IRONING WITHOUT CONTROL

JUUSO TOIKKA
Abstract. I extend Myersons (1981) ironing technique to more general ob-
jective functions. The approach is based on a generalized notion of virtual
surplus which can be maximized pointwise even when the monotonicity con-
straint implied by incentive compatibility binds. It is applicable to quasilinear
principal-agent models where the standard virtual surplus is weakly concave
in the allocation or appropriately separable in the allocation and type. No
assumptions on allocation rules are required beyond monotonicity.
1. Introduction
In quasilinear principal-agent models where the agents type is one-dimensional
and her payo function satises a single-crossing condition, optimal contracts are
found by maximizing expected virtual surplus subject to the constraint that allo-
cation be monotone in type (e.g., Salanie, 1997). Formally, the problem is one of
maximizing a real-valued functional

1
0
J((), )dF()
over the set of nondecreasing functions : X mapping types to allocations.
The standard approach to solving the problem when the monotonicity constraint
bindsknown as ironing because of the nature of the solutionis to apply op-
timal control theory (e.g., Fudenberg and Tirole, 1991). In this paper I build on
Myerson (1981) to develop an alternative method.
More specically, I rst observe that Myersons approach can be used in prob-
lems where the function J is appropriately separable in its arguments. For example,
this class of problems includes the price discrimination model of Mussa and Rosen
(1978) and the regulation model of Laont and Tirole (1986). I then extend the
technique to problems where J is weakly concave in its rst argument, i.e., where
the virtual surplus of each type exhibits weakly decreasing marginal returns in the
allocation. As discussed in Section 4, such concave problems include the aforemen-
tioned examples of separable problems as well as many models with interdependent
values that generally fail separability.
Heuristically, the approach is based on a generalized notion of virtual surplus
that takes into account total surplus, the agents information rent, and any distor-
tions the current type causes to other types through the monotonicity constraint.
As standard virtual surplus accounts only for the rst two, it fails to capture the
impact on the principals payo of global incentive compatibility constraints that
correspond to the monotonicity constraint. The generalized virtual surplus incorpo-
rates them through a convexication procedure due to Myerson (1981). It amounts
Date: February 19, 2011.
Acknowledgements to be added.
1
2 TOIKKA
to averaging the marginal contributions of types aected by the change in the al-
location of the current type. In separable problems, such as in Myersons auction
model, this can be done independently of the allocation. I show that the idea can
be extended to concave problems by applying the procedure to each allocation.
In either case, this results in a generalized virtual surplus which can be simply
maximized pointwise even when the monotonicity constraint binds.
The results in this paper improve on the standard approach of using optimal
control in that no restrictions are placed on admissible allocation rules beyond
monotonicity. To see how such restrictions come about in the standard approach, I
briey review how the maximum principle is applied to the problem.
1
The challenge
there is how to incorporate the monotonicity constraint, which can not be imposed
on the control variable. The trick is to let the allocation () be the state and take
its derivative

() as the control. Monotonicity is then simply the requirement

that the control be non-negative. However, this formulation is unsatisfactory as
it assumes that the allocation rule an endogenous objectis absolutely con-
tinuous.
2
For example, this assumption is violated in auctions and other trading
problems with linear utility, where optimal allocations are discontinuous. Further-
more, even in models where the optimal allocation is dierentiable, showing that
non-dierentiable rules oer no improvement requires formulating the problem in a
larger space where such rules are feasible and which is no longer covered by standard
results. In contrast, the approach in this paper allows for all monotone allocation
rules and thus provides a unied treatment of problems with or without jumps in
the optimal allocation. It oers a possible way of assessing whether the properties
assumed in the standard approach hold in applications.
Contemporaneous work by Hellwig (2009) also allows for jumps in the allocation
rule by extending the maximum principle to a class of problems with monotonicity
constraints while relaxing absolute continuity. He builds on the generalized maxi-
mum principle of Clarke (1976), which relies on non-smooth analysis. In contrast,
the approach in the present paper is elementary and self-contained.
Noldeke and Samuelson (2007) develop an approach to ironing without optimal
control, which is quite dierent from the present paper. Roughly put, they work
with the inverse of the allocation rule. The monotonicity constraint on the inverse
turns out to be non-binding more generally than the one on the allocation rule.
However, their approach requires virtual surplus to be strictly concave in the allo-
cation, which rules out jumps in the allocation rule. The approach in this paper
is more general for problems where the participation constraint binds only at an
extreme type, whereas Noldeke and Samuelson also consider models where it may
bind at intermediate types.
3
I set up the problem in the next section. I then review Myersons approach and
observe that it applies to separable problems in Section 3. The readers familiar
with the technique can proceed directly to Section 4 where I present the extension
to concave problems, which is the main contribution of the paper. An appendix
collects proofs omitted from the main text.
1
E.g., Fudenberg and Tirole (1991) or Jullien (2000). For more on optimal control, see, e.g.,
Leonard and van Long (1992), or the advanced treatment by Vinter (2010).
2
The state of the art in necessary conditions in the literature on optimal control requires the
state , also known as the arc, to be absolutely continuoussee Clarke (2011).
3
The approach presented here can be used in the latter type of problems once the types with
binding participation constraints have been identied.
IRONING WITHOUT CONTROL 3
2. The problem
Let X := [0, x] and := [0, 1]. Let J : X R, and let F : [0, 1] be a
cdf with density f := F

> 0.
4
Let M := { : X | is nondecreasing}. This
paper is concerned with problems of the form
5
(P) sup
M

1
0
J((), )dF()

.
The leading example is the expected virtual surplus maximization problem from
single-agent mechanism design. With that in mind I use the following terminology:
x X is an allocation, is a type, F is the principals belief about , J is the
virtual surplus function, and is an allocation rule. An allocation rule is optimal
if it attains the supremum in (P).
In the canonical principal-agent model virtual surplus takes the form
(2.1) J(x, ) = v(x, ) + u(x, )
1 F()
f()
u
2
(x, ),
where v and u are the principals and the agents payo functions, and u
2
(x, )
denotes the partial derivative of u with respect to its second argument. However,
as the methods apply more broadly, I take J as the primitive and provide sucient
conditions on u and v that deliver the required properties of J.
3. The separable case
It is instructive to start by reviewing the technique of Myerson (1981). I show
that it is applicable to problems that are separable in the following sense.
Denition 3.1. J : X R is separable if there exist functions a, b : X R
and k, l : R, where a is strictly increasing, such that for all (x, ) X ,
(3.1) J(x, ) = a(x)k() + b(x) + l().
The problem (P) is separable if J is separable.
Example 3.2 (Mussa and Rosen, 1978; Myerson, 1981). In the classic model of
monopolistic price discrimination of Mussa and Rosen (1978) the virtual surplus
from consumer of type takes the form
J(x, ) = x

1 F()
f()

c(x),
where x is quality and c(x) is the cost of producing it. Taking c(x) =
0
x and
interpeting x as the probability of sale gives the virtual surplus in a single-agent
version of the optimal auction problem of Myerson (1981).
4
It is straightforward to adapt the results to problems where X or (or both) is discrete.
Details are available from the author upon request. With a continuous type, it is without loss to
assume further that F is the uniform distribution. Namely, let F be a cdf on [, ] with F

> 0.
Let

:= F
1
and let

J(x, ) := J(x, F
1
()) for all x X. A change of variables then gives

J((), )dF() =

1
0
J((F
1
(q)), F
1
(q))dq =

1
0

J(

(q), q)dq.
5
For the problem to be well dened, the mapping J((), ) : R has to be integrable for
any M. A sucient conditions for this is that J : X R is measurable and bounded.
4 TOIKKA
Example 3.3 (Laont and Tirole, 1986). In a simplied version of the workhorse
model of regulation by Laont and Tirole (1986), a principal hires an agent to work
on a project that generates a benet b at an observable cost c(e, ) = e. Both the
parameter and eort e are the agents private information. The agents payo is
t (e), where t is a transfer and (x) is the cost of eort; the principals payo
is b c(e, ) t. The virtual surplus is
6
J(e, ) = b + e (e)
F()
f()

(e).
Letting x = e this satises Denition 3.1 as long as

> 0.
It can readily be veried that the canonical virtual surplus (2.1) is separable if
u is separable and values are private (i.e., v is independent of ).
Let J be separable with k : R as in Denition 3.1. For all q [0, 1], let
(3.2) h(q) := k(F
1
(q)),
and
(3.3) H(q) :=

q
0
h(r)dr.
Let convH be the convex hull of H on [0, 1] (see, e.g., Rockafellar, 1970). Dene
(3.4) G := convH.
Then G is the highest convex function on [0, 1] such that G H.
7
Since G is
convex, it is continuously dierentiable except possibly at countably many points.
Dene g : [0, 1] R as follows. For all q (0, 1) such that G

(q) exists, let

(3.5) g(q) := G

(q),
and extend g to all of [0, 1] by right-continuity. Finally, dene

k : R by
(3.6)

k() := g(F()).
Note that

k is nondecreasing by construction.
Armed with the function

k dened above, dene the generalized virtual surplus
(3.7)

J(x, ) := a(x)

k() + b(x) + l(),

where a, b and l are as in Denition 3.1. The generalized virtual surplus

J diers
from the virtual surplus J only in that k is replaced with

k. Dene the maximizer
correspondence : P(X), where P(X) is the power set of X, by
(3.8) () :=

x X |

J(x, ) = sup
yX

J(y, )

.
A function is a selection from if () () for all . By Topkis (1978),
a monotone (i.e., nondecreasing) selection exists if is nonempty-valued, since the
generalized virtual surplus

J has increasing dierences in (x, ) by construction.
Remark 3.4. J =

J if and only if k is nondecreasing. (Then H is convex so that
H = G, h = g and k =

k.) In that case the optimization problem in (3.8) reduces
to the standard practise of maximizing virtual surplus pointwise.
6
Take c to be the choice variable. Then payos are t ( c) and b c t, so the virtual
surplus is b c ( c)
F()
f()

( c). Using c = e we can write it in terms of eort.

7
I.e., convH(q) := min{H(q
1
) + (1 )H(q
2
) | (, q
1
, q
2
) [0, 1]
3
and q
1
+ (1 )q
2
= q}.
IRONING WITHOUT CONTROL 5
The eect of the monotonicity constraint is captured by the following condition.
Denition 3.5. : X has the pooling property if for all open intervals I ,
G(F()) < H(F()) for all I = is constant on I.
Remark 3.6. Since F, G and H are continuous, the pooling property requires to
be constant on any set where G F and H F dier.
The following theorem is a minor generalization of the single-agent version of
Myersons (1981) characterization of optimal auctions; the proof is presented in the
Supplementary material for completeness.
8
Theorem 3.7 (Myerson 1981). Let J be separable and let M. Then achieves
the supremum in (P) if and only if has the pooling property and () () a.e.
This result identies solutions to (P) with particular monotone selections from
. It implies the standard ironing result: Outside pooling intervals J =

J so the
solution there coincides with pointwise maximization of virtual surplus.
For the purposes of nding optimal allocation rules, the remarkable part of The-
orem 3.7 is given by two corollaries, which show that the transformed problem can
be solved by pointwise maximization even when the monotonicity constraint binds.
Corollary 3.8. Let J be separable. Assume () is non-empty and compact for
all , and let

() := max () and

() := min (). Then the selections

and

are monotone and achieve the supremum in (P).

Corollary 3.9. Let J be separable. Assume is single-valued except at countably
many points. Then any selection from attains the supremum in (P).
That is, the innite-dimensional constrained maximization problem (P) is equiv-
alent to the family of independent one-dimensional optimization problems in (3.8)
in the sense that the highest and the lowest optimal allocation rule can be found
by maximizing the generalized virtual surplus pointwise (without having to con-
sider the pooling property or the monotonicity of the selection). This is true of all
optimal allocation rules unless there is a non-degenerate region of types for which
multiple allocations maximize the generalized virtual surplus.
4. The concave case
I assume in this section that J is dened and once continuously dierentiable on
an open set containing X. This implies that the integral in problem (P) is well-
dened for any M. In order to simplify notation, I normalize the distribution
F to be uniform.
9
The results of this section require the virtual surplus J to be weakly concave in
the allocation x. Examples 3.2 and 3.3 satisfy this assumption provided that the
cost function c is weakly convex in Example 3.2 and we have

0 in Example
3.3. The following is an example of a non-separable concave problem.
8
Myerson only considers suciency, but necessity is straightforwardsee the Supplement.
9
As noted in footnote 4, this entails no loss of generality. Alternatively, the normalization
can be incorporated in the function h as in the separable case. Finally, if f is continuously
dierentiable, then it can simply be included in J by redening

J(x, ) := J(x, )f().
6 TOIKKA
Example 4.1 (Mussa and Rosen (1978) with interdependent values). Example 3.2
shows that the price discrimination model of Mussa and Rosen (1978) leads to a
separable problem. If the sellers cost depends directly on the buyers type (say,
because some types are more demanding and hence more costly to serve), then the
problem is in general no longer separable. The virtual surplus takes the form
J(x, ) = x

1 F()
f()

c(x, ),
which is weakly concave in x as long as the cost function c is weakly convex in x
for each . A version of the model is solved in section 4.4.
More generally, the canonical virtual surplus (2.1) is not separable even under
private values if the agents utility function is not separable. It will nevertheless be
concave in x if the social surplus v(x, ) + u(x, ) is weakly concave in x and the
agents marginal utility u
1
(x, ) is supermodular (i.e., u
112
0 so that changes in
marginal utility when x varies are weakly greater for higher types).
4.1. Denitions. I use essentially the same notation as in the separable case to
highlight the analogy between the constructions. The main dierence is that now
convexication has to be done allocation-by-allocation.
For all (x, ) X , let
(4.1) h(x, ) := J
1
(x, ),
and let
(4.2) H(x, ) :=

0
h(x, r)dr.
By assumption h is continuous (and hence integrable on {x} [0, ]) and thus
H(x, ) is continuously dierentiable on (0, 1) for any xed x. For all x X, let
(4.3) G(x, ) := convH(x, ).
G(x, ) is the highest convex function on [0, 1] such that G(x, ) H(x, ).
10
As
the convex hull of a dierentiable function, G(x, ) is continuously dierentiable on
(0, 1). Its derivative, denoted G
2
(x, ), is nondecreasing in . For all (0, 1), let
(4.4) g(x, ) := G
2
(x, ),
and extend g(x, ) to all of [0, 1] by continuity.
Dene the generalized virtual surplus

J : X R by
11
(4.5)

J(x, ) := J(0, ) +

x
0
g(s, )ds.
Dene the correspondence : P(X) by
(4.6) () :=

x X |

J(x, ) = sup
yX

J(y, )

.
Finally, let := { : X | is measurable}.
10
I.e., convH(x, ) = min{H(x,
1
)+(1)H(x,
2
) |
1
+(1)
2
= , (,
1
,
2
) [0, 1]
3
}.
11
I show below that the function g(, ) is continuous for all xed . It is thus bounded on the
compact set [0, x] and hence integrable.
IRONING WITHOUT CONTROL 7
Remark 4.2. Suppose J has increasing dierences (i.e., J
1
(x, ) is increasing in )
so that maximizing virtual surplus pointwise gives a solution that is nondecreasing
in . Then H(x, ) is convex (because H
2
(x, ) = J
1
(x, )). But then H = G and
h = g so that J =

J. Thus, under the standard sucient conditions for pointwise
maximization the generalized virtual surplus reduces to the virtual surplus.
Remark 4.3. If J is separable (but not necessarily concave), then the above con-
struction yields the same generalized virtual surplus (up to a constant) as the one
dened in (3.7) for the separable case (see the Supplementary material for details).
In this sense the above approach is a proper generalization of Myersons.
4.2. The results. With the denitions in place, I am now ready to state the main
results of the paper, the proofs of which are in the next subsection.
Theorem 4.4 (Values). If J is weakly concave in x, then
sup
M

1
0
J((), )d

= sup

1
0

J((), )d

.
That is, in order to nd the maximized expected value of the virtual surplus
function over all nondecreasing allocation rules in (P), it suces to maximize the
expected generalized virtual surplus over all (measurable) allocation rules. The
latter can be done pointwise. Indeed, by denition any selection from attains the
supremum on the right-hand side.
In terms of economics, Theorem 4.4 shows that if the agents utility function
in the canonical principal-agent model has the single-crossing property, then the
generalized virtual surplus captures all implications of incentive compatibility on
the principals problem. In this sense it generalizes the standard notion of virtual
surplus which only accounts for local incentive compatibility constraints.
It turns out that while any solution to (P) is (almost everywhere equal to) a
monotone selection from , there may be monotone selections form that are not
solutions to (P).
12
However, the smallest and largest optimal allocation rule can
be obtained by maximizing the generalized virtual surplus independently for each
type. In order to state this result, dene for all
(4.7)

() := max () and

() := min ().
Theorem 4.5 (Maximizers). Let J be weakly concave in x. Then

and

are
monotone and attain the supremum in (P). Furthermore, if M attains the
supremum in (P), then

() ()

() and () () a.e.
Note that this result also establishes the existence of a maximizer in problem
(P). If is single-valued except at countably many points, then all selections are
monotone and agree with

a.e. This gives the following analog of Corollary 3.9.

12
Maximizers can be characterized by a pooling property as in the separable case. A function
M has the generalized pooling property (gpp) if for all x X and all open intervals I ,

1
(x) I and G(x, ) < H(x, ) for all I = is constant on I,
where
1
is the generalized inverse dened in (4.8). It can be shown that if J is weakly concave
in x, then M attains the supremum in (P) if and only if has gpp and () () a.e.
Checking gpp is in general quite complicated. Furthermore, unlike in the separable case, the
result does not immediately characterize optimal pooling intervals since the intervals depend on
the allocation rule . For these reasons I do not pursue this direction further.
8 TOIKKA
Corollary 4.6. Let J be weakly concave in x. Assume is single-valued except at
countably many points. Then any selection from attains the supremum in (P).
Remark 4.7. The supplementary material contains an example showing that in gen-
eral weak concavity of J in x cannot be dropped from the assumptions in Theorems
4.4 and 4.5. This is because the construction here considers only small changes in
the allocation as it builds upon the derivative J
1
(x, ). Under concavity this is
sucient as then local optimality implies global optimality.
13
Remark 4.8. Since

J is weakly concave in x, the extremal selections satisfy

() = max {x X | g(x, ) 0} and

() = min {x X | g(x, ) 0} ,
where max = 0 and min = x.
Subsection 4.4 contains a fully solved example. While obtaining closed-form
solutions can be tedious, the results can be used to derive comparative statics
and other properties of the maximizers to problem (P). I illustrate this with two
applications: I rst argue that strict concavity of J in x is a sucient condition for
the maximizer to be unique and hence continuous. I then show that the solutions
have the familiar ironing property.
Suppose J is strictly concave in x. The argument in the proof of Lemma 4.11
can be adapted to show that then

J is also strictly concave in x. Hence (4.6)
is single-valued and thus continuous by the Maximum theorem. Thus there is a
unique, continuous optimal allocation rule. The general point illustrated here is
that it can be relatively easy to show that the convexication procedure preserves
certain properties of J which then translate to properties of solutions.
In order to show the ironing result, note rst that if J is weakly concave and
continuously dierentiable in x, so is

J (see Lemmas 4.10 and 4.11 below). Thus
rst-order conditions are necessary and sucient when maximizing

J(x, ). For
simplicity, suppose that for each there exists a unique interior allocation () that
maximizes the generalized virtual surplus

J(x, ). If () does not also maximize
the virtual surplus J(x, ), then by weak concavity of J in x we have
J
1
((), ) = h((), ) = 0 = g((), ) =

J
1
((), ).
But then H((), ) and G((), ) dier in some open neighborhood I of . Thus
g((), ) is at on I so that g((), ) = 0 for all I. Since we have assumed a
unique maximum for each type, this implies that () is assigned to all types in I.
That is, if the allocation () does not maximize virtual surplus from type , then
belongs to a pooling interval. Outside these intervals the solution coincides with
pointwise maximization of virtual surplus.
4.3. Proofs. I rst establish three technical lemmas, the rst two of which do
not require concavity. Their proofs are somewhat tedious but not particularly
illuminating, and hence relegated to the Appendix.
Lemma 4.9. The function g dened in (4.4) is continuous in x.
Lemma 4.10. The generalized virtual surplus

J dened in (4.5) is continuous,
continuously dierentiable in x, and has increasing dierences in (x, ).
13
Quasi- or pseudo-concavity does not suce as the construction here involves taking integrals
and the sum of two quasiconcave functions need not be quasiconcave.
IRONING WITHOUT CONTROL 9
Lemma 4.11. If the virtual surplus J is weakly concave in x, then the generalized
virtual surplus

J is weakly concave in x.
Given the properties of

J established in Lemma 4.10, the Maximum theorem
(Berge, 1963) implies that the correspondence dened by (4.6) is upper hemi-
continuous, nonempty, and compact-valued. As

J has increasing dierences, the
selections

and

dened in (4.7) exist and are nondecreasing.

It is convenient to introduce two more pieces of notation. For all M, dene
the generalized inverse
1
: X by
(4.8)
1
(x) := inf { : () x} ,
where inf = 1 by convention. Denote the dierence between the expected virtual
surplus and the expected generalized virtual surplus from allocation rule M by
() :=

1
0

J((), )

J((), )

d.
Since J(x, ) =

J(x, ) +

J(x, )

J(x, )

, the objective in (P) can be written as

1
0
J((), )d =

1
0

J((), )d + ().
The following two lemmas about () are the key to the proof.
Lemma 4.12. sup
M
() = 0.
Proof. For all M we have
() =

1
0

J((), )

J((), )

1
0

J(0, ) +

()
0
h(s, )ds

J(0, ) +

()
0
g(s, )ds

d
=

1
0

()
0
(h(s, ) g(s, )) dsd
=

x
0

1
(x)
(h(x, ) g(x, )) ddx
=

x
0

G(x,
1
(x)) H(x,
1
(x))

dx 0,
(4.9)
where the last equality follows since G(x, 1) = H(x, 1) for all x X. The inequality
is by denition of G. The constant function 0 achieves the bound.
Lemma 4.13. If J is weakly concave in x, then (

) = (

) = 0.
Given any M, let
A

:=

x (0, x) | G(x,
1
(x)) H(x,
1
(x)) < 0

.
Inspecting the last line in (4.9) it is seen that in order to prove Lemma 4.13 it
suces to show that A

and A

have at most countably many elements. In

showing this I use the following observation.
Claim. Let be a monotone selection from and let x A

. Then there exists

an open neighborhood U of
x
:=
1
(x) such that x () for all U.
10 TOIKKA
Proof of the Claim. Fix a monotone selection and x A

. Since is monotone,
we have lim
x
() x lim
x
(). Upper hemi-continuity of then implies
lim
x
() (
x
) and lim
x
() (
x
). But

J(,
x
) is concave by Lemma
4.11 and hence (
x
) is convex. Thus x (
x
). Since x is interior and

J(,
x
) is
continuously dierentiable in x by Lemma 4.10, we have the rst order condition

J
1
(x,
x
) = g(x,
x
) = 0.
As G(x, ) and H(x, ) are continuous, G(x,
x
) H(x,
x
) < 0 implies that there
is an open neighborhood U of
x
such that G(x, ) H(x, ) < 0 on U, and hence
g(x, ) is constant on U. Combined with the above rst order condition this gives

J
1
(x, ) = g(x, ) = g(x,
x
) = 0 for all U.
Since

J is concave in x, this implies that x () for all U.
Proof of Lemma 4.13. Let x A

. By the above Claim, there exists an open

neighborhood U of
x
:=
1

(x) such that x () for all U. Thus

() =
min () x for all U. But by denition of the inverse,

() x for all
>
x
. Thus

() = x for all U such that >

x
. But then x is a point of
discontinuity of
1

. Since
1

is monotone, there can be at most countably many

such points. Hence A

has at most countably many elements.

Let y A

. Again there is an open neighborhood V of

y
such that y ()
for all V . Thus

() = max () y for all V . But the denition of the

inverse implies

() < y for all <

y
, a contradiction. Thus A

is empty.
We are now ready to complete the proofs of the theorems.
Proof of Theorem 4.4. Since M , we have
sup

1
0

J((), )d

sup
M

1
0

J((), )d

sup
M

1
0

J((), )d + ()

= sup
M

1
0
J((), )d

,
where the second line is by Lemma 4.12 and the last by denition of ().
In the other direction, if J is weakly concave in x, we have
sup
M

1
0
J((), )d

1
0
J(

(), )d
=

1
0

J(

(), )d = sup

1
0

J((), )d

,
where the second line is by Lemma 4.13 and the denition of

.
Proof of Theorem 4.5. By inspection of the proof of Theorem 4.4,

attains the
supremum in (P). The same argument shows that

attains the supremum in (P).

Assume then that M attains the supremum in (P). Since () 0 by
Lemma 4.12, we have

1
0

J((), )d

1
0
J((), )d = sup
M

1
0
J((), )d

.
IRONING WITHOUT CONTROL 11
Thus () () a.e. by Theorem 4.4. The inequalities follow by (4.7).
4.4. An example. I conclude by presenting an example of a concave (but non-
separable) problem where the optimal allocation rule is discontinuous and hence it
can not be solved using the standard optimal control approach. Nevertheless, the
virtual surplus is weakly concave in the allocation and the technique developed in
this paper applies.
Consider a model of monopolistic price discrimination with interdependent values
as in Example 4.1. The buyers valuation is given by u(x, ) = x(
2
+
1
6
), and the
sellers cost is given by
c(x, ) =

0 if x < 1,
1
2
( + 1)(x 1)
2
if x 1.
That is, the sellers marginal cost is constant (and normalized to zero) for the rst
unit, and it is increasing thereafter. The type is distributed uniformly on the unit
interval. The set of feasible allocations is X = R
+
. Virtual surplus takes the form
J(x, ) =

x(
2
+
1
6
) (1 )2x if x < 1,
x(
2
+
1
6
) (1 )2x
1
2
( + 1)(x 1)
2
if x 1
= (3
2
2 +
1
6
)x
1
2
( + 1)(x 1)
2
1
{x1}
.
This specication captures a situation where higher types of the buyer have a
higher marginal willingness to pay for the good or service, but at the same time are
more expensive to serve than the lower types. One can think of x as the quality of
a restaurant meal or of some other service, where the customers who are willing to
pay the most for given quality are also the most demanding and require the most
attention from the sta. Because of these countervailing eects, ironing is needed:
Maximizing virtual surplus pointwise suggests the allocation rule

() =

3
2
2+
7
6
+1
if [0,
2

2
6
] [
2+

2
6
, 1],
0 if (
2

2
6
,
2+

2
6
).
In a sense then, the middle types are the worst. The highest types, while being
the most expensive to serve, have a high enough marginal willingness to pay to
be attractive to the seller. Similarly, the lowest types are attractive despite their
low willingness to pay since serving them is inexpensive. Note that because of the
initially constant marginal cost,

is discontinuous.
In order to apply Theorems 4.4 and 4.5, we use the denitions in (4.1)(4.5) to
construct the generalized virtual surplus. We have
h(x, ) =

3
2
2 +
1
6
if x < 1,
3
2
(x + 1) x +
7
6
if x 1,
and
H(x, ) =

2
+
1
6
if x < 1,

x+1
2

2
(x
7
6
) if x 1.
Note that, for x small enough, H(x, ) is initially concave and then convex. This
makes convexifying simple as it suces to iron out the initial concave part.
12 TOIKKA
Straightforward calculations give the convex hulls
G(x, ) =

1
12
if <
1
2
,

2
+
1
6
if
1
2
,
for all x < 1,
and
G(x, ) =

(
x
2
16

9
8
x +
53
48
) if <
x+1
4
,

x+1
2

2
(x
7
6
) if
x+1
4
,
for all x 1.
Thus we have
(4.10) g(x, ) =

1
12
if <
1
2
,
3
2
2 +
1
6
if
1
2
,
for all x < 1,
and
(4.11) g(x, ) =

x
2
16

9
8
x +
53
48
if <
x+1
4
,
3
2
(x + 1) x +
7
6
if
x+1
4
.
for all x 1.
Recall from Remark 4.8 that the maximal optimal allocation rule is given by

() = max {x X | g(x, ) 0} ,
where max = 0. Since g is weakly decreasing x, we immediately see from (4.10)
that

() = 0 for all <

2

2
6
. We may then note that the rst case in (4.11)
is negative. Hence the optimal allocation for
2

2
6
can be solved for from the
second case in (4.11). Thus we have

() =

0 if [0,
2+

2
6
),
3
2
2+
7
6
+1
if [
2+

2
6
, 1].
That is, the lowest types are excluded, then there is a jump from 0 to 1 at
2+

2
6
after which the solution coincides with pointwise maximization of standard virtual
surplus.
Appendix: Proofs of the technical lemmas
Proof of Lemma 4.9. The claim is established by showing successively that all the
functions dened in (4.1)(4.4) are continuous (in x).
By denition h = J
1
is continuous, since J is assumed once continuously dier-
entiable on an open set containing X .
For continuity of H, x (x, ) X . For any sequence converging to (x, ),
lim
n
H(x
n
,
n
) = lim
n

1
0
h(x
n
, r)1
{rn}
dr =

1
0
h(x, r)1
{r}
dr = H(x, ),
where the second equality follows by Lebesgues dominated convergence theorem
since

h(x
n
, r)1
{rn}

sup
(x,)X
|h(x, )| < for all (x
n
,
n
, r) X
2
by
continuity of h and compactness of its domain, and since for a.e. r , we have
lim
n
h(x
n
, r)1
{rn}
= h(x, r)1
{r}
by continuity of h. Thus H is continuous.
I then show that G is continuous in x. By denition,
G(x, ) = max{H(x,
1
) (1 )H(x,
2
) :
(,
1
,
2
) [0, 1]
3
and
1
+ (1 )
2
= },
IRONING WITHOUT CONTROL 13
Holding xed, the objective function in the above maximization problem is
continuous in (,
1
,
2
, x) and the feasible set is compact and independent of x.
Thus G is continuous in x by the Maximum theorem.
Consider then g. It is convenient to extend G to X D, where D R is
an open set containing [0, 1] = . Since G(, ) is continuous for all , and
G(x, ) is continuously dierentiable on (0, 1) and convex on [0, 1] for all x X,
the extension can be chosen such that G(, ) is continuous for all D, and
G(x, ) is continuously dierentiable and convex on D for all x.
14
We then have
g(x, ) = G
2
(x, ) for all (x, ) X .
Fix [0, 1]. Let x X and let (x
n
) be a sequence in X such that x
n
x. We
have
limsup
xnx
g(x
n
, ) = limsup
xnx
lim
t0
G(x
n
, + t) G(x
n
, )
t
= limsup
xnx
inf
t>0
G(x
n
, + t) G(x
n
, )
t
limsup
xnx
G(x
n
, + s) G(x
n
, )
s
s > 0 : + s D
=
G(x, + s) G(x, )
s
s > 0 : + s D,
where the rst equality is by denition of g, the second by convexity of G in , and
the last by continuity of G in x. Therefore,
limsup
xnx
g(x
n
, ) lim
s0
G(x, + s) G(x, )
s
= g(x, ).
An analogous argument gives
liminf
xnx
g(x
n
, )
G(x, ) G(x, s)
s
s > 0 : s D,
which implies
liminf
xnx
g(x
n
, ) lim
s0
G(x, ) G(x, s)
s
= g(x, ).
Thus lim
xnx
g(x
n
, ) = g(x, ) so that g(, ) is continuous at x. Since x and
were arbitrary, the claim follows.
Proof of Lemma 4.10. For any xed , the function

J is an integral function in x,
and the integrand g(, ) is continuous by Lemma 4.9. Hence

J is continuously
dierentiable in x for all .
For continuity, x (x, ) X and consider a sequence in X such that
(x
n
,
n
) (x, ). Note that g is nondecreasing and continuous in by construction,
and it is continuous in x by Lemma 4.9. Hence for all (s,
n
) X ,
< inf
yX
g(y, 0) g(s,
n
) sup
zX
g(z, 1) < .
14
For example, for all x X, let G(x, ) be ane on D\ [0, 1] with G
2
(x, ) = lim
0
G
2
(x, )
for all < 0 and G
2
(x, ) = lim
0
G
2
(x, ) for all > 1, where the ane parts are chosen such
that G(x, ) is continuous on D. By construction, G(x, ) so extended is convex and continuously
dierentiable. Continuity of G(, ) for all D \ [0, 1] follows from the continuity of G(, 0) and
G(, 1), since, e.g., |g(x, ) g(y, )| = |g(x, 0) g(y, 0)| for all < 0 and all (x, y) X
2
.
14 TOIKKA
We thus have
lim
n

J(x
n
,
n
) = lim
n
J(0,
n
) + lim
n

x
0
g(s,
n
)1
{sxn}
ds
= J(0, ) +

x
0
g(s, )1
{sx}
ds =

J(x, ),
where the second equality is by continuity of J and the Lebesgues dominated
convergence theorem as

g(s,
n
)1
{sxn}

max {|inf
yX
g(y, 0)| , |sup
zX
g(z, 1)|}
for all (s, x
n
,
n
) X
2
, and since for a.e. s X, lim
n
g(s,
n
)1
{sxn}
=
g(s, )1
{sx}
by continuity of g(s, ). Thus

J is continuous.
To show increasing dierences in (x, ), take any (x, x

) X
2
and (,

)
2
such that x

x and

. Then

J(x

)

J(x,

) =

x
g(s,

)ds

x
g(s, )ds =

J(x

, )

J(x, ),
where the inequality follows from g being nondecreasing in .
Proof of Lemma 4.11. The generalized virtual surplus

J is dierentiable in x by
Lemma 4.10. Recalling the denition from (4.5) we have

J
1
(x, ) = g(x, ). So it
suces to show that g is nonincreasing in x for any xed .
To this end, note rst that h(x, ) = J
1
(x, ) is nonincreasing in x for any
by the weak concavity of J in x. By (4.2) we have for all x

> x and

> ,
H(x

) H(x

, ) =

h(x

, r)dr

h(x, r)dr = H(x,

) H(x, ).
That is, H has increasing dierences in (x, ).
Suppose then to the contrary of g being nonincreasing in x that there exists
(x, x

,
o
) X
2
such that x

> x and g(x

,
o
) > g(x,
o
). Since g(x

, ) and
g(x, ) are continuous, it is without loss to assume that
o
is interior. Let L : [0, 1]
R be the unique ane function tangent to G(x, ) at
o
. Similarly, let L

: [0, 1] R
be the ane function tangent to G(x

, ) at
o
. Denote := G(x,
o
) G(x

,
o
).
Claim 1.
o
: H(x, ) H(x

, ) .
Proof of Claim 1. By denition G(x, ) is convex and lies everywhere below H(x, )
so that
H(x, ) G(x, ) L() for all .
If H(x

,
o
) = G(x

,
o
), we are done since then the rst of the above inequalities
(evaluated at
o
) implies H(x,
o
) H(x

,
o
) G(x,
o
) G(x

,
o
) = . Otherwise
H(x

,
o
) > G(x

,
o
). But then G(x

, ) is ane and coincides with L

in a neighbor-
hood of
o
. Moreover, the point (
o
, G(x

,
o
)) is a convex combination of two points
from the graph of H(x

, ) so that there exists

1
<
o
such that H(x

,
1
) = L

(
1
).
By construction
d
d
(L() L

()) = g(x,
o
) g(x

,
o
) < 0. This together with the
second of the above inequalities implies H(x,
1
) H(x

,
1
) L(
1
) L

(
1
) >
L(
o
) L

(
o
) = G(x,
o
) G(x

,
o
) = .
Claim 2. >
o
: H(x, ) H(x

, ) < .
IRONING WITHOUT CONTROL 15
Proof of Claim 2. Assume rst that H(x,
o
) = G(x,
o
). The dierence H(x, )
L

() is a continuously dierentiable function of with H(x,

o
)L

(
o
) = G(x,
o
)
G(x

,
o
) = . Moreover,
d
d
(H(x, ) L

()) |
=o
= g(x,
o
) g(x

,
o
) < 0 so that
there exists
2
>
o
such that > H(x,
2
) L

(
2
) H(x,
2
) H(x

,
2
) as we
wanted to show.
Assume then that H(x,
o
) > G(x,
o
). Then G(x, ) coincides with L in a
neighborhood of
o
and there exists
3
>
o
such that H(x,
3
) = L(
3
). But then
H(x,
3
) H(x

,
3
) L(
3
) L

(
3
) < L(
o
) L

(
o
) = .
We have arrived at the desired contradiction since Claims 1 and 2 are incompatible
with H having increasing dierences in (x, ). Hence g is nonincreasing in x
implying that

J is weakly concave in x.
References
Berge, C. (1963): Topological Spaces. Oliver & Boyd.
Clarke, F. (1976): The Maximum Principle under Minimal Hypothesis, SIAM
Journal of Control and Optimization, 14, 10781091.
(2011): A General Theorem on Necessary Conditions in Optimal Con-
trol, Discrete and Continuous Dynamical Systems, 29(2), 485503.
Fudenberg, D., and J. Tirole (1991): Game Theory. MIT Press.
Hellwig, M. F. (2009): A Maximum Principle for Control Problems with Mono-
tonicity Constraints, Working Paper, Max Blanck Institute.
Jullien, B. (2000): Participation Constraints in Adverse Selection Models,
Journal of Economic Theory, 93(1), 147.
Laffont, J., and J. Tirole (1986): Using Cost Observation to Regulate Firms,
Journal of Political Economy, 94(3), 614641.
L eonard, D., and N. van Long (1992): Optimal Control Theory and Static
Optimization in Economics. Cambridge University Press.
Mussa, M., and S. Rosen (1978): Monopoly and Product Quality, Journal of
Economic Theory, 18(2), 301317.
Myerson, R. B. (1981): Optimal Auction Design, Mathematics of Operation
Research, 6(1), 5873.
N oldeke, G., and L. Samuelson (2007): Optimal Bunching without Optimal
Control, Journal of Economic Theory, 134(1), 405420.
Rockafellar, R. T. (1970): Convex Analysis. Princeton University Press.
Salani e, B. (1997): The Economics of Contracts: A Primer. MIT Press.
Topkis, D. M. (1978): Minimizing a Submodular Function on a Lattice, Oper-
ations Research, 26(2), 305321.
Vinter, R. (2010): Optimal Control. Springer.
MIT Department of Economics
E-mail address: toikka@mit.edu