JOURNAL

OF ECONOMY

THEORY

30, 370-400

(1983)

Selling to Risk Averse Buyers with Unobservable Tastes*

STEVEN Department Received

A.

MATTHEWS

of Economics, Northwestern Evanston, Illinois 60201 March 12, 1982: revised August

University, 15, 1982

Schemes maximizing the expected profit from selling indivisible units to risk averse, heterogeneous buyers are studied. For the polar cases of an unlimited and a unit supply, optimal schemes are constructed for buyers who exhibit constant risk aversion and who differ only in their monetary evaluations. With a unit supply, the optimal scheme resembles a first-price auction in which buyers are paid to submit high bids, but must pay to submit low bids. The results are interpreted in terms of insurance provision and the use of risk to weaken incentive compatibility constraints. Journal of Economic Literature, Classification Numbers: 022, 026.

1.

INTRODUCTION

In order to practice perfect price discrimination, a monopoly must know the maximum price each customer is willing to pay. However, in many environments customers cannot be distinguished on the basis of their preferences. The monopolys task is then to practice so-called second degree price discrimination, i.e., to construct a choice setting in which customers willing to pay more are induced, by their own decisions, to pay greater prices. The first such environments to be studied were ones in which the empirical distribution of buyers tastes is known. Studies of monopoly in these environments, such as Spence [24], Goldman, Leland and Sibley [3], Mussa and Rosen [ 171, Roberts [22], and Mirman and Sibley [ 151, characterized profit-maximizing selling schemes within the class of schemes in which each buyer is presented the same nonuniform price schedule. This class of selling schemes is not the largest, e.g., it excludes schemes that result in random allocations. However, it can be shown that optimal nonuniform price schedules are optimal within any class of selling schemes if buyers are risk neutral.
*This paper has benefitted from my discussions with Roger Myerson and Andrew Postlewaite, who of course are not responsible for the shortcomings. Support from the National Science Foundation (Grant SES 81-07103) and from the Sloan Foundation is gratefully acknowledged.

370
0022.0531/83
Copyright All rights

$3.00

:P 1983 by Academx Press. Inc. of reproduction in any form reserved.

SELLING

TO RISK

AVERSE

BUYERS

371

A second environment studied more recently is one in which the monopoly does not know the empirical distribution of tastes, but instead has only probabilistic assessments. If buyers are risk neutral and the monopoly has constant costs, then again it is optimal to present each buyer with a nonuniform price schedule. But if the cost function is nonlinear, then average profit will depend on the total quantity purchased and hence be uncertain. In this case it may be better to use a selling scheme in which each buyers allocation depends jointly on the decisions of all buyers. The polar example of such an environment is one in which there is a binding quantity constraint. It is then infeasible to set a nontrivial price schedule, since there is positive probability that independently made purchase orders will exceed the capacity constraint. Consequently, the class of selling schemes must be enlarged in these environments to include auctionlike mechanisms. Optimal auctions have, been studied by Harris and Raviv [4,5], Myerson [ 191, Riley and Samuelson [20], and Maskin and Riley [9, lo]. These authors study an environment in which (a) the good is indivisible; (b) no buyer wants more than one unit; (c) buyers exhibit no income effects, i.e., each buyers (unobservable) willingness-to-pay for a unit is independent of his income: and (d) buyers are risk neutral. Then, given that buyers characteristics are independently distributed and that only one unit can be sold, the schemes that maximize expected profit are characterized. It turns out that the set of optimal schemes is quite large, containing, e.g., under a regularity condition, both first-price and second-price auctions with reserve prices.. The reason there are so many optimal auctions in this environment is that auctions that differ only in their payment functions, but that have the same expected payment functions, must be equivalent when all agents are risk neutral. This equivalency result cannot hold if buyers are risk averse. The lack of equivalency when buyers are risk averse is suggested by the wellknown result that, in this case, first-price auctions generate greater expected profit than second-price auctions [4, 8, 9, 11, 12, 14, 20, 261. In this paper, expected profit maximizing schemes are explicitly characterized for environments in which assumptions (a)-(c) are kept, but (d) is generalized to allow for risk averse buyers. In particular, buyers shall usually be assumed to have a constant measure of absolute risk aversion.
In a first-price (discriminatory) auction with reserve price. the high bidder obtains the unit and pays a price equal to his bid, provided it exceeds the (known) reserve price. A secondprice (competitive, Vickrey) auction with reserve price is the same, except that the price is the maximum of the reserve price and the second highest bid. *The fact that first- and second-price auctions generate the same expected profit in this environment was first observed by Vickrey [26]. However, Matthews [ 121 observes that a risk averse seller prefers the first-price auction, whereas Milgrom and Weber [ 141 show that a risk neutral seller prefers the second-price auction if buyers types are a!Iiliated rather than independent random variables.

312

STEVEN

A. MATTHEWS

Major consequences follow from this apparently minor generalization. First, when there is no quantity constraint, take-it-or-leave-it offer schemes, which are the optimal nonuniform price schedules in this environment, are not optimal in a broader class of selling schemes. Instead, the optimal scheme gives some buyers only a probability of obtaining a unit. The reason for this is that when buyers are risk averse, the expected cost of inducing them to reveal their evaluations can be decreased by allowing them to avoid. and therefore sometimes to bear, risk. The consequences of risk aversion are more startling when there is a unit quantity constraint. Then, even a small amount of buyer risk aversion results in the class of optimal auctions shrinking essentially to a unique auction. This auction does not resemble any standard auction. It can be described, for the case in which buyers are approximately risk neutral, as a first-price auction in which buyers must pay to submit low bids, but are paid to submit high bids. Buyers with high values are induced to submit bids nearly equal to their values. Independently, Maskin and Riley [ 111 have used a similar approach to study optimal auctions with risk averse buyers. They allow more general utility functions, but at the cost of not obtaining conditions that are both necessary and sufficient for an auction to be optimal. Their results are in accord with most of the qualitative properties of the optimal selling schemes found here. The paper is organized as follows. The environment and a very broad class of selling schemes are defined in Section 2. The assumption of constant absolute risk aversion is introduced in Section 3. The sellers maximization problem is formulated in Section 4. Its solution is presented in Section 5 for the case of no capacity constraint, and in Section 6 for the case of a unit capacity constraint. Summarizing remarks are contained in Section 7, and all proofs are contained in the Appendix.

2.

SELLING

SCHEMES

The environment contains n buyers and one seller of a good. The following assumptions are made: (i) the good is discrete; (ii) the seller is risk neutral and bears an opportunity cost c,, for each unit sold; (iii) no buyer wants more than one unit; (iv) the buyers exhibit no income effects, i.e., a buyer with evaluation 0 is willing to pay 0 dollars for one unit regardless of his income; (v) the buyers evaluations are realizations of independent random variables whose distribution is common knowledge; (vi) each
3 The desirability of incentive schemes yielding random outcomes suggested by, e.g., Riley and Zeckhauser [ 211 and Stiglitz 125 1. has previously been

SELLING

TO

RISK

AVERSE

BUYERS

373

buyers evaluation is observed only by himself; and (vii) all buyers have the same, increasing and concave expected utility functions u for money. Consequently, a buyer with evaluation 0, income I, and one unit of the good obtains utility u(Z + 0). After this section, u will be assumed to exhibit constant absolute risk aversion. A further technical assumption is that the buyers evaluations are identically distributed according to a cumulative distribution F. It is assumed that F has a continuous, positive density/on the support 10, 1 1. The sellers unit cost is assumedto satisfy 0 < c0 < 1. A selZing scheme fundamentally consists of a messageset Mi for each buyer, and an outcome function that maps an n-tuple of mesagesinto a (perhaps random) allocation 6. The behavior of buyers is described by a Nash (Bavesian) equilibrium, which is an n-tuple of functions (m:,..., m,*). Each m,? maps possibleevaluations for buyer i into the message set M;. The equilibrium property of these functions consists of mF(Bi) being the message that maximizes the expected utility of buyer i when he has the evaluation 1!9;, and every buyer j usesm,f to submit messages. This notion of a selling scheme is very general. incorporating even iterative schemeswhen the messagesets are defined broadly enough. Such generality is not required here, since any selling strategy, in any equilibrium, results in the same outcomes as a more tractable revelation selling scheme. The argument is the following: any schemewill, in a given equilibrium, result in an outcome a-(19,,...,t9,) when the buyers evaluations are 8, ,..., 8,. The function qe, ,..., f3,) can serve as the outcome function of another selling scheme in which each buyers messageset is the set of possible evaluations. Becausethe function Z(f?, ,..., 8,) is obtained as the composition of a selling scheme with one of its equilibria, the n-tuple of identity functions (mT(8,) = ei) is an equilibrium of the new scheme. We have therefore constructed a revelation schemein which truth-telling is an equilibrium that results in the same ourcomes as did the given equilibrium in the original
selling scheme.4

Henceforth, attention will be restricted to revelation schemes that are incentive compatible, i.e., schemesin which the n-tuple of identity functions is an equilibrium. The preceding argument indicates that this involves little loss of generality. However, having the feasible class of selling schemes depend upon utility functions does obscure comparison and implementation issues. Once optimal revelation schemeshave been bound, some attention
will be given to the question of their implementation via equivalent.

nonrevelation schemes. A revelation scheme must, most generally, specify a joint probability
Myerson [ 191 discusses the restriction to revelation selling strategies more fully, aptly naming the idea the Revelation Principle. The idea dates back at least to Gibbard 121 (for dominant strategy mechanisms) and is discussed generally in Il. 6, 181.

374

STEVEN

A. MATTHEWS

distribution for the buyers allocations as a function of reported evaluations. Denoting fi as the (possibly random) payment of buyer i, and ,Ci E (0, 1 } as the (possibly random) amount of the good received by buyer i, an outcome is a joint probability distribution G(x,, t, ,..., x,, t, 18, ,..., 0,). Any revelation scheme can be identified with such a function G. Two properties of a scheme G are of interest. Implementation is simplified and strategic interactions eliminated if each buyers allocation is independent of the other buyers actions. This occurs if G is separable in the sense of being of the form nl-, Gi(xi, ti 1ei). For example, a scheme in which each buyer is presented a price-quantity schedule is separable, whereas a first price auction is not. The second property of a scheme G that is of particular interest here is whether, given an n-tuple of reported evaluations, it places all probability on just one allocation (x,, t, ,..., x,. t,). If it does, then G is deterministic. Offering price-quantity schedules is a deterministic scheme. So is a first price auction, if the coin flip that resolves ties is ignored. Buyer i is only concerned with a few marginal distributions of a scheme G. Suppose that the scheme is incentive compatible, so that the expectations of buyer i over other buyers reports can be taken via the distribution F. Then buyer is expectations over the random elements in his environment are well-defined by the two distributions F and G, and his report di. When buyer i has the true evaluation Bi but reports e,, his expected utility is E[u(/3,,Ci - i;.) I SiI = Pr(2, = 1 119,)Elu(Bi - t;.) / 4,. -Yi = 11 +Pr(~i=OIBi)E[u(-t;)18i,xj=01. Using G and r, define (Ii = Pr(x, = 1 I oi), (2) (3) (4) (1)

Hi(pi,,. I di) = Pr(c < PiM.18j, ?j = 1). Ji(pi,l d,) = Pr(fi <pi,/lji, Ti = 0).

The probability Qi(di) is buyer is perceived probability of obtaining a unit. He perceives his random payment Fiw contingent upon obtaining (winning) a unit to be distributed according to Hi, and his random payment pi, contingent upon not obtaining (losing) a unit to be distributed according to Ji. Expression (1) indicates that buyer i is indifferent between any two incentive compatible schemes that result in the same Qi, Hi, and Ji functions.

SELLING

TO RISK

AVERSE

BUYERS

315

The sellers expected profits are

Therefore the seller, too, is indifferent between any two incentive compatible schemes that result in the same Qi, Hi, and Ji functions. If there is unlimited capacity, then the problem will be to find a scheme that maximizes (5) subject to the incentive compatibility constraint that si = 19~maximizes (l), and subject to the constraint that participation be voluntary, i.e., that expression (1) be no less than u(0) at &I,= oi. By the above arguments, therefore, attention can be restricted to finding the marginal distributions Qi, Hi, and Ji. Once these functions are found, then the simplest scheme consistent with them is the separable one that has buyer is allocation independent of the reports of other buyers. This scheme has Qi(si) as the actual probability of buyer i obtaining a unit, Hi(pi,,, 113,) as the actual distribution of his payment when he obtains a unit, and Ji(pi, ) 0;) as the actual distribution when he does not obtain a unit. A separable scheme will not suffice if fewer than n units can be sold. This is because in any (nontrivial) separable scheme, there is positive probability that each buyer will obtain a unit. Letting qi(e, ,..., 8,) = Pr(.fi = 1 18, ,..., 0,), then 0 < 6 qi(e, )..., 8,) < 1 ir (6)

expresses the capacity constraint of having no more than one unit to sell. With this additional constraint, the problem is to find the functions qi, J;, and Hi that maximize (5). where Qi and qi are related by

Once these functions are found, the simplest scheme consistent with them has the winning buyer determined as a function of all reports via the probabilities 4i(eL 2...>B,,), but the payments conditional upon who obtains the unit still determined independently via the distributions Ji and Hi. Because all buyers are ex anfe identical, only schemes that treat them symmetrically need be considered. Henceforth, the subscript i is deleted from the functions Qi, Hi, and Ji, and the three-tuple (Q, H, J) will be

376

STEVEN A.MATTHEWS

referred to as a scheme. If there is no capacity constraint, then (Q. H, .I) is feasible if it satisfies the incentive compatibility and voluntary participation constraints. If there is a unit capacity constraint, then feasibility also requires that Q be implementable, i.e., that functions qi exist such that Q and each qi are related via (7). The following lemma, proved in Maskin and Riley [9, 111, reduces the capacity constraint to a more tractable form. In words, it states that a given Q is implementable only if it implies that the probability of selling the unit to a buyer with an evaluation greater than 8 is no greater than 0. Suppose that Q: 10, 11+ [0, 11 is a function for which there exist functions qi[O, 11 --t [0, 11 (i = l,.... n) satis-ving (6) and (7). Then, for every 8,
LEMMA 1.

To incorporate the capacity constraint, (8) will be included as a constraint in the problem of finding a scheme (Q, H, J) to maximize expected profits. Once this scheme is found, it will be verified that its Q is indeed implementable, i.e., that the necessary qi functions exist. Regardless of the presence of a capacity constraint, one property of an optimal scheme can be obtained at the current level of generality. The following lemma states that attention can be restricted to schemesin which the payment contingent upon not obtaining a unit, fi,. is deterministic.
LEMMA 2. Suppose (Q, H,J) is a feasible scheme. Then a scheme (Q, H, .?), where j(p, ) 6) puts all probability on a point p,(0). can be constructed that is both feasible and yields no less expected profits than

(Q, H, J).
The proof of Lemma 2 relies upon replacing b, by its certainty equivalent. But the proof is not standard, since incentive compatibility must be shown to still hold. It is generally not possible to replace j,,. by a deterministic P,~(@ that both increasesexpected profit and maintains incentive compatibility. Because of Lemma 2, a three-tuple (Q, H,p,), where p, is a real-valued function of evaluations, can be referred to as a scheme.

3. CONSTANT

ABSOLUTE

RISK AVERSION

Henceforth it is assumedthat u exhibits constant absolute risk aversion R. In particular, the symbol u shall denote the function u(x) = (1 - emR)/R.

SELLINGTO

RISK AVERSE BUYERS

377

There are two reasons why this functional form makes the problem tractable. The first is that the multiplicative separability of u implies that g,,. can be made deterministic without destroying incentive compatibility.
LEMMA 3. Suppose (Q, H, p,) is feasible. Then a scheme (Q, I?, p,), where i?(p,. 18) puts all probability on a point p,.(B), can be constructed that is both feasible and yields no less expected profit than (Q, H, p,).

Attention can now be restricted to schemeswith deterministic payment functions pW and p,. However, rather than using the function p,?, it will be more convenient to use the bid function b defined by b(e) = P,,.(@- p,(d). The function p, will be denoted p, and called the bid submissionor entry fee function. A schemeshall be denoted (Q, 6, p). The interpretation now is that a buyer must pay the submission fee p(d) in order to submit a bid b(8). Submitting this bid gives the buyer, from his point of view, a probability Q(O) of obtaining a unit. In the event that he obtains a unit, he must additionally pay the amount that he bid, b(d). Given a scheme (Q, b, p), a buyer with evaluation 0 who reports 8 is paying p(d) to obtain the lottery that gives him O-b(8) dollars with probability Q(d), and zero dollars with probability 1 - Q(e). The second simplification due to assuming constant absolute risk aversion is that the buyers evaluation of the lottery is independent of the entry fee. Specifically, the buyers expected utility function,
@t 0) = Q(e^> U(e - P(d) - b(8)) + (1 - Q(8)) U(-p(8)),

is equivalent to the additively separablefunction u(d. 8) = v(e - b(8), Q(o)) - p(d), where v is defined by (9)

w(.v.Q>=-(l/R)log[l

-Q+Qu(ytl.

(10)

The amount W(J: Q) is just the certainty equivalent of a lottery offering y dollars with probability Q, and zero dollars with probability 1 - Q; it is equal to the expected value & minus a risk premium n( y, Q). The incentive compatibility constraint can now be made tractable by a technique based roughly upon that of Mirrlees [ 161. Given a revelation scheme (Q, 6, p), which then defines U(O, 0) by (9). define

v(e) = u(e, e).

(11)

378

STEVEN A.MATTHEWS

If the scheme is incentive compatible, then V(0) is the indirect utility function of a buyer with evaluation 19.Incentive compatibility is satisfied if and only if V(0) > U(f?, 0) for all possible 8 and 8.
LEMMA

4. A scheme(Q, b, p) is incentive compatible if and only if Q(8) eRb( 1-

Q(e)is nondecreasing in 8,
0

(12) (13)

P(e) = V(0) + J y,(z - b(z), Q(z)) dz.

4. THE SELLER'S PROBLEM If the seller uses a feasible scheme (Q, b, p), then the expected profit obtained from a buyer with evaluation 8 is P(o) + mm@> CO)

= V(e - b(o), !a?)) - v(e) i- Q(w@>

- Cd

(14)

Therefore the sellers expected profit is

nI: [de- bm em - wa + cwww)- co) 1 f(e)de.

(15)

The problem of the seller is to maximize (15) by choosing control functions Q and b. The constraints are those imposed by feasibility. The incentive compatibility constraint is embodied by (12) and by the equation of motion for the state variable V,

we) = de - wh em

with

V(0) 2 0.

(16)

Expression (16) which is derived from (13), also incorporates the voluntary participation constraint. Implementability requires at least the constraint

O<Q(@< 1.

(17)

Implementability further requires, in the case of a unit capacity constraint. that nonnegative functions qi exist satisfying (6) and (7). The problem is solved in the following way. Expression (15) shall be maximized subject to the constraints (16), (17), and, in the case of a unit capacity constraint,

y(e) 2 0,

(18)

SELLING

TO RISK

AVERSE

BUYERS

379

where Y is a state variable with an equation of motion

Y(e) = f(@lQ(@ - F(@- I

with

Y(1) = 0.

(19)

The constraints (18) and (19) are necessary, by Lemma 1, for Q to be implementable in the case of a unit capacity constraint. The solution (Q, b). when adjoined to an entry fee function defined by , ~(6) = 446 - b(@),

Q(6)) - v(e).

(201

constitutes a selling scheme (Q, b, p). This scheme must then be shown to be feasible by, first, showing that it satisfies the ignored constraint (12). Secondly, in the case of a unit capacity constraint, functions qi that implement Q must be shown to exist. Both tasks are facilitated by imposing the following regularity condition, which is satisfied if the density does not decrease rapidly. It is equivalent to the one used in Myerson [ 191 and in Maskin and Riley (91. (RC) For any R in an interval 10, R]. the function F(B) - 1 f(e)

46 c) = 4e-

cl +

is increasing in B at any (0. c) E [ 0, 1 ] *. Henceforth, (RC) is imposed and 0 < R < l? is assumed. For later use, we remark now that (RC) implies that the function

/q&c)=

u(e-C) ; f-y;ey 1 u(e- c)

< c. two and the

is also increasing in 8 whenever 8 > c. Both a and p are negative if 8 and a(0, c) < p(S, c) holds for all (0, c), strictly if 0 # c. Consequently, increasing functions on the unit interval are defined by a(&~), c) = 0 /3@(c),c)=O, with c <e(c) <g(c) < 1 if O<c < 1. The first step in maximizing (15) subject to (16~( 19) is to define Hamiltonian

F(O, K Y, Q. b, A,P) = [de - b, Q>- I + (b - c,,)Qlf(@ + G,(e - b, Q) + iuf(@lQ - F(U- 1,

and the Lagrangian L = P + vY. Notice that when Q and b are chosen to maximize ;i for fixed (8. V, Y, A,,u), the maximized Hamiltonian is concave

380

STEVEN

A. MATTHEWS

in V and Y. Therefore the following conditions, in conjunction (16t(l9), are sufficient as well as necessary for a solution.

with

A(8)=-$=f(B).
p(e) = - g = y(8),

A( 1) = 0,

1 continuous;

P(O) < 03

P(O) Y(O) = 0; (23)

if ,U is discontinuous at 13,then 8 is an entry or an exit point of an interval upon which Y = 0, and ,u(&) > ~(0); v(e) > 0 and e9 y(e) = 0; is continuous;

(24)
(25) (26)

278,
(Qw,

v(e), y(e), Q(e), b(e), n(e), m9)

b(e)) E -a;f5tyym xv,
O<QSl

v(e), y(e), a b, n(e), de)).

Condition (21) immediately implies A(e) = F(B) - 1. This makes sense. The amount -11(e) is the implicit cost, due to the incentive compatibility constraint (16), of increasing by one dollar the dollar-equivalent utility level V(0) of a buyer with evaluation 8. The incentive compatibility constraint requires that if one more dollar is to be given to a buyer with evaluation 8, then an extra dollar must also be given to the buyer if his evaluation is greater than 8. Hence, the additional expected cost of increasing V(0) by one dollar is 1 -F(8) = -A(8). Substituting 1 - F for A,, and letting c = co - ,LL and dividing by f(e), the important part of the Hamiltonian becomes ;%(e, Q, b, c) = ty(B - b, Q) + (b - c>Q - [ ;(;;)I v,(e - b.

Qh (27)

The cost c(0) = C, - ~(0) is the true opportunity cost of selling a unit to a buyer with evaluation B; it is the sum of the sellers personal opportunity cost co and the implicit cost -p(e) of not being able to also sell a unit to another buyer. Condition (26) can be replaced now by

(Q(e), b(O)) E j-l;g<yy, .P(e, Q. 6, cm.

OCQS I

(26)

5 See Theorem 8 in Seierstad piecewise continuous controls variables (V and u).

and Sydsaeter [23]. As is usual, attention is restricted to piecewise differentiable state (Q and b) and continuous,

SELLING TO RISK AVERSE BUYERS

381 unique if

The following lemma states that this maximization is virtually buyers are strictly risk averse.
LEMMA 5. Zf 0 < c < 1, then ,@(0, Q, b, c) -co < b < co and 0 < Q < 1 bv b(B, c) = c and

is

maximized

for

PC4 cl = /I(& c)- a(& c)

= 1.

863 Ge<8(c),
e 2 B(C).

(28)

Q = Q(., c) is the unique maximizing Q and, on v(c), 8(c)), b = c is the unique maximizing b.6 The function 0 increases in 6 and decreases in c if 8 E c8(4, &D
5. SOLUTION WITH UNLIMITED CAPACITY

By Lemma 5, the optimal scheme is determined once p(O) is determined. But p(B), the implicit cost due to a binding capacity constraint, is zero when there is no quantity constraint. Consequently, the sellers optimal schemein this case has been found.
THEOREM 1. With no quantity Q*(e) = @(e, co), b*(0) = c,. and

constraint,

an optimal

scheme is

P*(e) = v/(0- co, Q*(e)) -j:

WkZ - coT Q*(Z)) dZ.

(29)

Zf (Q, b, _p) is also optimal, then almost everywhe_re Q = Q*, p(t) = p*(B) for t3< e(c,), and b(B) = b*(B) for BE (@(co),@co)). Zf O> @co), then b(e) + P(e) = co + P*(&,)).
The scheme described in Theorem 1 could be implemented directly as a separable, nondeterministic revelation scheme. Alternatively, each buyer could be offered the increasing payment-probability schedule

p*(Q) = P*(Q* -I(Q))

(30)

from which to choose a probability Q of obtaining a unit. A buyer thus

6 It is an artifact of constant absolute risk aversion that b(0, c) = c. If buyers exhibited decreasing (increasing) absolute risk aversion, then a generalization of Lemma 5 would conclude that 6(0. c) > c (b(0, c) < c), where 6 is a component of the optimal scheme within the class of schemes with deterministic payment functions.

382

STEVEN A.MATTHEWS

purchases a probability Q for an amount P*(Q). Q can be viewed as reliability, or as the likelihood that the seller will have the good in stock. Upon paying for a Q, a buyer enters the store and finds a unit with probability Q. If he finds a unit, he purchases it at cost cO. Theorem 1 implies that an optimal scheme for risk averse buyers must subject them to risk. Hence, it must yield greater expected profit than the particular deterministic take-it-or-leave-it-offer scheme that is well-known (and will be shown) to be optimal when buyers are risk neutral. Consequently, since the deterministic scheme yields the same expected profit regardlessof how risk averse buyers are, greater profit can be made when the buyers are risk averse. The reason for this is that risk avoidance weakens the incentive compatibility constraint. In particular, as the following argument indicates, greater payments can be extracted from buyers with high values if buyers with low values are subjected to risk. Suppose there are only two buyers, with evaluations 8, < 02. Two schemesare to be constructed in which the expected payment of buyer 1 is the same x > 0. SchemeA imposes no risk on buyer 1, so that it simply involves selling a unit for the price x to any buyer claiming to have value 8,. Since buyer 2 can obtain this allocation as well, x is the largest amount that buyer 2 can be induced to pay in SchemeA. In SchemeB, on the other hand, buyer 1 receives the risky allocation (Q,, b,, pl), with 0 ( Q, < 1 and x = p, + Q, b,. By mimicking buyer 1 in SchemeB, buyer 2 receives the certainty equivalent which is less than the expected value Q, 8, - x because v(+b,,Q,)-~,, of risk aversion. Therefore, if y is the largest price that buyer 2 can be satisfies 8: - J = induced to in SchemeB, then y pay c/(0, - b,, Q,) - p, < Q,& -x. This implies y > x, so that the risky SchemeB extracts a greater payment from buyer 2 than does the deterministic SchemeA. The next result emphasizes the usefulness of risk aversion: as buyers become extremely risk averse, expected profit approaches that which could be obtained by perfect price discrimination.
COROLLARY 1. Assume R-= co. Then as R + 03, the optimal scheme in the case of unlimited capacity converges to the scheme(Q, b, p) defined bv bm(0) = CO,

p(t)

= 0, =e-co,

BGC,, e> co, two,

QYB) =

0,

= 1,

e> c,.

SELLING

TO

RISK

AVERSE

BUYERS

383 that is possible

Furthermore, the expected proJt converges to the maximum when information is perfect, n IL, (0 - co) f (8) de.

Returning to Theorem 1, there is a more direct and perhaps more insightful way of deriving it, based on methods used in some of the nonuniform pricing literature [3, 15, 17, 241. The first step is to show that if Q*, b*, and p* are numbers that maximize profit p + Q(b - c,,) subject to constraints w(e - b, Q> - P = V and

w,(e-b,Q>=

J,

where V and V are arbitrary numbers, then b* = c,,. Therefore, given any indirect utility V(S) that the seller must give to the buyers with an incentive compatible scheme, the optimal such scheme will have b*(8) = cO. Therefore an optimal feasible scheme will have b*(O) = c,, at least when 0 < Q*(e) < 1. Then a series of change-of-variable and reversing-ofintegration operations show that the expected profit from a scheme with b=co,

n1d [de - co5em - w)i f(e) de =?I

can be written as
Q(l)

de - co, e<e, -j-O w,(z - co, Q(z)>dz] f(e) de3

0

i0

wz(Q-(4

coq)[l

-F(Q-(s>)l

ds.

(31)

This expression can be maximized pointwise. Doing so yields a function Q - that has as its inverse Q*(e) = Q(0, co), which is increasing in 0 by (RC) and hence justifies the use of the inverse Q- in deriving (31). Expression (31) is interpretable. A buyer with evaluation 8= Q-(q) purchases the probability q of obtaining a unit. All buyers with evaluations greater than Q-(q) purchase probabilities greater than q. Thus 1 - F(Q-(4)) is the probability of selling at least an amount q of probability to an arbitrary buyer-the demand for the qth increment of probability is n(l - F(Q-l(q)). The maximum price that the marginal buyer B = Q-(q) would pay for an additional increment of probability above the qth is y,(Q-(4) - cO, q), since his profit in the event of obtaining a unit is Q-(q) - co. This price is what must be charged for the qth increment to ail buyers who purchase it. Thus (31) is the sum of the profits made on each increment of probability sold. Expression (3 1) is valid even for R = 0, in which case ty,(Q-(4) - co, q) is simply Q-(q) - co. Hence (3 1) is maximized by setting Q - (4) equal to

384

STEVEN A.MATTHEWS

whatever constant 19, maximizes (0 - c,)(l -F(8)). This results in a particular take-it-or-leave-it scheme.This schemeis equivalent to the take-itor-leave-it offering of the price 8,,, which is a scheme found by Riley and Zeckhauser [21] and implicity by Maskin and Riley [9] and by Myerson [ 191, to be an optimal auction when there is only one risk-neutral bidder. The following corollary formally states these results and also a continuity property.
COROLLARY 2. If there is no quantity constraint, then as R -+ 0, the optimal schemeconverges to the scheme (Q, b, p) deJned by b(8) = c,,.

PO(O)= 0, =eo-co,

ec e,, ew3,,

Q0(8> =
where

0,

= 1,

e< eo. e>e,,

c,)(l

8, E argrax (e -

- F(e)).

(32)

This scheme is equivalent to the take-it-or-leave-it scheme consisting of posting a price 8,, and it maximizes expected profits when buyers are risk neutral, regardless of (RC) holding.

6. SOLUTION WITH A UNIT CAPACITY

CONSTRAINT

The solution is far more complicated if no more than one unit can be sold. The implicit cost ,D is nontrivial and interactions between buyers must be allowed. A simple case is presented before the general result. One more function must first be defined. Let 8 be the minimal 8 > 0 satisfying e(0, co) = F(e)-, and define C: [19,1] + [co, 1] by
Q(e, qe))

= F(e)-

I.

(33)

Since 0 decreasesin c and e(0, 1) = 0, F is well-defined. By Theorem 1, a seller that had cost F((e)would, without a quantity constraint, assign a buyer with evaluation tJ the probability F(0)nP of obtaining a unit. It will be shown that if R is small then F is an increasing function. In this case, the opportunity cost co - ~(0) can be determined by
C, - p(e) = co, 8 < 80, e 2 80,

= c(e),

SELLING

TO

RISK

AVERSE

BUYERS

385

FIGURE

without violating the requirement that c, -,u(e) be nondecreasing (see (22) and (24)). Depicted in Figs. 1 and 2, the optimal scheme for this case has b(B) = co, = e*(e)

F(e), F(e)-1,

= Out co), =

e c 80, e 2 80, e < 80, e 2 80,

the only optimal

and p*(8) determined from (20). This scheme is virtually scheme when C is increasing.

6 / / / / :

/ o ,w

@(6,

FIGURE

386

STEVEN A.MATTHEWS

This scheme is feasible, since functions qi that implement Q can be found explicitly. Letting 8,,, = max(8, ,..., 0,) and m(0, ,..., 0,) be the number of buyers with evaluation 8,,,, define

Q*<eJ = m(t9, ,...,e,) qe,y- 1

ei = e,,, .

Because Q*(e) ( F(B)- for 0 ( 8, q: is a well-defined probability. It is easy to show that 0 < C q; < 1, and that

Q*(eJ = j: -. jy qi*(@, ,...,0,) [I f(fJ dei.

i#i

These qjr functions imply that a buyer obtains the unit only if his evaluation is the largest, and that even in this event, if this evaluation is lessthan 0 he has only a probability Q*(0,) of obtaining the unit. This schemefirst identifies the buyer with the maximum evaluation, and then either awards the unit to him if his evaluation is greater than 8, or if B < C?,treats him exactly as all buyers are treated by the scheme that is optimal when no quantity constraint exists. Before further discussion, let me present the general result.
THEOREM 2. Suppose I, = 10, e' ), I, = Iel,e2 1, z2 = (e2.e") ,..., I,, + , = 1 is a partition with 8 > 0 that satisfies the following criteria: 182K+?1

q(e*k) = qp+

1) ,

(k = o,.... K), (k = o,..., K), VB E Z2k (k = O,..., K),

(34) (35) (36)

F is increasing on I,,,
*Zk+l

[F(z)-

- e(z, F(e2k))] f(z) dz > 0

with equality holding in (36) for 0 = t12k (k = l,..., K). Define (Q*, b*, p*) by e*(e) = 003 wk)), = F(e)-1, b*(e) = qe*k), = q(e),
e E I2k

(k = O,..., K), (k = O,..., K). (k = O,..., K), (k = O,..., K). (38) (37)

e=2k+1
e E eEz2k+l 2,

P*(e) = de - b*(e), Q*(e)) -lo V,(Z - b*(z),

Q*(z)> dz.

(39)

SELLING

TO

RISK

AVERSE

BUYERS

387

(Q*, b*, p*> maximizes expected profit and is feasible. Futhermore, if (Q, b, p) is any feasible piecewise continuous scheme maximizing expected profit, then Q = Q*, p = p*, and, for all 0 E @(c,), l), b(B) = b*(8).
Then,

Remark 1. In the original version of this paper, it was only conjectured that the Q* in Theorem 2 is implementable. Its implementability now follows from results of Maskin and Riley [ 11, Theorem 7] and [ 131, which together imply that any nondecreasing function Q that satisfies constraint (8), such as Q*l can be implemented. Remark 2. Because the partition satisfying (34)-(36) is assumed to exist, Theorem 2 is a characterization rather than an existence theorem. However, F must be pathological for the partition not to exist. A more complicated example is shown in Figs. 3 and 4. Here there is a region in which C decreases, so that setting the cost cO -,u equal to C on this region would violate the restriction that cO - ,D be nondecreasing. The points O2 and r33 are found, loosely speaking, by letting b*(B) = C(B) hold for 0 > 8 until the function Q(., F(e)) has fallen enough to make the two shaded areas (weighted by the density) equal. After this point, B*, b* is held equal to the constant C(O) and Q*(.) is set equal to Q(., F((e*)) until the endpoint 0 of this irregular segment is reached. There are two qualitative features of the general scheme in Theorem 2 that are due to the buyers risk aversion. The first is that buyers with high evaluations are induced to bid nearly their evaluation, i.e., B - b*(0) + 0 as 0 + 1. This result follows from Theorem 2 because b* = C < 1 on ]e2 I, 1 ], and C + 1 as 8 + 1. But it can also be shown more generally whenever buyers are risk averse. No analagous result is obtained when buyers are risk neutral, as then only the expected payment x(0) = p(0) + Q(e) b(B), as well as the probability Q(O), are determined in an optimal scheme.

FIGURE

388

STEVEN

A. MATTHEWS

45

FIGURE

The second distinctive feature casts some light upon the first. The bid submission fee p*, while positive for buyers with modest evaluations, becomes negative for buyers with high evaluations. This can be seen from (39), since I!- 6*(e) --t 0 implies w(S-- 6*(o))+ 0. But again the result appears to hold generally when buyers are risk averse. What is an intuition for these features? Potentially, the largest payments can be extracted from the buyers with the greatest evaluations. But a buyer must be rewarded to induce him to reveal that this evaluation is high. One way of rewarding is to let the probability of obtaining a unit increase rapidly with the reported evaluation. This is the strategy utilized when a capacity constraint is absent. With this rewarding strategy the sellers profit is due to the increasing bid submission fee. The success of this strategy depends upon the perceived probability-of-winning function increasing rapidly enough that buyers with high evaluations do not bear too much risk, and hence pay only small bid submission fees. A capacity constraint therefore limits the use of this strategy, since then a buyers perceived probability of obtaining a unit cannot be made too large. When the constraint becomes binding, buyers with high evaluations are rewarded by lowering their entry fees, with the seller making the extra profit instead from an increasing bid function. By making the bid submission fee negative for buyers with high evaluations, the seller is selling them insurance against the event of not obtaining a unit. By increasing the insurance coverage with the reported evaluation, the expected value of the reward necessary to induce buyers with high evaluations to reveal themselves is minimized. Further light is shed upon the role of risk aversion by examining the

SELLING TO RISK AVERSE BUYERS

389

scheme when buyers are nearly risk neutral. To derive the limiting scheme. observe first that the function E is defined by
@e. F(e)) = w, p(e, w

F(8))- a(e,F(e))= F(e)- .

Rearranging and substituting in for a and ,fI yields C(0)) F(0) - 1 = U(@ - F(e)) - u(0 u(0 - C(8)) u(0 - F(e)) + f(e) [ Since u(y)=K(l -ePRP) and u(J~)/u(~)=R-(~~to y as R + 0, the function C converges to the function c(0) =e+

u(e-

F(e)))F(e)-

1) both converge

F(B) - 1
f(e)

The regularity condition (RC) implies co is an increasing function. Also, as R + 0 the points e(c,) and 8 both converge to a point B. defined by
c 0 = e 0 f (00) +

F(eo)

1
= cO(B,),

which is also the 8, = argmax, (0 - c,)(l - F(B)) of Corollary 2. Theorem 2 therefore implies that the optimal schemeconverges to the one presented in Corollary 3. This limiting scheme is optimal for risk neutral buyers, given VW
COROLLARY 3. As R + 0, the optimal scheme with a unit capacit) constraint converges to one defined bjl b(B) = c(8) and co(e)

= 0, = F(B)- ,

e<

8,.

e < 00
e< e> 8,. e,.

PO(e) = 0,
= max(z, 8,) dF(z)- - c(0) F(0)n- ,

In the limiting scheme of Corollary 3, a buyer with value 8 obtains the unit if and only if 19exceeds both 0, and the other buyers values. Buyers with high values still submit bids nearly equal to their values, since b(B) = co(B) + 1 as 8 + 1. Buyers with high values are still paid to participate, since p(B) + j: max(z, 19,)dF(z)- - 1 < 0 as 8 + 1. Buyers with values only slightly larger than 8, pay positive entry fees, since ~(0,) =

390

STEVEN A.MATTHEWS

(19,- CJ F(L),,)- I > 0. Any buyer with value greater than 8, has the total expected payment p(B) + b(B) Q(0) = li max(z, 0,) &(z)~ , which, if we call B. the sellers (reported) value, is the expectation of the largest of the other agents values conditional upon 0 being the largest value. This is the same expected payment made by risk neutral buyers in a first- or secondprice auction with reserve price 8, (see, e.g., 1201). There is a simple nonrevelation scheme that is equivalent to the one of Corollary 3. This scheme is a first price auction in which the entry fee depends upon the size of the bid. If the entry fee for a buyer who bids 6 is P(6) = p(bo-(b)), then no buyer with evaluations lessthan B. will enter the auction, and buyers with evaluations 8 > 0, will bid 6 = b(B) and pay the bid submission fee P*(b) = g(e). The final corollary identifies the limiting scheme as R + 00. It can be shown that as R --t co, c(0) converges to 8 and 8 converges to co (see the proof of Corollary 2). Hence Theorem 2 implies
COROLLARY 4. Assume if= co. Then as R + co, the sellers optimal scheme with a unit capacity constraint converges to (Q, b, p>, where p=(e)

= 0,
= F(e)+,

m,,
two,

b(8) = 0, and p*(8) = 0. Consequently, the sellers maximum profit approaches that obtainable by perfect price discrimination.

expected

7. SUMMARY AND DISCUSSION Initially, an environment was considered in which buyers have the same no-income-effects, risk averse utility function, but different private evaluations for the good. For a selling schemeto maximize expected profit in this environment, the payments made by buyers who do not obtain a unit must be nonstochastic. On the other hand, the payments made by buyers who do obtain units cannot generally be assumedto be nonstochastic. It was shown, however, that attention could be restricted to schemes with deterministic payment functions if buyers exhibit constant absolute risk aversion, which was subsequently a maintained assumption. One fundamental conclusion is that maximum expected profits are greater if the buyers are risk averse rather than risk neutral. If buyers are extremely risk averse, then expected profits from an optimal schemeapproximate those

SELLING

TO RISK

AVERSE

BUYERS

391

achievable by perfect price discrimination. This result can be explained from the revelation point of view: the desire of buyers with high values to avoid risk can be used to decrease the reward that is necessary to prevent them from mimicking buyers with lower values. In the case of unlimited capacity, an expected profit maximizing scheme presents the buyers with a nontrivial, nonuniform price schedule for the probability of acquiring a unit. Additionat payment is made by buyers who obtain a unit. Given that buyers exhibit constant absolute risk aversion, this conditional payment is equal to the sellers opportunity cost. In the case of a unit constraint, the expected profit maximizing scheme was found in reduced form. In other words, only an individuals perceived probability of obtaining the unit (Q*(Oi)), rather than his actual probability as a function of all buyers values (q,(B, ,..,, O,)), was constructed. The functions qi were, however, constructed for the case which necessarily occurs when buyers are only mildly risk averse. Then, the optimal scheme approximates a first price auction with both a reserve price and a bid submission fee, where the submission fee is positive for low bids and negative for high bids. As long as the buyers are not risk neutral, the optimal scheme is virtually unique, and it induces buyers with high values to submit bids approximately equal to their values. In the sense that their marginal utility of income does not depend upon obtaining a unit, buyers with high values are nearly perfectly insured. Schemes that perfectly insure all buyers do exist. When there is no capacity constraint, a take-it-or-leave-it offer scheme subjects no buyers to risk. Similarly, to sell a single unit, a first price auction found in ]20] involves paying buyers to submit according to a schedule that induces each one, as a dominant strategy, to bid his true value. Setting the reserve price at cO in these schemes results in both ex ante and ex post efficiency, since they subject only the risk neutral agent, the seller, to risk, and they award units to those who value them the most. The schemes that have been found here are also ex ante efficient or, more specifically, incentive-efficient in the sense of Myerson [18]. This is obvious, since switching to another feasible scheme must decrease the expected profit of the seller. It is somewhat paradoxicai, therefore, that they impose risk on the risk averse buyers. It is, of course, the incentive

In this auction a buyer who bids b is paid the amount 19 F(r)- dz, where r is the reserve price. * In our context, a scheme is incentive-eflcient if any other incentive compatible scheme either decreases the sellers expected profit or, for some evaluation 0, decreases the expected utility of some buyer having the evaluation 0. This efficiency notion incorporates the inform mational constraints of the environment. See Holmstrom 171. Myerson 1181, or Harris and Townsend [ 6 ] for original discussions of the concept.

392

STEVEN A.MATTHEWS

compatibility requirement that prevents Pareto dominating schemesfrom being feasible. Risk is imposed on a buyer only to decreasethe amount that must be paid to keep him from being mimicked by buyers with greater values. It is therefore unnecessary to impose much risk on a buyer with a high value, since it is unlikely that another buyer with a greater value exists. This explains why, in the optimal scheme with a unit capacity constraint, the difference between a buyers marginal utilities in the two states win and lose is approximately zero for buyers with high values. Thus the marginal conditions for ex ante optimality under complete information, which requires buyers to be perfectly insured, are satisfied for buyers with the highest values. This result is analogous to ones found in the optimal taxation (individuals with the highest ability have zero marginal tax rates) and nonuniform pricing (individuals with the highest demand curves pay a marginal price equal to marginal cost) literatures.
APPENDIX

Proof of Lemma 2.

Let -p,(O) be a buyers certainly equivalent for -j,:

4-P,(e)) = !_, et) Mt I 0

Let .$p, 10) then be the distribution putting all probability on p,(O). Since u is increasing and concave, p,(B) is greater than E(J?, 18). Therefore, replacing J with j yields greater expected profit if feasibility is maintained. A capacity constraint only involves Q, so that any existing capacity constraint is still satisfied. The expected utility of a buyer with evaluation 0 who reports d in the original schemeis Q(8) j ~(0 - t) dH(t ( 8) + (1 - Q(&) 1 u(-t) dl(t / 6 By the construction of p,(8), this expression is also equal to the expected utility of a buyer with evaluation 0 who reports 8 in the new scheme.Hence, since they are satisfied by the orginal scheme, the constraints of incentive compatibility and voluntary participation are also satisfied by the new scheme. It is therefore feasible. 1 Proof of Lemma 3. Let Z?(p,, 10) be the distribution that puts all probability on the p,.(B) defined by err,,. - LX eRdH(f 10). . --r,

SELLING

TO

RISK

AVERSE

BUYERS

393

Since p,(8) > E(d, / 0), (Q, H, pr), if it is feasible, yields greater expected profits than (Q, H, p,). Because Q is unchanged, the new scheme still satisfies any capacity constraint. In either scheme, the expected utility of a buyer with evaluation 8 who reports 0 is 1 - Q(g) FRRe I= dH(t 1d) - (1 - Q(8)) eRp. - cc! Hence the new scheme satisfies the incentive compatibility and voluntary participation constraints because the original scheme satisfies them. SO (Q, fi, JJ,) is feasible. 1 Proof of Lemma 4. Supose first that (Q, b, p) is incentive compatible. Let 0, > 8,) and let Qi = Q(si) and 6, = b(8,). Incentive compatibility implies V(ei) ~ U(Bj, ei) = v(e,) + v(Oi - bj. Qi) - w(ei - bi, Qi). Combining these two expressions results in (Al)

w(e,-b,,Q,>-vl(e,-b,,Q,),<w(B,-b,,Qz)-w(e,-b,.Q,).
This inequality proves (12), since it can be rearranged, utilizing (lo), to yield Q,eRbl
G

l-Q,

Q2eRbz. l-Q2

Because I,&J, Q) increases in y, expression (Al) with (i,j) = (2, 1) implies that V(0,) > V(t9,). Therefore, using (Al) again with (i, j) = (1, 2), 0 < VW Consequently, since 1 ( I= 1 + (1 Q)/Qu(y)

- Ue,) < ~(6 - b,,

Qz,- ~(0, - bz 3 QJ

0 < I(@,) - V(B,) < 6, - 8,. This shows that Y is absolutely continuous and hence equal to the integral of its derivative. Since V(e) - U(g, 8) > 0, with equality when B = 0, it is true that f9 E argFin Hence, whenever V(e) exists, v(d) Therefore expression = U,(& 0) = ty,(d - b(8), Q(d)). V(0) - I/(&

e).

(13) holds.

394

STEVEN

A. MATTHEWS

Now we must show that (12) and (13) imply Expression (13) implies V(e) = V(d) + jo yu,(z -b(z), ri Subtracting (Al)

incentive

compatibility.

Q(z)) dz. yields

with (Bj, ei) = (8, 0) from this expression = f+ {v,(z t7 - b(z),

V(B) - U(&@ Hence (A2) implies

Q(z)) - w,(z - W% Q@,,l dz.

v(e) - u@,e) = i,
+

l-Q(z) Q(z) ev(W

- 4

dz*

Because of (12), the expression in brackets is positive z > 0 (z < d). Therefore V(e) - U(o, 19)> 0, which compatibility. I Proof of Lemma 5. The derivatives of Lp are
v2(e by

(negative) only if proves incentive

~2(e,Q,b,c)=b+c+IC/,2(e-b,Q>

Q> w (e-b 9 12 w,(e-hQ)-Q+W9-1 &(S, Q, b, c) = -w, I(0 - b, Q> I w,,(e-b, Q) f(B)

Q)+ I

i.

Calculation

yields

i/lb

Q)= , _

Qu(Y)
Q + Qu(Y)

w2(y

I=

1_

U(Y)
Q + Qu(y)

VII(Y,Q)=

- RQ(l - Q>U(Y) ,,
-Q+Qu(y)]

v/lz(Y,

Q>= , , _

Qu;(;;,(y),

Therefore v/Jv/,~ = (w, - QYv,, from the definitions of a and /3, .e(& Q, 6, c) = b -c +

= Qu(Y) + (1 - Q) ~Y)/u(Y).

Hence3

{Qa(O,b) + (1 - Q>D(O. b)} v1,2(8 - b, Q>, (A3) e(B, Q, b, C) = -{Qa(& b) + (1 - Q>P(R b)l v,,(& b, Q>. 644)

Now, fix 8 and c, and let g(Q, 6) = &a(B, b) + (1 - Q)/3(0, b). Then, letting h(Q, b) =*(0, Q, b, c), (A3) and (A4) imply h,(Q, 6) = b - c +

SELLING

TO RISK

AVERSE

BUYERS

395

g(Q, b)

w,*(O - b, Q) and h2(Q, 6) = -g(Q, 6) w, ,(S - b, Q). Inspection of a and /.I indicates that g, < 0 everywhere, and g, < 0 unless b = 8. Also, if Q ( 1 then g(b, Q) = 0 has a unique solution b = b(Q) that is decreasing. Let b(1) E [-co, co) be the limiting value of b(Q). If 0 < Q < 1, (b-b(Q)) Jz,(Q, b) < 0 for b # b(Q), since wI1 < 0 and b(Q) decreases. Hence, b(Q) uniquely maximizes h for fixed 0 < Q < 1. Since u/r, = 0 if Q = 0 or Q = 1. any b maximizes h(0, .) and h(1, .). Now, dh(Q, b(Q))/dQ = h,(Q, b(Q)) = b(Q) - CT and dIz(Q, b(Q))/dQ* = b(Q) < 0, with strict inequality if b(Q) # 6. Hence h(Q, b(Q)) is strictly concave on [0, 11. Thus h(Q, b(Q)) has a unique maximizer Q*, and any (Q, b) maximizing h has Q = Q*. Also, if 0 < Q* < 1, then (Q*, b(Q*)) = (Q*, c) uniquely maximizes h. That Q* = Q(0, c) now follows by verifying that b(Q) < c for all Q if 0 < e(c), b(e, (8, c)) = c if e(c) < 8< e(c), and b(Q) > c for all Q if 0 > B(c). Verifying these is straightforward. If e(c) < 0 < g(c), then p(f?, c) > a((?, c). Therefore (RC) implies

Q (e a,(03 c> + (1- m c))P,(O9 c> >o 1 7c)= ew c> P(& c)- a(& c)
Hence Q increases in 19. Similarly, Q decreases because a,(8, c) < 0 and /3,(0, c) < 0. 1 in c for 0 E w(c), e(c))

Proof of Theorem 1. Without a capacity constraint, the constraints (18) and (19) can be dropped, and both p and q set to zero. Therefore c(0) = c0 -p(8) = c,,. Hence (26) and Lemma-5 imply that Q*(8) = Q(S, cO) is the unique optimizing Q, and that b*(d) = b(f?, co) = c, is an optimizing b and is the only optimizing b on @(c,), &c,)). Letting

satisfaction of (16), (17), (25) and (26) by (Q*, b*) is immediate. Expression (29) is obtained from (20) by substituting v*(e) for V(0), and is required for (Q*, b*, p*) to be incentive compatible. Since Q* is nondecreasing and b* is constant, the neglected constraint (12) is satisfied. Hence (Q*, b*, p*) is feasible. The uniqueness of Q* and b on @(c,), &c,,)) imply the uniqueness of p* on the same interval. For 19< @(c,), Q*(e) = 0 is unique-hence p*(e) < 0 is necessary by the voluntary participation constraint. Hence p*(B) = 0 for 0 < jY(c,,) is the unique optimizing p. For any 0 > @c,), p*(O) = p*(#(c,)) follows from (29). Also, incentive compatibility requires the total payment p(8) + b(B) to be constant on the set of f? for which Q(0) = 1. Hence b*(8) + p*(B) = co + p*(&c,)) is unique on [&co), 11. I

396 Proof of Corollary

STEVEN

A. MATTHEWS

1. For y > 0,
1 _ e-R~

lim u(y) = lim

R-m R-02

= 0,

Hence a(& c,,) + (F(B) - 1)/f(0) and P(0, c,) + cc as R + 03 for any 0 > co. Therefore @(c,) + c,, and e(c,) + 1 as R + co. Thus if 0 < co, Q*(e) = 0 = Q(a) for any R. If 0 > co, then lim Q*(8)=
R-CC

R-a

lim

iwt
p(& C,) -

co)
a(& Co) =

Thus Q* + Qm as R + m. Because Q*(e) = Q(& co),

i-~*(e)+~*(e)d(e-co)= [:;;-::;l[ l ,,I? I

if 0 E (@(co), @co)). Consequently, ft
v(e

if c0 < 0 < 1, then

- co, Q*(e))

= lim -

Rep 1 --eIf e= 1, then y/(0--co,Q*(8))=~(8-co, co< 8 < 1, then 1)=8-c,

R(B-C*l R(B-c,,) I =

e-c,,.

for every R. Also if

e*(e)w - co)

Q*(e) u(e - =0.

Co)

Hence, as 0 < y/1 < 1, the Dominated

Convergence

Theorem implies that

=e-co.

SELLING

TO

RISK

AVERSE

BUYERS

397

Hence p* + pO as R -+ co. Finally, lim j* P*(w(B)dB=f ,,

the convergence of expected profits (e-C,)f(@de ccl Theorem, since 0 < p* < pa

R-m

also follows from the Dominated for all R. 1

Convergence

Proof of Corollary 2. Because of (RC) there is a unique 8, maximizing (0 - c,)(l -F(8)). Both @(c,) and &c,) converge to 8, as R + 0. Hence Q(0, c,) + Q(0) as R -+ 0. Since w(y, Q) -+ yQ and c/i(y, Q) + Q as R --f 0, taking the limit in (29) yields P(e) = (e - co) Q(e) - 1 Q(z) dz.
0

Hence pO(B)=O-if e<e,. If e&e,, p0(8)=(e-c,)-(e-e,)=8,-~,. Therefore (Q*, b*, p*) + (Q, b, p) as R + 0. The scheme (Q, b, p) is obviously incentive compatible and equivalent to posting a take-it-or-leave-it price e,. Now assume buyers are risk neutral. Holding Q fixed, they are indifferent among all schemes (Q, b, p) having the same sum Qb + p. Hence, in finding an optimal scheme, b* = co = b can be set. The derivation of (3 1) is valid for R = 0. But now wz(y, Q) = y. so that (31) is maximized by a constant function Q-(q) = 0,, where 8, satisfies (32). This Q- is obviously nondecreasing, regardless of (RC). Hence the optimal scheme has any buyer with 82 B. purchasing a unit with probability one for the amount co + (0 - co) = 8,. This is equivalent to posting the price 0,. 1 Let V*(8) = w(0 - b*(8), Q*(e)) - p*(0), Y*(e) = and p*(e) = co - b*(B). Then (39) implies that V* satisfies (16) and Y* satisfies (19) by construction. Also by construction, (17) 0 < Q* < 1 holds. Expressions (36) and (37) imply (18) Y* > 0. Since b* is continuous p* is continuous and (23) vacuously holds. Letting A* = F - 1. (21) is saiistied. The Hamiltonian composed with the functions I*, Y*, Q*, b*. A*, and p* is continuous in 0 because each of these functions is continuous. Hence (25) holds. Notice that if c* is defined by c*(e) = co -p*(e), then b*(8) = c(0) = b(8, c*(0)) and Q*(e) = Q(e. c*(e)). Hence Lemma 5 implies that (26) Q* and b* maximize the Hamiltonian. It remains to show (22) and (24) hold for an appropriately defined q*. If If BEZZk+,, then b*(0) = c(0) > 0. Hence 8E I,,, then b*(8)=0. * -b* exists (except at endpoints of the intervals) and is nonpositive on P I Zk+ , , and zero on I,, . Let --I* be the right (left at 8= 1) derivative of p*. Proof of Theorem 2.

1; IF(z)- - Q*(z)1f(z) dz,

398

STEVEN

A. MATTHEWS

Then q* > 0. If r*(e) > 0, then BE IZk+ i and hence Y*(0) = 0. Thus (24) holds. Expression (22) holds because ,u* = -v* by construction, and p*(o) = co - b*(O) = co - F((e) = 0. We have shown that (Q*, b*, p*) satisfies the sufftcient conditions for the problem in Section 4. Thus it maximizes expected profit, provided it is feasible. It has been show: that b* is nondecreasing. From (37), Q*(0) is also nondecreasing, since Q and F are both nondecreasing in 8. Hence the neglected constraint (12) that Q*e*/(l - Q*) be nondecreasing is satisfied. Lemma 3 therefore implies (Q*, b*, p*) is incentive compatible. As mentioned in Remark 1, Q* is implementable by [ 11, Theorem 7, and 131, since it is nondecreasing and satisfies (8). Hence (Q*, b*, p*) is feasible. Now assume (Q, b, p) is another piecewise continuous scheme maximizing expected profit subject to the feasibility constraints. Then (Q, b) satisfies the constraints (16)-( 19), where V is derived from p in (20). Since (Q, b, p) results in the same expected profit as (Q*, b*, p*), (Q, b) must also maximize (15) subject to (16)-( 19). Thus A, ,D and r7 exist such that they and (Q, b) satisfy the necessary conditions (2 l )-(26). We first show that the multiplier ,D is continuous. If not, then (23) implies that 8 exists such that p(K) > p(e), and either Q(&) = F(8)- or Q(0+) = F(e)+. For i E { + , -), let ,u =p(B), Qi = Q(e), and b = b(ei). Becauseof (26), (Qi, b) maximizes .F(Q, b) = v/(e - 6, Q> + (b +pi[Q - F(e)-]. Continuity of the Hamiltonian requires Rt(Qt,b+)=R-(Qm,bP). If Q- = F(e)-, then ZP(Q-,b-)=Rt(Qm,bP). However, F(e)- < 1 and p- > pi imply 0 =,;;v;(Q-, b-) > P;(Q-, b&). + v,(e - b, Q)

Co)Q

O<

Thus we have the contradiction that Zt (Q , b ) > Pi (Q ~, b - ) = A similar argument leads to the contradiction ,a?(Q-, b-). ,P-(Q-, b-) >;It-(Q+, b+) =A?(Q+, b+) if Qt = F(e)-. Hence ,D must be continuous. Since Q(e) = Q(0, co -,u(@) and b(e) = co - ,u(e), Q and b are also continuous. Now we show ~(0) = 0. Note that (22) and (24) imply ,L-< 0. Hence, because Q(0, c) decreases in c, Q(e)= m%Co-P(e)>& Q<&Co>. ht Q(0, co) = 0 on a nondegenerate interval [O,e(c,)]. Hence Q(e) = 0 on [0, @(co)], which in turn implies Y(e) > 0 on 10,@(co)). Hence (22) implies p(0) = 0.

SELLING

TO

RISK

AVERSE

BUYERS

399

We show next that ZA= ,u*. First, notice that if ,u < 0 and ,u* < 0 on an interval I, then (22) and (24) imply Y = Y* = 0 on I, and hence Q = Q* = F- on I. Therefore, since Q(e) = Q(0, c,, -p(S)) and Q*(8) = Q(0, c0 -,u*(8)), and Q(8, c) decreases in c, .D = ,u* on I. Consequently, if we assume p # ,u* and let

then there exists an interval J= [s, t) upon which p is constant and ,u* is strictly decreasing and p < ,u*, or vice versa. We may assume fi < ,u* on J, as the argument is symmetric in ,U and ,u*. We can then let s be minimal, i.e., s = inf(8 1pu(e) =&t)}. We know s > 0, since .D is continuous, ~(0) = 0. and p(s) < p*(s) < 0. Hence ,D is strictly decreasing on a nondegenerate interval (r, s). Observe that if x > s, then Q(x) = 0(x, c0 - p(x)) < 0(x, c,, - p*(x)) = Q*(x), with the inequality strict for x E J. Hence, Y(B) = j [F(x)- 0 >
[F(X)-

- Q(x)] f(x) - Q*(x)] f(x)

dx dx= Y*(8) > 0

for any BE J. In particular, Y(s) > 0. But (22), (24) and p < 0 almost everywhere on (r, s) imply Y(s) = 0, a contradiction. Therefore ZI =,u*. But p =,D* and Lemma5 imply Q= Q*. Since 0 < Q*(0) < 1 for any e > e(c,), Lemma 5 also implies b(e)= b*(e) for e > #(c,). Incentive compatibility now implies p = p*. 1 Proof of Corollary 3. Since C+ co as R + 0, and co is increasing, we know immediately that Q* + Q and b* + b. If t9 < B,, then p(B) = 0 is implied by Q(e) = 0. If 0 > Bo, then integration by parts yields

p(e) = v(e - b(e), Q(e)) - 1 v,(z - b(z), Q(z)) dz @lb

= (e-c(e)) = -c(B) F(e)- F(B)+ - jR F(z)- 00 dz I

max(8,. z) dF(z)-.

400

STEVEN

A. MATTHEWS

REFERENCES 1. P. DASGUPTA. P. HAMMOND, AND E. MASKIN, 185-216. of voting schemes: The implementation of social choice rules,

Rev. Econ. Studies 46 (1979),

2. A. GIBBARD, Manipulation A general result, Econometrica 41 (1973). 587-602. 3. M. GOLDMAN, H. LELAND, AND D. SIBLEY, Optimal Nonlinear Prices. Bell Laboratories Technical Report, 1978. 4. M. HARRIS AND A. RAVIV, Allocation mechanisms and the design of auctions, Econometrica 49 (1981). 1477-1500. 5. M. HARRIS AND A. RAVIV. A theory of monopoly pricing schemes with demand uncertainty, Amer. Econ. Rev. 71 (1981), 347-365. 6. M. HARRIS AND R. TOWNSEND, Resource allocation under asymmetric information, Econometrica 49 (198 I), 33-64. 7. B. HOLMSTROM. On Incentives and Control in Organizations, Ph.D. dissertation. Stanford University. 1977. 8. C. HOLT, Competitive bidding for contracts under alternative auction procedures, J. Polit. Econ. 88 (1980) 433445. 9. E. MASKIN AND J. RILEY, Auctioning an Indivisible Object, JFK School of Government, DP No. 870. 1980. 10. E. MASKIN AND J. RILEY. Multi-unit Auctions. Price Discrimination and Bundling. UCLA DP No. 201, 1981. 11. E. MASKIN AND J. RILEY, Optimal Auctions with Risk Averse Buyers. Churchill College. Cambridge University, 1982. 12. S. MATTHEWS, Risk Aversion and the Efliciency of First and Second Price Auctions. Northwestern University, 1980. 13. S. MATTHEWS. On the Implementability of Reduced Form Auctions. Northwestern University, CMSEMS DP No. 560. 1983. 14. P. MILGROM AND R. J. WEBER, A theory of auctions and competitive bidding. Econometrica 50 (1982) 1089-l 122. 15. L. MIRMAN AND D. SIBLEY, Optimal nonlinear prices for multiproduct monopolies, Be!/ J.

Econ. 11 (1980), 659-670.

16. J. A. MIRRLEES, Optimal tax theory. a synthesis, J. Public Econ. 6 (1976) 327-358. 17. M. MLJSSA AND S. ROSEN, Monopoly and product quality. J. Econ. Theory 18 (1978), 301-317. 18. R. MYERSON. Incentive compatibility and the bargaining problem, Econometrica 47 (1979). 61-73. 19. R. MYERSON, Optimal auction design, kfarh. Op. Res. 6 (1981) 58-73. 20. J. RILEY AND W. SAMUELSON, Optimal auctions, Amer.-Econ. Rec. 71 (1981), 381-392. 21. J. RILEY AND R. ZECKHAUSER, Optimal selling strategies: When to haggle, when to hold firm, Quart. J. Econ.. in press. 22. K. ROBERTS, Welfare considerations of nonlinear pricing, Econ. J. 89 (1979). 66-83. 23. A. SEIERSTAD AND K. SYDSAETER, Sufficient conditions in optimal control theory. Internal. Econ. Rev. 18 (1977). 367-391. 24. A. M. SPENCE, Nonlinear prices and welfare, J. Public Econ. 8 (1977). 1-18. 25. J. E. STIGLITZ, Self-selection and pareto efficient taxation, J. Public Econ. 12 (1982). 2 13-240. 26. W. VZKREY, Counterspeculation. auctions and competitive sealed tenders, J. Finance 16 (1961), 8-37.