a

r
X
i
v
:
1
1
1
0
.
2
5
7
3
v
1


[
q
-
f
i
n
.
P
M
]


1
2

O
c
t

2
0
1
1
Optimal investment with intermediate
consumption and random endowment
Oleksii Mostovyi
Carnegie Mellon University,
Department of Mathematical Sciences,
5000 Forbes Avenue, Pittsburgh, PA, 15213-3890, US
(omostovy@andrew.cmu.edu)
Abstract
In this paper we study several classical problems of optimal in-
vestment with intermediate consumption and random endowment in
incomplete markets. We establish the key assertions of the utility
maximization theory assuming that both primal and dual value func-
tions are nite in the interiors of their domains as well as that random
endowment at maturity can be dominated by the terminal value of
a self-nancing wealth process. In order to facilitate verication of
these conditions, we present alternative, but equivalent conditions,
under which the conclusions of the theory hold.
1 Introduction
The problem of utility maximization in incomplete markets is one of the cen-
tral problems in mathematical nance. The theory was developed by He and
Pearson [8, 9], Karatzas, Lehoczky, Shreve, and Xu [14], Karatzas and Shreve
[12], Kramkov and Schachermayer [16, 17], Karatzas and

Zitkovic [13], and

Zitkovic [21]. In particular, Kramkov and Schachermayer [16, 17] considered

the problem of optimal investment from terminal wealth and found minimal
conditions on the utility function itself as well as the condition on the utility
function and on the market model, such that several key assertions of the
utility maximization theory hold.
1
In this paper we consider a problem for an agent, who in addition to
the initial wealth receives random endowment. The goal of such an agent
is to consume and invest in a way that maximizes his expected utility. In
complete market settings this problem is considered by Karatzas and Shreve
[12], see Chapter 4. Using replication argument, the authors were able to
reduce the problem to one without endowment. Since in incomplete markets
such replication might not be possible, alternative techniques were used. For
example, Cuoco [1] used martingale techniques to reformulate the dynamic
optimization problem as an equivalent static one. In Markovian settings a
possible approach is to use a Hamilton-Jacobi-Bellman equation for the value
function, see Due and Zariphopoulou [6] and Due, Fleming, Soner, and
Zariphopoulou [7]. Cvitanic, Schachermayer, and Wang [2] considered the
problem of optimal investment from terminal wealth under the presence of
random endowment in incomplete semimartingale market. Using the space
(L

of nitely additive measures as the domain of the dual problem they

were able to characterize the value function and the optimal terminal wealth
in terms of the solution to the dual problem.
In contrast to [2] Hugonnier and Kramkov [10] treated not only the initial
capital as the variables of the optimization problem, but also the number of
shares of random endowment. Although this lead to higher dimensionality
of the problem, it allowed to relax some technical assumptions. Stability of
this utility maximization problem was investigated by Kardaras and

Zitkovic
[15]. Karatzas and

Zitkovic [13] as well as

Zitkovic [21] extended the results
of Cvitanic, Schachermayer, and Wang [2] to include the intermediate con-
sumption. Introducing the stochastic clock process and assuming certain
regularity conditions on the utility stochastic eld, they solved several clas-
sical problems of utility maximization in one formulation.
Mostovyi [18] considered the problem of optimal investment with inter-
mediate consumption under the condition that both primal and dual value
functions are nite in their domains. It is shown in [18] that such conditions
are both necessary and sucient for the validity of the key conclusions of the
theory.
Present work extends the results of Mostovyi [18] to incorporate random
endowment process into the model. As in [10] we treat number of shares
of random endowment as the variables of the optimization problem. This
approach allows us to obtain the desired conclusions of the utility maximiza-
tion theory under the minimal conditions on the endowment process, market
model and the utility stochastic eld. Thus we assume that random endow-
2
ment at maturity can be dominated by the terminal value of a self-nancing
nonnegative wealth process as well as that both primal and dual functions
are nite in the interiors of their domains. In order to facilitate verication
of the latter conditions, we present an alternative, but equivalent criterion
in terms of the niteness of the value functions without random endowment.
The condition on the endowment can also be formulated in several equivalent
ways.
It turns out to be possible to establish some properties of the primal
and dual value functions on the boundaries of their domains, namely to
show upper semi-continuity of the primal value function u and lower semi-
continuity of dual value function.
The remainder of this paper is organized as follows. In section 2 we
describe the mathematical model and state our main results. In section 3 we
prove them.
2 Main Results
We consider a model of a nancial market with nite time horizon [0, T] and
zero interest rate. The price process S = (S
i
)
d
i=1
of the stocks is assumed
to be a semimartingale on a ltered probability space (, F, F, P) , where
the ltration F = (F
t
)
t[0,T]
satises the usual conditions. We assume that
there are non-traded contingent claims with a payment process F = (F
i
)
N
i=1
.
If q = (q
i
)
N
i=1
is the number of such claims then the cumulative payo of this
portfolio is given by
qF
N

i=1
q
i
F
i
=
_
N

i=1
q
i
F
i
t
_
t[0,T]
.
Thus, the random variable qF
t
stands for the cumulative amount of endow-
ment received by a holder of q such claims during the time interval [0, t].
Both processes S and F are given exogenously.
In order to consider several utility maximization problems in one formu-
lation, let us dene a stochastic clock as a nondecreasing, cadl ag, adapted
process that starts at zero and has a nite value at maturity T. For nor-
malization purposes we assume that this value equals to 1 (without loss of
generality). A stochastic clock induces a measure on [0, T].
Dene a portfolio as a quadruple (x, q, H, c), where the constant x is
the initial value of the portfolio, vector q gives the number of shares of illiquid
3
contingent claims, H = (H
i
)
d
i=1
is a predictable Sintegrable process that
species the amount of each stock in the portfolio, and c = (c
t
)
t[0,T]
is the
consumption rate, which we assume to be optional and nonnegative.
The wealth process V = (V
t
)
t[0,T]
of such a portfolio is dened as
V
t
= x +
_
t
0
H
s
dS
s

_
t
0
c
s
d
s
+qF
t
, t [0, T].
A portfolio with c 0 and q = 0 is called self-nancing. The collection
of nonnegative wealth processes of self-nancing portfolios with initial value
x > 0 is denoted by X(x), i.e.
X(x)
_
X 0 : X
t
= x +
_
t
0
H
s
dS
s
, t [0, T]
_
, x > 0.
A probability measure Q is an equivalent local martingale measure if Q is
equivalent to P and any X X(1) is a local martingale under Q. We denote
the family of equivalent local martingale measures by M and assume that
(2.1) M = .
This condition is essentially equivalent to the absence of arbitrage oppor-
tunities on the market without endowment and consumption, see Delbaen
and Schachermayer [3, 5] as well as Karatzas and Kardaras [11] for the exact
statements and further references.
To rule out doubling strategies in the presence of random endowment we
need to impose additional restrictions. Following Delbaen and Schachermayer
[4], we say that a nonnegative process in X(x) is maximal if its terminal value
cannot be dominated by that of any other process in X(x).
As in [4], we dene an acceptable process to be a process of the form
X = X

, where X

is nonnegative and X

is maximal. Following
Hugonnier and Kramkov [10], we denote by X(x, q) the set of acceptable
processes with initial value x whose terminal value dominates the random
payo qF
T
:
(2.2) X(x, q) {X : X is acceptable, X
0
= x and X
T
+qF
T
0} .
The set X(x, q) may be empty for some (x, q) R
N+1
. We are interested in
the values of x and q, for which X(x, q) = , and dene
(2.3) K int
_
(x, q) R
N+1
: X(x, q) =
_
.
4
It is proved Lemma 6 in Hugonnier and Kramkov [10] that
(2.4) clK =
_
(x, q) R
N+1
: X(x, q) =
_
,
where clK denotes the closure of the set K in R
N+1
.
As in

Zitkovic [21], we restrict our attention to the wealth processes with
nonnegative terminal values. Thus for each (x, q) clK we denote
(2.5)
A(x, q)
_
c = (c
t
)
t[0,T]
: c is nonnegative, optional,
and there exists X X(x, q) s.t. X
T

_
T
0
c
t
d
t
+qF
T
0
_
.
For the treatment of the utility maximization problem (2.7) we need the
following assumption on the endowment process.
Assumption 2.1. (F
i
T
)
i=1,...,N
are F
T
-measurable functions. There exists
a maximal nonnegative wealth process X

of a self-nancing portfolio, such

that
(2.6) X

T

N

i=1
|F
i
T
|.
Remark 2.2. Since in the denition (2.5) the endowment process F enters
only via its terminal value, it was natural to impose a regularity condition
on F
T
(and not on the whole F) as in the Assumption 2.1. This condition
is equivalent to the assumption on random endowment in Hugonnier and
Kramkov [10] (see equation (5)). As a result in order to check that (2.6) holds
one can use alternative equivalent conditions in the statement of Lemma 1
in [10]. By the same lemma under Assumption 2.1, (x, 0) K for all x > 0.
The preferences of an economic agent are modeled with a utility stochastic
eld dened on a positive real line. U : [0, T][0, ) R, a B([0, T])
F B([0, )) measurable function, where B(S) denotes the smallest
eld containing all open sets of a topological space S. We assume that
U satises the conditions below.
Assumption 2.3. For each (t, ) [0, T] the function x U(t, , x) is
increasing, strictly concave, continuously dierentiable, satises Inada con-
ditions:
lim
x0
U

(t, , x) = + and lim

x
U

(t, , x) = 0,
5
where U

(, , ) denotes the partial derivative with respect to the last ar-

gument, U(t, , 0) = lim
x0
U(t, , x). For each x 0, the stochastic process
U(, , x) is optional.
The agent can control investment and consumption. His goal is to do this
in a way that maximizes expected utility. For any (x, q) clK he wants to
compute the value function u, dened as:
(2.7) u(x, q) sup
cA(x,q)
E
__
T
0
U(t, , c
t
)d
t
_
.
If E
_
_
T
0
U

(t, , c
t
)d
t
_
= +for some c, to ensure that E
_
_
T
0
U(t, , c
t
)d
t
_
is well-dened, we let E
_
_
T
0
U(t, , c
t
)d
t
_
= . Here and below U

de-
notes the negative part of the stochastic eld U. Also, we set u(x, q)
for (x, q) (clK )
c
.
We are primarily interested in the following questions.
(i) Under what conditions on the market model and on the utility stochas-
tic eld U does the maximizer to the problem (2.7) exist for every
(x, q) {u > }?
(ii) What are the properties of the function u?
(iii) What is the corresponding dual problem?
We employ convex duality techniques to answer these questions and dene a
convex conjugate stochastic eld V as the pointwise Fenchel-Legendre trans-
form of the agents utility U:
V (t, , y) sup
x>0
(U(t, , x) xy) , (t, , y) [0, T] (0, ).
To construct the feasible set of the dual problem we dene the set L as
the relative interior of the polar cone K :
(2.8) L ri
_
(y, r) R
N+1
: xy +qr 0 for all (x, q) K
_
.
It is proved that L is an open set in R
N+1
if and only if for any q = 0 the
random variable qF
T
is non-replicable, see Lemma 7 in [10] for details.
6
By

Y we denote the set of cadlag densities of equivalent local martingale
measures:

Y
_
_
dQ
t
dP
t
_
t[0,T]
, Q M
_
,
and for each y > 0 we dene
(2.9)
Y (y) cl {Z : Z is cadlag adapted and
0 Z yY (d P) a.e. for some Y

Y
_
.
Here the closure is taken in the topology of convergence in measure (d P)
on the space of optional processes on the probability space. Now we are
ready to set the domain of the dual problem:
(2.10)
Y (y, r)
_
Y Y (y) : E
_
_
T
0
c
t
Y
t
d
t
_
xy +qr
for all (x, q) clK and c A(x, q)} , (y, r) clL;
and to state the dual optimization problem itself:
(2.11) v(y, r) inf
Y Y (y,r)
E
__
T
0
V (t, , Y
t
)d
t
_
, (y, r) clL.
Similarly to denition of the primal problem, if E
_
_
T
0
V
+
(t, , Y
t
)d
t
_
=
+, we let E
_
_
T
0
V (t, , Y
t
)d
t
_
= +. Here and below V
+
denotes the
positive part of the eld V. Also, let v(y, r) + for all (y, r) (clL)
c
.
The following theorem constitutes the main result of this work.
Theorem 2.4. Assume that (2.1), Assumptions 2.1 and 2.3 hold as well as
(2.12)
u(x, q) > for all (x, q) K and v(y, r) < for all (y, r) L.
Then
(i) The functions u, v are nitely valued on K and L respectively, u and
v satisfy biconjugacy relations:
(2.13)
u(x, q) = inf
(y,r)clL
{v(y, r) +xy +qr} , (x, q) clK ;
v(y, r) = sup
(x,q)clK
{u(x, q) xy qr} , (y, r) clL.
7
(ii) The function u is upper semi-continuous, u(x, q) < for all (x, q)
R
N+1
. For any (x, q) {u > } there exists a unique maximizer to
the problem (2.7).
The function v is lower semi-continuous, v(y, r) > for all (y, r)
R
N+1
. For any (y, r) {v < } there exists a unique minimizer to the
problem (2.11).
(iii) For any (x, q) K the subdierential of u at (x, q) is non-empty.
Moreover, (y, r) u(x, q) if and only if the following conditions hold:
(2.14)

Y
t
(y, r) = U

(t, , c
t
(x, q)) , t [0, T],
where

Y (y, r) and c(x, q) are optimizers to problems (2.11) and (2.7)
respectively;
(2.15) E
__
T
0

Y
t
(y, r) c
t
(x, q)d
t
_
= xy +qr;
(2.16) |v(y, r)| < .
Remark 2.5. Condition (2.12) might be dicult to verify. Let
(2.17) w(x) u(x, 0) = sup
cA(x,0)
E
__
T
0
U (t, , c
t
) d
t
_
, x > 0.
(2.18) w(y) inf
Y Y (y)
E
__
T
0
V (t, , Y
t
) d
t
_
, y > 0.
In Lemma 2.6 we show that instead of checking (2.12) one can verify an
equivalent condition:
(2.19) w(x) > for all x > 0 and w(y) < for all y > 0.
Note that (2.19) is precisely the condition that was used by Mostovyi [18] in
the statement of the main theorem.
Lemma 2.6. Let (2.1) and Assumptions 2.1, 2.3 hold. Then condition (2.12)
holds if and only if condition (2.19) holds.
8
3 Proofs
We start from a proposition that gives a useful characterization of the do-
mains of the primal and dual problems.
Proposition 3.1. The families (A(x, q))
(x,q)clK
and (Y (y, r))
(y,r)clL
de-
ned in (2.5) and (2.10) have the following properties:
(i) For any (x, q) K , the set A(x, q) contains a strictly positive constant
process. For any (x, q) clK a nonnegative optional process c belongs
to A(x, q) if and only if
(3.1)
E
__
T
0
c
t
Y
t
d
t
_
xy +qr for all (y, r) clL and Y Y (y, r).
(ii) For any (y, r) L the set Y (y, r) contains a strictly positive process.
For any (y, r) clL a nonnegative process Y belongs to Y (y, r) if and
only if
(3.2)
E
__
T
0
c
t
Y
t
d
t
_
xy +qr for all (x, q) clK and c A(x, q).
The proof of Proposition 3.1 is broken into several lemmas.
Lemma 3.2. Under the assumptions of Theorem 2.4 u is sub-conjugate of
v, i.e. for all (x, q) clK , (y, r) clL we have:
(3.3) u(x, q) v(y, r) +xy +qr.
As a result |u| < on K , |v| < on L. u < and v > everywhere
in R
N+1
.
Proof. Fix (x, q) clK , (y, r) clL. Take arbitrary c A(x, q), Y
Y (y, r). It follows from the denition of the conjugate eld V that
E
_
_
T
0
U(t, , c
t
)d
t
_
E
_
_
T
0
U(t, , c
t
)d
t
_
+xy +qr E
_
_
T
0
c
t
Y
t
d
t
_
E
_
_
T
0
V (t, , Y
t
)d
t
_
+xy +qr.
9
This implies inequality (3.3). The remaining assertions of the lemma follow
from inequality (3.3) inequality and the assumptions of Theorem 2.4.
As in [10] we dene the set P to be the set of points in the intersection
of L and the hyperplane y 1, that is,
(3.4) P
_
p R
N
: (1, p) L
_
.
Note that under the Assumption 2.1 it follows from Lemma 1 in [10] that
the set P is bounded.
Let M

be the set of equivalent local martingale measures Q, such that

the process X

(in the Assumption 2.1) is a uniformly integrable martin-

gale. According to Theorem 5.2 in Delbaen and Schachermayer [4], M

is
a nonempty, convex subset of M, which is dense in M with respect to the
variation norm.
For any p P denote M

(p) = {Q M

: E
Q
[F
T
] = p} . It follows from
Lemma 8 in [10] that (under the conditions (2.1) and Assumption 2.1) M

(p)
is non-empty for every p P and
(3.5)
_
pP
M

(p) = M

,
Lemma 3.3. Let p P. Then the cadlag density process of any Q M

(p)
belongs to Y (1, p).
Proof. Fix an arbitrary (x, q) clK , c A(x, q), and X X(x, q) such
that X
T
+ qF
T

_
T
0
c
t
d
t
0. Taking expectation under Q M

(p) and
using integration by parts formula we get:
x +qp E
Q
__
T
0
c
t
d
t
_
= E
__
T
0
dQ
t
dP
t
c
t
d
t
_
.
Lemma 3.4. Assume that the assumptions of Theorem 2.4 hold true. For
any (x, q) clK , a nonnegative optional process c belongs to A(x, q) if and
only if
(3.6) E
Q
__
T
0
c
t
d
t
_
x +qp for all p P and Q M

(p).
10
Proof. If c A(x, q) for (x, q) clK then the validity of (3.6) follows
from Lemma 3.3. Viceversa, let c be a nonnegative optional process such
that (3.6) holds. Denote
h
_
T
0
c
t
d
t
qF
T
, M max
1iN
|q
i
|.
Then h MX

T
and
(h) sup
QM

E
Q
[h] = sup
pP
sup
QM

(p)
E
Q
[h]
= sup
pP
sup
QM

(p)
_
E
Q
_
_
T
0
c
t
d
t
_
qp
_
x,
where in the second equality we used (3.5). Lemma 5 in [10] implies the
existence of an acceptable process X such that X
0
= (h) and X
T
h. It
follows that
X
T
+qF
T

_
T
0
c
t
d
t
0.
Therefore c A((h), q) A(x, q).
Proof of Proposition 3.1. We prove the item (i) rst. Take (x, q) K .
Since K is an open set there exist > 0 such that (x , q) K . Take
X X(x , q) then Z X + X(x, q) and
Z
T
+qF
T
=
_
T
0
d
t
.
Therefore the process that takes the constant value belongs to A(x, q).
Let c be a nonnegative optional process such that (3.1) holds. It follows
from Lemma 3.3 that the cadlag density process of any Q M

(p) is in
Y (1, p) for some p P. Consequently, c satises condition (3.6). It follows
from Lemma 3.4 that c A(x, q). This concludes the proof of the item (i).
To prove the assertion of the item (ii), let us observe that
aY (y, r) = Y (ay, ar) for all a > 0 and (y, r) L.
Therefore it suces to consider (y, r) = (1, p) for p P. Fix p P. By
Lemma 8 in [10] we deduce the existence of Q M

(p). By Lemma 3.3 the

11
cadlag density process
_
dQt
dPt
_
t[0,T]
is in Y (1, p). Since Q is equivalent to P
it follows that P
_
_
dQt
dPt
_
t[0,T]
is strictly positive for all t [0, T]
_
= 1.
For any (1, p) clL if Y Y (1, p), condition (3.2) follows from the
denition of the set Y (1, p). Conversely, let Y be a nonnegative process such
that (3.2) holds. Then
E
__
T
0
c
t
Y
t
d
t
_
1 for all c A(1, 0).
By Proposition 4.3 of [18] Y Y (1) (the set Y (y) is dened in (2.9)) is such
that (3.2) holds. I.e. Y Y (1, p).
Let L
0
= L
0
([0, T] , B([0, T]) F, d P) be the vector space of
optional process on the probability space (, F, F, P) . L
0
+
be the positive
orthant of L
0
. For notational clarity optional stochastic processes (c
t
)
t[0,T]
are identied with the corresponding elements c of L
0
. As in [20] we can
prove the following lemma.
Lemma 3.5. The function u is upper semi-continuous. For any (x, q)
{u > } there exists a unique maximizer to the problem (2.7). The func-
tion v is lower semi-continuous. For any (y, r) {v < } there exists a
unique minimizer to the problem (2.11).
Proof. Let (x
n
, q
n
)
n1
be a sequence in K converging to (x, q). Fix c
n

A(x
n
, q
n
) such that
E
__
T
0
U(t, , c
n
t
)d
t
_
u(x
n
, q
n
) 1/n, n 1.
Denote
x = sup
n1
|x
n
|, q = sup
n1
max
1iN
|q
n
i
|.
Then for any n 1 by denition of the set A(x
n
, q
n
) there exists X
n

X(x
n
, q
n
), such that
_
T
0
c
n
d
t
X
n
T
+q
n
F
T
( x X
n
0
) +X
n
T
+ qX

T
,
12
where in the last inequality we use Assumption 2.1. Therefore (c
n
)
n1

A( x+ qX

0
, 0). By Lemma A1.1 in [3] we deduce the existence of a sequence of
convex combinations c
n
conv (c
n
, c
n+1
, . . .) , n 1, and an optional process
c, such that ( c
n
)
n1
converges to c L
0
+
(d P) almost everywhere.
For any (y, r) clL and Y Y (y, r) using Fatous lemma we get:
E
__
T
0
c
t
Y
t
d
t
_
liminf
n
E
__
T
0
c
n
t
Y
t
d
t
_
xy +qr.
It follows from Proposition 3.1 that c A(x, q).
Thus, using uniform integrability of (U
+
(t, , c
n
t
))
n1
(here U
+
denotes
the positive part of the eld U), which follows from Lemma 3.12 in [18] and
Lemma 2.6 we get:
(3.7)
u(x, q) E
_
_
T
0
U(t, , c
t
)d
t
_

limsup
n
E
_
_
T
0
U(t, , c
n
t
)d
t
_
= limsup
n
u(x
n
, q
n
).
This implies upper semi-continuity of u. Moreover, by Lemma 3.2 u < ev-
erywhere in its domain. Consequently if u(x, q) > then taking (x
n
, q
n
) =
(x, y), n 1, and using (3.7) we deduce existence of a maximizer to the
problem (2.7). Uniqueness of the maximizer follows from strict concavity of
U(t, , ) for all (t, ) [0, T] .
The proof of corresponding assertions for the function v is similar.
For each (y, r) L dene the following sets:
(3.8) A(y, r) {(x, q) K : xy +qr 1} ,
(3.9)

C(y, r)
_
(x,q)A(y,r)
A(x, q),
(3.10) C(y, r) cl

C(y, r),
where closure is taken in the topology of convergence in measure (d P) .
For the proof of Theorem 2.4 we need the following lemma.
Lemma 3.6. Let (y, r) L,

C(y, r) and C(y, r) be given by (3.9) and (3.10)
respectively. Then
sup
g

C(y,r)
E
__
T
0
U(t, , xc
t
)d
t
_
= sup
gC(y,r)
E
__
T
0
U(t, , xc
t
)d
t
_
, x > 0.
13
Proof. For each x > 0 let us dene
(x) sup
g

C(y,r)
E
__
T
0
U(t, , xc
t
)d
t
_
, (x) sup
gC(y,r)
E
__
T
0
U(t, , xc
t
)d
t
_
.
Due to concavity of U(t, , ) both and are concave, and . If
(x) = for some x > 0 then, due to concavity, is innite for all x > 0. In
this case the assertion of the theorem is trivial. Also, it follows from Lemma
3.2 that (x) > for all x > 0. Therefore without loss of generality for
the remainder of this proof we will assume that is nite.
Fix x > 0 and g C(y, r). Let (g
n
)
n1
be a sequence in

C(y, r) that
converges to g (d P) almost everywhere. It follows from Lemma 3.2 that
for any > 0 there exists c

C(y, r) such that
E
__
T
0
U(t, , c
t
)d
t
_
> .
Therefore we have:
E
_
_
T
0
U(t, , xg
t
)d
t
_
E
_
_
T
0
U(t, , xg
t
+c
t
)d
t
_
liminf
n
E
_
_
T
0
U(t, , xg
n
t
+c
t
)d
t
_
(x +),
where the rst inequality is valid because U(t, , ) is increasing for all (t, )
[0, T], the second one follows from Fatous lemma, and the third one comes
from the fact that

C(y, r) is convex. Since is concave, it is continuous. As
a result
(x) = sup
gC(y,r)
E
__
T
0
U(t, , xg
t
)d
t
_
lim
0
(x +) = (x).
Proof of the Theorem 2.4. Concavity of the function u follows from
strict concavity of the function U(t, , ) for each (t, ) [0, T] . Fix
(y, r) L. Applying Lemma 3.6 and the denition of the set

C(y, r), we get
for each z > 0:
(3.11)
u(z) sup
cC(y,r)
E
_
_
T
0
U(t, , zc
t
)d
t
_
=
= sup
c

C(y,r)
E
_
_
T
0
U(t, , zc
t
)d
t
_
= sup
(x,q)zA(y,r)
u(x, q) > .
14
It follows from Proposition 3.1 that
Y Y (y, r) E
__
T
0
c
t
Y
t
d
t
_
1 for all c C(y, r).
We obtain that the sets C(y, r), Y (y, r) satisfy the assumption of Theorem
3.2 in Mostovyi [18]. Since v(y, r) < for all (y, r) L, using (3.11) we
get:
v(y, r) = sup
z>0
( u(z) z) = sup
(x,q)K
(u(x, q) xy qr) .
It follows from Lemma 3.5 that u and v are proper closed convex functions.
Therefore the latter equality implies the biconjugacy relations (2.13) (see
Rockafellar [19], Section 12).
The assertions of item (ii) in Theorem 2.4 follow from Lemma 3.5. To
prove the conclusions of item (iii), x an arbitrary (x, q) K and dene
B(x, q) {(y, r) L : xy +qr 1} .

D(y, r)

(y,r)B(x,q)
Y (y, r),
and denote by D(y, r) the closure of

D(y, r) in measure (d P) . Then for
all z > 0 we have
v(z) inf
Y D(y,r)
E
_
_
T
0
V (t, , zY
t
)d
t
_

inf
Y

D(y,r)
E
_
_
T
0
V (t, , zY
t
)d
t
_
= inf
(y,r)zB(x,q)
v(y, r) < .
By Proposition 3.1 we can deduce that
c A(x, q) E
__
T
0
c
t
Y
t
d
t
_
1 for all Y D(x, q).
Therefore the sets A(x, q) and D(x, q) satisfy the assumptions of Theorem
3.2 in [18]. Let us denote

Y
t
U

(t, , c
t
(x, q)) , t [0, T]; z E
__
T
0
c
t
(x, q)

Y
t
d
t
_
.
We can get from the same theorem that

Y zD(x, q) and that

Y is the
unique minimizer to the optimization problem:
E
__
T
0
V (t, ,

Y
t
)d
t
_
= inf
Y zD(x,q)
E
__
T
0
V (t, , Y
t
)d
t
_
.
15
It Follows from Fatous lemma and Proposition 3.1 that there exists (y, r)
clB(x, q), such that

Y Y (y, r). Here closure of B(x, q) is taken in R
N+1
.
For this (y, r) by Theorem 3.2 in [18] we have:
E
__
T
0
V (t, ,

Y
t
)d
t
_
= v(y, r) = u(x, q) xy qr.
Using conjugacy of the functions u and v together with Theorem 23.5 in [19]
we conclude that (y, r) u(x, q). This proves that u(x, q) = .
Fix (y, r) clL such that (2.14), (2.15), and (2.16) hold for some (x, q)
clK . Then using conjugacy of U and V we obtain:
0 = E
_
_
T
0
_
V (t, ,

Y
t
(y, r)) U(t, , c
t
(x, q)) + c
t
(x, q)

Y
t
(y, r)
_
d
t
_
= v(y, r) u(x, q) +xy +qr.
Since u and v and conjugate, it follows from Theorem 23.5 in [19] that (y, r)
u(x, q).
Conversely, let (y, r) u(x, q) for some (x, q) {u > }. By Lemma
3.5 u and v are closed functions. We also proved that they satisfy biconjugacy
relations (2.13). Consequently,
(3.12) u(x, q) +v(y, r) +xy +qr 0.
Combining this with Lemma 3.2 we deduce that |v(y, r)| < . By Lemma
3.5 this implies that there exists

Y (y, r), a unique minimizer to the problem
(2.11). As u(x, y) > by Lemma 3.5 we deduce that there exists c(x, q),
a unique maximizer to the problem (2.7). Using Proposition 3.1 we obtain
from (3.12)
E
_

_
T
0
_
V
_
t, ,

Y
t
(y, r)
_
+ c
t
(x, q)

Y
t
(y, r) U (t, , c
t
(x, q))
_
d
t

_
= E
_
_
T
0
_
V
_
t, ,

Y
t
(y, r)
_
+ c
t
(x, q)

Y
t
(y, r) U (t, , c
t
(x, q))
_
d
t
_
v(y, r) +xy +qr u(x, q) 0,
which gives (2.14) and (2.15).
Proof of Lemma 2.6. Assume that (2.19) holds. Fix (x, q) K . It
follows from Assumption 2.1 and by Lemma 1 in [10] that (x, 0) K for each
x > 0. Since K is an open cone, there exists a point (x
1
, q
1
) K , such that
(x, q) = (x
1
, q
1
) + (1 )(x
2
, 0)
16
for some (0, 1) and x
2
> 0. By Proposition 3.1 there exists a process
A(x
1
, q
1
) that is strictly positive and is equal to a constant, which we
denote by . Take c A(x
2
, 0), such that
(3.13) E
__
T
0
U(t, , (1 )c
t
)d
t
_
> .
Note that such a process c exists by assumption (2.19). Now,
+ (1 )c A(x, q).
Since U(t, , ) is increasing for all (t, ) [0, T] we obtain from (3.13):
u(x, q) E
_
_
T
0
U(t, , + (1 )c
t
)d
t
_

E
_
_
T
0
U(t, , (1 )c
t
)d
t
_
> .
To show that v(y, r) < for any (y, r) L dene the set
E = {(y, r) clL : v(y, r) < } .
By Theorem 3.2 in [18] for any x > 0 there exists c(x), the maximizer to
problem (2.17). Dene the process Y as: Y
t
= U

(t, , c
t
(x)), t [0, T]. Then
by the same theorem E
_
_
T
0
V (t, , Y
t
)d
t
_
< . Using Proposition 3.1 and
Theorem 3.2 in [18] we can deduce that there exists (y, r) clL, such that
Y Y (y, r). Thus E = . As x can be taken arbitrary large, we can show
that the closure of E in R
N+1
contains origin. Moreover, E is convex. Now,
for any (y, r) L take (y
1
, r
1
) L and (y
2
, r
2
) E , such that
(y, r) = (y
1
, r
1
) + (1 )(y
2
, r
2
)
for some (0, 1). By Proposition 3.1 there exists a strictly positive process
Y

Y (y
1
, r
1
). By construction of the set E there exists a process Y

Y (y
2
, r
2
), such that
E
__
T
0
V (t, , (1 )Y

t
)d
t
_
< .
17
Since V (t, , ) is decreasing for all (t, ) [0, T] we deduce
v(y, r) E
_
_
T
0
V (t, , Y

t
+ (1 )Y

t
)d
t
_
E
_
_
T
0
V (t, , (1 )Y

t
)d
t
_
< .
Conversely, if (2.12) holds then by Theorem 3.2 in [18] we have:
u(x, 0) = w(x) = inf
y>0
( w(y) +xy) .
Dene w(y) = inf
pP
v(y, yp), y > 0. Then from Theorem 2.4 we obtain:
u(x, 0) = inf
(y,r)L
(v(y, r) +xy) = inf
y>0
( w(y) +xy) .
Convexity of w, w, and w implies that w w. Therefore for any y > 0 we
get from (2.12):
w(y) = w(y) < .
The other assertion of (2.19) follows trivially.
References
[1] D. Cuoco. Optimal consumption and equilibrium prices with portfolio
constraints and stochastic income. J. Econom. Theory, 72:3373, 1997.
[2] J. Cvitanic, W. Schachermayer, and H. Wang. Utility maximization in
incomplete markets with random endowment. Finance Stoch., 5:259
272, 2001.
[3] F. Delbaen and W. Schachermayer. A general version of the fundamental
theorem of asset pricing. Math. Ann., 300:463520, 1994.
[4] F. Delbaen and W. Schachermayer. The banach space of workable
contingent claims in arbitrage theory. Ann. Inst. H. Pincare Statist.
Probab., 33:113144, 1997.
[5] F. Delbaen and W. Schachermayer. The fundamental theorem of asset
pricing for unbounded stochastic processes. Math. Ann., 312:215250,
1998.
18
[6] D. Due and T. Zariphopoulou. Optimal investment with undiversi-
able income risk. Math. Finance, 3:135148, 1993.
[7] D. Due, W. Fleming, M. Soner, and T. Zariphopoulou. Heding in
incomplete markets with hara utility. J. Econ. Dymaics and Control,
21:753782, 1997.
[8] H. He and N. D. Pearson. Consumption and portfolio policies with
incomplete markets and short-sale constraints: the nite-dimensional
case. Math. Finance, 1:110, 1991.
[9] H. He and N. D. Pearson. Consumption and portfolio policies with
incomplete markets and short-sale constraints: the innite-dimensional
case. J. Econom. Theory, 54:259304, 1991.
[10] J. Hugonnier and D. Kramkov. Optimal investment with random en-
dowment in incomplete markets. Ann. Appl. Probab., 14:845864, 2004.
[11] I. Karatzas and K. Kardaras. The numeraire portfolio in semimartingale
nancial models. Finance Stoch., 11:447493, 2007.
[12] I. Karatzas and S. Shreve. Methods of Mathematical Finance. Springer,
1998.
[13] I. Karatzas and G.

Zitkovic. Optimal consumption from investment
and random endowment in incomplete semi-martingale markets. Ann.
Probab., 31:18211858, 2003.
[14] I. Karatzas, J. P. Lehoczky, S. E. Shreve, and G. L. Xu. Martingale
and duality methods for utility maximization in an incomplete market.
SIAM J. Control Optim., 29:702730, 1991.
[15] K. Kardaras and G.

Zitkovic. Stability of the utility maximization prob-
lem with random endowment in incomplete markets. Math. Finance, 21:
313333, 2011.
[16] D. Kramkov and W. Schachermayer. The asymptotic elasticity of utility
functions and optimal investment in incomplete markets. Ann. Appl.
Probab., 9:904950, 1999.
19
[17] D. Kramkov and W. Schachermayer. Necessary and sucient conditions
in the problem of optimal investment in incomplete markets. Ann. Appl.
Probab., 13:15041516, 2003.
[18] O. Mostovyi. Necessary and sucient conditions in the problem of op-
timal investment with intermediate consumption. arXiv:1107.5852v1
[q-n.PM], 2011.
[19] R. T. Rockafellar. Convex Analysis. Princeton Univ. Press., 1970.
[20] P. Siorpaes. The relation between arbitrage free prices and marginal
utility based prices. Preprint, 2011.
[21] G.

Zitkovic. Utility maximization with a stochastic clock and an un-
bounded random endowment. Ann. Appl. Probab., 15:748777, 2005.
20