HTML from LaTeXML, with custom CSS/JS. The PDF is more accurate.

Single-Crossing Differences in Convex Environments111Previous versions of this paper were circulated under the title “Single-Crossing Differences on Distributions”. We thank Nageeb Ali, John Duggan, Federico Echenique, Mira Frick, Ben Golub, Ryota Iijima, Ian Jewitt, Alexey Kushnir, Shuo Liu, Daniele Pennesi, Jacopo Perego, Lones Smith, Bruno Strulovici, and various audiences for helpful comments.

Navin Kartik Department of Economics, Columbia University. E-mail: nkartik@gmail.com    SangMok Lee Department of Economics, Washington University in St. Louis. E-mail: sangmoklee@wustl.edu    Daniel Rappoport Booth School of Business, University of Chicago. E-mail: Daniel.Rappoport@chicagobooth.edu
Abstract

An agent’s preferences depend on an ordered parameter or type. We characterize the set of utility functions with single-crossing differences (SCD) in convex environments. These include preferences over lotteries, both in expected utility and rank-dependent utility frameworks, and preferences over bundles of goods and over consumption streams. Our notion of SCD does not presume an order on the choice space. This unordered SCD is necessary and sufficient for “interval choice” comparative statics. We present applications to cheap talk, observational learning, and collective choice, showing how convex environments arise in these problems and how SCD/interval choice are useful. Methodologically, our main characterization stems from a result on linear aggregations of single-crossing functions.


Keywords: monotone comparative statics, choice among lotteries, interval equilibria, aggregating single crossing

1 Introduction

1.1 Overview

Single-crossing properties and their implications for choices are at the heart of many economic models, as highlighted by Milgrom and Shannon (1994). Consider a utility function u⁢(a,θ), where a is the choice object and θ a preference parameter. In this paper, we completely characterize the structure of u⁢(⋅) when it has single-crossing differences in a rich—specifically, convex—environment. Before defining this property more precisely and explaining our results, we begin with some background and motivation for our work in a leading context: choice among lotteries.

Motivation.

In various applications of single crossing, choices have been restricted to deterministic outcomes when it would be desirable to accommodate lotteries. Consider Crawford and Sobel’s (1982) canonical cheap-talk model: a sender with private type θ∈Θ⊆ℝ chooses a costless message to send a receiver, who then takes a decision a. A single-crossing property on the sender’s utility u⁢(a,θ) ensures that any equilibrium is an “interval equilibrium”, i.e., Θ is partitioned into intervals that induce the same decision. This result is predicated on assumptions ensuring that in equilibrium, the sender can fully anticipate the decision induced by each message. However, in assorted contexts, one would like to model a sender who is uncertain about the receiver’s preferences; but then, from the sender’s perspective, each message would induce a lottery over decisions.

A challenge with extending single-crossing properties to choice among lotteries is that there is no natural order on the set of all lotteries. Moreover, it is not always apparent a priori what restrictions are reasonable on the set of feasible lotteries, in particular whether some form of stochastic dominance can order every choice set an agent may face. In the cheap-talk problem, the nature of the receiver’s preferences may well imply that a higher message (in equilibrium) induces lotteries that have both higher mean and higher variance. In other applications, the lotteries may be the result of still further interactions—e.g., they may represent continuations in dynamic strategic problems—that are intractable to structure ex ante.

An alternative to restricting the set of lotteries is to require that an agent’s utility difference between any pair of lotteries—or, more abstractly, any pair of choice objects—is single crossing in the agent’s preference parameter or type. (For any ordered set Θ, a function Θ↦ℝ is single crossing if its sign is monotonic.) It is this single-crossing property that we study, which we refer to as single-crossing differences (SCD). Our notion is closely related to that of Milgrom and Shannon (1994) for comparative statics, and the “order restriction” property Rothstein (1990) introduced for collective choice. As we explain in Section 2, SCD characterizes interval choice (Theorem 1), a fundamental property for applications. Loosely, interval choice says that given any choice set and any alternative in that set, if a low and a high type both find that alternative optimal, then so do all intermediate types. In the cheap-talk application, the sender’s alternatives are messages, and interval choice yields the aforementioned interval equilibrium desideratum.

A key question, then, is: for an expected-utility agent, which (von Neumann Morgenstern) utility functions assure SCD over lotteries? It is not enough that SCD holds over pure outcomes:

Example.

Let Θ=[−1,1], A={0,1,2}, and u⁢(a,θ)=a for a≠1 while u⁢(1,θ)=θ2+1/2. For any a,a′∈A, u⁢(a,θ)−u⁢(a′,θ) is single-crossing in θ as its sign does not depend on θ. But for G the uniform lottery over outcomes 0 and 2, the expected utility difference u⁢(1,θ)−𝔼a∼G⁢[u⁢(a,θ)]=θ2+1/2−(1/2)⁢2=θ2−1/2 is not single crossing in θ. Hence, u has SCD over pure outcomes but does not have SCD over lotteries.

SCD in Convex Environments.

In the foregoing example, SCD holds for “extreme” alternatives (i.e., pure outcomes) but fails over “in-between” alternatives (i.e., lotteries). We find this broader viewpoint valuable, as various economic settings have such richness in the alternatives. Accordingly, consider a utility function u⁢(a,θ), where a∈A is an action (i.e., a choice alternative) and θ∈Θ the ordered type. Say that the choice environment is convex if the set of functions {u⁢(a,⋅):Θ→ℝ}a∈A is convex.222That is, for all a,a′∈A and λ∈(0,1), there is a′′∈A such that for all θ∈Θ, u⁢(a′′,θ)=λ⁢u⁢(a,θ)+(1−λ)⁢u⁢(a′,θ).

Convex choice environments abound. The function u can be expected utility and A the set of lotteries over arbitrary outcomes. But, as explained by Example 2 in Section 3, u can also be rank-dependent utility (Quiggin, 1982). Or, as detailed in Example 3, A can be a product set representing different dimensions of the choice object, and u can be a multidimensional utility function with convex range. This class captures examples of deterministic settings in mechanism design, dynamic consumption streams, and choices over bundles of goods or products with multiple characteristics.

Our paper’s main result, Theorem 2, is a characterization of SCD in any convex environment. Theorem 2 establishes that in a convex environment, u has SCD if and only if

u⁢(a,θ)=g1⁢(a)⁢f1⁢(θ)+g2⁢(a)⁢f2⁢(θ)+h⁢(θ), (1)

where f1 and f2 are single-crossing functions that satisfy a ratio-ordering property we define in Section 3. Roughly speaking, ratio ordering requires that the relative importance placed on g1⁢(⋅) versus g2⁢(⋅) changes monotonically with type.333In particular, if either f1⁢(⋅) or f2⁢(⋅) is strictly positive, then ratio ordering reduces to saying that the ratio of the two functions is monotonic. More generally, Lemma 1 establishes that ratio ordering is necessary and sufficient for all linear combinations of two single-crossing functions to be single crossing. The idea is transparent when Θ has a minimum θ¯ and a maximum θ¯. Then, u having SCD is equivalent to the existence of a (type-dependent) representation u~⁢(a,θ) that satisfies

u~⁢(a,θ)=λ⁢(θ)⁢u~⁢(a,θ¯)+(1−λ⁢(θ))⁢u~⁢(a,θ¯),

where λ:Θ→[0,1] is increasing (Proposition 1). In other words, in such a convex environment, SCD is equivalent to being able to represent each type’s preferences by a utility function that is a convex combination of those of the extreme types, with higher types putting more weight on the highest type’s utility.

In the context of expected utility, there are canonical (von Neumann Morgenstern) functional forms that induce SCD over lotteries: in mechanism design and screening, u⁢((q,t),θ)=θ⁢q−t, where q∈ℝ is the quantity, t∈ℝ is the transfer, and θ∈ℝ is the agent’s marginal rate of substitution; in optimal delegation without transfers, u⁢((q,t),θ)=θ⁢q+g⁢(q)−t, where q∈ℝ is the allocation, t∈ℝ+ is money burning, and θ∈ℝ is the agent’s type (cf.  Amador and Bagwell, 2013); in communication/delegation and voting, u⁢(a,θ)=−(a−θ)2=2⁢θ⁢a−a2−θ2, where a∈ℝ is an outcome and θ∈ℝ is the agent’s bliss point. On the other hand, our characterization also makes clear that SCD is quite stringent. For example, within the class of power loss functions, only the quadratic loss function generates SCD over lotteries (Corollary 2). Outside of expected utility, we explain in Subsection 3.2 when discounted utility and Cobb-Douglas utility satisfy SCD in a convex environment; in particular, it holds for the simple two-good case of u⁢((x,y),θ)=θ⁢log⁡x+(1−θ)⁢log⁡y, where x,y∈ℝ++ are the quantities and θ parameterizes the marginal rate of substitution.

Applications.

Section 4 applies SCD in convex environments to three economic problems, highlighting the implications of interval choice. Among other things, the cheap-talk application with uncertain receiver preferences in Subsection 4.1 demonstrates concretely how choices from all lotteries emerge naturally. In Subsection 4.2’s observational-learning application, a convex environment stems from multidimensional utility. Subsection 4.3 considers collective choice over lotteries, showing how our results contribute to a long-standing question of whether equilibrium exists when political candidates can offer lottery platforms (Zeckhauser, 1969; Shepsle, 1972).

1.2 An Intuition

A key step towards Theorem 2’s characterization of SCD in convex environments is establishing that every type’s utility is an affine combination of two (type-independent) functions: Equation 1. We can provide a succinct intuition when Θ has a minimum θ¯ and a maximum θ¯. It suffices to show that there are three actions, a1, a2, and a3, such that any type θ’s utility from any action a satisfies

u⁢(a,θ)=λ1⁢(a)⁢u⁢(a1,θ)+λ2⁢(a)⁢u⁢(a2,θ)+λ3⁢(a)⁢u⁢(a3,θ), (2)

for some 𝝀⁢(a)≡(λ1⁢(a),λ2⁢(a),λ3⁢(a)) with ∑i=13λi⁢(a)=1. (Equation 1 follows by setting, for i=1,2, fi⁢(θ)=u⁢(ai,θ)−u⁢(a3,θ), gi⁢(a)=λi⁢(a), and h⁢(θ)=u⁢(a3,θ).) The desired 𝝀⁢(a) is the solution to

[u⁢(a,θ¯)u⁢(a,θ¯)1]=[u⁢(a1,θ¯)u⁢(a2,θ¯)u⁢(a3,θ¯)u⁢(a1,θ¯)u⁢(a2,θ¯)u⁢(a3,θ¯)111]⁢[λ1⁢(a)λ2⁢(a)λ3⁢(a)],

which exists at least when there are three actions for which the 3×3 matrix on the right-hand side is invertible. To interpret this matrix equation, note that in a convex environment, any convex combination of utilities from {a1,a2,a3,a} is the utility from some action. Hence, the equation says that one can find two distinct actions with utilities equal to convex utility combinations of {a1,a2,a3,a} such that the lowest and highest types are both indifferent between those two actions.444Take one utility combination to be that corresponding to the uniform distribution 𝑷=(1/4,1/4,1/4,1/4) on {a1,a2,a3,a} and the other corresponding to 𝑷+ε⁢(λ1⁢(a),λ2⁢(a),λ3⁢(a),−1) for any sufficiently small ε>0. By SCD, all types must be indifferent between these two actions, which amounts to Equation 2.

A second key step towards Theorem 2 is showing that the two type-independent utilities in Equation 1 must be ratio ordered. The intuition for this step is provided in Subsection 3.3.

1.3 Related Literature

We now offer a summary of related work, supplying additional details later.

Quah and Strulovici (2012) consider an expected-utility agent choosing under uncertainty about her preferences, which depend on some unknown “state”. They ask when single-crossing differences in the Milgrom and Shannon (1994) sense is preserved regardless of the state distribution. In our expected-utility application, we consider an agent who knows her preferences but chooses among lotteries. Although these are conceptually different questions, there is a mathematical connection in portions of our analysis. Quah and Strulovici’s (2012) question concerns when single crossing is preserved by positive linear combinations. On the other hand, our problem turns on arbitrary linear combinations preserving single crossing. As elaborated after Lemma 1, this explains the difference between Quah and Strulovici’s signed-ratio-monotonicity condition and our ratio-ordering condition. Moreover, there is no analog in their analysis to the linear dependence we deduce in Proposition 2, which is crucial to our main characterization’s functional form (5).

When Θ⊆ℝ, the utility specification u⁢(a,θ)=θ⁢g1⁢(a)+g2⁢(a) induces expected utility with SCD over lotteries; indeed, the expected-utility difference between any two lotteries is monotonic in θ. The usefulness of this utility specification (or slight variants) to structure choices among arbitrary lotteries has been highlighted by Duggan (2014), Celik (2015), and Kushnir and Liu (2018). In Subsection 5.1, we show how SCD preferences always have such a “monotonic differences” representation in convex environments that satisfy a reasonable additional condition. This is a striking consequence of convex environments, as we are not aware of any such result more generally.

In the operations research literature, there has been interest in functional forms for “multi-attribute utility functions” (e.g., Fishburn, 1974). When there are two attributes and the agent has expected-utility preferences, Abbas and Bell (2011) study a “one-switch condition” that is akin to SCD over lotteries on one attribute. In that setting, they offer a result related to our Proposition 1. However, they do not identify ratio ordering as the property that characterizes when single crossing is preserved under aggregation, which is a key contribution of our analysis (Lemma 1 and Proposition 2). Also novel to our paper are the comparative statics characterizations of SCD (Theorem 1 and Theorem 4) and our economic applications.

For choice among restricted sets of lotteries (which we wish to largely avoid, for reasons mentioned earlier), there are various prior results on the conditions for monotone comparative statics under expected-utility preferences. Standard restricted classes of lotteries include those ordered by first-order stochastic dominance (Topkis, 1978) and likelihood-ratio dominance (Karlin, 1968; Athey, 2002).555For first-order stochastic dominance, the requirement is that the utility function is supermodular; for likelihood-ratio dominance, it is log-supermodularity. Smith (2011) considers choice among arbitrary lotteries and their certainty equivalents.

Outside of expected utility on lottery spaces, we are not aware of any work highlighting SCD in convex environments. In our view, the lens of convex environments is a contribution of our paper.

2 Single-Crossing Differences and Interval Choice

Our analysis begins by formalizing comparative statics results that justify a notion of single-crossing differences without reference to an order over the choice space.

Let (Θ,≤) be a (partially) ordered set containing upper and lower bounds for all pairs.666A partial order—hereafter, also referred to as just an order—is a binary relation that is reflexive, anti-symmetric, and transitive (but not necessarily complete). An upper (resp., lower) bound of Θ0⊆Θ is θ¯∈Θ (resp., θ¯∈Θ) such that θ≤θ¯ (resp., θ¯≤θ) for all θ∈Θ0. While none of our results require any assumptions on the cardinality of Θ, the results in Section 3 are trivial when |Θ|<3. See an earlier version of this paper, Kartik et al. (2019, Appendix I), for how our results extend when (Θ,≤) is only a pre-ordered set, i.e., when ≤ does not satisfy anti-symmetry. We often refer to elements of Θ as types.

Definition 1.

A function f:Θ→ℝ is:

  1. 1.

    single crossing (resp., from below or from above) if sign⁡[f] is monotonic (resp., increasing or decreasing);777For x∈ℝ, sign⁡[x]=1 if x>0, sign⁡[x]=0 if x=0, and sign⁡[x]=−1 if x<0. A function h:Θ→ℝ is increasing (resp., decreasing) if θh>θl⟹h⁢(θh)≥h⁢(θl) (resp., h⁢(θh)≤h⁢(θl)); it is monotonic if it is either increasing or decreasing. An equivalent, and perhaps more familiar, definition of f being single crossing from below is (∀θ<θ′)⁢f⁢(θ)≥(>)⁢0⟹f⁢(θ′)≥(>)⁢0.

  2. 2.

    strictly single crossing if it is single crossing and there are no θ′<θ′′ such that f⁢(θ′)=f⁢(θ′′)=0.

Definition 2.

Given any set A, a function u:A×Θ→ℝ has:

  1. 1.

    single-crossing differences (SCD) if ∀a,a′∈A, the difference Da,a′⁢(θ)≡u⁢(a,θ)−u⁢(a′,θ) is single crossing in θ;

  2. 2.

    strict single-crossing differences (SSCD) if ∀a,a′∈A such that a≠a′, Da,a′⁢(θ) is strictly single crossing in θ.

Our definition of (S)SCD is related to but different from Milgrom and Shannon (1994), who stipulate that u:A×Θ→ℝ has (strict) single-crossing differences given an order ⪰ on A if for all a′≻a′′, Da′,a′′⁢(θ) is (strictly) single crossing from below. (Here, ≻ is the strict component of ⪰. Note that the (S)SCD terminology is due to Milgrom (2004).) We do not presume that A is ordered, but we consider differences for all pairs of elements of A. For an arbitrary complete order on A, our definition is weaker than Milgrom and Shannon’s (1994) because ours does not constrain the direction of single crossing; on the other hand, it can be checked that our definition implies theirs for a suitably chosen order on A that is largely complete.888The order, and the sense in which it is largely complete, is elaborated around Theorem 4 in Subsection 5.2. As established below, our notion characterizes related but distinct comparative statics from Milgrom and Shannon’s.

Our notion of SCD is also closely related to Rothstein’s (1990) notion of order restriction, which he introduced in the context of collective choice. Like SCD, order restriction does not presume an order on the choice space. Order restriction is more permissive insofar as the order on the type space can be chosen to generate SCD; it is more restrictive in requiring that order to be complete.

Interval Choice.

We say that Θ0⊆Θ is an interval if θl,θh∈Θ0 and θl<θm<θh imply θm∈Θ0. Let C:2A×Θ⇉A with C⁢(S,θ)⊆S for each S⊆A and θ∈Θ. We say that C has interval choice if {θ:a∈C⁢(S,θ)} is an interval for each S⊆A and a∈S. That is, interpreting C as a choice correspondence, the set of types choosing any alternative given any choice set is an interval. We say that u:A×Θ→ℝ strictly violates SCD if there are a,a′∈A and θl<θm<θh such that min⁡{Da,a′⁢(θl),Da,a′⁢(θh)}>0>Da,a′⁢(θm).

Theorem 1.

Let u:A×Θ→ℝ and Cu⁢(S,θ)≡arg⁢maxa∈S⁡u⁢(a,θ) for any S⊆A and θ.

  1. 1.

    If u has SCD, then the choice correspondence Cu has interval choice. If u strictly violates SCD, then Cu does not have interval choice.

  2. 2.

    If |Θ|≥3, then u has SSCD if and only if every selection from Cu has interval choice.

The intuition for the sufficiency of (S)SCD in Theorem 1 is straightforward. Regarding necessity, we note that a violation of SCD—as opposed to a strict violation—is compatible with the choice correspondence having interval choice: e.g., A={a′,a′′}, Θ={θl,θm,θh} with θl<θm<θh, and min⁡{Da′,a′′⁢(θl),Da′,a′′⁢(θh)}>0=Da′,a′′⁢(θm).999On the other hand, a strict violation of SCD is slightly stronger than needed: one could weaken its requirement to min⁡{Da′,a′′⁢(θl),Da′,a′′⁢(θh)}≥0>Da′,a′′⁢(θm). Our formulation with both inequalities being strict amounts to putting aside indifferences, which proves convenient for the applications in Section 4. In Part 2 of the theorem, if |Θ|=2 then any selection from any choice correspondence trivially has interval choice, yet u does not have SSCD when Da,a′⁢(θ)=0 for some a,a′ and all θ.

Monotone Comparative Statics.

Our choice space A is unordered. Intuitively, interval choice is intimately related to there being monotone comparative statics (MCS)—i.e., in some sense, higher types make higher choices—with respect to some complete order on the choice space. We formalize this connection in Subsection 5.2 by tying (S)SCD to MCS. In brief, given any order on A, we order subsets of A by the corresponding strong set order, and say that the function u has MCS if, for every choice set S⊆A and all types θh>θl, the higher type chooses a higher set: Cu⁢(S,θh)≥Cu⁢(S,θl). Roughly speaking, Theorem 4 in Subsection 5.2 shows that an order on A induces MCS if and only if u has SCD and the order is a refinement of a natural “SCD-order” generated by u.

3 Single-Crossing Differences in Convex Environments

This section characterizes single-crossing differences in “rich” environments. We now assume the existence of a strictly increasing real-valued function on (Θ,≤).101010That is, we assume ∃h:Θ→ℝ such that θ¯<θ¯⟹h⁢(θ¯)<h⁢(θ¯). This requirement is related to utility representations for possibly incomplete preferences (Ok, 2007, Chapter B.4.3). A sufficient condition is that Θ has a countable order dense subset, i.e., there is a countable set Θ0⊆Θ such that (∀θ¯,θ¯∈Θ∖Θ0)θ¯<θ¯⟹∃θ0∈Θ0⁢ s.t. ⁢θ¯<θ0<θ¯ (Jaffray, 1975, Corollary 1). The assumption only plays a technical role in establishing our characterization of strict SCD, i.e., in the second statement of Theorem 2. This requirement is satisfied, for example, when Θ is finite, or Θ⊆ℝn is endowed with the usual order. We assume the environment (A,Θ,u) is convex in the following sense:

 the set of functions ⁢{u⁢(a,⋅):Θ→ℝ}a∈A⁢ is convex. (⋆)

That is, a convex environment is rich enough insofar as for any pair of actions and any weighting, there is a third action that replicates the weighted sum of utilities of the original actions. We stress that convexity is in terms of utilities: we have not assumed any structure on A. However, if A is convex, then it is straightforward that (⋆) is assured by linearity of u in its first argument.

Example 1 (Expected Utility).

Consider an expected-utility agent who chooses among lotteries. There is a set of consequences X and the agent has utility v⁢(x,θ). Letting A≡Δ⁢X be the set of all finite-support lotteries over X, the agent’s utility from lottery P∈A is given by u⁢(P,θ)≡∫xv⁢(x,θ)⁢dP.111111We restrict attention to finite-support distributions throughout the paper for ease of exposition, as it guarantees that expected utility is well defined no matter the distribution and utility function. Nevertheless, we write integrals rather than summations when it simplifies notation regarding the domain of integration/summation. This is a case in which Condition (⋆) holds because u is linear in its first argument and A is convex.

Example 2 (Rank-Dependent Expected Utility).

Continuing with lotteries, non-expected utility environments can also be convex. By virtually the same logic as above for expected utility, it is sufficient for (⋆) that the utility from lottery P be given by u⁢(P,θ)≡∫xv⁢(x,θ)⁢d⁢(w∘P), with w:Δ⁢X→Δ⁢X an arbitrary reweighting function whose image is convex. For example, a standard formulation of rank-dependent utility (Quiggin, 1982; Diecidue and Wakker, 2001) corresponds to X≡{x1,…,xn}⊂ℝ with x1<⋯<xn, and a strictly increasing function w^:[0,1]→[0,1] satisfying w^⁢(0)=0 and w^⁢(1)=1 such that for any lottery P and consequence xi, the reweighting is given by (w∘P)⁢(xi)≡w^⁢(∑j=1ip⁢(xj))−w^⁢(∑j=1i−1p⁢(xj)), where p is the probability mass function of the lottery P. So long as w^ is continuous, w has a convex image as required; the image is simply Δ⁢X.

Example 3 (Multidimensional Utility).

While lotteries naturally induce a convex environment, they are not necessary. Another example is choice among multidimensional actions. Specifically, A≡A1×…×An⊆ℝn and for any 𝐚≡(a1,…,an), utility is given by u⁢(𝐚,θ)≡∑i=1ngi⁢(ai)⁢fi⁢(θ) for some pairs of functions (gi,fi)i=1n, with each gi having a convex image. That A is a product set and each gi has a convex image ensures (⋆). We refer to this specification as multidimensional utility.

Here are some economic contexts in which there is multidimensional utility. First, a consumer chooses among products with multiple characteristics or a bundle of multiple goods, denoted (a1,…,an). Each characteristic or good i with quality or quantity ai has a common value gi⁢(ai), but the tradeoff across characteristics/goods varies with the consumer’s preference parameter θ, as given by fi⁢(θ). Second, a designer uses an incentive-compatible direct mechanism φ:T→A that maps an agent’s private type t∈T≡{1,…,n} to an action φ⁢(t)∈A. The designer’s payoff is ∑tv⁢(φ⁢(t),t)⁢f⁢(t;θ), where v⁢(a,t) is the designer’s utility from allocating a to t, and f⁢(⋅;θ)∈Δ⁢T is a type distribution that depends on some parameter θ. Third, an agent chooses consumption ct in each period t∈T≡{1,…,n}. The agent’s present discounted value is ∑tv⁢(ct)⁢ρ⁢(t;θ), where ρ⁢(t;θ) is discount function parameterized by θ.

Example 4 (Experiments).

Our final example is one in which an expected-utility agent can only choose from a proper subset of lotteries. Let Ω be a finite set of “states”. We refer to Δ⁢Ω as the set of beliefs or posteriors and Δ⁢Δ⁢Ω as the set of experiments (with finite support). In Bayesian persuasion (Kamenica and Gentzkow, 2011) or, more broadly, information design (Bergemann and Morris, 2019), an expected-utility agent has preferences represented by v⁢(p,θ), where p∈Δ⁢Ω and θ is a preference parameter. The expected utility from experiment Q∈Δ⁢Δ⁢Ω is given by u⁢(Q,θ)≡∫pv⁢(p,θ)⁢dQ. If we consider all experiments, then this setting is a special case of Example 1. But given a prior p∗∈Δ⁢Ω, any experiment must in fact be Bayes-plausible, i.e., its distribution of posteriors must average to the prior p∗. So, given p∗, the agent can only choose an experiment in A≡{Q∈Δ⁢Δ⁢Ω:∫p∈Δ⁢Ωp⁢dQ=p∗}. Nevertheless, the linearity of u in its first argument and the convexity of A again imply (⋆). Indeed, based on Example 2, Condition (⋆) will hold even for rank-dependent expected-utility agents choosing among Bayes-plausible experiments.

3.1 The Characterization

Our characterization of SCD in convex environments (Theorem 2 below) requires the following definition.

Definition 3.

Let f1,f2:Θ→ℝ each be single crossing.

  1. 1.

    f1 ratio dominates f2 if

    (∀θl<θh) f1⁢(θl)⁢f2⁢(θh)≤f1⁢(θh)⁢f2⁢(θl),and (3)
    (∀θl<θm<θh) f1⁢(θl)⁢f2⁢(θh)=f1⁢(θh)⁢f2⁢(θl)⇔{f1⁢(θl)⁢f2⁢(θm)=f1⁢(θm)⁢f2⁢(θl),f1⁢(θm)⁢f2⁢(θh)=f1⁢(θh)⁢f2⁢(θm). (4)
  2. 2.

    f1 strictly ratio dominates f2 if Condition (3) holds with strict inequality.

  3. 3.

    f1 and f2 are (strictly) ratio ordered if either f1 (strictly) ratio dominates f2 or vice-versa.

Condition (3) contains the essential idea of ratio dominance and is what we focus on in the main text; Condition (4) only deals with some knife-edged cases that are discussed in Appendix B.1. The definition of strict ratio dominance does not make reference to Condition (4) because that condition is vacuous when Condition (3) holds with strict inequality.

We use the terminology “ratio dominance” because when f2 is a strictly positive function, Condition (3) is the requirement that the ratio f1⁢(θ)/f2⁢(θ) must be increasing in θ. Indeed, if both f1 and f2 are probability densities of random variables Y1 and Y2, then (3) says that Y1 stochastically dominates Y2 in the sense of likelihood ratios.121212From the viewpoint of information economics, think of θ as a signal of a state s∈{1,2}, drawn from the density f⁢(θ|s)≡fs⁢(θ). Condition (3) is Milgrom’s (1981) monotone likelihood-ratio property for f⁢(θ|s).

Condition (3) is a natural generalization of the increasing ratio property to functions that may change sign. To get a geometric intuition, suppose f1 strictly ratio dominates f2. Let 𝒇⁢(θ)≡(f1⁢(θ),f2⁢(θ)). For every θl<θh, f1⁢(θl)⁢f2⁢(θh)−f1⁢(θh)⁢f2⁢(θl)<0 implies that the vector 𝒇⁢(θl) moves to 𝒇⁢(θh) through a rescaling of magnitude and a clockwise—rather than counterclockwise—rotation (throughout our paper, a “rotation” must be no more than 180 degrees).131313To elaborate on the direction of rotation, recall that from the definition of cross product, (f1⁢(θl),f2⁢(θl),0)×(f1⁢(θh),f2⁢(θh),0)=∥𝒇⁢(θl)∥⁢∥𝒇⁢(θh)∥⁢sin⁡(r)⁢e3=(f1⁢(θl)⁢f2⁢(θh)−f1⁢(θh)⁢f2⁢(θl))⁢e3, where r is the counterclockwise angle from 𝒇⁢(θl) to 𝒇⁢(θh), e3≡(0,0,1), × is the cross product, and ∥⋅∥ is the Euclidean norm. If sin⁡(r)<0 (resp., sin⁡(r)>0), then 𝒇⁢(θl) moves to 𝒇⁢(θh) through a clockwise (resp., counterclockwise) rotation.

Refer to caption
(a) Condition (3) holds for θl<θm<θh.
Refer to caption
(b) Condition (3) fails both when Θ={θl,θm,θh} with θl<θm<θh, and when Θ={θl,θm,θ¯h} with θl<θm<θ¯h.
Figure 1: Geometric representation of Condition (3).

Hence, f1 and f2 are ratio ordered only if 𝒇⁢(θ) rotates monotonically as θ increases, either always clockwise or always counterclockwise.141414The preceding discussion establishes this point under the presumption of strict ratio ordering; however, because of the hypothesis in Definition 3 that f1 and f2 are single crossing and because of Condition (4), the conclusion holds for ratio ordering too. Furthermore, it can be confirmed that a monotonic rotation of 𝒇⁢(⋅) implies ratio ordering if there are no θ′ and θ′′ such that 𝒇⁢(θ′) and 𝒇⁢(θ′′) are collinear. It follows that the set {𝒇⁢(θ):θ∈Θ} must be contained in a closed half-space of ℝ2 defined by a hyperplane that passes through the origin: otherwise, there will be two pairs of vectors such that an increase in θ corresponds to a clockwise rotation in one pair and a counterclockwise rotation in the other. See Figure 1. In its panel 1(b), the non-monotonic rotation of 𝒇⁢(θ) is clear when Θ={θl,θm,θh} with θl<θm<θh, but it is also present when Θ={θl,θm,θ¯h} with θl<θm<θ¯h; in the latter case, the aforementioned half-space requirement is violated.

We can now state our main characterization. For brevity, we say that u has (S)SCD⋆ as shorthand for “In a convex environment, u has (S)SCD.” Furthermore, we say that A is minimal if there is no pair of utility-indistinguishable actions: ∀a,a′∈A, ∃θ such that Da,a′⁢(θ)≡u⁢(a,θ)−u⁢(a′,θ)≠0.

Theorem 2.

The function u:A×Θ→ℝ has SCD⋆ if and only if it takes the form

u⁢(a,θ)=g1⁢(a)⁢f1⁢(θ)+g2⁢(a)⁢f2⁢(θ)+h⁢(θ), (5)

with f1 and f2 each single crossing and ratio ordered. In addition, if A is minimal, then u has SSCD⋆ if and only if f1 and f2 are strictly ratio ordered.151515If A is not minimal, then so long as |Θ|>1, u violates SSCD⋆ because for some a and a’, Da,a′ is a zero function and hence not strictly single crossing. Nonetheless, the characterization applies when we consider utility-indistinguishable actions as equivalence classes and the corresponding utility function on those equivalence classes.

We make a number of observations to help interpret Theorem 2.

The theorem says that for u to have SCD⋆, it must be possible to write it in the form (5). Notice that given (5), for any a0,a1∈A, the function u⁢(a1,⋅)−u⁢(a0,⋅) is a linear combination of f1⁢(⋅) and f2⁢(⋅). Therefore, to rule out the possibility of the form (5), it is sufficient to find a0,a1,a2,a3∈A and θl<θm<θh such that the 3×3 matrix M≡[u⁢(ai,θj)−u⁢(a0,θj)]i∈{1,2,3},j∈{l,m,h} is invertible. This procedure is often useful to reject SCD⋆, as we illustrate subsequently in Corollary 2.

Given the functional form (5), not only is SCD⋆ assured by f1 and f2 each being single crossing and ratio ordered, but these properties are almost necessary.161616“Almost” excludes the case in which g1 and g2 are affinely dependent, i.e., g1=λ⁢g2+γ for some λ,γ∈ℝ. Intuitively, affine independence ensures that neither g1 nor g2 is dispensable in (5).

An asymmetry between a and θ in Equation 5 bears noting: the function h:Θ→ℝ does not have a counterpart function A↦ℝ. The reason is that whether the utility difference between two actions is single crossing could be altered by adding a function of a alone to u⁢(a,θ). On the other hand, adding a function of θ alone has no such effect because SCD is an ordinal property that is invariant to any (type-dependent) increasing transformation, or representation, of u, i.e., u~⁢(a,θ)≡m⁢(u⁢(a,θ),θ), where each m⁢(⋅,θ):ℝ→ℝ is strictly increasing.

In general, whether the convexity condition (⋆) holds can depend on which representation one chooses. But since SCD is ordinal, the scope of our analysis is in fact broader than it may seem: if a utility function satisfies SCD and some representation satisfies (⋆), then Theorem 2 applies to that representation.171717For example, given a product set A⊆ℝ+2 and θ∈[0,1], the Cobb-Douglas utility a1θ⁢a21−θ does not satisfy (⋆), but the representation θ⁢log⁡a1+(1−θ)⁢log⁡a2 is a multidimensional utility (Example 3) and therefore satisfies (⋆), indeed SCD⋆. We illustrate how this is useful in Subsection 3.2.3.

If u⁢(a,θ) has the form (5) with strictly positive functions f1 and f2, then up to a positive affine transformation (viz., subtracting h⁢(θ) and dividing by f1⁢(θ)+f2⁢(θ)),181818In general, a positive affine transformation of u⁢(a,θ) is any b⁢(θ)⁢u⁢(a,θ)+c⁢(θ) where b⁢(⋅)>0. Unlike arbitrary representations, positive affine transformations preserve condition (⋆). any type’s utility becomes a convex combination of two type-independent utility functions over actions, g1 and g2. Theorem 2’s ratio ordering requirement then simply says that the relative weight on g1 and g2 changes monotonically with the agent’s type. This idea underlies the following proposition.

Proposition 1.

Let Θ have both a minimum and a maximum (i.e., ∃θ¯,θ¯∈Θ such that (∀θ) θ¯≤θ≤θ¯), and the environment (A,Θ,u) be convex. Then, u has SCD⋆ if and only if u has a positive affine transformation u~ satisfying

u~⁢(a,θ)=λ⁢(θ)⁢u~⁢(a,θ¯)+(1−λ⁢(θ))⁢u~⁢(a,θ¯), (6)

with λ:Θ→[0,1] increasing. In addition, if A is minimal, then u has SSCD⋆ if and only if λ is strictly increasing.

Proposition 1 provides an economic interpretation of SCD⋆ as capturing preferences that monotonically shift weight from one extreme type’s to the other’s. We note, though, that even when Θ has extreme types, it could be easier to verify whether a given function has SCD⋆ using Theorem 2 because one does not have to search among the affine transformations allowed by Proposition 1.

We offer one additional interpretational comment. On the one hand, Theorem 2 and Proposition 1 indicate that SCD—while desirable for tractability (including interval choice), economic intuition, etc.—is a demanding property in a convex environment. On the other hand, because the functions g1 and g2 in Theorem 2 are arbitrary, SCD⋆ nevertheless allows for a broad economic landscape. Specifically, when one assumes SCD in any environment (convex or not), one generally has in mind that there is an underlying tradeoff—e.g., risk vs. expected return among lotteries, delay vs. total amount in payment streams, or price vs. product quality in markets—whose balance shifts monotonically with type. Our characterizations show that SCD⋆ is broad enough to capture such desiderata because the g1 and g2 functions in Theorem 2 can evaluate the tradeoff differently: e.g., compared to g2, the function g1 can be more risk averse, discount the future more, or be more price sensitive; furthermore, as highlighted by Proposition 1, higher types put more weight on one criterion.

3.2 Implications for Leading Examples

This subsection develops the implications of Theorem 2 for our leading examples. For brevity, we focus on the implications of SCD⋆, stating the SSCD⋆ counterpart only in Corollary 1.

3.2.1 Single-Crossing Expectational Differences

Suppose, following Example 1, that A≡Δ⁢X is a set of lotteries and u is the expected utility induced by v⁢(x,θ). We say that v has (strict) single-crossing expectational differences, or (S)SCED, if the expected utility function u has (S)SCD⋆.191919In general, an expected utility function u may not have SSCD⋆ simply because A≡Δ⁢X is not minimal. This arises, for example, when u is a mean-variance utility function and multiple lotteries have the same mean and variance. In such cases we consider utility-indistinguishable lotteries as equivalence classes and the corresponding utility function u~ defined on these equivalence classes of lotteries. We say v has SSCED if u~ has SSCD⋆. SCED is not implied by SCD or even supermodularity of v.202020For instance, given x,θ∈ℝ, any power loss function v⁢(x,θ)=−|x−θ|z is supermodular when z>1, but Corollary 2 below establishes that SCED fails for z≠2. Rather:

Corollary 1.

The von Neumann Morgenstern utility function v has (S)SCED if and only if it has the same form as u in Theorem 2.

The corollary’s proof is straightforward: if v satisfies the characterization, then u⁢(P,θ)=(∫xg1⁢(x)⁢dP)⁢f1⁢(θ)+(∫xg2⁢(x)⁢dP)⁢f2⁢(θ)+h⁢(θ) with f1 and f2 satisfying the conditions given in Theorem 2, so u has SCD⋆. Conversely, if u⁢(P,θ)≡∫xv⁢(x,θ)⁢dP has SCD⋆, and hence has the form in Theorem 2, then so does v because v⁢(x,θ)≡u⁢(δx,θ)=g1⁢(x)⁢f1⁢(θ)+g2⁢(x)⁢f2⁢(θ)+h⁢(θ), where δx denotes the degenerate lottery on x.

Corollary 1 is related to Abbas and Bell (2011, Theorem 1). They study expected utility over lotteries with some additional restrictions on the environment (e.g., Θ is finite and completely ordered, and preferences satisfy some substantive economic conditions). For that setting, their Theorem 1 states a similar result to the version of Corollary 1 that would obtain using Proposition 1 instead of Theorem 2.212121Abbas and Bell (2011) use the terminology of “one-switch independence”, which appears equivalent to SSCED up to one minor detail that can be set aside here (concerning the treatment of distinct lotteries that all types are indifferent over). Translated into our notation, they assert that a utility function v⁢(x,θ) has SSCED if and only if v⁢(x,θ)=f1⁢(θ)⁢g1⁢(x)+f2⁢(θ)⁢g2⁢(x)+h⁢(θ), with f1 strictly positive and f2/f1 strictly increasing. This can be seen from our Proposition 1 because, given its hypothesis of extreme types, it implies that v has SSCED if and only if v(x,θ)=b(θ)[g1(x)+λ(θ)(g2(x)−g1(x)]+c(θ), where b is strictly positive, λ is strictly increasing, and g1 and g2 are, respectively, monotonic transformations of v⁢(⋅,θ¯) and v⁢(⋅,θ¯).

Even aside from the ratio-ordering and single-crossing requirements in Theorem 2, the functional form (5) deserves emphasis: there are only two “interaction terms”, each of which is multiplicatively separable in a and θ. This means that preferences over lotteries must be summarized by two linear statistics: for any lottery P∈Δ⁢X, the statistics are ∫xg1⁢(x)⁢dP and ∫xg2⁢(x)⁢dP.222222The utility function v⁢(x,θ)=exp⁡(x⁢θ) cannot be written in the form of Equation 5 and hence does not have SCED. Indeed, its expected utility from an arbitrary lottery cannot be summarized by any finite number of linear statistics, let alone two. This point underlies the following corollary, which identifies quadratic loss as the unique power loss function that has SCED.

Corollary 2.

Let X=ℝ and Θ⊆ℝ with |Θ|≥3. A loss function v⁢(x,θ)=−|x−θ|z with z>0 has SCED if and only if z=2.

Under quadratic loss, preferences over lotteries are summarized by the lotteries’ first and second moments. We note that some non-power-loss generalizations of quadratic loss, such as v⁢(x,θ)=x⁢θ+g⁢(x)+h⁢(θ) with g:ℝ→ℝ, also satisfy SCED; these functional forms—or variants that also augment a quasilinear money burning component, which continues to preserve SCED—are used in the study of delegation with stochastic mechanisms or money burning (Amador and Bagwell, 2013; Kleiner, 2022).

SCED is useful in dynamic problems, as seen in Banks and Duggan (2006), Duggan (2014), Celik (2015), and Ali et al. (2022). It is thus noteworthy that:

Corollary 3.

Suppose v⁢(x,θ) has SCED. Then, denoting 𝐱∞≡(xt)t=0∞, so does the discounted utility function v~⁢(𝐱∞,θ)=∑t=0∞ρ⁢(t)⁢v⁢(xt,θ) for any ρ:{0,1,…}→ℝ.

We omit a proof as the result follows straightforwardly from Corollary 1. Note that Corollary 3 does not require exponential discounting. (Of course, implicitly ρ must ensure that v~⁢(⋅) is finite.)

3.2.2 Single-Crossing Rank-Dependent Expected Utility

Suppose an agent’s preferences over lotteries A≡Δ⁢X have a rank-dependent expected utility (RDEU) representation as described in Example 2, with underlying utility v⁢(x,θ).

Corollary 4.

An RDEU function has SCD⋆ if and only if the underlying utility v has the same form as u in Theorem 2.

We omit a proof because it is analogous to that for Corollary 1. An RDEU agent thus evaluates a lottery P according to two summary statistics: ∫xg1⁢(x)⁢d⁢(w∘P) and ∫xg2⁢(x)⁢d⁢(w∘P). The difference with expected utility is that these summary statistics are no longer linear in P. Instead the statistics reweight probabilities according to the original reweighting function.

3.2.3 Multidimensional Utility

Suppose, following Example 3, that an agent has a multidimensional utility function u⁢(𝒂,θ)≡∑i=1ngi⁢(ai)⁢fi⁢(θ).

Corollary 5.

A multidimensional utility function u has SCD⋆  if and only if

u⁢(𝒂,θ)=(∑i=1nλiI⁢gi⁢(ai))⁢fI⁢(θ)+(∑i=1nλiI⁢I⁢gi⁢(ai))⁢fI⁢I⁢(θ)+h⁢(θ), (7)

with fI and fI⁢I each single crossing and ratio ordered, and 𝛌I,𝛌I⁢I∈ℝn.

The interpretation is that a multidimensional utility has SCD⋆ when at most two “summary” dimensions matter. The values on these summary dimensions are weighted sums of the original values gi⁢(ai) over the primitive dimensions i=1,…,n. Higher types place relatively more weight on products/bundles that are more valuable on one of the two summary dimensions; recall Proposition 1.

As mentioned in Example 3, multidimensional utility can capture a consumer choosing among consumption bundles. A canonical specification is the Cobb-Douglas utility u⁢(𝒂,θ)=∏i=1naifi⁢(θ), where 𝒂∈ℝ+n is the consumption bundle and the vector of fi⁢(θ)≥0 parameterizes the consumer’s marginal rates of substitution (MRS). Note that an alternative representation is ∑i=1nfi⁢(θ)⁢log⁡(ai), which has the multidimensional form. Corollary 5 implies that for Cobb-Douglas utility to have SCD, it must be representable as

u⁢(𝒂,θ)=(∏i=1naiλiI)f⁢(θ)⁢(∏i=1naiλiI⁢I)1−f⁢(θ),

with f:Θ→[0,1] monotonic and 𝝀I,𝝀I⁢I≥0.232323More precisely, Corollary 5 yields the representation (∏i=1naiλiI)fI⁢(θ)⁢(∏i=1naiλiI⁢I)fI⁢I⁢(θ), where fI and fI⁢I are each single crossing and ratio ordered, and Proposition 1 yields the further simplification. The interpretation is that to satisfy SCD, the Cobb-Douglas utility must have “two layers”: the consumer first evaluates the θ-independent Cobb-Douglas value of two composite goods, and then trades off these composite goods according to a Cobb-Douglas utility function with MRS that is monotonic in θ. SCD guarantees that for any choice set (e.g., a budget set), the value of each composite good in the chosen consumption bundle changes monotonically (necessarily in opposite directions) in θ. A special case is the textbook example of two goods, say 1 and 2, and u⁢(𝒂,θ)=a1θ⁢a21−θ with θ∈[0,1]. In that case, we recover the textbook observation that given any budget set, the consumption of each good is monotonic in the MRS.

As another example, consider discounted utility over consumption streams. For a consumption stream 𝒄≡(ct)t=1T, the discounted utility ∑t=1Tct⁢ρ⁢(t,θ) is of the multidimensional form. Corollary 5 implies that the discounted utility has SCD⋆ only if, for any consumption stream 𝒄, ∑tct⁢ρ⁢(t,θ) is linearly generated by two functions of θ. By considering consumption streams that are positive only in a single period, we see that each ρ⁢(t,⋅) must in fact be generated by the same two functions of θ, i.e., ρ⁢(t,θ)=λtI⁢fI⁢(θ)+λtI⁢I⁢fI⁢I⁢(θ). When T≥3, such linear dependency does not hold for exponential discounting, i.e., when each type θ has a discount factor δθ such that ρ⁢(t,θ)=(δθ)t. Consequently, exponential discounting is incompatible with interval choice: given any three discount factors, there are consumption streams 𝒄 and 𝒄′ such that an agent with either low or high patience strictly prefers 𝒄 while an agent with intermediate patience strictly prefers 𝒄′.242424For example, let (ct)t=13=(1,0,6) and (ct′)t=13=(0,5,0). The difference u⁢(𝒄′,θ)−u⁢(𝒄,θ)=δθ⁢(1−5⁢δθ+6⁢δθ2) is strictly positive if and only if δθ∈(1/3,1/2). In fact, our analysis reveals that in general choice between an arbitrary pair of consumption streams will be monotonic in the time preference parameter θ only if ρ⁢(t,θ) has the aforementioned linear dependence. An example is linear time cost, say θ∈[θ¯,θ¯]⊂ℝ and ρ⁢(t,θ)=α−θ⁢t for some α>θ¯⁢T.252525The fact that SCD imposes strong restrictions on the nature of discounting can be related to difficulties in aggregating time preferences (Jackson and Yariv, 2015). Specifically, with three types (or agents), θ¯<P<θ¯, where “P” stands for Planner, SCD is very similar to a Pareto principle: whenever types θ¯ and θ¯ have the same preference over a pair of consumption streams, so should type P. Our discussion indicates that if types θ¯ and θ¯ are exponential discounters, then P cannot be, as was highlighted by Jackson and Yariv (2015). In fact, our result says that P’s preference must linearly depend on the other two types, echoing Harsanyi (1955).

3.2.4 Single-Crossing Expectational Differences over Experiments

Suppose, following Example 4, that Ω, Δ⁢Ω, and Δ⁢Δ⁢Ω are a set of states, posteriors, and experiments respectively. Given a prior p∗∈Δ⁢Ω, an agent can only choose a Bayes-plausible experiment: A≡{Q∈Δ⁢Δ⁢Ω:∫p∈Δ⁢Ωp⁢dQ=p∗}. We say that v⁢(p,θ) has single-crossing expectational differences over experiments (SCED-X) if the expected utility function u⁢(Q,θ)≡∫pv⁢(p,θ)⁢dQ over A has SCD⋆. Since A⊊Δ⁢Δ⁢Ω, SCED is sufficient for SCED-X but not necessary. SCED-X is, instead, characterized as follows.

Corollary 6.

For any full-support prior p∗, the function v has SCED-X if and only if

v⁢(p,θ)=g1⁢(p)⁢f1⁢(θ)+g2⁢(p)⁢f2⁢(θ)+∑ωv⁢(δω,θ)⁢p⁢(ω), (8)

with f1 and f2 each single crossing and ratio ordered.

As the environment is convex, Theorem 2’s characterization applies to the expected utility function u. Similar to the SCED characterization in Corollary 1, the form of the expected utility function u passes to the von Neumann Morgenstern utility function v.262626This is not as straightforward as in the case SCED, where one can appeal to degenerate lotteries on any posterior. Under SCED-X, the only posterior on which there can be a degenerate lottery is the prior. The difference is that the last term in Equation 8 can depend on the posterior p as well as the type θ, unlike the h⁢(θ) term in Equation 5. This is because the expected utility having the form in Theorem 2 only imposes, for SCED-X, that the expectation of that term is constant over lotteries that average to the prior, rather than over all lotteries. This means that the term can admit a linear dependence in p, as seen in Equation 8.

3.3 Aggregating Single-Crossing Functions

This subsection explains the logic behind Theorem 2; we focus on single crossing here, deferring strict single crossing to Supplementary Appendix SA.2. Owing to the convex environment, the central issue is when linear aggregations of functions are single crossing. Lemma 1 below shows that ratio ordering is the characterizing property when aggregating two functions; Proposition 2 then establishes that when aggregating more than two functions, no more than two can be linearly independent, which leads to Theorem 2.272727Real-valued functions f1,…,fn are linearly independent if (∀𝝀∈ℝn\{0}) ∑i=1nλi⁢fi is not a zero function, i.e., is not everywhere zero.

Lemma 1.

Let f1,f2:Θ→ℝ. The linear combination α1⁢f1⁢(θ)+α2⁢f2⁢(θ) is single crossing ∀𝛂∈ℝ2 if and only if f1 and f2 are each single crossing and ratio ordered.

Refer to caption
(a) Sufficiency of ratio ordering.
Refer to caption
(b) Necessity of ratio ordering, with θl<θm<θh.
Figure 2: Ratio ordering and single crossing of all linear combinations.

Here is the lemma’s intuition. For sufficiency, consider any linear combination α1⁢f1+α2⁢f2. Assume 𝜶∈ℝ2\{0}, to avoid triviality. The vector α defines two open half spaces ℝ𝜶,−2≡{𝒙∈ℝ2:𝜶⋅𝒙<0} and ℝ𝜶,+2≡{x∈ℝ2:𝜶⋅𝒙>0}, where ⋅ is the dot product; see Figure 22(a). As explained earlier after Definition 3, ratio ordering of f1 and f2 implies that the vector 𝒇⁢(θ)≡(f1⁢(θ),f2⁢(θ)) rotates monotonically as θ increases. If the rotation is from ℝ𝜶,−2 to ℝ𝜶,+2 (resp., from ℝ𝜶,+2 to ℝ𝜶,−2), then 𝜶⋅𝒇≡α1⁢f1+α2⁢f2 is single crossing only from below (resp., only from above). If ⋃θ∈Θ𝒇⁢(θ)⊆ℝ𝜶,−2 or ⋃θ∈Θ𝒇⁢(θ)⊆ℝ𝜶,+2, then 𝜶⋅𝒇 is single crossing both from below and above. Other cases are similar.282828Notice that this argument does not require either f1 or f2 to be single signed. By contrast, Abbas and Bell (2011, p. 769, in their last paragraph on “Necessity”) incorrectly claim that all linear combinations of two (strictly) single crossing functions are (strictly) single crossing only if one function is single signed.

To see why Condition (3) of ratio ordering is necessary, suppose the vector 𝒇⁢(θ) does not rotate monotonically. Figure 22(b) illustrates a case in which, for θl<θm<θh, 𝒇⁢(θl) rotates counterclockwise to 𝒇⁢(θm), but 𝒇⁢(θm) rotates clockwise to 𝒇⁢(θh). As shown in the figure, one can find 𝜶∈ℝ2 such that 𝒇⁢(θm)∈ℝ𝜶,−2 while both 𝒇⁢(θl),𝒇⁢(θh)∈ℝ𝜶,+2, which implies that 𝜶⋅𝒇 is not single crossing. See Appendix B.1 for why Condition (4) is necessary.

Lemma 1 relates to Quah and Strulovici (2012, Proposition 1). They establish that for any two functions f1 and f2 that are single crossing from below, α1⁢f1+α2⁢f2 is single crossing from below for all 𝜶∈ℝ+2 if and only if f1 and f2 satisfy a condition they call signed-ratio monotonicity (see Supplementary Appendix SA.5 for the definition). In general, ratio ordering is not comparable with signed-ratio monotonicity because we consider a different aggregation problem from Quah and Strulovici: (i) the input functions may be single crossing in either direction; (ii) the linear combinations involve coefficients of arbitrary sign; and (iii) the resulting combination can be single crossing in either direction. The example in the introduction highlights the importance of point (ii). There, θ∈[−1,1] and u⁢(a,θ)=(θ2+1/2)⁢11{a=1}+211{a=2}. Both f1⁢(θ)=θ2+1/2 and f2⁢(θ)=2 are positive functions (hence, single crossing from below), and so all positive linear combinations are also positive functions, but f1⁢(θ)−(1/2)⁢f2⁢(θ)=θ2−1/2 is not single crossing because f1 and f2 are not ratio ordered. If the input functions in Lemma 1 are restricted to be single crossing from below, then ratio ordering implies signed-ratio monotonicity.

Lemma 1 implies a characterization of likelihood-ratio ordering for random variables with strictly positive densities. While this likelihood-ratio ordering characterization does not appear to be well-known among economists, it is a special case of Karlin’s (1968) results on the variation diminishing property of totally positive functions. More generally, however, Lemma 1 differs from Karlin (1968) because it considers aggregations of functions that can change sign, whereas he only studies non-negative functions.292929Under some conditions, it is possible to transform the lemma’s aggregation problem into one of aggregating non-negative functions, and then apply Karlin’s (1968) results. However, the precise conditions are opaque (Kartik, Lee, and Rappoport, 2019, Appendix H.2); moreover, the procedure is somewhat involved and, in our view, less instructive than our direct proof of Lemma 1.

Theorem 2 requires an extension of Lemma 1 to more than two functions. Consider any set Z and f:Z×Θ→ℝ. We say that f is linear combinations SC-preserving if ∫zf⁢(z,θ)⁢dμ is single crossing in θ for every function μ:Z→ℝ with finite support.303030The notation ∫zf⁢(z,θ)⁢dμ should be read as ∑{z:μ⁢(z)≠0}f⁢(z,θ)⁢μ⁢(z).

Proposition 2.

Let f:Z×Θ→ℝ for some set Z. The function f is linear combinations SC-preserving if and only if there exist z1,z2∈Z and λ1,λ2:Z→ℝ such that

  1. 1.

    f⁢(z1,⋅):Θ→ℝ and f⁢(z2,⋅):Θ→ℝ are (i) each single crossing and (ii) ratio ordered, and

  2. 2.

    (∀z)⁢f⁢(z,⋅)=λ1⁢(z)⁢f⁢(z1,⋅)+λ2⁢(z)⁢f⁢(z2,⋅).

Proposition 2 says that a family of single-crossing functions {f⁢(z,⋅)}z∈Z preserves single crossing of all finite linear combinations if and only if the family is “linearly generated” by two single-crossing functions that are ratio ordered. In particular, given any three single-crossing functions, f1, f2, and f3, all their linear combinations will be single crossing if and only if there is a linear dependence in the triple, say λ1⁢f1+λ2⁢f2=f3 for some 𝝀∈ℝ2, and f1 and f2 are ratio ordered.

The sufficiency direction of Proposition 2 follows from Lemma 1, as does necessity of the “generating functions” being ratio ordered. The intuition for the necessity of linear dependence is as follows. Assume Θ is completely ordered. For any θ, let 𝒇⁢(θ)≡(f1⁢(θ),f2⁢(θ),f3⁢(θ)). If {f1,f2,f3} is linearly independent, then there exist θl<θm<θh such that {f⁢(θl),f⁢(θm),f⁢(θh)} spans ℝ3. Take any 𝜶∈ℝ3∖{0} that is orthogonal to the plane Sθl,θh that is spanned by 𝒇⁢(θl) and 𝒇⁢(θh), as illustrated in Figure 3. The linear combination 𝜶⋅𝒇 is not single crossing because (𝜶⋅𝒇)⁢(θl)=(𝜶⋅𝒇)⁢(θh)=0 while (𝜶⋅𝒇)⁢(θm)≠0.

Refer to caption
Figure 3: The necessity of linear dependence in Proposition 2.

While the necessity portion of Proposition 2 only asserts ratio ordering of the “generating functions”, Lemma 1 implies that if f:Z×Θ→ℝ is linear combinations SC-preserving, then for all z,z′∈Z, the pair f⁢(z,⋅):Θ→ℝ and f⁢(z′,⋅):Θ→ℝ must be ratio ordered.

Proof Sketch of Theorem 2.

We can now sketch the argument for Theorem 2. That its characterization is sufficient for SCD⋆ is straightforward from Lemma 1. For necessity, suppose as a simplification that, for some a0∈A, (∀θ) u⁢(a0,θ)=0.313131This is just a normalization, since u⁢(a,θ) has SCD⋆ if and only if u~⁢(a,θ)≡u⁢(a,θ)−u⁢(a0,θ) has SCD⋆. For any {a1,…,an}⊂A and (λ1,…,λn), we build on the Hahn-Jordan decomposition of (λ1,…,λn) to write the linear combination ∑i=1nλi⁢v⁢(ai,θ) as M⁢∑i=0n(p⁢(ai)−q⁢(ai))⁢u⁢(ai,θ), where p and q are probability mass functions on {a0,a1,…,an}, and M is a scalar.323232Let L≡∑i=1nλi. For i>0, set p′⁢(ai)≡max⁡{λi,0} and q′⁢(ai)≡−min⁡{λi,0}. If L≥0, set p′⁢(a0)=0 and q′⁢(a0)≡L; if L<0, set p′⁢(a0)≡−L and q′⁢(a0)≡0. Let M≡∑i=0np′⁢(ai)=∑i=0nq′⁢(ai). Finally, for all a∈{a0,…,an}, set p⁢(a)≡p′⁢(a)/M and q⁢(a)≡q′⁢(a)/M. (Unless ∑i=1nλi=0, we have ∑i=1np⁢(ai)≠∑i=1nq⁢(ai); the assumption that u⁢(a0,⋅)=0 permits us to assign all the “excess difference” to a0, as detailed in fn. 32.) Since the environment is convex there exist ap,aq∈A such that the linear combination is equal to M⁢(u⁢(ap,θ)−u⁢(aq,θ)), which is single crossing because u has SCD⋆ . Since every such linear combination is single crossing, Proposition 2 guarantees a′ and a′′ such that for all a, u⁢(a,⋅)=g1⁢(a)⁢u⁢(a′,⋅)+g2⁢(a)⁢u⁢(a′′,⋅), with u⁢(a′,⋅) and u⁢(a′′,⋅) each single crossing and ratio ordered.

Remark 1.

The proof of Theorem 2 is materially no different from that for an expected-utility environment. Indeed, any convex environment (A,Θ,u) can be transformed into another convex environment (A~,Θ,u~) where A~≡{u⁢(a,⋅)}a∈A is a convex subset of the vector space of functions Θ↦ℝ, and, under mild conditions, Choquet’s Theorem implies that each u~⁢(⋅,θ):A~→ℝ is an expected utility function. We develop the lens of convex environments for two reasons: (i) to highlight that our characterization is driven only by Condition (⋆), not other properties of expected utility such as linearity or continuity; and (ii) to cover additional economic settings, as in Examples 2–4.

4 Applications

This section illustrates the usefulness of our results in three applications. Interested readers may consult our earlier working paper, Kartik, Lee, and Rappoport (2019, Section 4.3), for another application to costly signaling.

4.1 Cheap Talk with Uncertain Receiver Preferences

There are two expected-utility players, a sender (S) and a receiver (R). The sender’s type is θ∈Θ, where Θ is ordered by ≤. After learning his type, S chooses a payoff-irrelevant message m∈M, where |M|>1. After observing m but not θ, R takes a decision x∈X. The sender’s von Neumann Morgenstern utility function is vS⁢(x,θ); the receiver’s is vR⁢(x,θ,ψ), where ψ∈Ψ is a preference parameter that is unknown to S when choosing m, and known to R when choosing x. Note that ψ does not affect the sender’s preferences. The variables θ and ψ are independently drawn from commonly-known probability distributions.

An example is Θ=[0,1], X=ℝ, 𝝍∈Ψ⊆ℝ2, vS⁢(x,θ)=−(x−θ)2 and vR⁢(x,θ,𝝍)=−(x−ψ1−ψ2⁢θ)2. Here the variable ψ1 captures the receiver’s “type-independent bias” and ψ2 captures the relative “sensitivity” to the sender’s type. If 𝝍 were commonly known and θ uniformly distributed, this would be the model of Melumad and Shibano (1991), which itself generalizes the canonical example from Crawford and Sobel (1982) that obtains when ψ1≠0 and ψ2=1.

We focus on (weak Perfect Bayesian) equilibria in which S uses a pure strategy, μ:Θ→M, and R plays a possibly-mixed strategy, α:M×Ψ→Δ⁢X.333333Our notion of equilibrium requires optimal play for every (not just almost every) type of sender. The restriction to pure strategies for the sender is for expositional simplicity. Given any α, every message m induces some lottery Pα⁢(m)∈Δ⁢X from the sender’s viewpoint. An equilibrium (μ,α) is: (i) an interval equilibrium if every message is used by an interval of sender types, i.e., if (∀m) {θ:μ⁢(θ)=m} is an interval; and (ii) sender minimal if for all on-the-equilibrium-path m≠m′, there is some θ such that 𝔼Pα⁢(m)⁢[vS⁢(⋅,θ)]≠𝔼Pα⁢(m′)⁢[vS⁢(⋅,θ)]. In other words, sender minimality rules out all sender types being indifferent between two distinct on-path messages.343434In Crawford and Sobel (1982) and Melumad and Shibano (1991), all equilibria are outcome equivalent to sender-minimal equilibria. More generally, all equilibria are sender minimal when there is a complete order over messages under which higher messages are infinitesimally more costly for all sender types.

Claim 1.

If vS has SSCED (Corollary 1) then every sender-minimal equilibrium is an interval equilibrium.

  • Proof.

    From the sender’s viewpoint, the lottery over the receiver’s decisions that is induced by any message (given any receiver strategy) is independent of θ because ψ and θ are independent. The result follows from Theorem 1, as sender-minimality implies that one can restrict attention to the sender choosing among lotteries that are utility-distinguishable (i.e., if P and Q are equilibrium lotteries, then DP,Q is not a zero function). ∎

Claim 1 relates to Seidmann (1990), who first considered an extension of Crawford and Sobel (1982) to sender uncertainty about the receiver’s preferences. His goal was to illustrate how such uncertainty could facilitate informative communication even when the sender always strictly prefers higher receiver decisions. Example 2 in Seidmann (1990) constructs a non-interval and sender-minimal equilibrium that is informative. Claim 1 clarifies that the key is a failure of SSCED.

The strict single-crossing property in standard cheap-talk models (e.g., Crawford and Sobel (1982) and Melumad and Shibano (1991)) not only yields interval equilibria, but it also implies that local incentive compatibility is sufficient for global incentive compatibility. This additional tractability also holds under SSCED. Let Θ={θi:i∈ℤ} such that θi<θj for i<j, and P:Θ→Δ⁢X be a candidate equilibrium outcome (i.e., the distribution of receiver’s choices that each sender type induces in equilibrium). Under SSCED, it is sufficient for sender incentive compatibility that (∀i∈ℤ) u⁢(P⁢(θi),θi)≥max⁡{u⁢(P⁢(θi−1),θi),u⁢(P⁢(θi+1),θi)}.

Besides being sufficient, (S)SCED is also necessary to guarantee interval cheap-talk equilibria. Say that vS strictly violates SCED if the expected utility function strictly violates SCD, i.e., if there are P,Q∈Δ⁢X and θl<θm<θh such that min⁡{DP,Q⁢(θl),DP,Q⁢(θh)}>0>DP,Q⁢(θm).

Claim 2.

Let Θ⊆ℝ, X≡ℝ, Ψ⊆ℝ2, and vR⁢(x,θ,𝛙)≡−(x−ψ1−ψ2⁢θ)2. If vS strictly violates SCED, then for some pair of distributions of θ and 𝛙 there is a non-interval equilibrium in which each sender type plays its unique best response.

  • Proof.

    Assume vS strictly violates SCED and let P and Q be the distributions and θl<θm<θh the types in that definition. In what follows, we can without loss assume |M|=2. So let M≡{m′,m′′} and consider the sender’s strategy

    μ⁢(θ)={m′if θ∈{θl,θh}m′′if θ=θm.

    Let Fθ be any distribution with support {θl,θm,θh} and θ′≡𝔼Fθ⁢[θ|θ∈{θl,θh}]≠θm. Then, the unique best response against μ for a receiver of type 𝝍≡(ψ1,ψ2) is

    α⁢(m′,𝝍)=ψ1+ψ2⁢θ′andα⁢(m′′,𝝍)=ψ1+ψ2⁢θm.

    It can be verified that there is a distribution F𝝍 such that, from the sender’s viewpoint, the message m′ leads to the distribution P and the message m′′ leads to the distribution Q, and so μ is the sender’s unique best response.353535Let x and y be random variables with distributions P and Q, respectively. Let F𝝍 be the distribution of a random vector 𝝍=(ψ1,ψ2) defined by [ψ1ψ2]≡[1θ′1θm]−1⁢[xy]. As 𝝍∼F𝝍 (i.e., 𝝍 has distribution Fψ), [1θ′1θm]⁢[ψ1ψ2]=[ψ1+ψ2⁢θ′ψ1+ψ2⁢θm] is stochastically equivalent to (x,y). Thus, α⁢(m′,𝝍)=ψ1+ψ2⁢θ′∼P and α⁢(m′′,𝝍)=ψ1+ψ2⁢θm∼Q. ∎

The particular specification of vR in Claim 2 is not critical; what is important is that there be enough flexibility to generate appropriate lotteries from the sender’s viewpoint using best responses for the receiver.363636In particular, the result in Claim 2 holds under the following more general assumptions: Θ,A⊆ℝ, the receiver’s preferences are represented by uR⁢(a,θ,ψ)=g1⁢(ψ,a)⁢θ+g2⁢(ψ,a), and for every θl<θh and a′,a′′∈A, there exists ψ∈Ψ such that a′∈arg⁢maxa⁡uR⁢(a,ψ,θl) and a′′∈arg⁢maxa⁡uR⁢(a,ψ,θh). The proof is very similar to that of Claim 2. For example, the result would also hold—more straightforwardly, but less interestingly—if the receiver were totally indifferent over all actions for some preference realization. On the other hand, if ψ∈ℝ and vR⁢(x,θ,ψ)≡−(x−θ−ψ)2, then (S)SCED is not necessary, because any pair of lotteries that the sender may face are ranked by first-order stochastic dominance. Strict supermodularity of vS⁢(x,θ) then guarantees that all sender-minimal equilibria are interval equilibria; however, strict supermodularity does not imply (S)SCED, as noted in Subsection 3.2.1.

In our cheap-talk application it is uncertainty about the receiver’s preferences that leads to the sender effectively choosing among lotteries over the receiver’s decisions. Similar results could also be obtained when the receiver’s preferences are known but communication is noisy, à la Blume et al. (2007).

4.2 Observational Learning with Multidimensional Utility

The classic sequential observational learning model (Banerjee, 1992; Bikhchandani et al., 1992) considers agents sequentially choosing products with uncertainty about product values but learning from predecessors’ choices.

For concreteness, suppose that students t=1,2,… sequentially purchase laptops before entering an engineering school. Each laptop is identified by its attribute vector 𝒂≡(a1,…,an)∈[0,1]n, which consists of processing power, memory, screen size, price, etc. The students are uncertain about the nature of the work required to complete the degree, which is captured by a state θ∈Θ⊂ℝ, where Θ is countable. The state affects how students weigh each attribute of a laptop. Specifically, following Example 3, students have a common multidimensional utility function u⁢(𝒂,θ)≡∑i=1ngi⁢(ai)⁢fi⁢(θ). Each gi has a convex range, which ensures a convex environment, i.e., Condition (⋆). Students are expected-utility maximizers.

Given a finite set of available laptops 𝒜⊆[0,1]n and a prior μ0 over the state, each student chooses a laptop using two information sources. Conditional on the state, student t obtains an independent private signal st about θ (e.g., her own impression from reading syllabi and talking to alumni). A canonical example is normal information: st∼𝒩⁢(θ,σ2), where σ>0 is a known constant. Each student also observes all predecessors’ choices.

The question is whether there is adequate learning, as defined by Kartik et al. (2022): no matter the prior μ0 and the choice set 𝒜, do students’ sequential choices eventually settle on the laptop they would choose if the state were known?373737Here is a precise definition. The full-information expected utility given a belief μ∈Δ⁢Θ is the expected utility under that belief if the state will be revealed before an action is chosen: U∗⁢(μ)≡∑θ∈Θmax𝒂∈𝒜⁡u⁢(𝒂,θ)⁢μ⁢(θ). Given a prior μ0 and a strategy profile σ, student t’s utility Ut is a random variable. Let 𝔼σ,μ0⁢[Ut] be student t’s ex-ante expected utility. There is adequate learning at a choice set 𝒜 if for every prior μ0 and every equilibrium strategy profile σ, 𝔼σ,μ0⁢[Ut]→U∗⁢(μ0) as t→∞. There is adequate learning if there is adequate learning at all choice sets. Those authors identify SCD as the necessary and sufficient condition on the utility function u⁢(𝒂,θ) for learning under normal information, and more generally, for information structures that have “directionally unbounded beliefs”.383838Roughly, an information structure has directionally unbounded beliefs if for every state there exist signals that provide arbitrarily-close-to relative certainty about that state vs. all lower states, and analogously (potentially different) signals for that state vs. all higher states. Under normal information, for any state, arbitrarily high signals deliver the former while arbitrarily low signals deliver the latter. In the current multidimensional utility environment, the following characterization obtains, which we state without proof.

Claim 3.

There is adequate learning under normal information if and only if u⁢(𝐚,θ) has SCD⋆, i.e., it has the form stated in Corollary 5.

The claim implies that adequate learning under normal information obtains if only if students’ preferences depend on at most two “summary attributes”, which are each linear combinations over the original value over attributes, i.e., linear combinations of the functions gi⁢(ai). For example, students may trade off “convenience”, which is an aggregate of price, weight, and size, with “performance”, which aggregates processing speed, memory, and storage. Learning requires that the convenience and performance aggregators do not vary with the state. A second implication derives from fI and fI⁢I in Equation 7 being ratio ordered. The interpretation is that for learning, students must value one summary attribute more as the state increases, e.g., a higher state indicates more computationally intensive work and hence a higher value of performance relative to convenience.

4.3 Collective Choice over Lotteries

Collective choice over lotteries manifests in many contexts: elections entail uncertainty about what policies elected politicians will implement; and a board of directors may view each candidate for CEO as a probability distribution over firm earnings. Zeckhauser (1969) first pointed out that pairwise-majority comparisons in these settings can be cyclical, even when comparisons over deterministic outcomes are not. Our results shed light on when such difficulties do not arise.

Consider a finite group of individuals indexed by i∈{1,2,…,N}, where for simplicity N is odd. The group must choose from a set of lotteries over X, where X is the space of outcomes (political policies, earnings, etc.) with generic element x. Each individual i has von Neumann Morgenstern utility function v⁢(x,θi), where θi∈Θ is a preference parameter or i’s type. We assume Θ is completely ordered; without further loss of generality, let Θ⊂ℝ and θ1≤⋯≤θN. The expected utility for an individual of type θ from lottery P is u⁢(P,θ)≡∫xv⁢(x,θ)⁢dP. We denote individual i’s preference relation over lotteries by ⪰i, with strict component ≻i.

Define the group’s preference relation, ⪰m⁢a⁢j, over lotteries P and Q by majority rule:

P⪰m⁢a⁢jQ⁢ if ⁢|{i:P⪰iQ}|≥N/2.
Claim 4.

If v has SCED (Corollary 1), then the group’s preference relation is transitive and coincides with that of individual (N+1)/2.

The claim follows from the arguments of Rothstein (1990) and Gans and Smart (1996), but we include a direct proof given how short it is.

  • Proof.

    Let M≡(N+1)/2. By SCD, (i) if P⪰MQ , then P⪰m⁢a⁢jQ because there cannot exist i<M<j such that Q≻iP and Q≻jP; analogously, (ii) if P≻MQ, then P≻m⁢a⁢jQ. Hence, ⪰m⁢a⁢j coincides with ⪰M. ∎

Claim 4 can be applied to a well-known problem in political economy (Shepsle, 1972). For simplicity, we now further assume the policy space is a finite set X⊂ℝ and there is a group of N voters whose types are their ideal points, i.e., {θi}=arg⁢maxx∈ℝ⁡v⁢(x,θi). There are two office-motivated candidates, L and R; each j∈{L,R} can commit to any lottery from some given set 𝒜j⊆Δ⁢X. A restricted set 𝒜j may capture various kinds of constraints; for example, Shepsle (1972) assumed the incumbent candidate could only choose degenerate lotteries. In our setting, what ensures the existence of an equilibrium, and which policy lotteries are offered in an equilibrium?393939More precisely: the two candidates simultaneously choose their lotteries, and each voter then votes for his preferred candidate (assuming, for concreteness, that a voter randomizes between the candidates with equal probability if indifferent). A candidate wins if he receives a majority of the votes. Candidates maximize the probability of winning. We seek a Nash equilibrium of the game between the two candidates.

Claim 4 implies that if voters’ utility functions v have SCED and if voter M≡(N+1)/2 is indifferent between her most-preferred lottery in 𝒜L and in 𝒜R, then there is a unique equilibrium: each candidate offers the best lottery for voter M; in particular, both candidates converge to δθM, the degenerate lottery on θM, if that is feasible for both. It bears emphasis that in this case policy convergence at the median ideal point is not driven by all voters being globally “risk averse” (e.g., v⁢(x,θ)=−(x−θ)2); rather, it is because SCED ensures the existence of a decisive voter whose most-preferred lottery is degenerate.404040An example may be helpful. Let X=[−1,1], Θ={−1,0,1}, and v⁢(x,θ)=x⁢θ+1/(|x|+1)+1. The corresponding functions f1⁢(θ)=θ and f2⁢(θ)=1 are each strictly single crossing from below and strictly ratio ordered. For all θ, v⁢(⋅,θ) is maximized at x=θ but is convex on some subinterval of X.

There is a sense in which SCED is necessary to guarantee that each candidate j will offer the median ideal-point voter’s most-preferred lottery from the feasible set 𝒜j. Suppose v⁢(x,θ) strictly violates SCED, i.e., there are P,Q∈Δ⁢X and θl<θm<θh such that min⁡{DP,Q⁢(θl),DP,Q⁢(θh)}>0>DP,Q⁢(θm). Then, if the population of voters is just {l,m,h} and 𝒜L=𝒜R={P,Q}, the unique equilibrium is for both candidates to offer lottery P, which is voter m’s less-preferred lottery.

5 Discussion

5.1 Single Crossing vs. Monotonicity

We have characterized when u:A×Θ→ℝ has SCD in a convex environment. SCD is an ordinal property. Analogous to the interest in monotonic functions rather than just single-crossing functions, one may also be interested in the following cardinal property, which we term monotonic differences (MD):

(∀a,a′∈A)⁢u⁢(a,θ)−u⁢(a′,θ)⁢ is monotonic in ⁢θ.

In a convex environment, we write MD⋆ analogously to SCD⋆.

Theorem 3.

The function u:A×Θ→ℝ has MD⋆ if and only if it takes the form

u⁢(a,θ)=g1⁢(a)⁢f1⁢(θ)+g2⁢(a)+h⁢(θ), (9)

where f1 is monotonic.

Suppose that an expected-utility agent with von Neumann Morgenstern utility v⁢(x,θ) is choosing among lotteries, so A≡Δ⁢X. We say that v has monotonic expectational differences (MED) if the expected utility function u has MD⋆. Analogous to the SCED characterization in Corollary 1, it is straightforward that the function v has MED if and only if it has the same characterization as given for u in Theorem 3. This MED characterization has largely been obtained by Kushnir and Liu (2018) in their study of the equivalence between Bayesian and dominant-strategy implementation.

A function u with SCD⋆ has MD⋆ when the function f2 in the form (5) is identically equal to one; f1 and f2 being ratio ordered is then equivalent to f1 being monotonic. In general, SCD⋆ functions need not have MD⋆; however, by Proposition 1, SCD⋆ functions have MD⋆  representations if Θ has both a minimum and a maximum.414141Absent this condition, an SCD⋆ function may not even have an MD⋆ representation. In the context of expected utility, an earlier version of our paper characterized precisely when such representations do not exist and provided an example (Kartik, Lee, and Rappoport, 2019, Appendix G). Duggan (2014, Section 4) also discusses the difficulty of finding SCED functions that do not have MED representations. We view this finding as unexpected; outside of convex environments, we are not aware of any general result on when SCD functions have MD representations.

5.2 Interval Choice and Monotone Comparative Statics

In Section 2, we showed that (S)SCD characterizes interval choice. There is a sense in which interval choice is intimately related to monotone comparative statics holding with respect to some order over the choice space. In light of Theorem 1, the connection is elucidated below by tying (S)SCD to monotone comparative statics.

Throughout this subsection, we consider an ordered set of alternatives, (A,⪰). To simplify exposition and avoid dealing with equivalence classes, we will focus on comparative statics for a function u:A×Θ→ℝ such that A is minimal (with respect to u), i.e., (∀a≠a′)⁢(∃θ)⁢u⁢(a,θ)≠u⁢(a′,θ). For any a,a′∈A, let a∨a′ and a∧a′ denote the usual join and meet respectively.424242Alternative a¯∈A is the join (or supremum) of {a,a′} if (i) a¯⪰a and a¯⪰a′, and (ii) if b⪰a and b⪰a′, then b⪰a¯. The meet (or infimum) of {a,a′} is defined analogously. Neither need exist. Given any S,S′⊆A, we say that S dominates S′ in the strong set order, denoted S⪰S⁢S⁢OS′, if for every a∈S and a′∈S′, (i) a∨a′ and a∧a′ exist, and (ii) a∨a′∈S and a∧a′∈S′. It can be verified that ⪰S⁢S⁢O is transitive on non-empty subsets of (A,⪰).

Definition 4.

u:A×Θ→ℝ has monotone comparative statics (MCS) on (A,⪰) if

(∀S⊆A)⁢(∀θl≤θh)arg⁢maxa∈S⁡u⁢(a,θh)⪰S⁢S⁢Oarg⁢maxa∈S⁡u⁢(a,θl).

Our definition of MCS is closely related to but not the same as Milgrom and Shannon (1994). We take (A,⪰) to be any ordered set while they require a lattice. We focus only on monotonicity of choice in θ but require the monotonicity to hold for every subset S⊆A; Milgrom and Shannon require monotonicity of choice jointly in the pair (θ,S), but this implicitly only requires choice monotonicity in θ to hold for every sublattice S⊆A.

Define binary relations ≻S⁢C⁢D and ⪰S⁢C⁢D on A as follows: a≻S⁢C⁢Da′ if Da,a′⁢(θ)≡u⁢(a,θ)−u⁢(a′,θ) is single crossing only from below; a⪰S⁢C⁢Da′ if either a≻S⁢C⁢Da′ or a=a′. It is clear that ⪰S⁢C⁢D is reflexive and anti-symmetric. If u:A×Θ→ℝ has SCD, then ⪰S⁢C⁢D is also transitive. Moreover, given SCD (and minimality), the ⪰S⁢C⁢D order is incomplete only over pairs with “dominance”: if a⋡S⁢C⁢Da′ and vice-versa, then either (∀θ)⁢Da,a′⁢(θ)>0, or (∀θ)⁢Da,a′⁢(θ)<0.

Theorem 4.

u:A×Θ→ℝ has monotone comparative statics on (A,⪰), where A is minimal, if and only if u has SCD and ⪰ is a refinement of ⪰S⁢C⁢D.

Our definition of SCD does not require an order on the set of alternatives, whereas MCS does. Theorem 4 says that SCD is necessary and sufficient for there to exist an order that yields MCS. Moreover, the theorem justifies viewing ⪰S⁢C⁢D as the prominent order for MCS: MCS does not hold with any order that either coarsens ⪰S⁢C⁢D or reverses a ranking by ≻S⁢C⁢D. The argument for necessity in Theorem 4 only makes use of binary choice sets. If SCD fails, then there is no order ⪰ for which there is choice monotonicity for all binary choice sets. If SCD holds, then choice monotonicity on all binary choice sets requires ⪰ to refine ⪰S⁢C⁢D. For sufficiency in Theorem 4, we show that for any choice set, ⪰S⁢C⁢D is a complete order among the set of alternatives that are chosen by some type; thereafter, we appeal to Milgrom and Shannon (1994, Theorem 4).

In Supplementary Appendix SA.1, we state and prove an analog of Theorem 4 for monotone selection, i.e., that every selection from the choice correspondence is increasing; similarly to Milgrom and Shannon (1994, Theorem 4’), this result uses SSCD rather than SCD.

6 Conclusion

This paper has characterized the class of utility functions that have SCD⋆, i.e., single-crossing differences in a convex environment (Theorem 2). Our notion of SCD does not presume an order on the choice space and is the appropriate notion for interval-choice comparative statics (Theorem 1). We have given a number of examples of convex environments, most notably expected utility, rank-dependent expected utility, and multidimensional utility, and discussed the implications of our characterization in these contexts. The applications in Section 4 illustrate how SCD⋆ is useful in economic problems.

While we have emphasized interval choice, it bears noting that there are other familiar implications of single-crossing properties that in fact rely only on our order-independent notion of (S)SCD; for example, SSCD is the key to guaranteeing that local incentive compatibility implies global incentive compatibility.434343We mean in a sense analogous to Carroll (2012, Proposition 4). While Carroll establishes his result by defining a single-crossing property with respect to some given order over alternatives, essentially the same logic applies with our order-independent notion of strict SCD (Definition 2).

We close by mentioning some avenues for future research. First, it may be of interest to characterize exactly when preferences have a utility representation that satisfies our convexity condition (⋆). Second, and relatedly, one may explore characterizations of SCD outside convex environments. In particular, we are intrigued by the question of when SCD preferences do not possess a representation that takes a similar form to that we have characterized.

Third, our results have direct bearing on problems in which all types of an agent face the same choice set. Such situations are natural. But consider a variation of the cheap-talk application (Subsection 4.1) in which the sender’s type is correlated with the receiver’s type. Even though the receiver’s type does not affect the sender’s payoff, different sender types will generally have different beliefs about the distribution of the receiver’s action that any message induces in equilibrium. Effectively, different sender types will be choosing from different choice sets. An approach that synthesizes the current paper’s with that of, for example, Athey’s (2002) may be useful for such problems.

Appendices

Organization.

Appendix A contains proofs for our comparative statics results (Theorem 1 and Theorem 4); Appendix B for aggregation of single crossing functions (Lemma 1 and Proposition 2) and our characterizations of SCD⋆ (Theorem 2 and Proposition 1); and Appendix C for our MD⋆ characterization (Theorem 3). Some additional material is in the online Supplementary Appendices.

A preliminary result.

Before turning to the proofs, we state a useful equivalence with the violation of single crossing; the result is obvious when (Θ,≤) is a completely ordered set but also applies when it is not.

Claim 5.

A function f:Θ→ℝ is not single crossing if and only if for some θl<θm<θh, either

sign⁡[f⁢(θl)]⁢<sign⁡[f⁢(θm)]⁢ and ⁢sign⁡[f⁢(θm)]>⁢sign⁡[f⁢(θh)],or (10)
sign⁡[f⁢(θl)]>sign⁡[f⁢(θm)]⁢ and ⁢sign⁡[f⁢(θm)]<sign⁡[f⁢(θh)]. (11)
  • Proof of Claim 5.

    The “if” direction of the claim is immediate. For the “only if” direction, suppose f:Θ→ℝ is single crossing neither from below nor from above:

    (∃θ1<θ2)sign⁡[f⁢(θ1)]<sign⁡[f⁢(θ2)], and (∃θ3<θ4)sign⁡[f⁢(θ3)]>sign⁡[f⁢(θ4)].

    Let Θ0≡{θ1,θ2,θ3,θ4} and θ¯ and θ¯ be upper and lower bounds of Θ0. If f⁢(θ¯)=f⁢(θ¯)=0, then (θl,θm,θh)=(θ¯,θ0,θ¯) for some θ0∈Θ0 with f⁢(θ0)≠0 satisfies either (10) or (11). So assume f⁢(θ¯)≠0; an similar argument applies if f⁢(θ¯)≠0. If f⁢(θ¯)<0, then (θl,θm,θh)=(θ1,θ2,θ¯) satisfies (10). If f⁢(θ¯)>0, then (θl,θm,θh)=(θ3,θ4,θ¯) satisfies (11). ∎

Appendix A Proofs for Comparative Statics

A.1 Proof of Theorem 1

Part 1. Suppose u has SCD, and consider any S⊆A, a′∈S, and θl,θh∈{θ:a′∈arg⁢maxa∈S⁡u⁢(a,θ)} with θl<θh. For any a′′∈S, Da′,a′′⁢(θl)≥0 and Da′,a′′⁢(θh)≥0, which imply that Da′,a′′⁢(θm)≥0 for all θm with θl<θm<θh. It follows that {θ:a′∈arg⁢maxa∈S⁡u⁢(a,θ)} is an interval.

If u strictly violates SCD, ∃a′,a′′∈A and θl<θm<θh such that min⁡{Da′,a′′⁢(θl),Da′,a′′⁢(θh)} >0>Da′,a′′⁢(θm). Clearly, {θ:a′∈arg⁢maxa∈{a′,a′′}⁡u⁢(a,θ)} is not an interval.

Part 2. Suppose |Θ|≥3. A function u:A×Θ→ℝ does not have SSCD when there exist a′,a′′∈A such that Da′,a′′ (and Da′′,a′) are not strictly single crossing. Alternatively, using Claim 5, u does not have SSCD if and only if (∃a′,a′′) (∃θl<θm<θh) Da′,a′′⁢(θl)≥0, Da′,a′′⁢(θm)≤0, and Da′,a′′⁢(θh)≥0. This condition is equivalent to (∃S⊆A⁢with⁢a′,a′′∈S) (∃θl<θm<θh) a′∈⋂θ∈{θl,θ}arg⁢maxa∈S⁡u⁢(a,θ) and Da′,a′′⁢(θm)≤0, which holds if and only if some selection from the choice correspondence Cu does not have interval choice.

A.2 Proof of Theorem 4

(⟹)

Suppose u:A×Θ→ℝ has MCS on A with some order ⪰. We first prove the following claim.

Claim 6.

For every a′,a′′∈A, if ∃θl<θh such that sign⁡[Da′,a′′⁢(θl)]<sign⁡[Da′,a′′⁢(θh)], then a′≻a′′.

  • Proof.

    Consider S={a′,a′′}. Since sign⁡[Da′,a′′⁢(θl)]≠sign⁡[Da′,a′′⁢(θh)], we have

    arg⁢maxa∈S⁡u⁢(a,θl)≠arg⁢maxa∈S⁡u⁢(a,θh).

    Thus, either (i) a′∈arg⁢maxa∈S⁡u⁢(a,θl) and a′′∈arg⁢maxa∈S⁡u⁢(a,θh), or (ii) a′′∈arg⁢maxa∈S⁡u⁢(a,θl) and a′∈arg⁢maxa∈S⁡u⁢(a,θh). Since u has MCS on (A,⪰), we have arg⁢maxa∈S⁡u⁢(a,θh)⪰S⁢S⁢Oarg⁢maxa∈S⁡u⁢(a,θl). Therefore, a′∧a′′∈arg⁢maxa∈S⁡u⁢(a,θl) and a′∨a′′∈arg⁢maxa∈S⁡u⁢(a,θh), which implies that either a′⪰a′′ or a′′⪰a′. Since a′≠a′′, we have either a′≻a′′ or a′′≻a′. If a′′≻a′, then a′′=a′∨a′′∈arg⁢maxa∈S⁡u⁢(a,θh), contradicting sign⁡[Da′,a′′⁢(θl)]<sign⁡[Da′,a′′⁢(θh)]. Thus, a′≻a′′. ∎

To show that u has SCD on A, suppose not, towards contradiction. Claim 5 implies there exist a′,a′′∈A and θl<θm<θh such that either

sign⁡[Da′,a′′⁢(θl)]<sign⁡[Da′,a′′⁢(θm)]andsign⁡[Da′,a′′⁢(θm)]>sign⁡[Da′,a′′⁢(θh)],or (12)
sign⁡[Da′,a′′⁢(θl)]>sign⁡[Da′,a′′⁢(θm)]andsign⁡[Da′,a′′⁢(θm)]<sign⁡[Da′,a′′⁢(θh)]. (13)

Given either (12) or (13), Claim 6 implies a′≻a′′ and a′′≻a′, a contradiction.

To show that ⪰ is a refinement of ⪰S⁢C⁢D, it suffices to show that

(∀a′,a′′∈A)a′≻S⁢C⁢Da′′⟹a′≻a′′, (14)

because both ⪰ and ⪰S⁢C⁢D are anti-symmetric. Take any a′,a′′∈A such that a′≻S⁢C⁢Da′′. As Da′,a′′ is single crossing only from below, ∃θl<θh such that sign⁡[Da′,a′′⁢(θl)]<sign⁡[Da′,a′′⁢(θh)]. Claim 6 implies a′≻a′′, which proves (14).

(⟸)

Suppose that u:A×Θ→ℝ has SCD, and ⪰ is a refinement of ⪰S⁢C⁢D. For any S⊆A, define Cu⁢(S)≡⋃θ∈Θarg⁢maxa∈S⁡u⁢(a,θ). It is clear that

(∀θ)arg⁢maxa∈S⁡u⁢(a,θ)=arg⁢maxa∈Cu⁢(S)⁡u⁢(a,θ).

We claim that Cu⁢(S) is completely ordered by ⪰S⁢C⁢D. To see why, take any pair a′,a′′∈Cu⁢(S) with a′≠a′′. As u has SCD, Da′,a′′ is single crossing in θ. As A is minimal, Da′,a′′ is not a zero function. Also, as a′,a′′∈Cu⁢(S), sign⁡[Da′,a′′] is not a constant function with value either 1 or -1. Thus, Da′,a′′ is single crossing either only from below, or only from above. It follows that a′≻S⁢C⁢Da′′ or a′′≻S⁢C⁢Da′.

Since ⪰ is a refinement of ⪰S⁢C⁢D, ⪰ coincides with ⪰S⁢C⁢D on Cu⁢(S), and the strong set orders generated by ⪰ and ⪰S⁢C⁢D on the collection of all subsets of Cu⁢(S) also coincide. By definition of ⪰S⁢C⁢D, when restricted to Cu⁢(S)×Θ, u satisfies Milgrom and Shannon’s (1994) single-crossing property in (a,θ) with respect to ⪰S⁢C⁢D and ≤.444444Milgrom and Shannon’s single-crossing property of u on Cu⁢(S)×Θ is equivalent to (∀a′≻S⁢C⁢Da′′)⁢(∀θl<θh)u⁢(a′,θl)≥(>)⁢u⁢(a′′,θl)⟹u⁢(a′,θh)≥(>)⁢u⁢(a′′,θh). As Cu⁢(S) is completely ordered by ⪰S⁢C⁢D, it follows from Milgrom and Shannon (1994, Theorem 4) that ∀θl<θh,

arg⁢maxa∈S⁡u⁢(a,θh)=arg⁢maxa∈Cu⁢(S)⁡u⁢(a,θh)⪰S⁢S⁢Oarg⁢maxa∈Cu⁢(S)⁡u⁢(a,θl)=arg⁢maxa∈S⁡u⁢(a,θl).

Appendix B Proofs for Aggregating Single-Crossing Functions and Main Characterizations

Before providing the proofs in this appendix, we first clarify Condition (4) in the definition of ratio dominance.

B.1 On the Definition of Ratio Dominance

Section 3 in the main text explained Condition (3) in the definition of ratio dominance. We impose Condition (4) to rule out cases in which, for some θl<θm<θh, either (i) 𝒇⁢(θl) and 𝒇⁢(θh) are collinear in opposite directions while 𝒇⁢(θm) is not, or (ii) 𝒇⁢(θl) and 𝒇⁢(θh) are non-zero vectors while 𝒇⁢(θm) is not. See Figure 4, wherein panel 4(a) depicts case (i) and panel 4(b) depicts case (ii). Note that Condition (3) is satisfied in both panels.

Refer to caption
(a) Failure of ⟹.
Refer to caption
(b) Failure of ⟸.
Figure 4: f1 and f2 are not ratio ordered because Condition (4) fails for θl<θm<θh.

It turns out that for all linear combinations of two single-crossing functions f1 and f2 to be single crossing, Condition (4) of ratio ordering is required; see the proof of Lemma 1. Figure 4 illustrates: in panel 1(a), (f1+f2)⁢(θl)=(f1+f2)⁢(θh)=0 while (f1+f2)⁢(θm)<0; in panel 1(b), (f1+f2)⁢(θl)>0 and (f1+f2)⁢(θh)>0 while (f1+f2)⁢(θm)=0. So in both cases, f1 and f2 are each single crossing but f1+f2 is not.

B.2 Proof of Lemma 1

When |Θ|≤2, the proof is trivial as all functions are single crossing and every pair of functions are ratio ordered. Hereafter, we assume |Θ|≥3.

(⟹)

It is clear that each function f1 and f2 is single crossing. We must show that f1 and f2 are ratio ordered.

To prove (3), we suppose towards contradiction that

(∃θl<θh)f1⁢(θl)⁢f2⁢(θh)<f1⁢(θh)⁢f2⁢(θl), and (∃θ′<θ′′)f1⁢(θ′)⁢f2⁢(θ′′)>f1⁢(θ′′)⁢f2⁢(θ′). (15)

Take any upper bound θ¯ of {θl,θh,θ′,θ′′}.

First, let 𝜶l≡(f2⁢(θl),−f1⁢(θl)). Then (𝜶l⋅𝒇)⁢(θl)=(f2⁢(θl),−f1⁢(θl))⋅(f1⁢(θl),f2⁢(θl))=0, and (𝜶l⋅𝒇)⁢(θh)>0. Thus, 𝜶l⋅𝒇 is single crossing from below and (𝜶l⋅𝒇)⁢(θ¯)>0.

Second, let 𝜶′≡(f2⁢(θ′),−f1⁢(θ′)). Then (𝜶′⋅𝒇)⁢(θ′)=0 and (𝜶′⋅𝒇)⁢(θ′′)<0. Thus, 𝜶′⋅𝒇 is single crossing from above and (𝜶′⋅𝒇)⁢(θ¯)<0.

Let 𝜶¯≡(f2⁢(θ¯),−f1⁢(θ¯)). It follows that

(𝜶¯⋅𝒇)⁢(θl)=(f2⁢(θ¯),−f1⁢(θ¯))⋅(f1⁢(θl),f2⁢(θl))=−(𝜶l⋅𝒇)⁢(θ¯)<0,(𝜶¯⋅𝒇)⁢(θ′)=−(𝜶′⋅𝒇)⁢(θ¯)>0, and (𝜶¯⋅𝒇)⁢(θ¯)=0.

Therefore, 𝜶¯⋅𝒇 is not single crossing, a contradiction.

To prove (4), take any θl<θm<θh.

First, we show that f1⁢(θl)⁢f2⁢(θh)=f1⁢(θh)⁢f2⁢(θl) implies f1⁢(θm)⁢f2⁢(θh)=f1⁢(θh)⁢f2⁢(θm) and f1⁢(θm)⁢f2⁢(θl)=f1⁢(θl)⁢f2⁢(θm). Assume f1 is not a zero function on {θl,θm,θh}, as otherwise the proof is trivial. Since f1 is single crossing, either f1⁢(θl)≠0 or f1⁢(θh)≠0. We consider the case of f1⁢(θh)≠0 (and omit the proof for the other case, as it is analogous). Let 𝜶h≡(f2⁢(θh),−f1⁢(θh)). Since 𝜶h⋅𝒇 is single crossing and (𝜶h⋅𝒇)⁢(θ)=0 for θ=θl,θh, it holds that (𝜶h⋅𝒇)⁢(θm)=f2⁢(θh)⁢f1⁢(θm)−f1⁢(θh)⁢f2⁢(θm)=0. It follows immediately that f1⁢(θm)⁢f2⁢(θh)=f1⁢(θh)⁢f2⁢(θm). As (f1⁢(θm),f2⁢(θm)) and (f1⁢(θh),f2⁢(θh)) are linearly dependent and (f1⁢(θh),f2⁢(θh)) is a non-zero vector, there exists λ∈ℝ such that fi⁢(θm)=λ⁢fi⁢(θh) for i=1,2. Thus,

f1⁢(θl)⁢f2⁢(θm)=λ⁢f1⁢(θl)⁢f2⁢(θh)=λ⁢f2⁢(θl)⁢f1⁢(θh)=f2⁢(θl)⁢f1⁢(θm).

Next, we show that if f1⁢(θl)⁢f2⁢(θm)=f1⁢(θm)⁢f2⁢(θl) and f1⁢(θm)⁢f2⁢(θh)=f1⁢(θh)⁢f2⁢(θm), then f1⁢(θl)⁢f2⁢(θh)=f1⁢(θh)⁢f2⁢(θl). Let 𝜶≡(f2⁢(θl)−f2⁢(θh),−f1⁢(θl)+f1⁢(θh)). It follows that

(𝜶⋅𝒇)⁢(θl) =(f2⁢(θl)−f2⁢(θh))⁢f1⁢(θl)−(f1⁢(θl)−f1⁢(θh))⁢f2⁢(θl)=f1⁢(θh)⁢f2⁢(θl)−f1⁢(θl)⁢f2⁢(θh),
(𝜶⋅𝒇)⁢(θh) =(f2⁢(θl)−f2⁢(θh))⁢f1⁢(θh)−(f1⁢(θl)−f1⁢(θh))⁢f2⁢(θh)=f1⁢(θh)⁢f2⁢(θl)−f1⁢(θl)⁢f2⁢(θh),and
(𝜶⋅𝒇)⁢(θm) =(f2⁢(θl)−f2⁢(θh))⁢f1⁢(θm)−(f1⁢(θl)−f1⁢(θh))⁢f2⁢(θm)=0.

As 𝜶⋅𝒇 is single crossing, it follows that (𝜶⋅𝒇)⁢(θl)=(𝜶⋅𝒇)⁢(θh)=0, as we wanted to show.

(⟸)

Assume that f1 and f2 are each single crossing. We prove the result for the case in which f1 ratio dominates f2; the other case is analogous. For any 𝜶∈ℝ2, we prove that 𝜶⋅𝒇 is single crossing. We may assume that 𝜶≠0, as the result is trivial otherwise.

Suppose, towards contradiction, that 𝜶⋅𝒇 is not single crossing. Claim 5 implies there exist θl<θm<θh such that either

sign⁡[(𝜶⋅𝒇)⁢(θl)]⁢<sign⁡[(𝜶⋅𝒇)⁢(θm)]⁢ and ⁢sign⁡[(𝜶⋅𝒇)⁢(θm)]>⁢sign⁡[(𝜶⋅𝒇)⁢(θh)],or (16)
sign⁡[(𝜶⋅𝒇)⁢(θl)]>sign⁡[(𝜶⋅𝒇)⁢(θm)]⁢ and ⁢sign⁡[(𝜶⋅𝒇)⁢(θm)]<sign⁡[(𝜶⋅𝒇)⁢(θh)]. (17)

First, we consider the case in which 𝒇⁢(θ)≡(f1⁢(θ),f2⁢(θ)) for all θ∈{θl,θm,θh} are non-zero vectors. Take any θ1,θ2∈{θl,θm,θh} such that θ1<θ2. As f1 ratio dominates f2, by Condition (3), 𝒇⁢(θ1) moves to 𝒇⁢(θ2) in a clockwise rotation with an angle less than or equal to 180 degrees. Let r12 be the clockwise angle from 𝒇⁢(θ1) to 𝒇⁢(θ2). The vector 𝜶≠0 defines a partition of ℝ2 into ℝ𝜶,+2≡{𝒙∈ℝ2:𝜶⋅𝒙>0}, ℝ𝜶,02≡{𝒙∈ℝ2:𝜶⋅𝒙=0}, and ℝ𝜶,−2≡{𝒙∈ℝ2:𝜶⋅𝒙<0}. In both cases (16) and (17), both f⁢(θl) and f⁢(θh) are not in the same part of the partition that 𝒇⁢(θm) belongs to. Thus, rl⁢m>0 and rm⁢h>0. On the other hand, both 𝒇⁢(θl) and 𝒇⁢(θh) are in the same closed half-space, either ℝ𝜶,+2∪ℝ𝜶,02 or ℝ𝜶,−2∪ℝ𝜶,02, and 𝒇⁢(θm) is in the other closed half-space, either ℝ𝜶,−2∪ℝ𝜶,02 or ℝ𝜶,+2∪ℝ𝜶,02, respectively. Thus, rl⁢h≥180. Since Condition (3) implies rl⁢h≤180, it follows that rl⁢h=180. Hence, 𝒇⁢(θl) and 𝒇⁢(θm) are linearly independent (0<rl⁢m<180), and similarly for 𝒇⁢(θm) and 𝒇⁢(θh). However, 𝒇⁢(θl) and 𝒇⁢(θh) are linearly dependent (rl⁢h=180). This contradicts (4).

Second, suppose either 𝒇⁢(θl)=0 or 𝒇⁢(θh)=0. We provide the argument assuming 𝒇⁢(θl)=0; it is analogous if 𝒇⁢(θh)=0. Under either (16) or (17), 𝒇⁢(θm)≠0. By Condition (4), 𝒇⁢(θm) and 𝒇⁢(θh) are linearly dependent. In particular, because 𝒇⁢(θm)≠0, there exists a unique λ∈ℝ such that 𝒇⁢(θh)=λ⁢𝒇⁢(θm). Under either (16) or (17), λ≤0, which contradicts the hypothesis that f1 and f2 are single crossing.

Last, suppose 𝒇⁢(θl)≠0, 𝒇⁢(θm)=0, and 𝒇⁢(θh)≠0. By Condition (4), 𝒇⁢(θl) and 𝒇⁢(θh) are linearly dependent. Hence, there exists a unique λ∈ℝ such that 𝒇⁢(θl)=λ⁢𝒇⁢(θh). Under either (16) or (17), λ>0, which contradicts the hypothesis that f1 and f2 are single crossing.

B.3 Proof of Proposition 2

The result is trivial if |Z|=1 and it is equivalent to Lemma 1 if |Z|=2, so we may assume |Z|≥3. The proof is also straightforward if all functions f⁢(z,⋅):Θ→ℝ are multiples of one function f⁢(z1,⋅), i.e., if there is z1 such that (∃λ:Z→ℝ)(∀z)f(z,⋅)=λ(z)f(z1,⋅). Thus, we further assume there exist z′,z′′ such that f⁢(z′,⋅):Θ→ℝ and f⁢(z′′,⋅):Θ→ℝ are linearly independent.

(⟸)

Assume f⁢(z1,⋅) and f⁢(z2,⋅) are (i) each single crossing and (ii) ratio ordered, and that there are functions λ1,λ2:Z→ℝ such that (∀z) f⁢(z,⋅)=λ1⁢(z)⁢f⁢(z1,⋅)+λ2⁢(z)⁢f⁢(z2,⋅). Then, for any function μ:Z→ℝ with finite support,

∫zf⁢(z,θ)⁢dμ =∫z(λ1⁢(z)⁢f⁢(z1,θ)+λ2⁢(z)⁢f⁢(z2,θ))⁢dμ=∑i=1,2(∫zλi⁢(z)⁢dμ)⁢f⁢(zi,θ),

which is single crossing in θ by Lemma 1.

(⟹)

Take any z1,z2∈Z such that f1⁢(⋅)≡f⁢(z1,⋅) and f2⁢(⋅)≡f⁢(z2,⋅) are linearly independent. Then, by Lemma 1, f1 and f2 are each single crossing and ratio ordered, as their linear combinations are all single crossing.

For every θ′,θ′′, let

Mθ′,θ′′≡[f1⁢(θ′)f2⁢(θ′)f1⁢(θ′′)f2⁢(θ′′)].

We first prove the following claim:

Claim 7.

There exists θl<θh such that rank⁡[Mθl,θh]=2.

  • Proof of Claim 7.

    As f1 and f2 are linearly independent, there exists θ0 such that f2⁢(θ0)≠0. Let λ≡−f1⁢(θ0)f2⁢(θ0). Then, for some θλ, f1⁢(θλ)+λ⁢f2⁢(θλ)≠0 and rank⁡[Mθ0,θλ]=2.

    The proof is complete if θ0>θλ or θ0<θλ. If not, take a lower and upper bound, θ¯ and θ¯, of {θ0,θλ}. Then rank⁡[Mθ¯,θ¯]=2. For otherwise, there exists 𝜶∈ℝ2\{0} such that Mθ¯,θ¯⁢𝜶=0. As θ0 and θλ are between θ¯ and θ¯, and α1⁢f1+α2⁢f2 is single crossing, we have Mθ0,θλ⁢𝜶=0, which contradicts rank⁡[Mθ0,θλ]=2. ∎

Now take any z∈Z, the function fz⁢(⋅)≡f⁢(z,⋅), and θl,θh in Claim 7. As rank⁡[Mθl,θh]=2, the system

[fz⁢(θl)fz⁢(θh)]=[f1⁢(θl)f2⁢(θl)f1⁢(θh)f2⁢(θh)]⁢[λ1λ2] (18)

has a unique solution 𝝀∈ℝ2. We will show that fz=λ1⁢f1+λ2⁢f2.

Suppose, towards contradiction, there exists θ𝝀 such that

fz⁢(θ𝝀)≠λ1⁢f1⁢(θ𝝀)+λ2⁢f2⁢(θ𝝀). (19)

Let θ¯ and θ¯ respectively be a lower and an upper bound of {θl,θh,θ𝝀}. If rank⁡[Mθ¯,θ¯]<2, there is 𝝀′∈ℝ2\{0} such that λ1′⁢f1⁢(θ)+λ2′⁢f2⁢(θ)=0 for θ=θ¯,θ¯. As λ1′⁢f1+λ2′⁢f2 is single crossing, we have λ1′⁢f1⁢(θ)+λ2′⁢f2⁢(θ)=0 for θ=θl,θh, which contradicts rank⁡[Mθl,θh]=2.454545The function λ1′⁢f1+λ2′⁢f2 must be single crossing because we can consider μ:Z→ℝ such that μ⁢(z1)=λ1′, μ⁢(z2)=λ2′, and μ⁢(z)=0 for any z≠z1,z2. We use similar reasoning subsequently.

If, on the other hand, rank⁡[Mθ¯,θ¯]=2, the system

[fz⁢(θ¯)fz⁢(θ¯)]=[f1⁢(θ¯)f2⁢(θ¯)f1⁢(θ¯)f2⁢(θ¯)]⁢[λ1′λ2′]

has a unique solution 𝝀′∈ℝ2. As fz−λ1′⁢f1−λ2′⁢f2 is single crossing,

fx⁢(θl) =λ1′⁢f1⁢(θl)+λ2′⁢f2⁢(θl)andfx⁢(θh)=λ1′⁢f1⁢(θh)+λ2′⁢f2⁢(θh), and (20)
fx⁢(θ𝝀) =λ1′⁢f1⁢(θ𝝀)+λ2′⁢f2⁢(θ𝝀). (21)

(20) implies that 𝝀′ solves (18). As the unique solution to (18) was 𝝀, it follows that 𝝀′=𝝀. But then (19) and (21) are in contradiction. Therefore, there exist λ1,λ2:Z→ℝ such that

(∀z,θ)f⁢(z,θ)=λ1⁢(z)⁢f⁢(z1,θ)+λ2⁢(z)⁢f⁢(z2,θ).

B.4 Proof of Theorem 2’s SCD⋆ Characterization

Here we only prove Theorem 2’s characterization of SCD⋆. The proof for its SSCD⋆ characterization is deferred to Supplementary Appendix SA.2.

(⟸)

Suppose that u⁢(a,θ)=g1⁢(a)⁢f1⁢(θ)+g2⁢(a)⁢f2⁢(θ)+h⁢(θ), with f1,f2:Θ→ℝ each single crossing and ratio ordered. Then, for any a,a′∈A, Da,a′⁢(θ)=(g1⁢(a)−g1⁢(a′))⁢f1⁢(θ)+(g2⁢(a)−g2⁢(a′))⁢f2⁢(θ), which is single crossing by Lemma 1.

(⟹)

Assume, without loss of generality, that |A|≥2. Take any a0∈A, and define A′≡A∖{a0}. Define f:A×Θ→ℝ as f⁢(a,θ)≡u⁢(a,θ)−u⁢(a0,θ), which, for every a, is single crossing in θ.

We will show that, for every function μ′:A′→ℝ with finite support, ∫a∈A′f⁢(a,θ)⁢dμ′ can be represented as a multiple of the difference between two convex utility combinations ∫au⁢(a,θ)⁢dP and ∫au⁢(a,θ)⁢dQ. Since the environment is convex, there exist aP,aQ∈A such that u⁢(aP,θ)=∫au⁢(a,θ)⁢dP and u⁢(aQ,θ)=∫au⁢(a,θ)⁢dQ for all θ. Since the utility difference u⁢(aP,θ)−u⁢(aQ,θ) is single crossing, so is ∫a∈A′f⁢(a,θ)⁢dμ′. The result then follows from Proposition 2.

For any function μ′:A′→ℝ with finite support, we define a function μ:A→ℝ as an extension of μ′:

μ⁢(a0)≡−∑{a:μ′⁢(a)≠0}μ′⁢(a), and ⁢(∀a∈A′)⁢μ⁢(a)≡μ′⁢(a).

In a sense, we let a0 absorb the function values on A′. In particular, note that

∑{a:μ⁢(a)≠0}μ⁢(a)=μ⁢(a0)+∑{a:μ′⁢(a)≠0}μ⁢(a)=0.

We construct the Hahn-Jordan decomposition (μ+,μ−) of μ. That is, we define functions μ+,μ−:A→ℝ+ by (∀a∈A) μ+⁢(a)≡max⁡{μ⁢(a),0} and μ−⁢(a)≡−min⁡{μ⁢(a),0}. These are both functions with finite support, and μ=μ+−μ−. Let M≡∑{a:μ⁢(a)≠0}μ+⁢(a)=∑{a:μ⁢(a)≠0}μ−⁢(a). If M=0, pick an arbitrary P∈Δ⁢A with finite support and let Q=P. If M>0, define P,Q∈Δ⁢A with probability mass functions p,q such that for any a∈A,

p⁢(a)=μ+⁢(a)Mandq⁢(a)=μ−⁢(a)M.

Note that P and Q have finite support. Since the environment is convex, there exist aP,aQ∈A such that u⁢(aP,θ)=∫au⁢(a,θ)⁢dP and u⁢(aQ,θ)=∫au⁢(a,θ)⁢dQ for all θ. It follows that

∫a∈A′f⁢(a,θ)⁢dμ′ = ∫a∈Af⁢(a,θ)⁢dμ(because f⁢(a0,θ)=0)
= ∫a∈Au⁢(a,θ)⁢dμ−u⁢(a0,θ)⁢μ⁢(A)
= ∫a∈Au⁢(a,θ)⁢dμ+−∫a∈Au⁢(a,θ)⁢dμ− (as μ⁢(A)=0)
= M⋅(u⁢(aP,θ)−u⁢(aQ,θ)),

which is single crossing.

Thus, if u has SCD⋆, then f:A′×Θ→ℝ is linear combinations SC-preserving. By Proposition 2, there exist a1,a2∈A′ and λ1,λ2:A′→ℝ such that (i) f⁢(a1,θ) and f⁢(a2,θ) are each single crossing and ratio ordered, and (ii) (∀a∈A′) f⁢(a,⋅)=λ1⁢(a)⁢f⁢(a1,⋅)+λ2⁢(a)⁢f⁢(a2,⋅). Hence, there exist functions g1,g2:A→ℝ with g1⁢(a0)=g2⁢(a0)=0 such that (∀a∈A) f⁢(a,⋅)=g1⁢(a)⁢f⁢(a1,⋅)+g2⁢(a)⁢f⁢(a2,⋅), or equivalently,

(∀a,θ)u⁢(a,θ)=g1⁢(a)⁢f⁢(a1,θ)+g2⁢(a)⁢f⁢(a2,θ)+u⁢(a0,θ).

B.5 Proof of Proposition 1

We prove Proposition 1’s result about SCD⋆ and omit an analogous proof for the result about SSCD⋆. If |Θ|≤2, then the proof is trivial, so assume |Θ|≥3. The “if” direction of the result follows directly from Theorem 2: if u has a positive affine transformation u~ of the form in Proposition 1, then u, as a positive affine transformation of u~, has SCD⋆.

For the “only if” direction, take any u that has SCD⋆. Following the form given in Theorem 2, a positive affine transformation of u is

u~⁢(a,θ)=g1⁢(a)⁢f1⁢(θ)+g2⁢(a)⁢f2⁢(θ),

where f1,f2:Θ→ℝ are each single crossing and ratio ordered.

First, we consider the case in which 𝒇⁢(θ¯) and 𝒇⁢(θ¯) are linearly dependent.464646This case can be ignored in the proof for SSCD⋆, because if f1 and f2 are strictly ratio ordered, then 𝒇⁢(θ¯) and 𝒇⁢(θ¯) must be linearly independent. Assume, with a positive affine transformation of u~, that the length of the vector 𝒇⁢(θ)≡(f1⁢(θ),f2⁢(θ)) in ℝ2 is either 0 or 1 for every θ. If 𝒇⁢(θ¯)=𝒇⁢(θ¯)=0, then because f1 and f2 are each single crossing, we have (∀θ) f1⁢(θ)=f2⁢(θ)=0 and (∀a,θ) u~⁢(a,θ)=0. We can easily now rewrite u~ in the form (6), with λ:Θ→[0,1] increasing.

Suppose 𝒇⁢(θ¯)≠0; we omit the analogous proof for the case of 𝒇⁢(θ¯)≠0. By Condition (4) of ratio ordering, for every θ, the vector 𝒇⁢(θ)∈ℝ2 is linearly dependent on 𝒇⁢(θ¯). As (∀θ) ∥𝒇⁢(θ)∥∈{0,1}, there exists λ:Θ→{−1,0,1} such that (∀θ) 𝒇⁢(θ)=λ⁢(θ)⁢𝒇⁢(θ¯). Note that λ is increasing because f1 and f2 are each single crossing. If either λ⁢(θ¯)=0 (and so (∀a)⁢u~⁢(a,θ¯)=0) or λ⁢(θ¯)=1 (and so (∀θ)⁢λ⁢(θ)=1), then

u~⁢(a,θ)=λ⁢(θ)⁢u~⁢(a,θ¯)+(1−λ⁢(θ))⁢u~⁢(a,θ¯),

with the last term equal to zero. If, on the other hand, λ⁢(θ¯)=−1, then

u~⁢(a,θ) =λ⁢(θ)⁢u~⁢(a,θ¯)=λ⁢(θ)+12⁢u~⁢(a,θ¯)+λ⁢(θ)−12⁢(−u~⁢(a,θ¯))
=λ⁢(θ)+12⁢u~⁢(a,θ¯)+(1−λ⁢(θ)+12)⁢u~⁢(a,θ¯).

Next, suppose that the vectors 𝒇⁢(θ¯),𝒇⁢(θ¯)∈ℝ2 are linearly independent, so the angle between the vectors is strictly less than 180 degrees. As f1 and f2 are ratio ordered, for each θ there exists α⁢(θ),β⁢(θ)∈ℝ+ such that

𝒇⁢(θ)=α⁢(θ)⁢𝒇⁢(θ¯)+β⁢(θ)⁢𝒇⁢(θ¯),

or equivalently,

u~⁢(a,θ)=α⁢(θ)⁢u~⁢(a,θ¯)+β⁢(θ)⁢u~⁢(a,θ¯).

By Condition (4), 𝒇⁢(θ)≠0, which implies that α⁢(θ)+β⁢(θ)>0. A positive affine transformation of dividing u~⁢(⋅,θ) by α⁢(θ)+β⁢(θ) results in the form (6), where λ⁢(θ)≡α⁢(θ)α⁢(θ)+β⁢(θ)∈[0,1].

To prove that the function λ is increasing, take θ1,θ2 such that θ¯≤θ1≤θ2≤θ¯. To reduce notation below, let αi≡α⁢(θi) and βi≡β⁢(θi) for i=1,2. We must show that α1α1+β1≤α2α2+β2, or equivalently that α1⁢β2≤α2⁢β1. Suppose f1 ratio dominates f2; the other case is analogous. Then f1⁢(θ1)⁢f2⁢(θ2)≤f1⁢(θ2)⁢f2⁢(θ1), and hence

(α1⁢f1⁢(θ¯)+β1⁢f1⁢(θ¯))⁢(α2⁢f2⁢(θ¯)+β2⁢f2⁢(θ¯))≤(α2⁢f1⁢(θ¯)+β2⁢f1⁢(θ¯))⁢(α1⁢f2⁢(θ¯)+β1⁢f2⁢(θ¯)),

or equivalently,

(α1⁢β2−α2⁢β1)⁢(f1⁢(θ¯)⁢f2⁢(θ¯)−f1⁢(θ¯)⁢f2⁢(θ¯))≤0.

Note that f1⁢(θ¯)⁢f2⁢(θ¯)−f1⁢(θ¯)⁢f2⁢(θ¯)>0 because f1 ratio dominates f2, and f⁢(θ¯) and f⁢(θ¯) are linearly independent. Hence, α1⁢β2≤α2⁢β1.

Appendix C Proofs for MD⋆

The proof of Theorem 3 requires analogs of Lemma 1 and Proposition 2 for monotonic functions; these results are stated next but their proofs are deferred to Supplementary Appendix SA.4.

C.1 Aggregating Monotonic Functions

Lemma 2.

Let f1,f2:Θ→ℝ be monotonic functions. The linear combination α1⁢f1⁢(θ)+α2⁢f2⁢(θ) is monotonic ∀𝛂∈ℝ2 if and only if either f1 or f2 is an affine transformation of the other, i.e., there exists 𝛌∈ℝ2 such that either f2=λ1⁢f1+λ2 or f1=λ1⁢f2+λ2.

We say that f:Z×Θ→ℝ is linear combinations monotonicity-preserving if ∫zf⁢(z,θ)⁢dμ is a monotonic function of θ for every function μ:Z→ℝ with finite support.

Proposition 3.

Let f:Z×Θ→ℝ for some set Z. The function f is linear combinations monotonicity-preserving if and only if there exist z′∈Z and λ1,λ2:Z→ℝ such that (i) f⁢(z′,⋅) is monotonic, and (ii) (∀z)⁢f⁢(z,⋅)=λ1⁢(z)⁢f⁢(z′,⋅)+λ2⁢(z).

C.2 Proof of Theorem 3

(⟸)

We omit the proof as it is similar to the proof of Theorem 2 in Appendix B.4.

(⟹)

The proof is trivial if (∀a,θ) u⁢(a,θ)=0, so assume there exists a0 such that u⁢(a0,⋅):Θ→ℝ is not a zero function. Define f:A×Θ→ℝ by f⁢(a,θ)≡u⁢(a,θ)−u⁢(a0,θ). Note that (∀a) f⁢(a,θ) is a monotonic function of θ.

Let A′≡A\{a0}. As in the proof of Theorem 2 in Appendix B.4, for every μ′:A′→ℝ with finite support, there exist convex utility combinations ∫au⁢(a,θ)⁢dP and ∫au⁢(a,θ)⁢dQ, where P and Q have finite support, such that ∫a∈A′f⁢(a,θ)⁢dμ′ is monotonic if and only if ∫a∈Au⁢(a,θ)⁢dP−∫a∈Au⁢(a,θ)⁢dQ is monotonic. Since the environment is convex, the latter utility difference is indeed monotonic, and so ∫a∈A′f⁢(a,θ)⁢dμ′ is monotonic. By Proposition 3, there exist a′∈A\a0 and λ1,λ2:A\{a0}→ℝ such that (∀a,θ) f⁢(a,θ)=λ1⁢(a)⁢f⁢(a′,θ)+λ2⁢(a). Hence, there exist functions g1,g2:A→ℝ with g1⁢(a0)=g2⁢(a0)=0 such that f⁢(a,θ)=g1⁢(a)⁢f⁢(a′,θ)+g2⁢(a), or equivalently, u⁢(a,θ)=g1⁢(a)⁢f⁢(a′,θ)+g2⁢(a)+u⁢(a0,θ).

References

  • A. E. Abbas and D. E. Bell (2011) One-switch independence for multiattribute utility functions. Operations Research 59 (3), pp. 764–771. Cited by: §1.3, §3.2.1, footnote 21, footnote 28.
  • S. N. Ali, N. Kartik, and A. Kleiner (2022) Sequential veto bargaining with incomplete information. Note: unpublished Cited by: §3.2.1.
  • M. Amador and K. Bagwell (2013) The theory of optimal delegation with an application to tariff caps. Econometrica 81 (4), pp. 1541–1599. Cited by: §1.1, §3.2.1.
  • S. Athey (2002) Monotone comparative statics under uncertainty. Quarterly Journal of Economics 117 (1), pp. 187–223. External Links: Document Cited by: §1.3, §6.
  • A. V. Banerjee (1992) A simple model of herd behavior. The Quarterly Journal of Economics 107 (3), pp. 797–817. Cited by: §4.2.
  • J. S. Banks and J. Duggan (2006) A social choice lemma on voting over lotteries with applications to a class of dynamic games. Social Choice and Welfare 26 (2), pp. 285–304. External Links: Link Cited by: §3.2.1.
  • D. Bergemann and S. Morris (2019) Information design: a unified perspective. Journal of Economic Literature 57 (1), pp. 1–57. Cited by: Example 4.
  • S. Bikhchandani, D. Hirshleifer, and I. Welch (1992) A theory of fads, fashion, custom, and cultural change as informational cascades. Journal of Political Economy 100 (5), pp. 992–1026. Cited by: §4.2.
  • A. Blume, O. J. Board, and K. Kawamura (2007) Noisy talk. Theoretical Economics 2 (4), pp. 395–440. Cited by: §4.1.
  • G. Carroll (2012) When are local incentive constraints sufficient?. Econometrica 80 (2), pp. 661–686. External Links: ISSN 00129682, 14680262, Link Cited by: footnote 43.
  • G. Celik (2015) Implementation by gradual revelation. RAND Journal of Economics 46 (2), pp. 271–296. External Links: Document, Link Cited by: §1.3, §3.2.1.
  • V. Crawford and J. Sobel (1982) Strategic information transmission. Econometrica 50 (6), pp. 1431–1451. Cited by: §1.1, §4.1, §4.1, §4.1, footnote 34.
  • E. Diecidue and P. P. Wakker (2001) On the intuition of rank-dependent utility. Journal of Risk and Uncertainty 23 (3), pp. 281–298. External Links: Link Cited by: Example 2.
  • J. Duggan (2014) Majority voting over lotteries: conditions for existence of a decisive voter. Economics Bulletin 34 (1), pp. 263–270. Cited by: §1.3, §3.2.1, footnote 41.
  • P. C. Fishburn (1974) Von neumann-morgenstern utility functions on two attributes. Operations Research 22 (1), pp. 35–45. Cited by: §1.3.
  • J. S. Gans and M. Smart (1996) Majority voting with single-crossing preferences. Journal of Public Economics 59 (2), pp. 219–237. External Links: Document Cited by: §4.3.
  • J. C. Harsanyi (1955) Cardinal welfare, individualistic ethics, and interpersonal comparisons of utility. Journal of Political Economy 63 (4), pp. 309–321. Cited by: footnote 25.
  • M. O. Jackson and L. Yariv (2015) Collective dynamic choice: the necessity of time inconsistency. American Economic Journal: Microeconomics 7 (4), pp. 150–78. External Links: Document, Link Cited by: footnote 25.
  • J. Jaffray (1975) Semicontinuous extension of a partial order. Journal of Mathematical Economics 2 (3), pp. 395–406. External Links: Document, Link Cited by: footnote 10.
  • E. Kamenica and M. Gentzkow (2011) Bayesian persuasion. American Economic Review 101 (6), pp. 2590–2615. Cited by: Example 4.
  • S. Karlin (1968) Total positivity. Vol. I, Stanford University Press, Stanford, California. Cited by: §1.3, §3.3, footnote 29.
  • N. Kartik, S. Lee, T. Liu, and D. Rappoport (2022) Beyond unbounded beliefs: how preferences and information interplay in social learning. Note: unpublished Cited by: §4.2.
  • N. Kartik, S. Lee, and D. Rappoport (2019) Single-crossing differences on distributions. Note: unpublished Cited by: §4, footnote 29, footnote 41, footnote 6.
  • A. Kleiner (2022) Optimal delegation in a multidimensional world. Note: unpublished Cited by: §3.2.1.
  • A. Kushnir and S. Liu (2018) On the equivalence of bayesian and dominant strategy implementation for environments with nonlinear utilities. Economic Theory 67 (3), pp. 617–644. Cited by: §1.3, §5.1.
  • N. D. Melumad and T. Shibano (1991) Communication in settings with no transfers. RAND Journal of Economics 22 (2), pp. 173–198. Cited by: §4.1, §4.1, footnote 34.
  • P. R. Milgrom (1981) Good news and bad news: representation theorems and applications. Bell Journal of Economics 12 (2), pp. 380–391. Cited by: footnote 12.
  • P. R. Milgrom (2004) Putting auction theory to work. Cambridge University Press. Cited by: §2.
  • P. Milgrom and C. Shannon (1994) Monotone comparative statics. Econometrica 62 (1), pp. 157–180. Cited by: §A.2, §1.1, §1.1, §1.3, §2, §5.2, §5.2, §5.2, footnote 44.
  • E. A. Ok (2007) Real analysis with economic applications. Vol. 10, Princeton University Press. Cited by: footnote 10.
  • J. K. Quah and B. Strulovici (2012) Aggregating the single crossing property. Econometrica 80 (5), pp. 2333–2348. Cited by: §1.3, §3.3.
  • J. Quiggin (1982) A theory of anticipated utility. Journal of Economic Behavior & Organization 3 (4), pp. 323–343. External Links: Document, Link Cited by: §1.1, Example 2.
  • P. Rothstein (1990) Order restricted preferences and majority rule. Social Choice and Welfare 7 (4), pp. 331–342. External Links: Link Cited by: §1.1, §2, §4.3.
  • D. J. Seidmann (1990) Effective cheap talk with conflicting interests. Journal of Economic Theory 50 (2), pp. 445–458. Cited by: §4.1.
  • K. A. Shepsle (1972) The strategy of ambiguity: uncertainty and electoral competition. American Political Science Review 66 (2), pp. 555–568. Cited by: §1.1, §4.3.
  • L. Smith (2011) Frictional matching models. Annual Review of Economics 3 (1), pp. 319–338. External Links: Document, https://doi.org/10.1146/annurev-economics-111809-125046, Link Cited by: footnote 5.
  • D. M. Topkis (1978) Minimizing a submodular function on a lattice. Operations Research 26 (2), pp. 305–321. Cited by: §1.3.
  • R. Zeckhauser (1969) Majority rule with lotteries on alternatives. Quarterly Journal of Economics 83 (4), pp. 696–703. Cited by: §1.1, §4.3.
HTML from LaTeXML, with custom CSS/JS. The PDF is more accurate.