CONTRACTING WITH UNKNOWN TECHNOLOGIES

Transcription

1 CONTRACTING WITH UNKNOWN TECHNOLOGIES NEMANJA ANTIĆ Princeton University PRELIMINARY See for the latest version. A. I study contracting with moral hazard when the agent has available a known (baseline) production technology but the principal thinks that the agent may also have access to other technologies, and maximizes his worst-case expected utilities under those possible technologies. All Pareto-efficient contracts take the form of participating preferred equity, a mixture of debt and equity. The nature of the contract depends on the most unproductive (in terms of stochastic dominance) technology that the principal thinks might be available to the agent. As this lower-bound technology becomes worse, the efficient contracts approach equity, generalizing existing work on robust contracting. When the lower-bound technology approaches the baseline technology, efficient contracts approach debt, providing more robust foundations for the classic financial contracting model. 1. I Can the theory of optimal contracting explain the financial contracts that we observe? The classic model of security design with moral hazard shows that debt contracts are efficient as long as contracts are restricted to be monotonic in cash flows (Innes, 199). Date: November 21, 214. Key words and phrases. Security design, Robust mechanism design, Maxmin principal. I would like to thank Marco Battaglini, Ben Brooks, Sylvain Chassang, Olivier Darmouni and Stephen Morris for invaluable discussions and advice. Thanks to seminar audiences at Princeton and Cornell for helpful comments. 1

2 2 NEMANJA ANTIĆ However, the efficient contract without this restriction is strikingly unrealistic 1. This leaves open the question of what are good microfoundations for a complete theory of financial contracting. Relative to the classical literature, the present paper relaxes the assumption that at the time of contracting the principal (or investor) knows exactly the set of technologies available to the agent (or entrepreneur) to convert effort into profits. Instead, I assume that the principal knows two things: (i) a specific baseline technology which will be available to the agent, and (ii) a lower-bound technology that yields less surplus (total profits) than any other technology. She evaluates other possible technologies with a maxmin criterion. I show that in this robust contracting setting monotonic contracts emerge because the principal is concerned that the agent might have access to a technology that exploits any non-monotonicity. I show that debt is an efficient robust contract when the lower-bound technology is close to the baseline technology. In this case, debt provides the best incentives for the agent to work hard by leaving all profits to him after a certain threshold. However, I show that equity contracts are efficient when the principal fears that arbitrarily bad technologies could be realized. Intuitively, equity financing guarantees that the agent will not chose a technology that excessively hurts the principal, since the incentives of the two are perfectly aligned. In this case, maxmin considerations dominate the value of providing incentives, consistent with the results of the recent literature on robust contracting, e.g., Chassang (213) and Carroll (214). In fact, the shape of efficient contracts changes as the lower-bound improves towards the baseline technology: it is equity for arbitrarily bad lower-bounds, participating preferred equity (a mixture of debt and equity) for intermediate cases, and is debt when the lower-bound is close to the baseline. The provision of incentives to the agent plays an increasingly important role as the Knightian uncertainty of the principal diminishes. My framework therefore reconciles two key economic forces that determine contracts in practice: incentives and robustness concerns. 1 The Pareto optimal contract is a live-or-die contract, where the principal gets paid the entire profit up to some cut-off level (lives) and gets zero above that level (dies).

3 CONTRACTING WITH UNKNOWN TECHNOLOGIES 3 The present paper makes four contributions to the literature. I first show that in a general robust contracting framework debt contracts are efficient, in one extreme case. As an intermediary step, I provide an ambiguity foundation for the monotonicity assumption commonly made in the security design literature. 2 Secondly, I show that in another extreme case equity is efficient, in line with the robust contracting literature. Third, I show that in a general environment efficient contracts take the form of participating preferred equity: a mixture of debt and equity, including both as special cases. The difference between the worst-case and baseline technology is the key simple parameter that determines whether the optimal contract is debt, equity or a mixture of the two. Finally, I prove a technical result: in sufficiently rich maxmin contracting environments, it is without loss of generality to focus on contracts which are lower semicontinuous. There is no need for strong ex-ante restrictions on the set of allowable contracts; 3 and this technical result justifies the use of simple constructive techniques. The rest of the paper is organized as follows: Section 2 presents a simple example with an application to financial contracting; Section 3 defines the model and makes some remarks about the MLRP; Section 4 makes initial general observations which are applied throughout the analysis which follows; Section 5 considers the "smallest ambiguity" extreme case and shows the Pareto optimality of debt; section 6 considers the largest possible ambiguity case and shows the efficient contract is simple equity; Section 7 provides general results that encompass the preceding observations and shows that in general participating preferred equity is optimal; Section 8 concludes. 2. F C E To provide a preview of the results and introduce key concepts, I develop a financial example and contrast the classic problem of contracting with moral hazard with the robust contracting model introduced in this paper. The example highlights the efficiency 2 This type of monotonicity assumption has been used by an array of authors, including DeMarzo & Duffie (1999), Matthews (21), Biais & Mariotti (25), DeMarzo (25), DeMarzo, Kremer & Skrzypacz (25), Inderst & Mueller (26), Axelson (27), Poblete & Spulber (212) and Dang, Gorton and Holmstrom (212). 3 For example, Carroll (214) assumes contracts are continuous.

4 4 NEMANJA ANTIĆ of the live-or-die contract in the classic model and shows that in a robust contracting environment equity and debt contracts are Pareto optimal, depending on the environment. An entrepreneur has an idea for a project and goes to a venture capital investor to get financing. The investor will provide seed funding for the project, but it is then up to the entrepreneur to put in effort to ensure the success of the project. To fix ideas, suppose that the possible profit realizations are π {, π, π}: either no profit, low proft π = 1 2 or high profit π = 1. Once a contract is agreed, the entrepreneur decides how much effort to put into the project, with more effort leading to better profit distributions. In particular, the entrepreneur can choose effort e [, 1] and the technology available to him produces random profit given by the PMF: if π = (2.1) f (π e) = 1 (1 e) if π = π. 2 1 (1 + e) if π = π 2 Observe that higher effort levels make higher profit outcomes more likely. Effort carries a utility cost for the entrepreneur, given by c (e) = 1 2 e2. A contract specifies how the profit is split between the investor and entrepreneur; it is a function, B, which maps profit realizations π B (π) [, π] to the dollars of profit given to the investor. 4 The entrepreneur gets the remaining π B (π) dollars. In this simple case, the contract is just two numbers B (π) = b [, 1 2] and B (π) = b [, 1], since if zero profit is realized neither party can get paid. We assume that the investor requires a return of 1 8 subject to this requirement. and the goal is to identify the best contract for the entrepreneur, C M In the textbook financial contracting model (Innes, 199) it is assumed that the investor perfectly knows the agent s technology at the ex-ante contracting stage. In this case, the 4 Observe that B (π) [, π], so that the entrepreneur cannot commit to paying the investor more than the entire profit and the investor is not liable for more than the initial investment.

5 problem we want to solve is: CONTRACTING WITH UNKNOWN TECHNOLOGIES 5 max max b,b e { f (π e) s.t. f (π e ) b + f (π e ) b 1 8, ( ) 1 2 b + f (π e) ( 1 b ) 12 } e2 where e is the entrepreneurs choice of effort for a given contract, which in this example is: e = , where := b b. Given this effort choice, the above problem becomes: ( 3 max b, ) ( ) ( b ) 4 (1 b) 1 ( ) 2 2 ( 3 s.t ) ( 5 4 b ) 4 ( + b) 1 8, or after simplifying: max b, s.t b 1 ( ) b The solution to this program has = 1 2 and hence the optimal contract is: b = 1 2, b =. This is called a live-or-die contract in the literature, because it has the feature the investor gets all of the profit if profit is less than 1 (the investor lives) and gets paid nothing if 2 profit is more than 1 (the investor dies).5 2 This type of contract is unrealistic and is extremely sensitive to the investor knowing exactly the technology available to the entrepreneur: if the entrepreneur had a slightly better technology than f, there would be a higher chance of π being realized, in which case the investor gets b =. Live-or-die contracts therefore are not robust to assumption that the investor perfectly knows the production technology ex-ante. 5 If we added the ad-hoc assumption that, the solution to the above program has = and is a simple debt contract.

6 6 NEMANJA ANTIĆ R F C I develop a model in which the investor does not know the production technology at the time of contracting. Instead assume the investor knows two things: (i) a baseline technology and (ii) a lower-bound CDF. The baseline technology will be available to the agent for sure, but there could be other, unknown technologies also available, as long as these are better (in terms of first-order stochastic dominance) than the lower-bound. The investor has maxmin (ambiguity averse) preferences: she wants to be sure of getting payoff 1 under every possible unknown technology that the agent might choose. 8 For the purposes of the example, let the baseline technology, denoted f, be defined as in expression 2.1. We will consider what happens for different choices of the lower-bound. Intuitively, worse lower-bounds imply larger sets for the principal to minimize over and thus more ambiguity. Largest Ambiguity. Suppose initially that the lower-bound CDF, denoted G, is arbitrarily bad, e.g., a dirac mass on, denoted δ. The investor thinks that any technology better than G could be available in addition to f. This is the case where the investor faces the largest ambiguity possible. Interestingly, it s not all doom-and-gloom for the investor. If the entrepreneur was choosing from the set of technologies A := f δ, then he would not be expected to use technology δ if f gives him more utility. In particular, the investor knows that whatever technology available, in order to be utilized it must give the entrepreneur at least as much utility as using technology f. This level of utility, denoted υ (B), is: ( ) 1 υ (B) = max f (π e) e 2 b + f (π e) ( 1 b ). The investor still requires a return of 1 8 and evaluates contract B as follows: min f (π e ) b + f (π e ) b f ( ) 1 s.t. f (π e ) 2 b + f (π e ) ( 1 b ) 1 2 (e ) 2 υ (B), where e is the effort the agent would choose if given technology f. Let V P (B f ) be the value function of the above program. Note that the constraint in the minimization

7 CONTRACTING WITH UNKNOWN TECHNOLOGIES 7 problem is relaxed if e is smaller 6 and thus the worst-case occurs when e =. We can therefore re-write the principal s problem as follows: V P (B f ) = min f (π) b + f (π) b f ( ) 1 s.t. f (π) 2 b + f (π) ( 1 b ) υ (B). This program has a linear objective and constraint and therefore has corner solutions unless b = 1 b, in which case any f that satisfies the constraint is a minimizer and 2 V P (B f ) = υ(b)b. If b > 1 υ(b) b, then f (π) = and f (π) = and still V 1 b 2 1 b P (B f ) = 1. However, since b > 1 b > this contract cannot be best for the entrepreneur 8 2 υ(b)b 1 b as b could be decreased at no loss to the investor. A similar argument applies if b < 1 2 b. Therefore b = 1 b, and hence an equity contract that simply gives the investor a b 2 ( ) percentage stake in the project is optimal. In this case b = =.161, since then υ (B) =.6512 and V P (B f ) = 1 8. Now, note that when the entrepreneur has no other technology than f, he gets utility.6512 and puts in effort e =.297. Small Ambiguity. Suppose now that the lower-bound is much better, e.g., G = (, 1 2, 1) = F ( ). In this case V P (B f ) is given by: min F G f (π) b + f (π) b ( ) 1 s.t. f (π) 2 b + f (π) ( 1 b ) υ (B). If b < b, the solution to the minimization problem has f (π) = and f (π) = υ(b) 1 b. This can never be optimal since we could decrease b (to the level of b) and the investor s payoff would be unaffected. Such a change in the contract is obviously good for the entrepreneur, since he gets a larger share if profit π is realized. This argument says that it is without loss of generality to consider monotonic contracts, i.e., focus on contracts for which b b, or. 6 This will be shown formally in lemma 4.2.

8 8 NEMANJA ANTIĆ We first want to find υ (B), i.e., the entrepreneur s utility given contract B when he is using technlogy f. This is the solution to the problem: { ( ) 1 max f (π e) e 2 b + f (π e) ( 1 b ) 12 } e2, which as we saw before has optimal effort choice e = 1 1, where = b b, and thus: 4 2 υ (B) = ( b). 32 The principal s value function given contract B is therefore given by the program: V P (B f ) = min f (π) b + f (π) (b + ) F G s.t. f (π) ( ) 1 2 b + f (π) (1 b ) ( b). 32 We find that if 1 2 (the relevant range): V P (B f ) = 1 ( ) b. 16 Thus we can now compute the best contract for the entrepreneur if he only ends up having technology f available. max max b, e s.t { f 1 (π e) ( ) + b = 1 8 ( ) 1 2 b + f 1 (π e) ( 1 b ) 12 } e2 and. The solution of the above has = and b = b = 1 8, i.e., the contract is B (π) = min ( π, 1 8). This is the discrete version of a debt contract, where the repayment level is 1. Note that 8 an equity contract (and many others) were still available to the entrepreneur to choose (as there are a lot of contracts that provide enough return for the investor). However, debt is optimal. Under the debt contract the chosen effort level is 1 4 >.297 (the effort under the previously derived equity contract) and the entrepreneur gets utility.6563 > In the large ambiguity case, however this debt contract, B (π) = min ( π, 1 8), is not available to the entreprenur, as the investor would only get utility υ(b)b = 5 1 < 1 from this 1 b contract. While other debt contracts are possible in the large ambiguity case, they are not optimal for the agent.

9 CONTRACTING WITH UNKNOWN TECHNOLOGIES 9 The above example has shown that when ambiguity is large, equity contracts are very appealing because of their robustness properties. On the other hand, when ambiguity is relatively small debt contracts are most efficient since they provide better incentives for the agent to undertake effort. 3. M I develop a moral hazard model where the agent may have technologies which are unknown to the principal at the ex-ante contracting stage. A principal (she) contract with an agent (he), who is to take a costly, private action which will randomly produce a publicly observable profit outcome π [, π] =: Π. More formally, an action is a pair (e, F ) [, e] (Π), where e [, e] is interpreted as a level of effort, F is a cumulative distribution function (CDF) over profit outcomes and (Π) is the set of Borel measures over Π, which we endow with the topology of weak convergence. The function mapping effort levels to utility cost for the agent, c: [, e] R +, is common knowledge, strictly increasing and convex. We normalize c so that c () =. A technology for the agent is a method for converting effort into random profit outcomes, i.e., a technology is a function F : [, e] (Π). Instead of writing (F (e)) (π) we write F (π e). Since functions can be represented by their graphs, we can think of technology F as the graph of F : Γ (F ) = {(e, F ( e)) [, e] (Π) : e [, e]}, that is, technology F is simply a set of actions (where effort levels are not repeated). We assume that F is continuous in e and satisfies a stochastic concavity property (Jewitt, 1988; Athey, 2). These technical assumptions are sufficient to guarantee the existence of solutions to the agent s problem and are common in the classic moral hazard literature. Where it causes little confusion we will abuse notation and denote Γ (F ) by F. The textbook models of moral hazard, starting with the classic paper by Holmström (1979), assume that there is a single profit technology, F, which is common knowledge. This literature requires further assumptions on the technology to deliver general results; in particular, these papers assume that F satisfies the monotone likelihood ratio property

10 1 NEMANJA ANTIĆ (MLRP). MLRP is a natural regularity condition on the profit technology which formalizes the idea that more effort should lead to better profit distributions: it assumes that higher effort results in better distributions over profit outcomes. Consistent with this literature, we will impose that each technology F satisfies the MLRP. Note however, that we will need a more general version of the MLRP than is frequently used as we want to allow for minimization over a rich set of measures and in particular measures which do not have densities. The general definition of the MLRP, due to Athey (22), and a discussion is given at the end of this section. Definition 1 (Technology). A technology is F : [, e] (Π), a continuous map from effort levels to distributions over profit, such that F satisfies the monotone likelihood ratio order in e and F satisfies stochastic concavity, i.e., for all π, F (π e) dπ is concave in e. I consider a robust moral hazard problem in which the assumption that there is a single common knowledge profit technology, F, is relaxed. In particular, the principal knows that some baseline technology F is available to the agent, but there could be other, unknown, profit technologies also available. This robust contracting assumption is a version of the assumption made by Carroll (214). 7 On top of the baseline technology F, we assume that the principal knows a lower-bound CDF, 8 G, such that any realized technology (first-order) stochastically dominates G. Let the set of all possible technologies be: D G := { } F (Π) [,e] : F satisfies MLRP, Γ (F ) compact, F ( e) G for all e. Note that if G = δ, then the constraint holds trivially for any CDF F. Note that as G approaches F, the Knightian uncertainty of the principal is diminishing. We will consider the problem a generic lower-bond CDF, G. 9 7 We will discuss the precise relationship in section 5. 8 We could assume that the principal knows a lower-bound technology. As we will see, the relevant bound for the principal s worst-case analysis is profit distribution the agent can costlessly induce. As such, we can replace this assumption by a lower-bound technology. If the technology is sufficiently unproductive (a lower-bound on how effort gets converted into marginal benefit in terms of profit distributions), the analysis is unchanged. 9 In what follows we do not need to impose any assumptions on G, except for in the proof of lemma 4.1. For that lemma, it is sufficient (but not necessary, in fact much weaker conditions could be given,

11 CONTRACTING WITH UNKNOWN TECHNOLOGIES 11 A contract, B : [, π] R +, specifies the payment made to the principal as a function of the realized profit. We assume B is measurable with respect to the Lebesgue σ-algebra (i.e., the completion of the Borel σ-algebra) and B (π) [, π] for all π (i.e., the investor s liability is limited to the initial investment and the entrepreneur s liability is limited to his entire profit). The agent is a risk-neutral expected utility maximizer: given the set of technologies available to him, A = {F, F 1,..., F N } D G, and a contract, B, he solves: (3.1) sup (π B (π)) df (π e) c (e), (e,f ) Γ(A) where Γ (A) = Γ (F ) Γ (F 1 )...Γ (F N ) is a set of actions representing the union of the possible actions under (or graphs of) the various available technologies. We let V A (B A) denote the value function of the above. Note that we assume A is a finite subset of D G, this is ensures that Γ (A) is compact, given our previous assumptions. Even after the regularity assumptions we have made, note that the supremum in the above problem may not be attained unless we further restrict the set of permissible contracts B. Although this is standard in the literature, e.g., Carroll (214) assumes B is continuous, one of the technical results in this paper is that it is without loss of generality to assume B is lower semicontinuous, which gives that the supremum in equation 3.1 is attained. Thus, it will make sense to talk about the arguments which maximize the agent s utility, A (B A) Γ (A). Principals are extremely ambiguity averse about the potential technologies available to the agent, but are risk-neutral with respect to risks they understand. In particular, the principal s utility: V P (B F ) = inf inf AF (e,f ) A (B A) B (π) df (π e), depending on the contract) that G has a bounded derivative on (, π]; denote the bound on G by K <. Note that this still allows for non-differentiability at, so that G can be δ for example.

12 12 NEMANJA ANTIĆ subject to knowing A D G. The assumption that the principal is getting the worst possible outcome when the agent is indifferent is largely inconsequential, since the worstcase A will usually have a single minimizing action. 1 Furthermore, when we show that restricting to lower semicontinuous contracts is without loss of generality, we will have that the infimum above is attained and therefore we may think of it as a minimum. We want to characterize Pareto efficient contracts in this environment. We say that contract B is efficient for technology set A if B such that: V P (B F ) V P (B F ), and V A (B A) V A (B A), with at least one of the above inequalities strict. By varying the outside options of the parties, we hope to get a sense of outcomes under different possible market structures, i.e., a monopolist agent and competitive principals, a monopolist principal and competitive agents, etc. One motivation for looking for Pareto optimal contracts is a central planner who wants to impose efficient outcomes in these markets. It is not immediate how a Pareto problem should be posed in this case since the agent perfectly knows the technology set A, while the principal faces Knightian uncertainty and is not aware of the realization of A. The idea is to give the agent any "extra" utility that results from the realized A, while satisfying a robust utility constraint for principal. As such, given a specific technology set, A, we want to solve for the Pareto frontier, 11 given by the following problem: (3.2) max B V A (B A) s.t. V P (B F ) R, 1 Brooks (214) makes the same assumption as above, while Carroll (214) assumes the agent maximizes the principal s utility when indifferent. The only instance in which the above is consequential is when we have a contract B and a baseline technology F, such that at the lowest effort level under F the agent is obtaining the maximum possible profit he can get given B. Carroll (214) rules these out by requiring contracts to be "eligibile". I make the assumption above predominantly because it avoids special cases and streamlines proofs. 11 Note that the notion of a Pareto frontier in the textbook setting also makes reference to a specific technology; in that case there is a single technology which is common knowledge.

13 CONTRACTING WITH UNKNOWN TECHNOLOGIES 13 where R [, R max ] denotes the location on the frontier and R max is the maximum payment the principal can be guaranteed (the point at which the agent s participation constraint binds). The reverse problem makes less sense as it assumes that excess utility from the unrealized set A is going to the principal, who does not even express a preference over this set. The key difference between a Pareto problem and a decentralized version of the above is that we are assuming away the possibility of screening or signaling. We focus on the centralized problem in the paper and I will discuss ways of decentralizing the model in section 8. The decentralization involves the agent proposing a set of contracts, from which, if the principal accepts, he can later choose any contract. This has the flavor of reverse convertible bonds/equity, where the issuer has the right to convert the contract given to the investor in some pre-agreed way. Since our A is very general, and in particular does not inherit the MLRP from individual technologies, 12 we will typically need to assume some additional structure to be able to solve the above Pareto problem. In problems of this type in the classical literature, starting with Holmström (1979), without the MLRP assumption we cannot hope to provide general results. The same thing is true in the robust contracting problem, unless the robustness of the principal s preferences is simplifying the problem significantly. While this is indeed true in the largest ambiguity case, we want to consider what happens when we place limits on the principal s ambiguity. Therefore, most of the results in the paper will assume that the agent is choosing from an MLRP set of technologies, i.e., the case where A can be represented by some F which satisfies the MLRP. One sufficient assumption that guarantees this is that there order on technologies i, such that they respect the MLRP in this order, i.e., for A =F F 1 F 2... F N we could assume that there is a reordering of the set {, 1,..., N}, denoted by r, such that for each i there exists an e i e i 1 such 12 We will describe this in detail in section 6.

14 14 NEMANJA ANTIĆ that: where e =. F r(i 1) (π e) F r(i) (π e) for e < e i, F r(i 1) (π e i ) MLRP F r(i) (π e i ), and F r(i 1) (π e) F r(i) (π e) for e > e i, In summary: the key features of the above assumptions is that (1) there is common knowledge of a lower-bound CDF and a baseline technology that the agent can choose and (2) we will characterize solutions to the Pareto problem, as stated in program 3.2, and mostly focus on the case where the agent is choosing from an MLRP set of technologies A A MLRP In this section, I make some basic remarks regarding a key assumption underlying most classical moral hazard models, including that of Holmström (1979) and Innes (199) the monotone likelihood ratio property (MLRP). The simplest version considers a family of CDFs, indexed by e, i.e., F (π e), which is twice-differentiable with respect to both π and e (as is the case in Innes (199) and most existing models). In this case, the monotone likelihood ratio property (MLRP) states that: ( ) fe (π e), π f (π e) where f is the density of F. A slightly more general definition of the MLRP, but still requiring the existence of densities, is that the likelihood ratio: f (π e H ) f (π e L ), is non-decreasing for any e H e L. An equivalent way to state this is to assume that f is log-supermodular, i.e., for all π H π L and e H e L : f (π H e H ) f (π H e L ) f (π L e H ) f (π L e L ).

15 CONTRACTING WITH UNKNOWN TECHNOLOGIES 15 Recall that a non-negative function defined on a lattice, h: X R is log-supermodular if, for all x, y X, h (x y) h (x y) h (x) h (y). definition, we can also treat f as the PMF if the measure is discrete. Note that in this version of the However, we want to allow for general distributions in the present model e.g., distributions which involve mixtures of continuous and discrete parts. As such, we work with general probability measures from the outset and require a general MLRP. The idea is to provide a similar definition using Radon-Nikodym derivatives instead of densities, however we need to be careful to ensure the absolute continuity condition in the Radon-Nikodym theorem is satisfied. This exact problem is addressed by Athey (22), who gives the right generalization of the MLRP (see definition A1). We now recount this definition, specialized to our setting. For any e L < e H R +, define a carrying measure as follows: C (π e L, e H ) = 1 2 F (π e L) F (π e H). Importantly, note that both F ( e L ) and F ( e H ) are absolutely continuous with respect to C ( e L, e H ). We say that a family of CDFs, F, satisfies the monotone likelihood ratio property (MLRP) if for any e L < e H, the Radon Nikodym derivative h (π, e) : (π, e) df (π e) dc(π e L,e H ) is log-supermodular for C-a.e. (π, e), where e {e L, e H }. To give a little intuition for this, consider the special case of differentiable CDFs. We have that: df (π e) dc (π e L, e H ) df (π e) /dπ = dc (π e L, e H ) /dπ = f (π e) 1 f (π e 2 L) + 1f (π e 2 H) f (π e) = 2 f (π e L ) + f (π e H ). The MLRP states that the Radon-Nykodym derivative above is log-supermodular, or: df (π H e H ) df (π L e L ) dc (π H e L, e H ) dc (π L e L, e H ) df (π H e L ) df (π L e H ) dc (π H e L, e H ) dc (π L e L, e H ).

16 16 NEMANJA ANTIĆ We write F ( e H ) MLR case reduces to: F ( e L ) if the above holds. Note that in the differentiable CDF f (π H e H ) f (π L e L ) f (π H e L ) + f (π H e H ) f (π L e L ) + f (π L e H ) f (π H e L ) f (π L e H ) f (π H e L ) + f (π H e H ) f (π L e L ) + f (π L e H ) f (π H e H ) f (π L e L ) f (π H e L ) f (π L e H ) f (π H e H ) f (π H e L ) f (π L e H ) f (π L e L ), which is one of the standard definitions given above. 4. P A This section makes some key observations, which will greatly simplify the proofs of the major results. While some results in this section may be of independent interest, the section may be skipped in its entirety on first reading. The key results of this section are: Lemma 4.1 which allows us to consider lower semicontinuous contracts without loss of generality, so that maximizers in the agent s problem and minimizers in the principal s problem exist; Lemma 4.2 which allows for a simpler representation of the principal s problem given by equation 4.2; and Theorem 4.3 which shows that only monotonic contracts are robust. The first observation is that in finding the principal s worst-case scenario we can, without loss of generality, assume this occurs with zero effort from the agent (at least in the limit, if an argmin does not exist). Since the only guarantee the principal has is that the agent is getting at least the utility guaranteed by F, i.e., V A (B F ), if the infimum limiting effort level for minimizing technology F 1 was not, but e >, we can construct a new technology as follows: F F1 1 (π e + e ) if e [, e e ] (π e) = F 1 (π e e ) if e (e e, e].

17 CONTRACTING WITH UNKNOWN TECHNOLOGIES 17 Clearly, V A (B F 1 ) > V A (B F 1 ) V A (B F ), thus the agent s constraint is not violated. Also note that F 1 which is preserved by limits. is an MLRP family, since the MLRP is a continuous property Secondly, when solving for V P (B F ) it suffices to consider A such that A (B A) = 1. We could simply take an alternative A which removes the technology that leads to multiple optimal choices for the agent, and the principal would be weakly worse-off. In the case where the argmax of the agent s problem does not exist, the same argument to sequences attaining the supremum for the agent. Thus, it is without loss of generality to think of the principal s preferences as: V P (B F ) = inf B (π) df (π e), subject to (e, F ( e)) A (B A). AF Lastly, we observe that one can without loss of generality, restrict attention to lower semicontinuous contracts. Since Innes (199) finds that optimal contracts are not continuous (his live-or-die contract is not continuous, given our definitions), we do not wish to restrict our analysis to purely continuous contracts in the moral hazard problem presented above (for example, Carroll (214) assumes continuous contracts). However, continuity, as well as some weaker versions of it, ensures that the infimum in the optimization problem of the principal is attained, which simplifies the analysis significantly (and is very useful when constructing worst-case scenarios in subsequent proofs). The next result shows that we can restrict attention to a class of contracts in which representative elements are lower semicontinuous. Let B denote the lower semicontinuous hull of B, i.e., B is the greatest lower semicontinuous function majorized by B. Lemma 4.1. We have that V P (B F ) = V P ( B F ) and: ( ) V P B F = min A A B (π) df (π), subject to F ( e) = A (B A), where we consider A D G, for some worst-case CDF G, where G is differentiable on (, π] with G K <.

18 18 NEMANJA ANTIĆ, 1 lim = 1 F 4.1. Proof idea for lemma 4.1. We first prove that the minimum problem is well-defined for lower semicontinuous contracts. This follows from a generalization of a classic theorem by Tonelli in the calculus of variations, as stated, for example, in Zeidler (1985) theorem 38.B, also known as the generalized Weierstrass theorem. The intuition behind the assertion that V P (B F ) = V P ( B F ) is represented in figure 4.1. The figure plots both CDFs and contracts on the same axis, assuming π = 1. The curve in red is the lower bound, or worst-case, CDF G, and the 45 line is in dashed yellow. A proposed contract, B, is in green and note that B is not lower semicontinuous. The infimum sequence of CDFs, represented in blue, puts mass on π ever closer to.4, as figure 4.1 shows. However, we cannot shift the mass all the way to.4, since this limiting CDF would result in a higher payoff to the principal. It should be clear that when the limiting CDF is considered with the lower semicontinuous hull of B (which in this case just involves moving the point at π =.4 down) we obtain the same payoff as the infimum of CDFs. The significance of this lemma is then also clear we are able to look at a single minimizing CDF (the limiting CDF) instead of a sequence.

19 CONTRACTING WITH UNKNOWN TECHNOLOGIES 19 Lemma 4.1 therefore shows that replacing a contract by it s lower semicontinuous hull results in the same solution to the principal s problem. This is also always weakly better for the agent, thus there is no loss of generality in focusing on lower semicontinuous contracts. For any contract, B, let B denote the lower semicontinuous hull of B; that is, B is the greatest lower semicontinuous function majorized by B. We say that contracts B and B are equivalent if B = B, and write B B. We can then define an equivalence { class, as follows [B] = B R [,π] + : B B }. Thus the lemma implies that it is without loss of generality consider B R [,π] + / and in particular we may take B to be lower semicontinuous. Hence, we have that the principal s problem can be written as follows: (4.1) max F A,e [,e] (π B (π)) df (π e) c (e), and denote by V A (B A) and A (B A) the value function and argmax of the above, respectively. Note that these are well-defined since π B (π) is upper semicontinuous. The principal who faces unknown technologies can still bound her payoff. In particular she has: The knowledge that the agent will not choose something worse than he was getting under F, the baseline technology, and The knowledge that all technologies must dominate the worst case G. These lead to a representation of the principal s preferences which makes plain the Gilboa-Schmeidler maxmin preference of the principal, since she is minimizing over a set of measures. Lemma 4.2. We have that V P (B F ) is the solution to: (4.2) min B (π) df (π), subject to (π B (π)) df V A (B F ). F G Furthermore, if B (π) and π B (π) are monotonic the constraint above holds as an equality. The lemma is a generalization of similar observations made in theorem 1 in Chassang (213) and lemma 2.2 in Carroll (214). In a moral hazard setting both of these papers

20 2 NEMANJA ANTIĆ find that the principal can essentially only bound her utility by the knowledge that the agent will not choose a worse outcome than what he is guaranteed under the known technology. Madarász and Prat (214) exploit a similar argument in a screening setting. The main differences between my proof and earlier literature arises from complications when there is a non-trivial lower-bound (when G δ ) and the assumption that the principal fears getting the worst possible outcome when the agent is indifferent. 13 We can generally think of the constraint in program 4.2 as tight. Holmström (1979) and Shavell (1979) point out that the monotonicity of π B (π) follows directly from the definition of the MLRP for optimal B, and we will show next that it is without loss of generality to focus on monotone B R M C Given the above preliminaries, the key assertion of this subsection is that robustness considerations lead to monotonic contracts. The intuition for this is that a principal facing a non-monotonic contract will assume that a productive technology which exploits the non-monotonicity will be available to the agent and therefore disregard any non-monotonic aspects of the contract. Theorem 4.3. For any G and any non-monotonic contract B (π) there exists a monotonic contract B m (π) such that: inf B m (π) dfm A (π) = inf B (π) df A (π), A D G A D G subject to Fm A A A (B m A) and F A A A (B A), i.e., the principal is indifferent between the two contracts, and B (π) B m (π), i.e., the agent s prefer the monotonic contract. The intuition for the above result is given in figure 4.1. The idea is that if a principal is offered a non-monotonic contract, contract B in figure 4.1, she would discount the non-monotonic part, since in the worst-case analysis she thinks that nature will endow the agent with a good technology which puts no mass on the non-monotonic part. 13 The latter assumption allows us to state the lemma without reference to "eligible" contracts.

21 CONTRACTING WITH UNKNOWN TECHNOLOGIES 21, 1 = 1 F 4.1. Proof idea for theorem 4.3. Since F (π e) = G (π) for all e, the two ways the principal can bound her payoff, the lower-bound and the agent s utility under the reference technology, are one and the same. In particular, if any technology better than G was available, the principal could only improve her payoff since contracts are monotonic. Thus the principal s worst-case in this extreme of the model is simply that only G is available. Note however that ambiguity still has a role: it is critical in proving theorem 4.3, which says that robust contracts are monotonic. Aside from this however, the key concern is the provision of incentives to the agent, as in the textbook model. 5. S A, F ( e) = G This section and the next consider the two extreme cases of the model and build intuition for the results. This section considers the smallest ambiguity case, where the lower-bound (G) and reference (F ) technologies are the same. I will show that debt contracts are optimal in this extreme. We will subsequently analyze the largest ambiguity case and with the intuition of these extremes proceed to the general results.

22 22 NEMANJA ANTIĆ The main result of this section will be the optimality of debt contracts when the agent is choosing from an MLRP set. The proof idea is the same as in Innes (199). We show that a non-debt contract induces a lower effort choice than a debt contract which gives the principal the same payoff, and that this is below the first-best level of effort. There are several complications in this version: we need to generalize the argument to allow for non-differentiability of CDFs and we need to be careful since ambiguity considerations are important when we are replacing contracts R M The model we have presented is sufficiently well-behaved. This section makes some general remarks that could be skipped on first reading. Remark 1. For any bounded, continuous function φ, in e. φ (π) df (π e) is continuous Since F (π ) is continuous in e, this observation follows directly from the portmanteau theorem. The first-best effort level would be the amount of effort chosen if the agent owned the firm. This is of course not feasible due to limited liability assumption. Note that the first-best problem, if the agent is using technology F, is: { } π df (π e) R c (e). max e Remark 2. The above has a unique solution with a positive effort level e >. This is true since c is strictly convex in e and F (π e) dπ is convex in e for all π. By Athey (22), the latter condition implies that π df (π e) is concave in e, which is sufficient for the existence and uniqueness of a solution. Under these assumptions, the first-order condition 15 for this problem is: ( π ) π df (π e) c e (e) =. e 14 In particular, replacements need to be done with respect to the worst-case of the principal, as opposed to the commonly known technology. 15 The display expression is assuming differentiability with respect to e, which is a special case of our model. In general the idea is exactly the same and the special case is shown in this case for simplicity.

23 CONTRACTING WITH UNKNOWN TECHNOLOGIES 23 This equation is strictly decreasing in e and is positive by assumption for e =, thus by the mean value theorem there is a unique e which solves the above first-order condition. Through a similar argument we see that contracts, like equity and debt lead to unique solutions for the agent s problem. In particular, the following has a unique solution: max e (π min (π, z)) df (π e) c (e), for z. Denote the solution by e (z) and note e (z) is continuous in z, by Berge s maximum theorem R The main result of this section is that Pareto optimal contracts take the form of debt. Theorem 5.1. For any A D G solution to: where the agent is choosing from an MLRP set, a max B V A (B A), subject to V P (B F ) R, is B D z (π) := min (π, z) for some z [, π]. The proof of the theorem goes by showing that when a monotonic contract is replaced by an appropriately chosen debt contract the agent is induced to put in more effort because of the MLRP. This is a key property of the MLRP and is summarized in the following lemma. Lemma 5.2. Let φ (π) be a function such that φ (π) for π π B, φ (π) for π π B and either: (1) φ (π) df (π e L) =, or (2) φ (π) df (π e L) and φ (π) decreasing for π π B. Then, for any e H > e L and any MLRP family F, we have that φ (π) df (π e L) φ (π) df (π e H). This is a generalization of lemma 1 from Innes (199). The proof technique is similar, but needs to take care of technical difficulties arising from the non-existence of densities.

24 24 NEMANJA ANTIĆ The lemma is key in the proof of the main theorem, since it says that replacing generic monotone contracts by debt contracts implies higher marginal benefits of effort. Note that the inequality in the above theorem holds strictly if the MLRP is strict. This observation will be useful for the uniqueness result that follows. Corollary 5.3. The repayment level, z, in the optimal contract, B D z increasing in R and decreasing in G. 16 (π) = min (π, z), is The above corollary follows since the level of debt is chosen so as to guarantee the principal the required utility R under the worst-case scenario where G and only G is available. This implies that the level of repayment z is increasing in R. Furthermore if G G, the level of repayment required under G would be greater than under G. Corollary 5.4. Debt is the unique solution to the above problem if R (, R max ), F satisfies strict MLRP and G has full support N E To demonstrate why the efficiency question is interesting, let us consider a simple example. Let Π = [, 1], e = [, 1], c (e) = 1 1 e2 and G = F = U [, 1]. Fix a level of principal utility R. As discussed, the worst-case scenario for the principal is that only the (constant) technology F is available to the agent 17. The principal is thus indifferent between many contracts. In particular, the principal is indifferent between an equity and debt contract defined as follows: Bα E (π) = απ, with α = 2R Bz D (π) = min (π, z), with z = 1 1 2R, and, since: 1 B D z (π) dg (π) = 2R 1 π dπ = R, 16 If we think of potential G CDFs as being ordered by stochastic dominance. 17 This is because contracts have to be monotonic and the agent gets his "promised" utility under G.

25 Agent's Optimal Utility and: 1 B D z (π) dg (π) = CONTRACTING WITH UNKNOWN TECHNOLOGIES 25 = 1 1 2R π dπ + ( ) R + 2 ( 1 ) ( ( 1 2R 1 G 1 )) 1 2R ( 1 1 2R) 1 2R = 1 1 2R R + 2R + 1 2R 1 = R. Consider now an agent with the following technology set A: F (π e) = π e+1, for e [, 1]. Note that this is an MLRP technology set and that F (π e) G (π) for all e. Figure 5.1 plots the utilities of the agent under the two contracts above, given different possible reservation utilities of the principal R..7 Agent's U Equity.6 Agent's U Debt Principal's Guaranteed Utility, R F 5.1. Numerical example illustrating Pareto Efficiency

26 26 NEMANJA ANTIĆ We see in figure 5.1 that although the principal is indifferent between the contracts, the agent clearly prefers the debt contract for all R (, R max ). Note that in this case R max = 1/2. When R = or R = R max the debt and equity contracts are the same they either award all profit to the agent or principal. 6. L A, G = δ We now consider the case where the lower-bound CDF is trivial. In this extreme version of the model the maxmin aspect of the principal s preferences really restricts what is achievable and we do not need to make further assumptions about A. As such we can consider general sets A D, and in particular we do not need to assume that the agent is choosing from an MLRP set of technologies. Theorem 6.1. For any A D, a solution to: max B V A (B A), subject to V P (B F ) R, is B α (π) = απ for some α [, 1], i.e., a linear/equity contract. The intuition for this proof is that an extremely uncertain principal places a huge premium on having preferences perfectly aligned with the agent, which is what happens when the contract is linear. Even if there are efficiency gains from providing stronger incentives for the agent at the upper end of profit outcomes, as is the case when A is an MLRP set, this benefit is over-ridden by the principal s pessimism. The proof of theorem 6.1 is illustrated in figure 6.1. The left-hand panel gives the intuition for why contracts have to be (weakly) convex. In particular, consider a concave contract B (in green). In performing her worst-case analysis, the principal is wants to find the worst way (for her) that the agent can gain exactly the utility guaranteed by F, υ := V A (B F ). Given that the set of CDFs she can minimize over is unrestricted, she will put mass on just two points: there will be a lot of mass on, since this gives her no payoff, and just enough mass on the point which minimizes B(π), i.e., the point π which minimizes what the principal gets relative to what the agent gets. In this case

27 CONTRACTING WITH UNKNOWN TECHNOLOGIES 27 $,. 1 $,. 1 Ratio minimized ` = 1 Optimal contract is (weakly) convex. Optimal contract is linear. = 1 F 6.1. Proof idea for theorem 6.1. "just enough" means to make the agent choose this constructed CDF (at zero effort cost) over whatever was optimal in F. This worst-case CDF is illustrated by F B in the figure. Now, consider replacing B by the lower convex hull, B c. Note that at the worstcase the principal is indifferent between B and B c. Furthermore, since B c is linear, it satisfies a "no-weak-point" constraint, so that the minimizing CDF for the principal is any CDF which delivers the required utility to the agent including F B. This replacement therefore makes the principal no worse off, but makes the agent weakly (and generally strictly) better off. The right-hand panel in figure 6.1 provides intuition for why contracts have to be linear. In particular, consider the principal s worst-case analysis when faced with a convex contract B, where the agent is guaranteed some level of utility υ. Jensen s inequality implies that the worst-case scenario is a dirac distribution δ π at the lowest level of profit which gives the agent exactly utility υ. One can replace B by a linear contract B α that goes through (π, B (π )) and we again note that the principal is no worse off. It is not immediate that the agent likes this replacement however, since there is an interval, [, π ], on which B α > B. The agent does like this replacement however since the agent s average payoff under whatever technology he was choosing from A is at least υ, it cannot be the case that the agent is putting much mass on [, π ] relative to the mass this CDF

28 28 NEMANJA ANTIĆ puts on [π, π]. Another application of Jensen s inequality ensures that this replacement indeed gives the agent higher utility (and strictly higher if the agent s chosen distribution is not δ π ). We say that A has full support, if for all F i A and e [, e], supp(f i ( e)) = [, π]. Corollary 6.2. Equity is the unique solution to the above problem if R (, R max ) and A has full support. The equity contract is the unique efficient contract if the agent s technologies have sufficiently large support R C (214) The robust contracting framework of Carroll (214) maps directly to the largest ambiguity case analyzed above. One difference is that Carroll (214) focuses on unknown actions, as opposed to technologies, and does not require MLRP. Our choice of focusing on technologies is inspired by the classical literature on contract theory which imposes natural restrictions such as the MLRP. However, as we noted in the model section, if we assume nothing about how these technologies are inter-related there is no bite to the MLRP assumption. In particular, an action in Carroll s setup can be converted to a technology as follows. Let (F, e) be an action available to the agent in Carroll s model. We can define an MLRP technology, from which (F, e) will be chosen if it dominates the zero action, as follows: δ F i ( e if e < e ) = F if e e. Note that F i, as defined above, satisfies the generalized MLRP. 18 The key difference is that Carroll (214) solves the principal-optimal problem. The main result is presented below. 18 Note however that F i is not continuous in e (although this could be easily modified through a standard mollifier construction) and that F i fails stochasic concavity. Both of these assumptions on technologies could be dropped without affecting any results in this section.