Econometrica, Vol. 90, No.

6 (November, 2022), 2795–2819

THE CONVERSE ENVELOPE THEOREM

LUDVIG SINANDER
Department of Economics and Nuffield College, University of Oxford

I prove an envelope theorem with a converse: the envelope formula is equivalent to

a first-order condition. Like Milgrom and Segal’s (2002) envelope theorem, my result
requires no structure on the choice set. I use the converse envelope theorem to extend
to general outcomes and preferences the canonical result in mechanism design that any
increasing allocation is implementable, and apply this to selling information.
KEYWORDS: Envelope theorem, first-order condition, mechanism design.

1. INTRODUCTION
ENVELOPE THEOREMS are a key tool of economic theory, with important roles in con-
sumer theory, mechanism design, and dynamic optimization. In blueprint form, an en-
velope theorem gives conditions under which optimal decision-making implies that the
envelope formula holds.
In textbook accounts,1 the envelope theorem is typically presented as a consequence
of the first-order condition. The modern envelope theorem of Milgrom and Segal (2002),
however, applies in an abstract setting in which the first-order condition is typically not
even well-defined. These authors therefore rejected the traditional intuition and devel-
oped a new one.
In this paper, I reestablish the intuitive link between the envelope formula and the
first-order condition. I introduce an appropriate generalized first-order condition that is
well-defined in the abstract environment of Milgrom and Segal (2002), then prove an
envelope theorem with a converse: my generalized first-order condition is equivalent to
the envelope formula. This validates the habitual interpretation of the envelope formula
as “local optimality,” and clarifies our understanding of the envelope theorem.
The converse envelope theorem proves useful for mechanism design. I use it to estab-
lish that the implementability of all increasing allocations, a canonical result when out-
comes are drawn from an interval of R, remains valid when outcomes are abstract. I apply
this result to the problem of selling information (distributions of posteriors).
The setting is simple: an agent chooses an action x from a set X to maximize f (x t),
where t ∈ [0 1] is a parameter. The set X need not have any structure. A decision rule
is a map X : [0 1] → X that assigns an action X (t) to each parameter t. A decision
rule X is associated with a value function VX (t) := f (X (t) t), and is called optimal iff
VX (t) = maxx∈X f (x t) for every parameter t.
The modern envelope theorem of Milgrom and Segal (2002) states that, under a reg-
ularity assumption on f , any optimal decision rule X induces an absolutely continuous

Ludvig Sinander: [email protected]

I am grateful to Eddie Dekel, Alessandro Pavan, and Bruno Strulovici for their guidance and support. This
work has profited from the close reading and insightful comments of Gregorio Curello, Eddie Dekel, Roberto
Saitto, Quitzé Valenzuela-Stookey, and four anonymous referees, and from comments by Piotr Dworczak,
Matteo Escudé, Daniel Gottlieb, Elliot Lipnowski, Benny Moldovanu, Ilya Segal, and audiences at Caltech,
Northwestern, Oxford, the Bonn Winter Theory Workshop, the Kansas Workshop in Economic Theory, and
the Southeast Theory Festival.
1
For example, Mas-Colell, Whinston, and Green (1995, §M.L).
© 2022 The Author. Econometrica published by John Wiley & Sons Ltd on behalf of The Econometric Society.
Ludvig Sinander is the corresponding author on this paper. This is an open access article under the terms of
the Creative Commons Attribution-NonCommercial-NoDerivs License, which permits use and distribution in
any medium, provided the original work is properly cited, the use is non-commercial and no modifications or
adaptations are made.
2796 LUDVIG SINANDER

value function VX which satisfies the envelope formula

 
VX (t) = f2 X (t) t for a.e. t ∈ (0 1)

The familiar intuition is as follows. The derivative of the value VX is


d    

VX (t) = f X (t + m) t  + f2 X (t) t 
dm m=0

where the first term is the indirect effect via the induced change of the action, and the
second term is the direct effect. Since X is optimal, it satisfies the first-order condition
d
dm
f (X (t + m) t)|m=0 = 0, which yields the envelope formula. Indeed, a decision rule X
satisfies the envelope formula if and only if it satisfies the first-order condition for a.e.
t ∈ (0 1).
The trouble with this intuition is that since the action set X is abstract (with no linear
or topological structure), the derivative dmd
f (X (t + m) t)|m=0 is ill-defined in general.
To restore the equivalence of the envelope formula and first-order condition, I first
introduce a generalized first-order condition that is well-defined in the abstract environ-
ment. The outer first-order condition is the following “integrated” variant of the classical
first-order condition:
 t 
d   
f X(s + m) s ds = 0 for all r t ∈ (0 1)
dm r m=0

I then prove an envelope theorem with a converse: under a regularity assumption on f , a

decision rule X satisfies the envelope formula if and only if it satisfies the outer first-order
condition and induces an absolutely continuous value function VX . The “only if” part is a
novel converse envelope theorem.
In §4, I apply the converse envelope theorem to mechanism design. There is an agent
with preferences over outcomes y ∈ Y and payments p ∈ R. Her preferences are indexed
in “single-crossing” fashion by t ∈ [0 1], and this taste parameter is privately known to her.
A canonical result is that if Y is an interval of R, then all (and only) increasing allocations
Y : [0 1] → Y can be implemented incentive-compatibly by some payment schedule P :
[0 1] → R.
I use the converse envelope theorem to extend this result to a large class of ordered out-
come spaces Y , maintaining general (nonquasilinear) preferences. The argument runs as
follows: fix an increasing allocation Y : [0 1] → Y . To implement it, choose a payment
schedule P : [0 1] → R to make the envelope formula hold. Then by the converse en-
velope theorem, the outer first-order condition is satisfied, which means intuitively that
(Y P) is locally incentive-compatible. The single-crossing property of preferences ensures
that this translates into global incentive-compatibility.
I apply this implementability theorem to study the sale of information. The result im-
plies that any Blackwell-increasing information allocation is implementable. I argue fur-
ther that if consumers can share their information with each other, then only Blackwell-
increasing allocations are implementable.

1.1. Related Literature

Envelope theorems entered economics via the theories of the consumer and of the firm
(Hotelling (1932), Roy (1947), Shephard (1953)), were systematized by Samuelson (1947)
 THE CONVERSE ENVELOPE THEOREM 2797

under “classical” assumptions, and were developed in greater generality by, for example,
Danskin (1966, 1967), Silberberg (1974), and Benveniste and Scheinkman (1979). Mil-
grom and Segal (2002) pointed out that classical-type assumptions were extraneous, and
proved an envelope theorem without them. Subsequent refinements were obtained by,
for example, Morand, Reffett, and Tarafdar (2015) and Clausen and Strub (2020).2 “Con-
verse” envelope theorems are almost absent from this literature, but appear in textbook
presentations (e.g., Mas-Colell, Whinston, and Green (1995, §M.L)).
The outer first-order condition appears to be novel. It bears no clear relationship to any
of the standard derivatives for nonsmooth functions.

2. SETTING AND BACKGROUND

In this section, I introduce the environment, the Milgrom–Segal (2002) envelope theo-
rem, and the classical envelope theorem and converse.

NOTATION: We will be working with the unit interval [0 1], equipped with the Lebesgue
σ-algebra and the Lebesguemeasure. The Lebesgue integral will  r be used
 tthroughout. For
t
r < t in [0 1], we will write r for the integral over [r t], and t for − r . L1 will denote
the space of integrable functions [0 1] → R, that is, those that are measurable and have
finite integral. We will write fi for the derivative of a function f with respect to its ith argu-
ment. Some important definitions and theorems are collected in Appendix A.1, including
Lebesgue’s fundamental theorem of calculus and the Vitali convergence theorem.

2.1. Setting
An agent chooses an action x from an arbitrary set X . Her objective is f (x t), where
t ∈ [0 1] is a parameter (or “type”).3

DEFINITION 1: A family {φx}x∈X of functions [0 1] → R is absolutely equicontinuous iff

the family of functions
  
 φx (t + m) − φx (t) 
t → sup 
x∈X m 
m>0

is uniformly integrable.4

Our only assumptions will be that the objective varies smoothly, and (uniformly) not
too erratically, with the parameter.

BASIC ASSUMPTIONS: f (x ·) is differentiable for every x ∈ X , and the family {f (x ·)}x∈X
is absolutely equicontinuous.

2
See also Oyama and Takenawa (2018).
3
If instead the parameter lives in a normed vector space, then the analysis applies unchanged to path deriva-
tives (as Milgrom and Segal (2002, footnote 7) point out).
4
The name “absolute equicontinuity” is inspired by the AC–UI lemma in Appendix A.1, which states that
absolute continuity of a continuous φ is equivalent to uniform integrability of the “divided-difference” family
{t → [φ(t + m) − φ(t)]/m}m>0 . As the term suggests, an absolutely equicontinuous family is equicontinuous,
and its members are absolutely continuous functions; this is proved in Appendix A.2.
2798 LUDVIG SINANDER

REMARK 1: An easy-to-check sufficient condition for absolute equicontinuity is as fol-

lows: f (x ·) is absolutely continuous for each x ∈ X , and there is an  ∈ L1 such that
|f2 (x t)| ≤ (t) for all x ∈ X and t ∈ (0 1). (This is the assumption that Milgrom and
Segal (2002) use in their envelope theorem.) An even stronger sufficient condition is that
f2 be bounded.

EXAMPLE 1: Let X = [0 1] and f (x t) = xt. The basic assumptions are satisfied since
f2 (x t) = x exists and is bounded.


A decision rule is a map X : [0 1] → X that prescribes an action for each type. The
payoff of type t from following decision rule X is denoted VX (t) := f (X (t) t).

DEFINITION 2: A decision rule X satisfies the envelope formula iff

 t
 
VX (t) = VX (0) + f2 X (s) s ds for every t ∈ [0 1]
0

Equivalently (by Lebesgue’s fundamental theorem of calculus), X satisfies the envelope

formula iff VX is absolutely continuous and
 
VX (t) = f2 X (t) t for a.e. t ∈ (0 1)
A decision rule X is called optimal iff at every parameter t ∈ [0 1], X (t) maximizes
f (· t) on X . The modern envelope theorem is as follows.

MILGROM–SEGAL ENVELOPE THEOREM: Under the basic assumptions, if X is optimal,

then it satisfies the envelope formula.

This follows from the main theorem (§3.2 below), so no proof is necessary. It is actually
a slight refinement of Theorem 2 in Milgrom and Segal (2002), as these authors impose
the sufficient condition in Remark 1 rather than absolute equicontinuity.
t
EXAMPLE 1—Continued: The envelope formula requires that X (t)t = 0 X for every
t
t ∈ [0 1], or equivalently X(t) = t −1 0 X for all t ∈ (0 1]. Thus the decision rules that
satisfy the envelope formula are precisely those that are constant on (0 1]. This includes
all optimal decision rules (which set X = 1 on (0 1]), as well as pessimal ones (which
choose 0 on (0 1]).


2.2. Classical Envelope Theorem and Converse

The textbook version of the envelope theorem, which has a natural and intuitive con-
verse, holds under additional topological and convexity assumptions.

CLASSICAL ASSUMPTIONS: The action set X is a convex subset of Rn , the action derivative
f1 exists and is bounded, and only Lipschitz continuous decision rules X are considered.

The classical assumptions are strong. Most glaringly, the Lipschitz condition rules out
important decision rules in many applications. In the canonical auction setting, for in-
stance, the revenue-maximizing mechanism is discontinuous (Myerson (1981)).5

5
Even when the classical assumptions are relaxed as much as possible, unless f is trivial, X still has to satisfy
a strong continuity requirement; see Appendix A.7.
 THE CONVERSE ENVELOPE THEOREM 2799

EXAMPLE 1—Continued: X = [0 1] is a convex subset of R, and f1 (x t) = t exists and

is bounded. If we restrict attention to Lipschitz continuous decision rules X : [0 1] →
[0 1], then the classical assumptions are satisfied.


Given a Lipschitz continuous decision rule X, suppose that type t considers taking the
action X (t + m) intended for another type. The map m → f (X (t + m) t) is differentiable
a.e. under the classical assumptions,6 so we may define a first-order condition.

DEFINITION 3: A decision rule X satisfies the first-order condition a.e. iff


d  
f X(t + m) t  = 0 for a.e. t ∈ (0 1)
dm m=0

The first-order condition a.e. requires that almost no type t can secure a first-order
payoff increase (or decrease) by choosing an action X (t + m) intended for a nearby type
t + m. It does not say that there are no nearby actions that do better (or worse).

CLASSICAL ENVELOPE THEOREM AND CONVERSE: Under the basic and classical as-
sumptions, a Lipschitz continuous decision rule satisfies the first-order condition a.e. iff it
satisfies the envelope formula.

The proof, given in Appendix A.7, shows that the envelope formula demands precisely
that VX (t) = f2 (X(t) t) for a.e. t ∈ (0 1), which is equivalent to the first-order condition
a.e. by inspection of the differentiation identity

d    
VX (t) = f X (t + m) t  + f2 X (t) t 
dm m=0

EXAMPLE 1—Continued: A Lipschitz continuous decision rule X is differentiable a.e.,

so satisfies the first-order condition a.e. iff

d 
X(t + m)t  = X  (t)t = 0 for a.e. t ∈ (0 1)
dm m=0

This requires that X  = 0 a.e. We saw that the envelope formula demands that X be con-
stant on (0 1]. For Lipschitz continuous decision rules X, both conditions are equivalent
to constancy on all of [0 1].


3. MAIN THEOREM
In this section, I define the outer first-order condition and state my envelope theorem
and converse.

3.1. The Outer First-Order Condition

Without the classical assumptions (§2.2), the “imitation derivative”

d  
f X (t + m) t 
dm m=0

6
Since f (· t) is differentiable, and X is differentiable a.e. since it is Lipschitz continuous.
2800 LUDVIG SINANDER

need not exist, in which case the first-order condition is ill-defined. To circumvent this
problem, we require a novel first-order condition.

DEFINITION 4: A decision rule X satisfies the outer first-order condition iff

 t 
d   
f X(s + m) s ds = 0 for all r t ∈ (0 1)
dm r m=0

As an intuitive motivation, suppose that types s ∈ [r t] deviate

t by choosing X(s + m)
rather than X(s). The aggregate payoff to such a deviation is r f (X (s + m) s) ds, and the
outer first-order condition says (loosely) that local deviations of this kind are collectively
unprofitable.

EXAMPLE 1—Continued: For any decision rule X that is a.e. constant at some k ∈
[0 1], the outer first-order condition holds:
 t   t 
d  d 

X (s + m)s ds = k s ds = 0 for all r t ∈ (0 1)
dm r m=0 dm r m=0

Conversely, any decision rule that is not constant a.e. violates the outer first-order condi-
tion.


As we shall see, the outer first-order condition is well-defined even when the classical
assumptions fail. When they do hold, the outer first-order condition coincides with the
first-order condition a.e.

HOUSEKEEPING LEMMA: Under the basic and classical assumptions, the outer first-order
condition is equivalent to the first-order condition a.e.

PROOF: Fix a Lipschitz continuous decision rule X : [0 1] → X . The family

    
f X(t + m) t − f X (t) t
t →
m m>0

is convergent a.e. as m ↓ 0 by the classical assumptions, and is uniformly integrable by

Lemma 4 in Appendix A.6. Hence by the Vitali convergence theorem, for any r t ∈ (0 1),
 t   t 
d    d  
f X(s + m) s ds = f X (s + m) s  ds
dm r m=0 r dm m=0

The left-hand side (right-hand side) is zero for all r t ∈ (0 1) iff the outer first-order
condition (first-order condition a.e.) holds.7 Q.E.D.

The term “outer” is inspired by this argument. By taking the differentiation operator
outside the integral, we change nothing in the classical case, and ensure existence beyond
the classical case.
As its name suggests, the outer first-order condition is necessary (but not sufficient) for
optimality. The following is proved in Appendix A.5.

7
For the right-hand side, this relies on the following basic fact (e.g., Proposition 2.23(b) in Folland (1999)):
t
for φ ∈ L1 , we have φ = 0 a.e. iff r φ = 0 for all r t ∈ (0 1).
 THE CONVERSE ENVELOPE THEOREM 2801

NECESSITY LEMMA: Under the basic assumptions, any optimal decision rule X satisfies
the outer first-order condition, and has VX (t) := f (X (t) t) absolutely continuous.

3.2. Envelope Theorem and Converse

My main result characterizes the envelope formula in terms of the outer first-order
condition.

ENVELOPE THEOREM AND CONVERSE: Under the basic assumptions, for a decision rule
X : [0 1] → X , the following are equivalent:
(i) X satisfies the outer first-order condition
 t 
d   
f X (s + m) s ds = 0 for all r t ∈ (0 1)
dm r m=0

and VX (t) := f (X(t) t) is absolutely continuous.

(ii) X satisfies the envelope formula
 t
 
VX (t) = VX (0) + f2 X (s) s ds for every t ∈ [0 1]
0

The implication (i) =⇒ (ii) is an envelope theorem with weak (purely local) assump-
tions; the Milgrom–Segal and classical envelope theorems in §2 are corollaries. The impli-
cation (ii) =⇒ (i) is the converse envelope theorem, which entails the classical converse
envelope theorem in §2.2.
The absolute-continuity-of-VX condition in (i) ensures that f (X(·) t) does not behave
too erratically near t. A characterization of this property is provided in Appendix A.4.

EXAMPLE 1—Continued: We saw that a decision rule satisfies the envelope formula
iff it is constant on (0 1] (p. 2798), and satisfies the outer first-order condition iff it is
constant a.e. (p. 2800). Thus the envelope formula implies the outer first-order condi-
tion. For the other direction, observe that an a.e. constant X for which VX (t) = X (t)t is
(absolutely) continuous must in fact be constant on (0 1], though not necessarily at zero.


In the classical case (§2.2), our proof relied on the differentiation identity

d    
VX (t) = f X (t + m) t  + f2 X (t) t 
dm m=0

or (rearranged and integrated)

   t
t
d    

f X(s + m) s  ds = VX (t) − VX (r) − f2 X (s) s ds
r dm m=0 r

To pursue an analogous proof, we require an “outer” version of this identity in which dif-
ferentiation and integration are interchanged on the left-hand side. The following lemma,
proved in Appendix A.3, does the job.
2802 LUDVIG SINANDER

IDENTITY LEMMA: Under the basic assumptions, if VX is absolutely continuous, then for
all r t ∈ (0 1),
 t   t
d     
f X(s + m) s ds = VX (t) − VX (r) − f2 X (s) s ds (I )
dm r m=0 r

where both sides are well-defined.

The left-hand side of (I ) is zero for all r t ∈ (0 1) iff the outer first-order condition
holds. The right-hand side is zero for all r t ∈ (0 1) iff the envelope formula holds.8
Therefore, we have the following.

PROOF OF THE ENVELOPE THEOREM AND CONVERSE: Suppose that the outer first-
order condition holds and that VX is absolutely continuous. Then the identity lemma
applies, so the outer first-order condition implies the envelope formula.
Suppose that the envelope formula holds. Then VX is absolutely continuous by
Lebesgue’s fundamental theorem of calculus. Hence the identity lemma applies, so the
envelope formula implies the outer first-order condition. Q.E.D.

4. APPLICATION TO MECHANISM DESIGN

A key result in mechanism design is that, provided the agent’s preferences are “single-
crossing,” all and only increasing allocations are implementable. While the “only” part is
straightforward, the “all” part has substance. Existing theorems of this sort require that
outcomes be drawn from an interval of R or that the agent have quasilinear preferences.
In this section, I use the converse envelope theorem to extend this result to abstract
spaces of outcomes, without requiring quasilinearity. I then apply it to the problem of
selling information, showing that all (and only) Blackwell-increasing information alloca-
tions are implementable (and robust to collusion).

4.1. Environment and Existing Results

There is a partially ordered set Y of outcomes. A single agent has preferences over
outcomes y ∈ Y and payments p ∈ R represented by f (y p t), where the type t ∈ [0 1] is
privately known to the agent.9 We assume that f (y · t) is strictly decreasing and onto R
for all y ∈ Y and t ∈ [0 1].
A direct mechanism is a pair of maps Y : [0 1] → Y and P : [0 1] → R that assign
an outcome and a payment to each type. A direct mechanism (Y P) is called incentive-
compatible iff no type strictly prefers the outcome–payment pair designated for another
type:
   
f Y (t) P (t) t ≥ f Y (r) P (r) t for all r t ∈ [0 1]
By a revelation principle, it is without loss of generality to restrict attention to incentive-
compatible direct mechanisms. An allocation Y : [0 1] → Y is called implementable iff
there is a payment schedule P : [0 1] → R such that (Y P) is incentive-compatible.10 An

8
For the “only if” part, if right-hand side is zero for all r t ∈ (0 1), then it is zero for all r t ∈ [0 1] since VX
and the integral are continuous, yielding the envelope formula.
9
All of the analysis carries over to the case of multiple agents with independent types.
10
Adding an individual rationality constraint does not change our results below.
 THE CONVERSE ENVELOPE THEOREM 2803

increasing allocation is one that provides higher types with larger outcomes (in the partial
order on Y ).
Preferences f are called single-crossing iff higher types are more willing to pay to in-
crease y ∈ Y . The details of how this is formalized vary from paper to paper. We are
interested in the following type of result.

THEOREM SCHEMA: If Y and f are “regular” and f is “single-crossing,” then any increas-
ing allocation is implementable.

The first result of this kind was obtained by Mirrlees (1976) and Spence (1974) un-
der the assumptions that Y is an interval of R and that f has the quasilinear form
f (y p t) = h(y t) − p. Maintaining quasilinearity, the result was extended to multidi-
mensional Euclidean Y by Matthews and Moore (1987) and García (2005),11 and may
be further extended to arbitrary Y via a standard argument. (That argument relies crit-
ically on quasilinearity; see the Online Supplemental Material, Appendix S.1 (Sinander
(2022)).) With Y an interval of R, the result was obtained without quasilinearity by Gues-
nerie and Laffont (1984) under classical assumptions,12 and by Nöldeke and Samuelson
(2018) assuming only that f is (jointly) continuous.
I shall extend the result to a wide class of outcome spaces Y , without imposing quasilin-
earity. I formulate notions of “regularity” and “single-crossing” in the next section, then
establish the implementability of increasing allocations in §4.3.

4.2. Regularity and Single-Crossing

Recall that a subset C ⊆ Y is called a chain iff it is totally ordered.

DEFINITION 5: The outcome space Y is regular iff it is order-dense-in-itself, countably

chain-complete and chain-separable.13

In words, Y must be “rich” (first two assumptions) and “not too large” (final assump-
tion). Many important spaces enjoy these properties, including Rn with the usual (prod-
uct) order, the space of finite-expectation random variables (on some probability space)
ordered by “a.s. smaller,” and the space of distributions of posteriors updated from a
given prior ordered by Blackwell informativeness. I prove these assertions and give fur-
ther examples in the Online Supplemental Material, Appendix S.2.

DEFINITION 6: The payoff f is regular iff (a) the type derivative f3 exists and is bounded,
and f3 (y · t) is continuous for each y ∈ Y and t ∈ [0 1], and (b) for every chain C ⊆ Y ,
f is jointly continuous on C × R × [0 1] when C has the relative topology inherited from
the order topology on Y .1415

11
Results of this type have been used to study sequential screening (e.g., Courty and Li (2000), Battaglini
(2005), Eső and Szentes (2007), and Pavan, Segal, and Toikka (2014)).
12
These authors restricted attention to piecewise continuously differentiable allocations; Milgrom (2004,
Theorem 4.2) generalized to piecewise absolutely continuous allocations.
13
A set A partially ordered by  is order-dense-in-itself iff for any a < a in A, there is a b ∈ A such that
a < b < a . B ⊆ A is order-dense in C ⊆ A iff for any c < c  in C, there is a b ∈ B such that c  b  c  . A is
chain-separable iff for each chain C ⊆ A, there is a countable set B ⊆ A that is order-dense in C. A is countably
chain-complete iff every countable chain in A with a lower (upper) bound in A has an infimum (a supremum)
in A.
14
The order topology on Y is the one generated by the open order rays {y  ∈ Y : y  < y} and {y  ∈ Y : y < y }
for each y ∈ Y , where < denotes the strict part of the order on Y .
15
It is sufficient, but unnecessarily strong, to assume joint continuity on Y × R × [0 1].
2804 LUDVIG SINANDER

The joint continuity requirement corresponds to Nöldeke and Samuelson’s (2018) regu-
larity assumption. By demanding in addition that the type derivative exist and be bounded,
I ensure that when this model is embedded in the general setting of §2.1 by letting
X := Y × R, the basic assumptions are satisfied. The converse envelope theorem is thus
applicable.16
It remains to formalize “single-crossing,” the idea that higher types are more will-
ing to pay to increase y ∈ Y . Under the classical assumptions, this is captured by the
Spence–Mirrlees condition, which demands that for any increasing Y : [0 1] → Y and any
P : [0 1] → R (both Lipschitz continuous), for any type s ∈ (0 1), the marginal gain to
mimicking

d  
f Y (s + m) P (s + m) s + n 
dm m=0
1718
be single-crossing in n. To extend this definition beyond the classical case to general
outcomes Y (and non-Lipschitz mechanisms (Y P)), I replace the (typically ill-defined)
marginal mimicking gain with its “outer” version.

DEFINITION 7: f satisfies the (strict) outer Spence–Mirrlees condition iff for any increas-
ing Y : [0 1] → Y , any P : [0 1] → R and any r < t in (0 1),
 t 
d   
n → f Y (s + m) P (s + m) s + n ds
dm r 
m=0

is (strictly) single-crossing, where d/dm denotes the upper derivative.19

The difference from the classical Spence–Mirrlees condition is merely technical: the
interpretation is the same, namely that on the margin, higher types have a greater willing-
ness to pay for increasing the outcome y ∈ Y . It is worth noting, however, that whereas
the classical Spence–Mirrlees condition is (nearly) ordinal,20 the outer Spence–Mirrlees
condition is not.

4.3. Increasing Allocations Are Implementable

IMPLEMENTABILITY THEOREM: If Y and f are regular and f satisfies the outer Spence–
Mirrlees condition, then any increasing allocation is implementable.

The proof is in Appendix B.1. The idea is as follows. Take any increasing allocation
Y : [0 1] → Y . By the existence lemma in Appendix B.1.1, there exists a payment schedule

16
The continuity of f3 (y · t) plays a technical role in the proof; see footnote 21 below.
17
Given T ⊆ R, a function φ : T → R is called single-crossing iff for any t < t  in T , φ(t) ≥(>) 0 implies
φ(t ) ≥(>) 0, and strictly single-crossing iff φ(t) ≥ 0 implies φ(t  ) > 0.

18
An equivalent definition of the Spence–Mirrlees condition requires instead that the slope
f1 (y p t)/|f2 (y p t)| of the agent’s indifference curve through any point (y p) ∈ Y × R be increasing in t.
See Milgrom and Shannon (1994, Theorem 3) for a proof of equivalence.
19
The upper derivative of φ : [0 1] → R at t ∈ (0 1) is dm d
φ(t + m)|m=0 := lim supm→0 [φ(t + m) − φ(t)]/m.
Nothing changes in the sequel if the upper derivative is replaced with the lower (defined with a lim inf), or with
any of the four Dini derivatives.
20
Precisely: if f satisfies this condition, then so does φ ◦ f for any differentiable and strictly increasing
transformation φ : R → R.
 THE CONVERSE ENVELOPE THEOREM 2805

P : [0 1] → R such that (Y P) satisfies the envelope formula.21 By the converse envelope
theorem, it follows that (Y P) is locally incentive-compatible in the sense that it satisfies
the outer first-order condition. The outer Spence–Mirrlees condition ensures that local
incentive-compatibility translates into global incentive-compatibility.
The argument for the final step actually applies only to allocations Y that are suitably
continuous. But the regularity of Y ensures (via a lemma in Appendix B.1.2) that any
increasing Y can be approximated by a sequence of continuous and increasing (hence
implementable) allocations.
Given two mild additional assumptions, the payment rule implementing a given increas-
ing allocation is in fact unique, and may be computed constructively via Picard’s method;
see Appendix B.1.1.
The implementability theorem admits a standard converse when Y is a chain (e.g., an
interval of R), proved in Appendix B.2.

PROPOSITION 1: If Y and f are regular, f satisfies the strict outer Spence–Mirrlees condi-
tion, and Y is a chain, then all and only increasing allocations are implementable.

4.4. Selling Information

In this section, I apply the implementability theorem to selling informative signals.
Here, the outcomes Y are distributions of posterior beliefs—a space very different from
an interval of R. I show that all Blackwell-increasing information allocations are imple-
mentable, and that only these are implementable if agents are able to share information
with each other.
There is a population of agents with types t ∈ [0 1], a finite set  of states of the world,
and a set A of actions. A type-t agent earns payoff U (a ω t) if she takes action a ∈ A in
state ω ∈ , so her expected value at belief μ ∈ () is

V (μ t) := sup U (a ω t)μ(ω)

a∈A
ω∈

Assume that the type derivative V2 exists and is bounded, and that V2 (· t) is continuous
for each t ∈ [0 1].22

EXAMPLE 2: Each agent is tasked with announcing a probabilistic forecast a ∈ A :=

() of the state ω ∈ . Ex post, the public’s assessment of an agent’s quality as a fore-
caster is some function of the forecast a and realized state ω (a scoring rule); for concrete-
ness, a(ω)/ a 2 , where · 2 denotes the Euclidean norm.23 Each agent attaches some im-
portance t ∈ [0 1] to being considered a good forecaster, so that U (a ω t) = ta(ω)/ a 2 .
Agents are expected-utility maximizers.

21
This is where the continuity of f3 (y · t) is used: the existence lemma requires it.
22
This is slightly stronger than assuming that the underlying type derivative U3 has the same properties; see,
for example, Milgrom and Segal (2002, Theorem 3) for sufficient conditions.
23
More generally, any bounded and strictly proper scoring rule will do. See, for example, Gneiting and
Raftery (2007) for an introduction to proper scoring rules.
2806 LUDVIG SINANDER

It is easily verified that an agent with belief μ ∈ () optimally announces forecast
a = μ. Her value is therefore
tμ(ω)
V (μ t) = μ(ω) = t μ 2 
ω∈
μ 2

By inspection, V2 (μ t) = μ exists, is bounded, and is continuous in μ.


2

Agents share a common prior μ0 ∈ int (). Before making her decision, an agent
observes the realization of a signal (a random variable correlated with ω), and forms a
posterior belief according to Bayes’s rule. Since the signal is random, the agent’s posterior
is random; write y for its distribution (a Borel probability measure on ()). The agent’s
expected payoff under a signal that induces posterior distribution y, if she makes payment
p ∈ R, is

f (y p t) := g V (μ t)y(dμ) p 
()

where g : R → R is jointly continuous, possesses a bounded derivative g1 that is contin-

2

uous in p, and has g(v ·) strictly decreasing and onto R for each v ∈ R. The payoff f
is regular: f3 exists, is bounded, and is continuous in p, and I verify the joint continuity
property in the Online Supplemental Material, Appendix S.4.
A Borel probability measure y on () is the distribution of posteriors induced by
some signal exactly if its mean () μy(dμ) is equal to μ0 .24 Write Y for the set of all
mean-μ0 distributions of posteriors, and order it by Blackwell informativeness: y  y  iff
 
v dy ≤ v dy 
() ()

for every continuous and convex v : () → R.25 I show in the Online Supplemental Ma-
terial, Appendix S.2 that the outcome space Y is regular.
Assume that f satisfies the strict outer Spence–Mirrlees condition. An information al-
location is a map Y : [0 1] → Y that assigns to each type a distribution of posteriors. By
the implementability theorem, we have the following.

PROPOSITION 2: Every increasing information allocation is implementable.

The converse is false. In particular, there are implementable allocations that assign
some types t < t  Blackwell-incomparable information. But any such information alloca-
tion is vulnerable to collusion, as agents of types t and t  would benefit by sharing their

24
The “only if” direction is trivial. Conversely, a y with mean μ0 is induced by a ()-valued signal whose
distribution conditional on each ω ∈  is

1
π(M|ω) = μ(ω)y(dμ) for each Borel-measurable M ⊆ ()
μ0 (ω) M
This construction is due to Blackwell (1951), and used by Kamenica and Gentzkow (2011).
25
 A Blackwell-less informative distribution of posteriors is precisely one that yields a lower expected payoff
V (μ t)y(dμ) no matter what the underlying action set A or utility U (· · t). This is because V (· t) is contin-
uous and convex for any A and U , and any continuous and convex v can be approximated by V (· t) for some
A and U .
 THE CONVERSE ENVELOPE THEOREM 2807

information.2627 Call an allocation sharing-proof iff no two types are assigned Blackwell-
incomparable information.

PROPOSITION 3: An information allocation is implementable and sharing-proof if and

only if it is increasing.

The proof is in Appendix B.3.

APPENDIX A: THEORY (§2 AND §3)

A.1. Mathematical Background
Two operations are important in this paper: writing a function as the integral of its
derivative, and interchanging limits and integrals. The former is permissible precisely for
absolutely continuous functions:

DEFINITION 8: A function φ : [0 1] → R is absolutely continuous iff for each ε > 0,

there is a δ > 0 such that for any finite collection {(rn  tn )}Nn=1 of disjoint intervals of [0 1],
N N
n=1 (tn − rn ) < δ implies n=1 |φ(tn ) − φ(rn )| < ε.

Absolute continuity implies continuity and differentiability a.e., but the converse is
false. Absolute continuity is implied by Lipschitz continuity.

LEBESGUE’S FUNDAMENTAL THEOREM OF CALCULUS28 : Let φ be a function [0 1] → R.

The following are equivalent:
(i) φ is absolutely continuous. t
(ii) There is a ψ ∈ L1 such that φ(t) = φ(0) + 0 ψ for every t ∈ [0 1].
(iii) φ is differentiable a.e., its (a.e.-defined) derivative φ belongs to L1 , and φ(t) = φ(0) +
t
0
φ for every t ∈ [0 1].

As for interchanging limits and integrals, uniform integrability is the key.

DEFINITION 9: A family  ⊆ L1 is uniformly integrable iff

 for each ε > 0, there is δ > 0
such that for any open T ⊆ [0 1] of measure <δ, we have T |φ| < ε for every φ ∈ .

VITALI CONVERGENCE THEOREM29 : Let {φn}n∈N be a uniformly

t integrable
t sequence in L1
converging a.e. to φ : [0 1] → R. Then φ ∈ L , and limn→∞ r φn = r φ for all r t ∈ [0 1].
1

(Lebesgue’s dominated convergence theorem is a corollary.)

Absolute continuity and uniform integrability are closely related.

26
This holds no matter how the underlying signals giving rise to the posterior distributions Y (t) and Y (t  )
are correlated with each other. For by a standard embedding theorem (e.g., Theorem 7.A.1 in Shaked and
Shanthikumar (2007)), Y (t)  Y (t  ) is necessary (as well as sufficient) for there to exist a probability space on
which there are random vectors with laws Y (t) and Y (t  ) such that the latter is statistically sufficient for the
former.
27
Both agents benefit strictly provided V (· t) and V (· t  ) are strictly convex.
28
See, for example, Folland (1999, §3.5, p. 106) for a proof.
29
For a proof and a partial converse see, for example, Royden and Fitzpatrick (2010, §4.6).
2808 LUDVIG SINANDER

AC–UI LEMMA—Fitzpatrick and Hunt (2015): Let φ be a continuous function

[0 1] → R. The following are equivalent:
(i) φ is absolutely continuous.
(ii) The “divided-difference” family {t → [φ(t + m) − φ(t)]/m}m>0 is uniformly integrable.

A.2. Housekeeping for Absolute Equicontinuity (§2.1, p. 2797)

The following lemma justifies the name “absolute equicontinuity,” and is used in Ap-
pendix A.5 below to prove the necessity lemma (§3.1, p. 2801).

LEMMA 1: An absolutely equicontinuous family {φx}x∈X is uniformly equicontinuous, and

each of its members φx is absolutely continuous.

PROOF: Let {φx}x∈X be absolutely equicontinuous. Then for every x ∈ X , {t → [φx (t +

m) − φx (t)]/m}m>0 is uniformly integrable, and hence φx is absolutely continuous by the
AC–UI lemma in Appendix A.1.
It follows that for any r < t in [0 1],
 t    t 
     φx (s + m) − φx (s) 
supφx (t) − φx (r) = sup φx  = suplim ds
x∈X x∈X r x∈X m↓0 r m
 t 
 φx (s + m) − φx (s) 
≤ sup sup ds
x∈X m>0 r m
 t  
 φx (s + m) − φx (s) 
≤ sup sup  ds
m 
m>0 r x∈X

where the first equality holds by Lebesgue’s fundamental theorem of calculus, and the
second holds by the Vitali convergence theorem.
Fix an ε > 0. By the absolute equicontinuity of {φx}x∈X , there is a δ > 0 such that when-
ever t − r < δ, the right-hand side of the above inequality is <ε, and thus supx∈X |φx (t) −
φx (r)| < ε. So {φx}x∈X is uniformly equicontinuous. Q.E.D.

A.3. Proof of the Identity Lemma (§3.2, p. 2802)

We use the results in Appendix A.1. We shall focus on the limit m ↓ 0, omitting the
symmetric argument for m ↑ 0.30 For t ∈ [0 1) and m ∈ (0 1 − t], write

VX (t + m) − VX (t)
φm (t) :=
m
       
f X (t + m) t + m − f X(t + m) t f X (t + m) t − f X (t) t
= + 
m
  m
 
=:ψm (t) =:χm (t)

30
Since the argument below relies on absolute equicontinuity, the omitted argument requires uniform inte-
grability of {m}m<0 := {t → supx∈X |(f (x t + m) − f (x t))/m|}m<0 . This follows from absolute equicontinuity
and the observation that m (t) = −m (t + m).
 THE CONVERSE ENVELOPE THEOREM 2809

Fix r t ∈ (0 1). Note that

 t  t 
d   
lim χm = f X (s + m) s ds
m↓0 r dm r m=0
t
whenever the limit exists. Our task is to show that { r χm}m>0 is convergent as m ↓ 0 with
limit
 t
 
VX (t) − VX (r) − f2 X (s) s ds
r

{ψm}m>0 need not converge a.e. under the basic assumptions.31 But
   
f X (t) t − f X (t) t − m
ψm (t) :=

m
converges pointwise to t → f2 (X(t) t), and by a change of variable,
 t  t+m  t  t+m  r+m  t
ψm = ψm = ψm + ψm − ψm = ψm + o(1)
r r+m r t r r

where the bracketed terms vanish as m ↓ 0 because {ψm}m>0 is uniformly integrable by the
basic assumptions.
By absolute continuity of VX and the AC–UI lemma in Appendix A.1, {φm}m>0 is uni-
formly integrable and converges a.e. to VX as m ↓ 0. Since {ψm}m>0 is uniformly integrable
and converges pointwise to t → f2 (X(t) t), it follows that
 t  t  t
 
lim χm = lim [φm − ψm ] = lim φm − ψm
m↓0 r m↓0 r m↓0 r
 
t
  t
  
= lim φm − ψm = VX (s) − f2 X (s) s ds
r m↓0 r

where the third equality holds by the Vitali convergence theorem. Since the last expres-
t
sion is well-defined, this shows { r χm}m>0 to be convergent as m ↓ 0. And because VX is
absolutely continuous, the value of the limit is
 t  t
 
lim χm = VX (t) − VX (r) − f2 X (s) s ds
m↓0 r r

by Lebesgue’s fundamental theorem of calculus. Q.E.D.

A.4. A Characterization of Absolute Continuity of the Value

The following lemma characterizes the absolute-continuity-of-VX condition that ap-
pears in the main theorem (§3.2, p. 2801). Apart from its independent interest, it is needed
for the proofs in Appendices A.5 and A.6 below.

31
This remains true even under much stronger assumptions. For example, equidifferentiability of
{f (x ·)}x∈X is not enough: a counterexample is X = [0 1], f (x t) = (t − x)1Q (x) and X(t) = t. (Here,
1Q (x) = 1 if x is rational and = 0 otherwise.) In this case, ψm (t) = 1Q (t + m), which is nowhere convergent as
m ↓ 0.
2810 LUDVIG SINANDER

LEMMA 2: Under the basic assumptions, the following are equivalent:

(i) VX (t) := f (X(t) t) is absolutely continuous.
(ii) The family {χm}m>0 is uniformly integrable, where
   
f X (t + m) t − f X (t) t
χm (t) := 
m

In the classical case, (ii) is imposed (it follows from the classical assumptions, by
Lemma 4 in Appendix A.6 below). In the modern case, (i) arises within the theorem.
Both are clearly joint restrictions on f and X.32

PROOF: Define {φm}m>0 and {ψm}m>0 as in the proof of the identity lemma (Ap-
pendix A.3). {ψm}m>0 is uniformly integrable by the basic assumption of absolute equicon-
tinuity. By the AC–UI lemma in Appendix A.1, (i) is equivalent to {φm}m>0 being uni-
formly integrable.
Suppose that {χm}m>0 is uniformly integrable, and fix ε > 0. Let δ > 0 meet the ε/2-
challenge for both {ψm}m>0 and {χm}m>0 ; then for any open T ⊆ [0 1] of measure <δ and
any m > 0, we have
  
ε ε
|φm | ≤ |ψm | + |χm | < + = ε
T T T 2 2
showing that {φm}m>0 is uniformly integrable.
An almost identical argument establishes that uniform integrability of {φm}m>0 implies
uniform integrability of {χm}m>0 . Q.E.D.

A.5. Proof of the Necessity Lemma (§3.1, p. 2801)

LEMMA 3: If {f (x ·)}x∈X is absolutely equicontinuous, then the value VX (t) := f (X (t) t)
of any optimal X : [0 1] → X is absolutely continuous.

PROOF: Let X be optimal. Then for any r < t in [0 1) and m ∈ (0 1 − t],
  t+m    
1 1 r+m   t VX (s + m) − VX (s) 
 VX − VX  =  ds
m m r m
t r
 t   t
 VX (s + m) − VX (s) 
≤  
 m  ds ≤ Dm 
r r

where
 
 f (x s + m) − f (x s) 
Dm (s) := sup 

x∈X m
Fix an ε > 0. The absolute equicontinuity of {f (x ·)}x∈X provides that {Dm}m>0 is uni-
formly integrable, so that there is a δ > 0 such that for any open T ⊆ [0 1] of measure

32
As emphasized by Milgrom and Segal (2002), however, any optimal X satisfies (i) provided f satisfies the
basic assumptions. See Appendix A.5 below for a proof.
 THE CONVERSE ENVELOPE THEOREM 2811

<δ, we have T Dm < ε/2 for every m > 0. Thus for any finite collection {(rn  tn )}Nn=1 of
disjoint open intervals of [0 1] whose union T has measure <δ, we have
  tn +m   
N
1 1 rn +m 
 V − V X ≤ Dm < ε/2 for every m > 0
m X
m rn
n=1 tn T

VX is (uniformly) continuous since {f (x ·)}x∈X is uniformly equicontinuous by Lemma 1

in Appendix A.2.33 Thus letting m ↓ 0 yields
N
 
VX (tn ) − VX (rn ) ≤ ε/2 < ε
n=1

by the mean-value theorem, showing VX to be absolutely continuous. Q.E.D.

PROOF OF THE NECESSITY LEMMA: Let X be optimal, and fix r < t in [0 1]. VX is ab-
solutely continuous by Lemma 3. Define φrt : [−r 1 − t] → R by
 t
 
φrt (m) := f X (s + m) s ds
r

for each m ∈ [−r 1 − t].34 φrt (0) exists by the identity lemma (§3.2, p. 2802). To show that
it is zero, observe that for any s ∈ (r t) and m ∈ (0 min{s 1 − s}], optimality requires
       
f X (s + m) s − f X (s) s f X (s − m) s − f X (s) s
≤0≤ 
m −m
Integrating over (r t) and letting m ↓ 0 yields φrt (0) ≤ 0 ≤ φrt (0). Q.E.D.

A.6. A Lemma Under the Classical Assumptions

The following result is used in the proof of the housekeeping lemma (§3.1, p. 2800),
as well as in the proof of the classical envelope theorem and converse in Appendix A.7
below.

LEMMA 4: Fix a decision rule X : [0 1] → X , and let

   
f X (t + m) t − f X (t) t
χm (t) := 
m
(i) Under the basic and classical assumptions, {χm}m>0 is uniformly integrable.
(ii) Under the basic assumptions, the following are equivalent:
(a) {χm}m>0 is uniformly integrable and convergent a.e. as m ↓ 0.
(b) VX (t) := f (X(t) t) is absolutely continuous, and the derivative dm d
f (X (t +
m) t)|m=0 exists for a.e. t ∈ (0 1).

33
For any ε > 0, the uniform equicontinuity of {f (x ·)}x∈X delivers a δ > 0 such that |t − r| < δ implies
|VX (t) − VX (r)| ≤ supx∈X |f (x t) − f (x r)| < ε.
34
The map s → f (X(s + m) s) is integrable because |f (X(s + m) s)| ≤ |VX (s)| + |f (X(s + m) s) −
f (X(s) s)|, where the former term is continuous, and the latter is integrable by Lemma 2 in Appendix A.4.
2812 LUDVIG SINANDER

PROOF: For (i), write K for the vector of nonnegative constants that bounds f1 , and
L ≥ 0 for the Lipschitz constant of X. Let · 2 denote the Euclidean norm. For any t ∈
[0 1) and m ∈ (0 1 − t], writing xω := (1 − ω)X(t) + ωX (t + m) for ω ∈ [0 1], we have
by the Cauchy–Schwarz inequality that
  1 
      
χm (t) =  1 f1 (xω  t) · X (t + m) − X (t) dω
m
0

1 1    
≤ f1 (xω  t)  × X (t + m) − X (t)  dω ≤ 1 K 2 × Lm = K 2 L
2 2
m 0 m

Thus {χm}m>0 is uniformly bounded, hence uniformly integrable.

For (ii), absolute continuity of VX is equivalent to uniform integrability of {χm}m>0 by
Lemma 2 in Appendix A.4, and a.e. existence of dm d
f (X (t + m) t)|m=0 is definitionally
equivalent to a.e. convergence of {χm}m>0 . Q.E.D.

A.7. Proof of the Classical Envelope Theorem and Converse (§2.2)

PROOF: Fix a Lipschitz continuous decision rule X : [0 1] → X . By Lemma 4 in Ap-
pendix A.6, VX (t) := f (X (t) t) is absolutely continuous, hence differentiable a.e. The
map r → f (X (r) t) is differentiable a.e. by the classical assumptions, and t → f (X (r) t)
is differentiable by the basic assumptions. Hence the a.e.-defined derivative of VX obeys
the differentiation identity

d    

VX (t) = f X (t + m) t  + f2 X (t) t for a.e. t ∈ (0 1)
dm m=0

It follows that the first-order condition a.e. is equivalent to

 
VX (t) = f2 X (t) t for a.e. t ∈ (0 1)

which in turn is equivalent to the envelope formula by Lebesgue’s fundamental theorem

of calculus. Q.E.D.

By inspection, the proof requires precisely absolute continuity of VX (so that the en-
velope formula can be satisfied) and a.e. existence of dm d
f (X (t + m) t)|m=0 (so that the
first-order condition a.e. is well-defined). Part (ii) of Lemma 4 in Appendix A.6 therefore
tells us that the classical assumptions can be weakened to uniform integrability and a.e.
convergence of {χm}m>0 , and no further. For f nontrivial, the uniform integrability part
involves a strong continuity requirement on X.35

35
For example, consider X = [0 1], f (x t) = x and X(t) = 1[r1] , where r ∈ (0 1). Then given m > 0, we
have χm (t) = 1/m for all t ∈ [r − m r]. Suppose toward a contradiction that {χm}m>0 is uniformly integrable,
and let δ > 0 meet the ε-challenge for ε ∈ (0 1); then for all m ∈ (0 δ/2), we have
 r+δ/2  r
|χm | ≥ |χm | = m/m = 1 > ε
r−δ/2 r−m

which is absurd. This example clearly generalizes: the gist is that uniform integrability of {χm}m>0 is incompat-
ible with nonremovable discontinuities in X unless f is trivial.
 THE CONVERSE ENVELOPE THEOREM 2813

APPENDIX B: APPLICATION (§4)

B.1. Proof of the Implementability Theorem (§4.3, p. 2804)
We state two lemmata in §B.1.1–§B.1.2, then prove the theorem in §B.1.3.

B.1.1. Solutions of the Envelope Formula

In the first step of the argument in §B.1.3 below, we are given an allocation Y , and
wish to choose a payment schedule P such that (Y P) satisfies the envelope formula. The
following asserts that this can be done.

EXISTENCE LEMMA: Assume that for all (y t) ∈ Y × [0 1], f (y · t) is strictly decreasing,
continuous and onto R. Further assume that the type derivative f3 exists and is bounded, and
that f3 (y · t) is continuous for all (y t) ∈ Y × [0 1]. Then for any k ∈ R and any allocation
Y : [0 1] → Y such that t → f (Y (t) p t) and t → f3 (Y (t) p t) are Borel-measurable
for every p ∈ R, there exists a payment schedule P : [0 1] → R such that (Y P) satisfies the
envelope formula with VYP (0) = k.

REMARK 2: The following corollary may prove useful elsewhere: suppose in addition
that Y is equipped with some topology such that f (· p t) and f3 (· p t) are Borel-
measurable and f3 (y p ·) is continuous. Then for any Borel-measurable allocation Y :
[0 1] → Y , there is a payment schedule P such that (Y P) satisfies the envelope formula.

The existence lemma is immediate from the following abstract result by letting
φ(p t) := f (Y (t) p t) and ψ(p t) := f3 (Y (t) p t).

LEMMA 5: Let φ and ψ be functions R × [0 1] → R. Suppose that φ(· t) is strictly de-
creasing, continuous, and onto R for every t ∈ [0 1], and that ψ is bounded with ψ(· t) con-
tinuous for every t ∈ [0 1]. Further assume that φ(p ·) and ψ(p ·) are Borel-measurable for
each p ∈ R. Then for any k ∈ R, there is a function P : [0 1] → R such that
 t
   
φ P (t) t = k + ψ P (s) s ds for every t ∈ [0 1]
0

PROOF: Since φ(· t) is strictly decreasing and continuous, it possesses a continuous

inverse φ−1 (· t), well-defined on all of R since φ(R t) = R. We may therefore define a
function χ : R × [0 1] → R by
 
χ(w t) := ψ φ−1 (w t) t for each w ∈ R and t ∈ [0 1]

χ(· t) is continuous since ψ(· t) and φ−1 (· t) are, χ is bounded since ψ is, and χ(w ·)
is Borel-measurable since ψ(· t) is continuous and ψ(p ·) and φ−1 (w ·) are Borel-
measurable.
Fix k ∈ R. Consider the integral equation
 t
 
W (t) = k + χ W (s) s ds for t ∈ [0 1]
0

where W is an unknown function [0 1] → R. Since χ(· t) is continuous and χ(w ·)

bounded and Borel-measurable, there is a local solution by Carathéodory’s existence the-
2814 LUDVIG SINANDER

orem;36 call it V . By boundedness of χ and a comparison theorem,37 V can be extended

to a solution on all of [0 1].
Now define P (t) := φ−1 (V (t) t). For every t ∈ [0 1], it satisfies
 t  t
     
φ P (t) t = V (t) = k + χ V (s) s ds = k + ψ P (s) s ds
0 0
Q.E.D.

UNIQUENESS COROLLARY: Under the hypotheses of the existence lemma, if in addition

{ 3 (y · t)}(yt)∈Y ×[01] is Lipschitz equicontinuous38 and the monotonicity of f (y · t) is uni-
f
form in the sense that for some M > 0,
   
f (y p t) − f y p  t ≥ M p − p for any p < p in R y ∈ Y and t ∈ [0 1]

then there is exactly one payment schedule P such that (Y P) satisfies the envelope formula
with VYP (0) = k, and this payment schedule may be computed via Picard’s method.

PROOF: Again let φ(p t) := f (Y (t) p t) and ψ(p t) := f3 (Y (t) p t), and return to
the proof of Lemma 5. The additional assumptions ensure, respectively, that {ψ(· t)}t∈[01]
and {φ−1 (· t)}t∈[01] are Lipschitz equicontinuous. In follows that {χ(· t)}t∈[01] is Lipschitz
equicontinuous, so that (the Picard operator is a contraction, and thus) the integral equa-
tion has a unique solution to which Picard iteration converges in the sup norm.39 Q.E.D.

B.1.2. Continuous Approximation of Increasing Maps

The second step of the argument §B.1.3 below relies on approximating an increasing
map [0 1] → Y by continuous and increasing maps. This is made possible by the follow-
ing.

APPROXIMATION LEMMA: Let Y be regular, and let Y be an increasing map [0 1] → Y .

The image Y ([0 1]) may be embedded in a chain C ⊆ Y with inf C = Y (0) and sup C = Y (1)
that is order-dense-in-itself, order-complete, and order-separable.40 Furthermore, there exists a
sequence (Yn )n∈N of increasing maps [0 1] → C , each with Yn = Y on {0 1}, such that when
C has the relative topology inherited from the order topology on Y , Yn is continuous for each
n ∈ N, and Yn → Y pointwise as n → ∞.

The (rather involved) proof is in the Online Supplemental Material, Appendix S.3.

B.1.3. Proof of the Implementability Theorem

Fix an increasing Y : [0 1] → Y . Embed its image Y ([0 1]) in the chain C ⊆ Y deliv-
ered by the approximation lemma in Appendix B.1.2, and equip C with the relative topol-
ogy inherited from the order topology on Y . We henceforth view Y as a function [0 1] →
C , and (with a minor abuse of notation) view f and f3 as functions C × R × [0 1] → R.

36
See, for example, Theorem 5.1 in Hale (1980, Chapter 1).
37
See, for example, Theorem 2.17 in Teschl (2012).
38
That is, there is an L ≥ 0 such that f3 (y · t) is L-Lipschitz for every (y t) ∈ Y × [0 1].
39
See, for example, Theorem 5.3 in Hale (1980, Chapter 1).
40
C ⊆ Y is order-complete iff every subset with a lower (upper) bound has an infimum (supremum), and
order-separable iff it has a countable order-dense subset.
 THE CONVERSE ENVELOPE THEOREM 2815

We seek a payment schedule P : [0 1] → R such that the direct mechanism (Y P) is
incentive-compatible. We do this first (step 1) under the assumption that Y is continuous,
then (step 2) show how continuity may be dropped.
Step 1: Suppose that Y is continuous. By preference regularity and the existence lemma
in Appendix B.1.1,41 there exists a payment schedule P : [0 1] → R such that the envelope
formula holds with (say) VYP (0) = 0:
 t
 
VYP (t) = f3 Y (s) P (s) s ds for every t ∈ [0 1]
0

This P must be continuous since Y , f and VYP are continuous and f (y · t) is strictly
monotone.42 We will show that (Y P) is incentive-compatible. t
Write U (r t) := f (Y (r) P (r) t) for type t’s mimicking payoff, and φrt (m) := r U (s +
m s) ds for the collective payoff of types [r t] ⊆ (0 1) from “mimicking up” by m. Clearly,
U is a continuous function [0 1]2 → R, and thus φrt : [−r 1 − t] → R is also continuous.
Note that VYP (t) ≡ U (t t).
The model fits into the abstract setting of §2.1 by letting X := C × R and X(t) :=
(Y (t) P (t)), and the basic assumptions are satisfied since f3 exists and is bounded. We
may thus invoke the converse envelope theorem (p. 2801): since (Y P) satisfies the enve-
lope formula, it must satisfy the outer first-order condition:
 t 

d 
U (s + m s) ds = 0 for all r  < t  in (0 1)
dm r  
m=0

Given r < t in (0 1), writing Dφrt (s ) := dm

d
φrt (s + m)|m=0 for the upper derivative,
the outer Spence–Mirrlees condition yields for each n ∈ (0 r) that
 t−n 

d 
0≤ U (s + m s + n) ds
dm r−n 
m=0
 t 

d 
= U (s + m − n s) ds = Dφrt (−n)
dm r 
m=0

which is to say that Dφrt ≥ 0 on (−r 0). Since φrt is continuous, it follows that φrt is
increasing on [−r 0].43 A similar argument shows that φrt is decreasing on [0 1 − t].

41
The measurability hypothesis in the existence lemma is satisfied because f (· p t), f3 (· p t), and Y are
continuous, and f (y p ·) and f3 (y p ·) are Borel-measurable (the former being continuous, and the latter
a derivative). (To complete the argument for measurability, deduce that r → f (Y (r) p t) is continuous and
that t → f (Y (r) p t) is Borel-measurable, so that (r t) → f (Y (r) p t) is (jointly) Borel-measurable, and
thus t → f (Y (t) p t) is Borel-measurable. Similarly for f3 .)
42
Suppose not: tn → t but limn→∞ P (tn ) = P (t). Then the continuity of Y and f and the strict monotonicity
of f (y · t) yield a contradiction with the continuity of VYP :
     
VYP (tn ) = f Y (tn ) P (tn ) tn → f Y (t) lim P (tn ) t = f Y (t) P (t) t = VYP (t)
n→∞

43
This is a standard result; see, for example, Bruckner (1994, §11.4, p. 128).
2816 LUDVIG SINANDER

It follows that for any r < t in [0 1] and m ∈ [−r 1 − t],

 t
 
U (s s) − U (s + m s) ds = φrt (0) − φrt (m) ≥ 0
r

Thus for every m ∈ [0 1], we have

U (s s) − U (s + m s) ≥ 0 for a.e. s ∈ [0 1] ∩ [−m 1 − m]
Since s → U (s s) = VYP (s) and s → U (s + m s) are continuous for any m ∈ [0 1], it
follows that for every m ∈ [0 1],
U (s s) − U (s + m s) ≥ 0 for every s ∈ [0 1] ∩ [−m 1 − m]
which is to say that (Y P) is incentive-compatible.
Step 2: Now drop the assumption that Y is continuous. By regularity of Y and the
approximation lemma in Appendix B.1.2, there exists a sequence (Yn )n∈N of continuous
and increasing maps [0 1] → C converging pointwise to Y , each of which satisfies Yn = Y
on {0 1}. At each n ∈ N, Step 1 yields a Pn : [0 1] → R such that such that (Yn  Pn ) is
incentive-compatible and satisfies the envelope formula with VYn Pn (0) = 0.
The sequence (VYn Pn )n∈N is Lipschitz equicontinuous44 by the envelope formula and the
boundedness of f3 . It is furthermore uniformly bounded, due to its Lipschitz equicontinu-
ity and the fact that VYn Pn (0) = 0 for every n ∈ N. Thus, by the Arzelà–Ascoli theorem,45
we may assume (passing to a subsequence if necessary) that (VYn Pn )n∈N converges point-
wise. Then (Pn )n∈N converges pointwise;46 write P : [0 1] → R for its limit.
By continuity of f , Un (r t) := f (Yn (r) Pn (r) t) converges to U (r t) := f (Y (r) P (r) t)
for all r t ∈ [0 1]. Each of the incentive-compatibility inequalities Un (t t) ≥ Un (r t) is
preserved in the limit n → ∞, ensuring that (Y P) is incentive-compatible. Q.E.D.

B.2. Converse to the Implementability Theorem (§4.3, p. 2804)

In this Appendix, we provide a partial converse to the implementability theorem, and
use it to prove Proposition 1 (p. 2805). We shall use the partial converse again in Ap-
pendix B.3 below to prove Proposition 3 (p. 2807).
Letting  denote the partial order on Y , we say that an allocation Y : [0 1] → Y is non-
decreasing iff there are no t ≤ t  in [0 1] such that Y (t  ) < Y (t). In other words, Y (t) and
Y (t  ) could either be ranked as Y (t)  Y (t  ), or they could be incomparable. Increasing
maps are nondecreasing, but the converse is false except if Y is a chain.

PROPOSITION 1 : If f is regular and satisfies the strict outer Spence–Mirrlees condition,

then only nondecreasing allocations are implementable.

PROOF OF PROPOSITION 1 (P. 2805): By the implementability theorem, any increasing

allocation is implementable. By Proposition 1 , any implementable allocation is nonde-
creasing, hence increasing since Y is a chain. Q.E.D.

44
That is, there is an L ≥ 0 such that VYn Pn is L-Lipschitz for every n ∈ N.
45
For example, Theorem 4.44 in Folland (1999).
46
Clearly f (Yn (t) infm≥n Pm (t) t) = supm≥n f (Yn (t) Pm (t) t) ≤ supm≥n VYm Pm (t) for any t ∈ [0 1], and
thus f (Y (t) p t) ≤ V (t), where p := lim infn→∞ Pn (t) and V (t) := limn→∞ VYn Pn (t). Similarly, V (t) ≤
f (Y (t) p  t), where p := lim supn→∞ Pn (t). Thus f (Y (t) p t) ≤ f (Y (t) p  t), which rules out p < p since
f (Y (t) · t) is strictly decreasing.
 THE CONVERSE ENVELOPE THEOREM 2817

The proof of Proposition 1 relies on two lemmata. The first is a “nondecreasing” com-
parative statics result.47

LEMMA 6: Let X and T be partially ordered sets, and let f be a function X × T → R.

Call a decision rule X : T → X optimal iff f (X (t) t) ≥ f (x t) for all x ∈ X and t ∈ T . If
f has strictly single-crossing differences,48 then every optimal decision rule is nondecreasing.

PROOF: Write  and , respectively, for the partial orders on X and on T . Let
X : T → X be optimal, and suppose toward a contradiction that there are t ≺ t  in T
such that X (t  ) < X (t). Since X(t) is optimal at parameter t, we have f (X (t  ) t) ≤
f (X(t) t). Because t ≺ t  and X(t  ) ≺ X(t), it follows by strictly single-crossing differ-
ences that f (X (t  ) t  ) < f (X (t) t  ), a contradiction with the optimality of X (t  ) at pa-
rameter t  . Q.E.D.

LEMMA 7: If f is regular and satisfies the (strict) outer Spence–Mirrlees condition, then for
any price schedule π : Y → R, the map (y t) → f (y π(y) t) has (strictly) single-crossing
differences.

PROOF: Fix y < y  in Y , p, p in R and t < t  in [0 1]. Define a mechanism (Y P) :
[0 1] → Y × R by (Y (s) P (s)) := (y p) for s ≤ t and (Y (s) P (s)) := (y   p ) for s > t,
and fix r r  ∈ (0 1) with r < t < r  . Clearly for n ∈ {0 t  − t},
 r 
d   
f Y (s + m) P (s + m) s + n ds
dm r 
m=0
 t−m  r 
    
d 
= f (y p s + n) ds + f y  p  s + n ds 
dm r t−m 
m=0
   
= f y  p  t + n − f (y p t + n)
If f satisfies the outer Spence–Mirrlees condition, then the left-hand side is single-
crossing in n, and thus f (y   p  t) − f (y p t) ≥(>) 0 implies f (y   p  t  ) − f (y p t  ) ≥
(>) 0. Similarly for the strict case. Q.E.D.

PROOF OF PROPOSITION 1 : Let Y : [0 1] → Y be implementable, so that (Y P) is

incentive-compatible for some payment schedule P : [0 1] → R. Define a price sched-
ule π : Y ([0 1]) → R by π ◦ Y = P; it is well-defined because by incentive-compatibility
and strict monotonicity of f (y · t), Y (r) = Y (r  ) implies P (r) = P (r  ). Define a function
φ : Y ([0 1]) × [0 1] → R by φ(y t) := f (y π(y) t). Take any t ∈ [0 1] and y ∈ Y ([0 1]),
and observe that there must be an r ∈ [0 1] with Y (r) = y. Then since (Y P) is incentive-
compatible,
       
φ Y (t) t = f Y (t) π Y (t)  t = f Y (t) P (t) t
   
≥ f Y (r) P (r) t = f y π(y) t = φ(y t)

47
Such results are dimly known in the literature, but rarely seen in print. Exceptions include Quah and
Strulovici (2007, Proposition 5), Anderson and Smith (2021), and Curello and Sinander (2022).
48
A function φ : X × T → R has (strictly) single-crossing differences iff t → φ(x  t) − φ(x t) is (strictly)
single-crossing for any x < x in X , where < denotes the strict part of the partial order on X . (“Single-crossing”
was defined in footnote 17 on p. 2804.)
2818 LUDVIG SINANDER

Since y ∈ Y ([0 1]) and t ∈ [0 1] were arbitrary, this shows that Y is an optimal decision
rule for objective φ. Since φ has strictly single-crossing differences by Lemma 7, it follows
by Lemma 6 that Y is nondecreasing. Q.E.D.

B.3. Proof of Proposition 3 (§4.4, p. 2807)

Any increasing Y : [0 1] → Y is implementable by the implementability theorem (§4.3,
p. 2804), and clearly sharing-proof. For the converse, let Y : [0 1] → Y be implementable
and sharing-proof, and fix t < t  ; then either Y (t)  Y (t  ) or Y (t  ) < Y (t) since Y is
sharing-proof, and it cannot be the latter because Y is nondecreasing by Proposition 1 in
Appendix B.2. Q.E.D.

REFERENCES
ANDERSON, AXEL, AND LONES SMITH (2021): “The Comparative Statics of Sorting,” Working paper, 14 Jun
2021. [2817]
BATTAGLINI, MARCO (2005): “Long-Term Contracting With Markovian Consumers,” American Economic Re-
view, 95, 637–658. [2803]
BENVENISTE, LAWRENCE M., AND JOSÉ A. SCHEINKMAN (1979): “On the Differentiability of the Value Func-
tion in Dynamic Models of Economics,” Econometrica, 47, 727–732. [2797]
BLACKWELL, DAVID (1951): “Comparison of Experiments,” in Berkeley Symposium on Mathematical Statistics
and Probability, Vol. 2, ed. by Jerzy Neyman. Berkeley, CA: University of California Press, 93–102. [2806]
BRUCKNER, ANDREW (1994): Differentiation of Real Functions (Second Ed.). CRM Monographs. Providence,
RI: American Mathematical Society. [2815]
CLAUSEN, ANDREW, AND CARLO STRUB (2020): “Reverse Calculus and Nested Optimization,” Journal of
Economic Theory, 187, 105019. [2797]
COURTY, PASCAL, AND HAO LI (2000): “Sequential Screening,” Review of Economic Studies, 67, 697–717.
[2803]
CURELLO, GREGORIO, AND LUDVIG SINANDER (2022): “The Comparative Statics of Persuasion,” Working
paper, 14 Apr 2022. [2817]
DANSKIN, JOHN M. (1966): “The Theory of Max–Min, With Applications,” SIAM Journal on Applied Mathe-
matics, 14, 641–664. [2797]
(1967): The Theory of Max–Min and Its Application to Weapons Allocation Problems. Berlin: Springer.
[2797]
ESŐ, PÉTER, AND BALÁZS SZENTES (2007): “Optimal Information Disclosure in Auctions and the Handicap
Auction,” Review of Economic Studies, 74, 705–731. [2803]
FITZPATRICK, PATRICK M., AND BRIAN R. HUNT (2015): “Absolute Continuity of a Function and Uniform
Integrability of Its Divided Differences,” American Mathematical Monthly, 122, 362–366. [2808]
FOLLAND, GERALD B. (1999): Real Analysis: Modern Techniques and Their Applications (Second Ed.). Pure and
Applied Mathematics. New York, NY: Wiley. [2800,2807,2816]
GARCÍA, DIEGO (2005): “Monotonicity in Direct Revelation Mechanisms,” Economics Letters, 88, 21–26.
[2803]
GNEITING, TILMANN, AND ADRIAN E. RAFTERY (2007): “Strictly Proper Scoring Rules, Prediction, and Esti-
mation,” Journal of the American Statistical Association, 102, 359–378. [2805]
GUESNERIE, ROGER, AND JEAN-JACQUES LAFFONT (1984): “A Complete Solution to a Class of Principal-
Agent Problems With an Application to the Control of a Self-Managed Firm,” Journal of Public Economics,
25, 329–369. [2803]
HALE, JACK K. (1980): Ordinary Differential Equations (Second Ed.). Malabar, FL: Krieger. [2814]
HOTELLING, HAROLD (1932): “Edgeworth’s Taxation Paradox and the Nature of Demand and Supply Func-
tions,” Journal of Political Economy, 40, 577–616. [2796]
KAMENICA, EMIR, AND MATTHEW GENTZKOW (2011): “Bayesian Persusasion,” American Economic Review,
101, 2590–2615. [2806]
MAS-COLELL, ANDREU, MICHAEL WHINSTON, AND JERRY R. GREEN (1995): Microeconomic Theory. Oxford:
Oxford University Press. [2795,2797]
MATTHEWS, STEVEN, AND JOHN MOORE (1987): “Monopoly Provision of Quality and Warranties: An Explo-
ration in the Theory of Multidimensional Screening,” Econometrica, 55, 441–467. [2803]
 THE CONVERSE ENVELOPE THEOREM 2819

MILGROM, PAUL (2004): Putting Auction Theory to Work. Cambridge: Cambridge University Press. [2803]
MILGROM, PAUL, AND ILYA SEGAL (2002): “Envelope Theorems for Arbitrary Choice Sets,” Econometrica, 70,
583–601. [2795,2797,2798,2805,2810]
MILGROM, PAUL, AND CHRIS SHANNON (1994): “Monotone Comparative Statics,” Econometrica, 62, 157–180.
[2804]
MIRRLEES, JAMES A. (1976): “Optimal Tax Theory: A Synthesis,” Journal of Public Economics, 6, 327–358.
[2803]
MORAND, OLIVIER, KEVIN REFFETT, AND SUCHISMITA TARAFDAR (2015): “A Nonsmooth Approach to En-
velope Theorems,” Journal of Mathematical Economics, 61, 157–165. [2797]
MYERSON, ROGER B. (1981): “Optimal Auction Design,” Mathematics of Operations Research, 6, 58–73. [2798]
NÖLDEKE, GEORG, AND LARRY SAMUELSON (2018): “The Implementation Duality,” Econometrica, 86, 1283–
1324. [2803,2804]
OYAMA, DAISUKE, AND TOMOYUKI TAKENAWA (2018): “On the (Non-)Differentiability of the Optimal Value
Function When the Optimal Solution Is Unique,” Journal of Mathematical Economics, 76, 21–32. [2797]
PAVAN, ALESSANDRO, ILYA SEGAL, AND JUUSO TOIKKA (2014): “Dynamic Mechanism Design: A Myersonian
Approach,” Econometrica, 82, 601–653. [2803]
QUAH, JOHN K.-H., AND BRUNO STRULOVICI (2007): “Comparative Statics With the Interval Dominance
Order: Some Extensions,” Working paper. [2817]
ROY, RENÉ (1947): “La distribution du revenu entre les divers biens,” Econometrica, 15, 205–225. [2796]
ROYDEN, HALSEY LAWRENCE, AND PATRICK M. FITZPATRICK (2010): Real Analysis (Fourth Ed.). Upper Sad-
dle River, NJ: Prentice Hall. [2807]
SAMUELSON, PAUL A. (1947): Foundations of Economic Analysis. Cambridge, MA: Harvard University Press.
[2796]
SHAKED, MOSHE, AND J. GEORGE SHANTHIKUMAR (2007): Stochastic Orders. Springer Series in Statistics.
New York, NY: Springer. [2807]
SHEPHARD, RONALD WILLIAM (1953): Cost and Production Functions. Princeton, NJ: Princeton University
Press. [2796]
SILBERBERG, EUGENE (1974): “A Revision of Comparative Statics Methodology in Economics, or, How to Do
Comparative Statics on the Back of an Envelope,” Journal of Economic Theory, 7, 159–172. [2797]
SINANDER, LUDVIG (2022): “Supplement to ‘The Converse Envelope Theorem’,” Econometrica Supplemental
Material, 90, https://doi.org/10.3982/ECTA18119. [2803]
SPENCE, MICHAEL (1974): “Competitive and Optimal Responses to Signals: An Analysis of Efficiency and
Distribution,” Journal of Economic Theory, 7, 296–332. [2803]
TESCHL, GERALD (2012): Ordinary Differential Equations and Dynamical Systems. Graduate Studies in Mathe-
matics. Providence, RI: American Mathematical Society. [2814]

Co-editor Alessandro Lizzeri handled this manuscript.

Manuscript received 14 February, 2020; final version accepted 26 June, 2022; available online 15 July, 2022.