000,000 with certainty and the binary gamble x that pays $3,000,000 with probability 0.98 and nothing otherwise. IA3 says that she will be indifferent between $2,000,000 with certainty and ð0:005Þx þ ð0:995Þa; where a is a sure thing for $2,000,000. However, the convex combination has pr: 0:0001 for $0; pr: 0:995 for $2; 000; 000; pr: 0:0049 for $3; 000; 000; and Paula says she prefers this to $2,000,000 because there is only one chance in 10,000 that it returns $0 and 49 chances in 10,000 that it returns $1,000,000 more than the expected outcome of $2,000,000. The second reason is that IA3 reduces RDU to EU. Theorem 1. If IA3 holds for the RDU representation, then p is the identity. The proofs of Theorem 1 and subsequent results are in the appendix. 3. BECI basics Our fundamental result for BECI is Theorem 2. BECI holds for the RDU representation if and only if (1) holds. We show next that the strict preference companion of BECI follows from it and the RDU representation. Lemma 1. Suppose BECI and the RDU representation hold. If xAX2 and bAI then xgb ) 12x þ 12bgb; and bgx ) bg12x þ 12b: Although (1) allows a variety of shapes for p; it implies special conditions between pð12Þ and the beha- lim t-0 1 pð1 tÞ ¼ a: 1 pð1 2tÞ For example, given d40; there exists an e40 such that 1 þ pðtÞ pð12 þ tÞA½1 a d; 1 a þ d for all toe; so pðtÞ=pð2tÞA½1 a d; 1 a þ d for all toe: It follows by taking d towards 0 that the limit defined in the condition near 0 exists and equals 1 a: Both limits are obviously in ð0; 1Þ and, because of their dual natures, their sum equals 1. When p is near 0, the condition near 0 says that pðp2Þ is approximately ð1 aÞpðpÞ; when p is near 1, the condition near 1 says that 1 pð1þp 2 Þ is approximately a½1 pðpÞ: We say that p is concave fconvexg near 0 if for some e40; pðp2Þ412pðpÞfpðp2Þo12pðpÞg for all 0opoe: Similarly, p is concave fconvexg near 1 if for some e40; 1þp 1 1 1 pð1þp 2 Þo2½1 pðpÞf1 pð 2 Þ42½1 pðpÞg for all 1 eopo1: Most studies of pwfs conclude that ao12 and suggest that pðpÞ ¼ p for p in the neighborhood of 0.3–0.4. Many, including Wu and Gonzalez (1996, 1999), Prelec (1998), Abdellaoui (2000) and Bleichrodt and Pinto (2000), also support the hypothesis that p begins with a strictly concave segment ½pðlx þ ð1 lÞyÞ4lpðxÞ þ ð1 lÞpðyÞ; 0olo1; xay followed by a strictly convex segment ½pðlx þ ð1 lÞyÞolpðxÞ þ ð1 lÞpðyÞ; 0olo1; xay; hence that p is concave near 0 and convex near 1. The latter behavior is impossible under BECI in the RDU context unless a ¼ 12: Lemma 2. Suppose BECI and the RDU representation hold. If pð12Þo12 then p is concave near 0 and concave near 1. If pð12Þ412 then p is convex near 0 and convex near 1. Lest one conclude that this dooms BECI as a viable descriptive condition, we emphasize that the conditions near 0 and 1 apply only to p’s extremes and do not invalidate the phenomena of subjective overvaluation of small probabilities and subjective undervaluation of large probabilities. For example, if a is near 0.4 and the ratios in the conditions are near their limit values within D.E. Bell, P.C. Fishburn / Journal of Mathematical Psychology 47 (2003) 244–258 249 0.01 of the extremes, then p near 0 and 1 may look like the parts of curve 1 in Fig. 1 near 0 and 1. In general, the intervals that contain 0 and 1 and over which p’s curvature does not reverse can be arbitrarily short. An important consequence of BECI’s even-chance feature is that p’s arguments in (1), namely p=2; p and ð1 þ pÞ=2; span exactly half of ½0; 1: If p is specified on a length-12 interval in ½0; 1 then (1) effectively determines p on all of ½0; 1: We comment on the construction of p for such a specification in the next several paragraphs, then conclude the section with cautionary examples and a final comment on extension by (1). Suppose first that p is specified on ½0; 12,f1g with p continuous and increasing on ½0; 12; pð0Þ ¼ 0 and pð1Þ ¼ 1: We assume that p satisfies the condition near 0. Because 2tp123tp14; we first extend p to ð12o34; and p thus extended is continuous at 12 because it satisfies the condition near 0. The next step extends p to ð34; 78 using (2) for 14otp38 because 2tp343tp38: Each succeeding step covers half the interval between the right end of the interval thus far covered and 1. The procedure is similar but two-sided if p is first specified on ½l; l þ 12 with 0olo12: An extension step in either direction covers half the remaining interval in that direction. Upward extension can use (2), downward extension (3), or modified expressions equivalent to (1). Our next lemma shows that the upward extension from ½0; 12 satisfies some of the properties presumed earlier for p: However, it need not satisfy all such properties, including monotonicity. As explained earlier, we therefore use another designation, w; instead of p: The oddity of Example 3 is compounded in the next example. For convenience there, w is defined to be constant over subintervals of ½0; 12: This violates strict monotonicity on ½0; 12; but essentially the same results will be obtained if the constant pieces are modified by giving them small positive slopes. Lemma 3. Suppose w : ½0; 1-R,fN; Ng is increasing and continuous on ½0; 12; has 0 ¼ wð0Þowð12Þ ¼ aowð1Þ ¼ 1; and satisfies (1) in place of p: Then w increases on fð2k 1Þ=2k : k ¼ 1; 2; y; g and Although we have emphasized extension of w to ½0; 1 from a length-12 interval, complete extension can proceed from specification of w on disjoint intervals with total length slightly greater than 14 provided that certain precautions are taken. Suppose in particular that w is specified on ½e; 2e,½12; 34 with small e40; and consider (1) for w in the form k 2 1 lim w ¼ 1: k-N 2k Unfortunately, the nice behavior of w on fð2k 1Þ=2k g by extension through (1) from w on ½0; 12 does not necessarily extend to other properties of p: Example 3. Suppose wðpÞ ¼ 3p2 for 0ppp12; where the constant 3 is chosen to satisfy the condition near zero. Using (2), we have wðpÞ ¼ 32ð1 2p þ 2p2 Þ for 12ppp34; so wð0:75Þ ¼ 15 16 ¼ 0:9375: But, by (2) for w; wð0:3Þ wð0:8Þ ¼ 1 þ wð0:3Þ wð0:6Þ 0:27 ¼ 1 þ 0:27 ¼ 0:9238; 0:78 so w decreases over parts of ½0; 1: Example 4. Let w on ½0; 12 be the continuous four-piece function defined by 2p 1 4 for 0ppp18; for 18ppp14; 5 ; 4p 34 for 14ppp16 1 2 for 5 1 16ppp2: When w is extended by (2), we obtain the picture shown in Fig. 2 for wðpÞ from p ¼ 0 to p ¼ 0:904: Fig. 2a shows a decrease in w near 0.6. In the extension of Fig. 2b, w goes negative near a¼0:80646 to a local ’ 1 minimum of wð13 Þ ¼ and remains negative until 16 2 b¼0:81392: Because w ¼ 0 at a and b; singularities ’ 0 0 0 occur in the ð78; 15 16 region at a and b ; where a ¼ 0 ð1 þ aÞ=2 and b ¼ ð1 þ bÞ=2: Other singularities occur whenever w crosses the abscissa. If we define the value of w at a singularity by w’s ‘limit’ as the singularity is approached from the left, then wða0 Þ ¼ N and wðb0 Þ ¼ þN: wð2tÞ ¼ wðtÞ : 1 þ wðtÞ wð12 þ tÞ When t ¼ e; this constrains the values of wð2eÞ; wðeÞ and wð12 þ eÞ: Assuming consistency at these values, the formula extends w to ½2e; 4e; then to ½4e; 8e; y; and also to ½e=2; e; then to ½e=4; e=2; y . Continuity would require wðtÞ-0 as t-0 for extension downward, and consistency near 12 for extension upward. Once w is determined on ½0; 34; its upward extension to ½34; 78; ½78; 15 16; y would proceed as before. Because of alignment/consistency problems in extensions based on intervals of length less than 12; we focus henceforth on extensions based on ½0; 12 with an exception noted in Section 5 that is based on ½0:125; 0:625: D.E. Bell, P.C. Fishburn / Journal of Mathematical Psychology 47 (2003) 244–258 250 These functions and their extensions to ½12; 1 are shown in Fig. 3. Each of (4a) through (4b) satisfies the condition near 0 and has wð0Þ ¼ 0: The first has wð12Þ ¼ 1 21=2 ¼0:29289 and the others have wð12Þ ¼ 12: Although ’ their full extensions to ½0; 1 are not typical of shapes for p in the literature, we will remedy this in the next section. Our sufficient condition is set within the following context. 4. Increasing extensions Although Examples 3 and 4 show that extension of w by (1) from its specification on ½0; 12 can grossly violate properties presumed earlier for p; such violations need not occur. In this section, we establish a sufficient condition under which extension by (1) from an increasing and continuous w on ½0; 12 is increasing and continuous on ½0; 1: The condition is then illustrated for the following four w’s on ½0; 12: pffiffiffi pffiffiffi wðpÞ ¼ ð 2 1Þ p; ð4aÞ wðpÞ ¼ 0:37p þ 0:8p2 þ 0:92p3 ; ð4bÞ wðpÞ ¼ e1=2 pep ðe1=2 ¼0:60653Þ; ’ ð4cÞ wðpÞ ¼ e1=2 pep ðe1=2 ¼1:64872Þ: ’ ð4dÞ Axiom 1. w on ½0; 12 into R is continuous, increasing, satisfies the condition near 0 and has wð0Þ ¼ 0 and wð12Þo1: In addition, w is differentiable on ð0; 12Þ: The condition is based solely on the behavior of w and w0 on ð0; 12Þ: Let w0 ðf ðtÞÞ ¼ dwðf ðtÞÞ=dt ¼ ½dwðf ðtÞÞ= 1 1 0.9 0.9 0.8 0.8 0.7 0.7 0.6 0.6 0.5 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0 (a) 0 0.2 0.4 0.6 0.8 1 0 (b) 0 0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1 1 1 0.9 0.9 0.8 0.8 0.7 0.7 0.6 0.6 0.5 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0 (c) 0 0.2 0:5 0.4 0:5 Fig. 3. (a) wðxÞ ¼ ð2 1Þ x x expð0:5 xÞ for xp0:5: 0.6 0.8 0 (d) 0 1 2 3 for xp0:5: (b) wðxÞ ¼ 0:37x þ 0:8x þ 0:92x for xp0:5: (c) wðxÞ ¼ x expðx 0:5Þ for xp0:5: (d) wðxÞ ¼ D.E. Bell, P.C. Fishburn / Journal of Mathematical Psychology 47 (2003) 244–258 df ðtÞ½df ðtÞ=dt; and let GðtÞ ¼ wðtÞw0 ð2tÞ w0 ðtÞwð2tÞ½1 wð2tÞ ; wðtÞ½1 wð2tÞfwð2tÞ wðtÞ½1 wð2tÞg 1 0oto : 4 Axiom 1 implies that 14wð2tÞ4wðtÞ4wðtÞ½1 wð2tÞ; so the denominator of G is positive when the axiom holds. Also let 2t þ 2k1 1 ak ðtÞ ¼ 2k 1 for 0ptp ; k ¼ 1; 2; y : 4 Thus, a1 ðtÞ ¼ t; a2 ðtÞ ¼ ð2t þ 1Þ=4; y; and a1 ð14Þ ¼ a2 ð0Þoa2 ð14Þ ¼ a3 ð0Þoa3 ð14Þ ¼ a4 ð0Þo?o12 with 2akþ1 ðtÞ ¼ 12 þ ak ðtÞ for k ¼ 1; 2; y : Finally, let sk ðtÞ ¼ w0 ðak ðtÞÞ 1 for 0oto ; k ¼ 1; 2; y : wðak ðtÞÞ 4 Because 0oak ðtÞo12 and w is increasing on ½0; 12 when Axiom 1 holds, the axiom implies that sk ðtÞ40 in all cases. Theorem 3. Suppose w : ½0; 1-R satisfies Axiom 1. If N X 1 ð5Þ sk ðtÞoGðtÞ for all 0oto ; 4 k¼2 and if w satisfies (1) in place of p; then w is continuous and increasing on ½0; 1Þ; and limp-1 wðpÞ ¼ 1: Thus, when Axiom 1 and (5) hold, the continuous extension of w from ½0; 12 to ½0; 1 via (1) has the properties assumed initially for p and satisfies BECI in the context of the RDU representation. Although Example 3 and a slight modification of Example 4 show that (5) can easily fail when Axiom 1 holds, many ‘reasonable’ w that satisfy Axiom 1 do satisfy (5). The hypotheses of Theorem 3 imply that w has left and right derivatives at 12; 34; 78; y; but they do not imply that w is differentiable there. This is evident in Fig. 3, where the extensions by (1) of (4a)–(4d) have kinks at 12; 3 4; y: The extensions apparently increase over ½0; 1; but Theorem 3 allows rigorous proofs. We conclude this section with a summary for pffiffiffieach case. pffiffiffi (4a): wðpÞ ¼ c p; c ¼ 2 1; on ½0; 12: Differentiation and simplification give 1 1 GðtÞ ¼ pffiffi ; ¼ pffiffi t½1 c2 2t t½1 0:34315t sk ðtÞ ¼ 1 2t þ 2k1 1 for kX2: 251 Within the range of 0 to 14 for t; we have N N X X 1 1 max sk ðtÞ ¼ sk ð0Þ ¼ 1 þ þ þ ?o1 t 3 7 2 k¼2 1 1 þ þ?¼2 2 4 1 o 2:1877 ¼ G ¼ min GðtÞ: t 4 þ Hence (5) holds and, by Theorem 3, w extends by (1) to a continuous and increasing function on ½0; 1: (4b): wðpÞ ¼ ap þ bp2 þ cp3 ; ða; b; cÞ ¼ ð0:37; 0:80; 0:92Þ; on ½0; 12: Evaluation of GðtÞ shows that it is minimized near t ¼ 0:05 and that GðtÞ45:6 for all 0oto14: We have ! 1 a þ 2bak ðtÞ þ 3cak ðtÞ2 sk ðtÞ ¼ k1 : 2 aak ðtÞ þ bak ðtÞ2 þ cak ðtÞ3 This decreases in t; so max sk ðtÞ ¼ sk ð0Þ: Computations give s2 ð0Þ þ s3 ð0Þ þ s4 ð0Þo4:7: For kX5; sk ð0Þo 4=2k1 ¼ 1=2k3 ; so s5 ð0Þ þ s6 ð0Þ þ ?o12: Therefore (5) holds: N X max sk ðtÞo5:2o5:6omin GðtÞ: k¼2 (4c): wðpÞ ¼ cpep ; c ¼ e1=2 ; on ½0; 12: Here GðtÞ ¼ 2et ½1 þ 2cð1 þ tÞe2t : ½1 2cte2t ½2et 1 þ 2cte2t This is minimized at t ¼ 0; where Gð0Þ ¼ 4:4261: Moreover, 1 2k sk ðtÞ ¼ k1 1 þ ; 2 2t þ 2k1 1 so max sk ðtÞ ¼ sk ð0Þ: We have N X 1 1 1 max sk ðtÞ ¼ 1 þ 2 1 þ þ þ þ ? 3 7 15 k¼2 1 1 1 o1 þ 2 1 þ þ þ 3 7 15 1 1 þ þ? þ 2 16 32 ¼ 4:3357o4:4261 ¼ min GðtÞ; so (5) holds. (4d): wðpÞ ¼ cpep ; c ¼ e1=2 ; on ½0; 12: Here 4cð1 tÞe2t 2 ; ð1 2cte2t Þð2 et þ 2cet Þ 1 2k1 2t þ 1 sk ðtÞ ¼ k1 k1 k ¼ 2; 3; y : 2 2 þ 2t 1 GðtÞ ¼ GðtÞ P decreases from about 4.59 at t ¼ 0 to 1.56 at t ¼ 14; and N 2 sk ðtÞ decreases from PN3.72 at t ¼ 0 to1 1.43 at t ¼ 14: GðtÞ remains above 2 sk ðtÞ over ½0; 4; so (5) holds. D.E. Bell, P.C. Fishburn / Journal of Mathematical Psychology 47 (2003) 244–258 252 5. Comparisons 6. Smooth BECI transformations In this section, we note that (1) accommodates pwfs that are quite similar to pwfs proposed by others. We then prove that essentially none of those functions satisfies (1) or BECI unless it is the identity. Fig. 1a shows a pwf that satisfies (1) approximately by a series of curve-fitting adjustments. Our fit of the curve to (1) fixed pðpÞ ¼ p for p ¼ 0:375 and worked outward from ½0:125; 0:625: The curve of Fig. 1a is not perfectly smooth (cf. Fig. 3) but is continuous and increasing. We constructed it to minimize differentiability kinks and simulate pwfs that appear in recent studies. It is concave near 1 (cf. Lemma 2) but otherwise resembles the smooth concave–convex standards, an example of which is Fig. 1b. We consider six such standard forms for p on ½0; 1: We conclude our technical analyses with observations on differentiability of pwfs that satisfy BECI in the RDU context. We focus first on differentiability at 12 and observe later that this implies differentiability at other key extension points, namely 34; 78; y: Our initial result at 1 2 is based on the following strengthening of Axiom 1. a pðpÞ ¼ ebðln pÞ ; pðpÞ ¼ pðpÞ ¼ p 0oao1; b40; g ½pg þ ð1 pÞg 1=g apg apg ; þ ð1 pÞg a pðpÞ ¼ eðb=aÞð1p Þ ; pðpÞ ¼ pb ; ð6aÞ ; g40; ð6bÞ a; g40; ð6cÞ aa0; b40; ð6eÞ b40; pðpÞ ¼ ð1 a ln pÞb ; ð6dÞ a40; b40: ð6fÞ All six are included in Prelec (1998). Prelec (1998) and Luce (2001) provide axioms for (6a). Earlier references include Tversky and Kahneman (1992) for (6b) and Tversky and Fox (1994) for (6c). When p is not the identity, none of the six satisfies (1), so all violate BECI for the RDU representation. Moreover, all but (6c) violate the condition near 0 unless p is the identity, and (6c) reduces to the identity if it satisfies both the condition near 0 and the condition near 1. Theorem 4. The only cases of (6a)–(6f) that satisfy the condition near 0 are (6b) and (6c) with pðpÞ ¼ p for all pA½0; 1; and (6c) with a ¼ 2g 1: If (6c) satisfies the condition near 0 and the condition near 1; then p is the identity. It follows from differentiability results in the next section that (6a)–(6f) cannot satisfy BECI except in special cases. Our proof of this in the appendix focuses on the more general condition near 0. An obvious corollary of Theorem 4 is that if C is a set of conditions that implies one of (6a)–(6f) within the RDU representation, then either C and BECI are incompatible or they jointly reduce RDU to EU. Axiom 2. Axiom 1 holds, w satisfies (1) and is increasing on ½0; 1; and w is twice differentiable near 0 with w0 ð0Þ and w00 ð0Þ finite and w0 ð0Þ40: Axiom 2 does not apply to certain forms of w like (4.1) that do not have finite derivatives at 0. We denote the limit of w0 ðtÞ as t approaches 12 from below (from þ above) by w0 ð12 Þ½w0 ð12 Þ: Lemma 4. Suppose w satisfies Axiom 2. Then w is differentiable at 12 if and only if 00 1 3 1 w ð0Þ : ð7Þ ¼ w0 ð0Þ þ w w0 2 4 2 w0 ð0Þ We continue by involving the third derivative near 0 to conclude that wð12Þ ¼ 12 when w00 ð0Þa0: This conclusion does not presume differentiability at 12: Lemma 5. Suppose w satisfies Axiom 2 and has a third derivative near 0 with w000 ð0Þ finite. Then 1 1 00 2 w 2 w ð0Þ ¼ 0: The techniques of the proofs of Lemmas 4 and 5 can be extended to determine conditions for higher-order derivatives at 12: We outline only the second-order result. Lemma 6. Suppose Axiom 2 holds, w is twice differentiable on ð0; 12Þ; is differentiable at 12; and has third and fourth derivatives near 0 with w000 ð0Þ and w0000 ð0Þ finite. Then w is second differentiable at 12 if and only if 4 1 1 w000 ð0Þ 00 1 00 w ð2 Þ ¼ w ð0Þ þ w 2 3 6 w0 ð0Þ 00 2 w ð0Þ 3 2w 12 : 2 w0 ð0Þ When wð12Þ ¼ 12; which is true by Lemma 5 when w ð0Þa0; the preceding equality simplifies to 1 1 w000 ð0Þ 1 w00 ð0Þ 2 w00 : ¼ w00 ð0Þ þ 2 2 w0 ð0Þ 2 w0 ð0Þ 00 Under Axiom 2, it is easily seen that w is differentiable under extension from ½0; 12 to ½0; 1 within the open intervals ð12; 34Þ; ð34; 78Þ; y: Hence, the question of differentiability everywhere concerns only differentiability at D.E. Bell, P.C. Fishburn / Journal of Mathematical Psychology 47 (2003) 244–258 1 3 7 2; 4; 8; y . Lemma 4 addressed differentiability at 12: Our next result shows that differentiability at 12 implies differentiability at the other points in question. 253 d ¼ 8ð1 aÞ 4b 2c: ground, the present paper might be viewed as an attempt to see how much of EU’s independence axiom can be retained without destroying the integrity of the RDU viewpoint and the behaviors it represents. Our contention is that BECI accomplishes these ends. A subtheme of the paper is that several roads lead from RDU to EU, including independence axioms stronger than BECI, and BECI in conjunction with strict concavity–convexity of the pwf p or other conditions that lead to parametric forms for p: Another route from RDU to EU is a contextual uncertainty condition in Bell and Fishburn (2000). But BECI by itself does not have this effect. The major technical focus of the paper is the functional equation (1) that is equivalent to BECI in the RDU setting. We have seen that (1) accommodates a wide range of shapes for p but can lead to unusual and even bizarre possibilities when applied under extension to ½0; 1 from specification on a length 12 interval in ½0; 1 such as ½0; 12: Nevertheless it can, with careful assessment and to a reasonable approximation, emulate weighting functions that have emerged as promising descriptors of probability distortion. Is BECI a viable descriptive condition? We have suggested that it may be, but remain open on the point. In any event, BECI is empirically testable, and we look forward to assessment of its descriptive tenability. When 0oao1; w is S-shaped; when 1oao1:5; w is inverse S-shaped. Appendix Theorem 5. Suppose Axiom 2 holds and w is differentiable at 12: Then w is differentiable everywhere. The proof method for Theorem 5 extends easily to all degrees of differentiability. For example, if w is assumed to be twice differentiable on ½0; 12 and at 12; then it is twice differentiable at all bk and is therefore twice differentiable everywhere. Figs. 1c and d are examples of differentiable w functions that satisfy BECI. Fig. 1c takes w as a quintic polynomial on ½0; 12; and Fig. 1d has a linear plus exponential form on ½0; 12: We discuss the latter in Example 6, preceded by comments on quartic polynomials in Example 5. Example 5. Suppose w satisfies (1) with wðpÞ ¼ ap þ bp2 þ cp3 þ dp4 on ½0; 12: Then w is continuous and differentiable on ½0; 1 if and only if the coefficients satisfy b ¼ 12fð3 þ 9aÞ þ ½ð3 þ 9aÞ2 þ 48að1 aÞða þ 2Þ1=2 g; c ¼ 4b 2b=a þ 16ð1 aÞ; Example 6. Suppose w satisfies (1) with wðpÞ ¼ ap þ bð1 ecp Þ on ½0; 12: Then w is continuous and twice differentiable if and only if b ¼ ð1 cec=2 =4Þ=½2ð1 ec=2 Þ c; Proof of Theorem 1. We assume that IA3 holds for the RDU representation. With aoc in I; assume without loss of generality that uðaÞ ¼ 0 and uðcÞ ¼ 1: Let xAX2 have probabilities 1 p for a and p for c: Then Epu ðxÞ ¼ pðpÞ: a ¼ 1 2bð1 e c=2 Þ: The cases of 5pcp10 are well-behaved. Moreover, w is S-shaped when co3:18 and inverse S-shaped when c43:18: The curve for c ¼ 8 is graphed in Fig. 1d. 7. Discussion The RDU model in this paper, which follows Quiggin’s (1993) formulation, is a generalization of EU that accounts for subjective weighting or distortion of probabilities in evaluation of gambles. Although often presented as a descriptive theory, RDU has normative overtones such as complete ordering and reduction of compound lotteries to holistic gambles. The main normative principle eschewed by RDU is the independence axiom of EU that rules out the probability-distortion phenomenon. Against this back- Continuity assures that uðbÞ ¼ pðpÞ for some aoboc; and with xBb; IA3 gives pðpÞ ¼ uðbÞ ¼ Epu ðlx þ ð1 lÞbÞ ¼ ½pðlp þ 1 lÞ pðlpÞuðbÞ þ pðlpÞ: Letting t ¼ lp þ 1 l; the equality for IA3 can be written as p p ð1 tÞ 1p pðpÞ ; 0opptp1; po1: ¼ ½1 pðtÞ 1 pðpÞ For any fixed b41 and probabilities p; q; s and t such that pps; ppt and qpminfs; t=bg ¼ t=b; set p p q ¼ ð1 sÞ and bq ¼ ð1 tÞ : 1p 1p The preceding equality implies, after cancellation of the pðpÞ=½1 pðpÞ terms, that pðqÞ 1 pðsÞ ¼ : pðbqÞ 1 pðtÞ D.E. Bell, P.C. Fishburn / Journal of Mathematical Psychology 47 (2003) 244–258 254 By varying p and taking q0 ¼ ð1 sÞp0 =ð1 p0 Þ and bq0 ¼ ð1 tÞp0 =ð1 p0 Þ; we obtain pðq0 Þ=pðbq0 Þ ¼ ½1 pðsÞ=½1 pðtÞ: Because t can be arbitrarily close to 1 here, we have 0 pðqÞ pðq Þ ¼ for all positive q; q0 o1=b: pðbqÞ pðbq0 Þ Equality also holds by continuity for q ¼ 1=b; so pðqÞ 1 ¼p pðbqÞ b for all 0oqp1=b: Now take q ¼ 1=b2 ; 1=b3 ; y to obtain n 1 1 p n ¼ p b b for n ¼ 1; 2; y : Let a ¼ pð12Þ: When 1obo2 and bn ¼ 2m for positive integers mon; we have k ( 1=n )k 1 1 1 p km=n ¼ p m=n ¼ p m ¼ akm=n : 2 2 2 Because fkm=n : 1pmon; k ¼ 1; 2; yg is dense in ½0; 12; it follows from continuity that 1 p w ¼ aw for all wX1: 2 Equivalently, pðpÞ ¼ alog2 ð1=pÞ for all 0ppp12: Given 0opptp12; substitution in the equation that concludes the opening paragraph and rearrangement give 1 alog 1t 1 1 alog t ¼ a 1 log 1p 1a 1 log p ; all logs base 2: Take t ¼ 12 and p ¼ 14 in this equation to get log 3 1 1 : 1þ ¼ a a This holds when a ¼ 12 and for no other a40: for wX0; derivatives show that 1 þ w intersects wlog 3 only at w ¼ 2: We conclude that pðpÞ ¼ p for all 0ppp1=2: Eq. (1) shows that this extends uniquely under BECI to the identity. & Proof of Theorem 2. Assume that the RDU representation holds. Suppose BECI also holds, aoboc are outcomes in I; xAX2 has probabilities 1 p for a and p for c; and xBb: Then Epu ðxÞ ¼ uðbÞ; i.e., uðbÞ ¼ uðcÞpðpÞ þ uðaÞ½1 pðpÞ: By BECI, 12x þ 12bBb and therefore Epu ð12x þ 12bÞ ¼ uðbÞ; i.e., p p 1þp uðcÞp þ uðbÞ p p 2 2 2 1þp þ uðaÞ 1 p ¼ uðbÞ: 2 Substitution here for uðbÞ from the preceding equation and rearrangement give ½uðcÞ uðaÞ p 1þp p ½1 pðpÞ ¼ 0: pðpÞ 1 p 2 2 Because uðcÞ uðaÞ40; the expression in braces equals 0 and gives the equation of (1) for 0opo1: Under the original suppositions, including xBb; the derivation’s steps are reversible. Since (1) holds trivially for pAf0; 1g; the proof is complete. & Proof of Lemma 1. Suppose x has probabilities 1 p for a and p for c with aoc in I: Clearly, xgb ) boc: Suppose xgb and apboc: With p p 1 1 1þp Epu x þ b ¼ uðcÞp þ uðbÞ p p 2 2 2 2 2 1þp þ uðaÞ 1 p ; 2 we use (1) to obtain 1 1 x þ bgb 2 2 1 1 3Epu x þ b 4uðbÞ 2 2 p 1þp 3½uðaÞ uðbÞ 1 p 4½uðbÞ uðcÞp 2 2 3½uðaÞ uðbÞ½1 pðpÞ4½uðbÞ uðcÞpðpÞ 3Epu ðxÞ4b 3xgb: If xgb and boaoc; then clearly Epu ð12x þ 12bÞ4uðbÞ; so þ 12bgb: The proof for bgx is similar. & 1 2x Proof of Lemma 2. Given the initial hypotheses, suppose a ¼ pð12Þo12: Let b be such that 1 a4b412: By the condition near 0, pðpÞ=pð2pÞ4b for all sufficiently small p40; so p is concave near 0. By the condition near 1, ½1 pð1 pÞ=½1 pð1 2pÞo1 b for all sufficiently small p40; so p is concave near 1. The proofs for a412 are similar. & D.E. Bell, P.C. Fishburn / Journal of Mathematical Psychology 47 (2003) 244–258 Proof of Lemma 3. Given the lemma’s hypotheses, let ak ¼ 2k1 1 ; 2k k ¼ 2; 3; y; rk ¼ wða2 Þwða3 Þ?wðak Þ; 1 wðakþ1 Þ½1 wð2akþ1 Þ w þ akþ1 ¼ 1 2 wð2akþ1 Þ wðakþ1 Þ½1 wð12 þ ak Þ : ¼1 wð12 þ ak Þ It follows that Because wðakþ1 Þoa; and aowð12 þ ak Þ by previous steps, it follows that wð12 þ ak Þowð12 þ akþ1 Þ; so w increases on fð2k 1Þ=2k : k ¼ 1; 2; yg: & Proof of Theorem 3. Assume that w : ½0; 1-R satisfies Axiom 1, (1) and (5). By (2) for w; k ¼ 2; 3; y : Note that 2akþ1 ¼ 12 þ ak ¼ ð2k 1Þ=2k : By (2), wð12Þ wðtÞ½1 wð2tÞ ; w 12 þ t ¼ 1 wð2tÞ w0 ¼ a; wð1Þ½1 wð12Þ 3 ð1 aÞr2 ; ¼1 w ¼1 4 1 4 a wð2Þ wð3Þ½1 wð34Þ 7 ð1 aÞr3 w ¼1 ; ¼1 8 3 8 a ð1 aÞr2 wð4Þ and, by continuation, k 2 1 ð1 aÞrk w ; ¼1 k 2 a ð1 aÞðr2 þ r3 þ ? þ rk1 Þ 1 wðtÞw0 ð2tÞ w0 ðtÞwð2tÞ½1 wð2tÞ þ t ¼ ; 2 ½wð2tÞ2 wða2 ðtÞÞ½1 wð12 þ tÞ w 12 þ a2 ðtÞ ¼ 1 ; wð12 þ tÞ ¼ Because w increases on ½0; 12; we have 0ork oak1 ; so rk -0 as k-N: In addition 1 0ptp : 4 wða2 ðtÞÞw0 ð12 þ tÞ w0 ða2 ðtÞÞwð12 þ tÞ½1 wð12 þ tÞ ½wð12 þ tÞ2 wða2 ðtÞÞ½1 wð12 þ tÞ wð12 þ tÞ ( ) wðtÞw0 ð2tÞ w0 ðtÞwð2tÞ½1 wð2tÞ s2 ðtÞ wð12 þ tÞ½1 wð12 þ tÞwð2tÞ2 wða2 ðtÞÞ½1 wð12 þ tÞ wð12 þ tÞ wðtÞw0 ð2tÞ w0 ðtÞwð2tÞ½1 wð2tÞ s2 ðtÞ wðtÞ½1 wð2tÞfwð2tÞ wðtÞ½1 wð2tÞg wða2 ðtÞÞ½1 wð12 þ tÞ ½GðtÞ s2 ðtÞ; ¼ wð12 þ tÞ rk e ¼ k¼2 a e for some o 1a 1 0oto ; 4 Then, for 0oto14; w0 12 þ a2 ðtÞ ¼ N X 1 0ptp ; 4 so w0 ð12 þ tÞ40 because the numerator of w0 ð12 þ tÞ equals the numerator of G: The condition near 0 ensures that w is continuous at p ¼ 12; so w is continuous and increasing on ½0; 34: For the next extension step to ð34; 78; (2) and 2a2 ðtÞ ¼ 1 1 2 þ a1 ðtÞ ¼ 2 þ t give k ¼ 3; 4; y : r2 þ r3 þ ? þ rk1 o 255 e40; so a ð1 aÞðr2 þ r3 þ ? þ rk1 Þ4ð1 aÞe for all k: It follows that wðð2k 1Þ=2k Þo1 for all k; and wðð2k 1Þ=2k Þ-1 as k-N: Now 1 3 ð1 aÞ 1 w w ow 3ao1 2 4 a 4 1 3w oa; 4 so wð12Þowð34Þ: For kX3; 1 1 w þ ak ow þ akþ1 2 2 wðakþ1 Þ½1 wð12 þ ak Þ 1 3w þ ak o1 2 wð12 þ ak Þ 1 3wðakþ1 Þow þ ak : 2 so w0 ð12 þ a2 ðtÞÞ40 by (5). Hence, w is continuous and increasing on ½0; 78: By the proof of Lemma 3, we have wð12 þ a2 ðtÞÞo1 for 0ptp14: For the following induction, the same proof gives wð12 þ ak ðtÞÞo1 for 0ptp14 for k ¼ 3; 4; y . We omit t to simplify notation. For kX2; (2) gives w wðakþ1 Þ½1 wð12 þ ak Þ ¼ 1 ; þ a kþ1 2 wð12 þ ak Þ 1 1 0ptp ; 4 w0 ð12 þ akþ1 Þ ¼ wðakþ1 Þw0 ð12 þ ak Þ w0 ðakþ1 Þwð12 þ ak Þ½1 wð12 þ ak Þ wð12 þ ak Þ2 D.E. Bell, P.C. Fishburn / Journal of Mathematical Psychology 47 (2003) 244–258 256 1 for 0oto : We claim that for kX1; 4 wðakþ1 Þ½1 wð12 þ ak Þ w0 12 þ akþ1 ¼ Q wð12 þ ak Þ kj¼2 wð12 þ aj Þ " # j1 kþ1 Y X 1 G s2 sj wð2 þ ai Þ : j¼3 i¼2 Pkj¼2 wð12 With the understanding that þ aj Þ ¼ 1 when k ¼ 1; this is true for k ¼ 1 by the preceding paragraph. Suppose it is true for k ¼ 1; y; m: Then (6a). Let q ¼ lnð1=pÞ: Then a pðpÞ a ¼ eb½ðqln 2Þ q : pð2pÞ As p-0; q-N; and for any constant c and fixed 0oao1 it is known that lim ½ðq cÞa qa ¼ 0: q-N Therefore, limp-0 ½pðpÞ=pð2pÞ ¼ 1: But 1 pð12Þa1; so (6.1) never satisfies the condition near 0. w0 ð12 þ amþ2 Þ ¼ ¼ wðamþ2 Þw0 ð12 þ amþ1 Þ w0 ðamþ2 Þwð12 þ amþ1 Þ½1 wð12 þ amþ1 Þ ( wðamþ2 Þ wð12 þ amþ1 Þ2 wð12 þ amþ1 Þ2 " #) j1 mþ1 Y X wðamþ1 Þ½1 wð12 þ am Þ Q G s2 sj wð12 þ ai Þ 1 wð12 þ am Þ m j¼2 wð2 þ aj Þ i¼2 j¼3 smþ2 wðamþ2 Þ½1 wð12 þ amþ1 Þ wð12 þ amþ1 Þ ( ) P Q wðamþ2 Þ½1 wð12 þ amþ1 Þ wðamþ1 Þ½1 wð12 þ am Þ½G s2 sj Q smþ2 ¼ wð12 þ amþ1 Þ ½1 wð12 þ amþ1 Þwð12 þ amþ1 Þwð12 þ am Þ ( ) Pmþ1 Qj1 wðamþ2 Þ½1 wð12 þ amþ1 Þ G s2 j¼3 sj i¼2 wð12 þ ai Þ ¼ smþ2 Qmþ1 1 wð12 þ amþ1 Þ j¼2 wð2 þ aj Þ " # j1 mþ2 Y X wðamþ2 Þ½1 wð12 þ amþ1 Þ 1 ¼ 1 G s2 sj wð2 þ ai Þ ; Q 1 wð2 þ amþ1 Þ mþ1 j¼3 i¼2 j¼2 wð2 þ aj Þ so it is true also for k ¼ m þ 1: At this point in the induction process, we know that w is continuous and increasing on ½0; 12 þ amþ1 ð14Þ ¼ ½0; 12 þ amþ2 ð0Þ with 0owð12 þ aj Þo1 for jpm þ 1: Thus, by (5) and 0owð12 þ aj Þo1 for jpm þ 1; we have G4 N X k¼2 sk 4s2 þ mþ2 X j¼3 sj j1 Y wð12 þ aj Þ; i¼2 so w0 ð12 þ amþ2 ðtÞÞ40 for 0oto14: It follows by induction that w extended is continuous and increasing on ½0; 1Þ: In addition, wðpÞ-1 as p-1 by the proof of Lemma 3. & Proof of Theorem 4. As noted earlier, we assume that p is an increasing function from ½0; 1 onto ½0; 1: This is true for each of (6a) through (6f). We focus on the condition near 0, i.e., pðpÞ ¼ 1 pð12Þ; lim p-0 pð2pÞ and show that it fails except for noted special cases. (6b). We have 1=g pðpÞ 1 ð2pÞg þ ð1 2pÞg 1 ¼ g - g g g pð2pÞ 2 2 p þ ð1 pÞ as p-0: Moreover, pð12Þ ¼ 1=ð2gþ1=g1 Þ; and lim½pðpÞ=pð2pÞ ¼ 1 pð12Þ if and only if therefore ð21=g1 Þð2g 1Þ ¼ 1: This holds for g40 if and only if g ¼ 1; and when g ¼ 1; p is the identity. (6c). As in the preceding case, lim½pðpÞ=pð2pÞ ¼ 1=2g : We also have pð12Þ ¼ a=ða þ 1Þ; so the condition near 0 holds if and only if a ¼ 2g 1: Because 1 pð1 pÞ 1 að1 2pÞg þ ð2pÞg 1 ¼ lim g lim ¼ g; p-0 1 pð1 2pÞ p-0 2 2 að1 pÞg þ pg the condition near 1 requires 1=2g ¼ a=ða þ 1Þ; i.e., a ¼ 1=ð2g 1Þ: It follows that the conditions near 0 and near 1 hold if and only if a ¼ g ¼ 1; in which case p is the identity. D.E. Bell, P.C. Fishburn / Journal of Mathematical Psychology 47 (2003) 244–258 It follows that w is differentiable at holds. & (6d). Here a pðpÞ a ¼ eðb=aÞ½p ð2pÞ ; aa0; b40: pð2pÞ If a40 then lim½pðpÞ=pð2pÞ ¼ 1: If ao0 then lim½pðpÞ=pð2pÞ ¼ 0: Because 0opð12Þo1; the condition near 0 never holds. (6e). We have pðpÞ=pð2pÞ ¼ 1=2b and pð12Þ ¼ 1=2b ; so 1=2b ¼ 1 1=2b 32b1 ¼ 13b ¼ 1; and p is the identity when b ¼ 1: (6f). Let q ¼ lnð1=pÞ: Then pðpÞ ð1 a ln pÞb ¼ ¼ pð2pÞ ð1 a ln 2pÞb 1 þ aq a ln 2 1 þ aq 257 1 2 if and only if (7) Proof of Lemma 5. Because the numerator and denominator of (A.1) both approach 0 as t-0; we can use l’Hospital’s rule on (A.1) to conclude that w0 ð12 þ tÞ ¼ fw000 ðtÞwð2tÞ½wð2tÞ 1 þ w00 ðtÞ½2wð2tÞw0 ð2tÞ w0 ð2tÞ þ 2w00 ðtÞwð2tÞw0 ð2tÞ þ 2w0 ðtÞw0 ð2tÞ2 þ 2w0 ðtÞwð2tÞw00 ð2tÞ þ w0 ðtÞw00 ð2tÞ þ wðtÞw000 ð2tÞg=f2w0 ð2tÞ2 þ 2wð2tÞw00 ð2tÞg for t-0: When t-0; this reduces to 1 w00 ð0Þ þ : w0 ð12 Þ ¼ w0 ð0Þ þ 4 w0 ð0Þ b : As p-0; q-N; so limp-0 ½pðpÞ=pð2pÞ ¼ 1: Because pð12Þ40; the condition near 0 never holds. & This and the final equation in the proof of Lemma 4 yield w00 ð0Þ ¼ ½3 4wð12Þw00 ð0Þ; which is tantamount to the conclusion of Lemma 5. & Proof of Lemma 4. Assume Axiom 2. By definition, w is þ differentiable at 12 if and only if w0 ð12 Þ ¼ w0 ð12 Þ: The 0 1 value of w ð2 Þ is obtained from the specification of w on þ ½0; 12 as in Axiom 1. Our evaluation of w0 ð12 Þ is based on (2) for w in the form Proof of Lemma 6 (Outline). When the first expression for w0 ð12 þ tÞ in the proof of Lemma 4 is differentiated in the usual way, we get wðtÞ½wð2tÞ 1 w 12 þ t ¼ 1 þ wð2tÞ þ 2w0 ðtÞw0 ð2tÞwð2tÞ 2wðtÞw0 ð2tÞ2 g=wð2tÞ3 : for 0oto12: According to our convention for derivatives following the statement of Axiom 1, the limits of w0 ð2tÞ and w00 ð2tÞ as t-0 are 2w0 ð0Þ and 4w00 ð0Þ; respectively. The preceding equation yields w0 1 w0 ðtÞwð2tÞ½wð2tÞ 1 þ wðtÞw0 ð2tÞ þt ¼ : 2 wð2tÞ2 Because the numerator and denominator both approach 0 as t-0; we use l’Hospital’s rule to get w0 ð12 w00 ðtÞwð2tÞ½wð2tÞ 1 þ 2w0 ðtÞwð2tÞw0 ð2tÞ þ wðtÞw00 ð2tÞ þ tÞ ¼ 2wð2tÞw0 ð2tÞ ðA:1Þ w00 ðtÞ½wð2tÞ 1 w00 ð2tÞ wðtÞ ¼ w ðtÞ þ þ 2w0 ð2tÞ 2w0 ð2tÞ wð2tÞ 0 w00 ð12 þ tÞ ¼ fw00 ðtÞ½wð2tÞ3 wð2tÞ2 þ w00 ð2tÞwðtÞwð2tÞ ðA:2Þ As t-0; the numerator and denominator both go to 0, so we use l’Hospital’s rule and find that the differentiated numerator and differentiated denominator both go to 0 as t-0: We therefore use l’Hospital’s rule again. The ratio of the twice-differentiated numerator to the twice-differentiated denominator of the preceding form of w00 ð12 þ tÞ is then evaluated as t-0 with the aid of the þ condition near 0 and wðjÞ ð2tÞ-2j wðjÞ ð0Þ to obtain w00 ð12 Þ equal to the right side of the displayed equation of Lemma 6. Because w is second differentiable at 12 if þ and only if w00 ð12 Þ ¼ w00 ð12 Þ; the lemma’s conclusion follows. & Proof of Theorem 5. Let bk ¼ ð2k 1Þ=2k for k ¼ 1; 2; y . By (1) for w with ð1 þ pÞ=2 ¼ bkþ1 þ t; we have wðbkþ1 þ tÞ ¼ 1 wð12bk þ tÞ½1 wðbk þ 2tÞ wðbk þ 2tÞ for all t in a small open interval that contains 0. Differentiation gives w0 ð12bk þ 1Þ wðbk þ 2tÞ 0 w ðbk þ 2tÞwð12bk þ tÞ þ : wðbk þ 2tÞ2 w0 ðbkþ1 þ tÞ ¼ w0 ð12bk þ tÞ for t-0: The condition near 0 says wðtÞ=wð2tÞ-1 wð12Þ as t-0; so (A.2) yields þ w0 ð12 Þ 00 3 1 w ð0Þ ¼ w ð0Þ þ wð2Þ 0 : 4 w ð0Þ 0 that Suppose w is differentiable at bj for jpk: Then w is differentiable on ½0; bkþ1 Þ; so the right side of the 258 D.E. Bell, P.C. Fishburn / Journal of Mathematical Psychology 47 (2003) 244–258 preceding expression is continuous in t for t in a small open interval that contains 0, and it follows that w is differentiable at bkþ1 : Since we assume that w is differentiable at b1 ¼ 12; induction on k implies that w is differentiable at all bk : & References Abdellaoui, M. (2000). Parameter-free elicitation of utility and probability weighting functions. Management Science, 46, 1497–1512. Abdellaoui, M. (2002). A genuine rank-dependent generalization of the von Neumann–Morgenstern expected utility theorem. Econometrica, 70, 717–736. Allais, M. (1953). Le comportement de l’homme rationnel devant le risque: Critique des postulats et axiomes de l’école américaine. Econometrica, 21, 503–546. Allais, M. (1979). The so-called Allais paradox and rational decisions under uncertainty. In M. Allais, & O. Hagen (Eds.), Expected utility hypotheses and the Allais paradox (pp. 437–681). Dordrecht: Reidel. Bell, D. E., & Fishburn, P. C. (2000). Utility functions for wealth. Journal of Risk and Uncertainty, 20, 5–44. Bleichrodt, H., & Pinto, J. L. (2000). A parameter-free elicitation of the probability weighting function in medical decision analysis. Management Science, 46, 1485–1496. Edwards, W. (1961). Behavioral decision theory. Annual Review of Psychology, 12, 473–498. Fishburn, P. C. (1982). The foundations of expected utility. Dordrecht: Reidel. Gonzalez, R., & Wu, G. (1999). On the shape of the probability weighting function. Cognitive Psychology, 38, 129–166. Herstein, I. N., & Milnor, J. (1953). An axiomatic approach to measurable utility. Econometrica, 21, 291–297. Jensen, N. E. (1967). An introduction to Bernoullian utility theory. I. Utility functions. Swedish Journal of Economics, 69, 163–183. Luce, R. D. (1991). Rank- and sign-dependent linear utility models for binary gambles. Journal of Economic Theory, 53, 75–100. Luce, R. D. (2001). Reduction invariance and Prelec’s weighting functions. Journal of Mathematical Psychology, 45, 167–179. Luce, R. D., & Fishburn, P. C. (1991). Rank- and sign-dependent linear utility models for finite first-order gambles. Journal of Risk and Uncertainty, 4, 25–59. Luce, R. D., & Fishburn, P. C. (1995). A note on deriving rankdependent utility using additive joint receipts. Journal of Risk and Uncertainty, 11, 5–16. Luce, R. D., & Raiffa, H. (1957). Games and decisions. New York: Wiley. Marschak, J. (1950). Rational behavior, uncertain prospects, and measurable utility. Econometrica, 18, 111–141; Errata, 1950, p. 312. Prelec, D. (1998). The probability weighting function. Econometrica, 66, 497–527. Quiggin, J. (1982). A theory of anticipated utility. Journal of Economic Behavior and Organization, 3, 323–343. Quiggin, J. (1993). Generalized expected utility theory: The rankdependent model. Dordrecht: Kluwer. Raiffa, H. (1968). Decision analysis: Introductory lectures on choice under uncertainty. Reading, MA: Addison–Wesley. Safra, Z., & Segal, U. (2001). Rank-dependent preferences without ranking axioms. Journal of Mathematical Economics, 35, 547–562. Samuelson, P. A. (1952). Probability, utility, and the independence axiom. Econometrica, 20, 670–678. Schmidt, U., & Zank, H. (2001). A new axiomatization of rankdependent expected utility with tradeoff consistency for equally likely outcomes. Journal of Mathematical Economics, 35, 483–491. Tversky, A., & Fox, C. R. (1995). Weighting risk and uncertainty. Psychological Review, 102, 269–283. Tversky, A., & Kahneman, D. (1992). Cumulative prospect theory: An analysis of decision under uncertainty. Journal of Risk and Uncertainty, 5, 297–323. Tversky, A., & Wakker, P. P. (1995). Risk attitudes and decision weights. Econometrica, 63, 1255–1280. von Neumann, J., & Morgenstern, O. (1944). Theory of games and economic behavior. Princeton: Princeton University Press. Wakker, P. P., & Tversky, A. (1993). An axiomatization of cumulative prospect theory. Journal of Risk and Uncertainty, 7, 147–176. Wu, G., & Gonzalez, R. (1996). Curvature of the probability weighting function. Management Science, 42, 1676–1690. Wu, G., & Gonzalez, R. (1999). Nonlinear decision weights in choice under uncertainty. Management Science, 45, 74–85.
© Copyright 2025 Paperzz