Optimal Impulse Control of Portfolios

Author(s): Jerome F. Eastham and Kevin J. Hastings

Source: Mathematics of Operations Research, Vol. 13, No. 4 (Nov., 1988), pp. 588-605
Published by: INFORMS
Stable URL: http://www.jstor.org/stable/3689945 .
Accessed: 07/10/2011 07:38

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact [email protected].

INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Mathematics of Operations
Research.

http://www.jstor.org
MATHEMATICS OF OPERATIONS RESEARCH
Vol. 13, No. 4, November 1988
Printed in U.S.A.

OPTIMAL IMPULSE CONTROLOF PORTFOLIOS*

JEROMEF. EASTHAMtANDKEVINJ. HASTINGSt
An investorhas the opportunityof holdingsharesin n riskyassetsandone nonriskyassetat
every time in a fixedinterval[t, T]. The riskyassetsare governedby a stochasticdifferential
equation.At randominstantsof his choicehe mayintervenein orderto rebalancehis portfolio
and consume a nonnegativeamountof money. Fixed and variabletransactionscosts are
incurredupon intervention.At time T all remainingwealthis consumed.The solutionto the
problem of maximizingtotal utility of consumptionis given by way of quasi-variational
inequalitiesfor the value function.With probabilityone the investoronly intervenesfinitely
many times. Indicationof the solutionof the quasi-variationalinequalitiesin the case of one
risky asset with log-normalprices is given, togetherwith a descriptionof a discretization
procedure.

0. Introduction. We considera portfolio-consumption problemwith the following

features.There are n riskyassets whose pricesfluctuatein continuoustime according
to Ito processeswithjumps,and thereis a nonriskyasset whosevalueincreasesstrictly
with time. An investorbeginswith a portfolioof sharesin these assets,and is allowed,
at some randominstantsin time, to rebalancehis portfolio,and at the same time to
consumewealth. In orderto intervene,a positivefixed cost plus nonnegativevariable
costs must be paid. There is a finite time horizon T, at which all wealth is to be
consumed. The problemis to choose a sequenceof interventiontimes(0o ) and a
sequence of re-initializedportfolios (S ) to maximizethe investor'sexpected total
utility of all consumptions.We thereforehave a problemof impulse control over a
finite time interval.
From Merton'simportant1971paper[13] to a recentpaperby Lehoczky,Sethi,and
Shreve [10], many authorshave examinedvariousportfolio-consumptionmodels. In
the overwhelmingmajorityof these, the methods of continuous-timedynamic pro-
gramming are employed to obtain a continuous consumption rate and portfolio
allocation function that maximizetotal expectedutility of consumption.Intervention
occurs continuously;hence transactioncosts cannot be treated.The requirementof
continuous interventionwithout cost raises naturalobjections.The work that comes
closest to ours in this senseis that of Magilland Constantinides[12].In this paperthe
risky assets are of the log-normaltype, and the riskless asset earns interest at a
constant rate. The variablecosts are proportionalto the value of the transaction,and
the fixed cost is not present. The control variables are the portfolio vector and
consumption,and the authorsare mostly interestedin the decreasingabsolute risk
aversion utility. The method of solution is to employ dynamic programmingon a
system with perturbeddiffusioncoefficients.It is concludedthat the optimalportfolio
is piecewiseconstantand some computationalformulasare obtained.

*ReceivedMarch28, 1985;revisedFebruary3, 1986.

AMS1980subjectclassification. 60K99.
93E20;Secondary:
Primary:
Main:Stochastic
IAOR1973subjectclassification. processes; Financial.
CrossReferences:
OR/MS Index 1978 subjectclassification.Primary:572 Probability:Stochasticmodels; Secondary:200
Finance:Portfolio.
Key words.Portfoliomanagement,consumptionutility,stochasticcontrol,optimalimpulsecontrol.
Universityof Delaware.
*Knox College.
588
0364-765X/88/1304/0588$01.25
Copyright ? 1988, The Institute of Management Sciences/Operations Research Society of America
OPTIMAL IMPULSE CONTROL OF PORTFOLIOS 589

We have modeled the problem as an impulse control problem, thus only piece-
wise constant portfolios are permissible.This allows us to draw from the work of
Bensoussanand Lions [3] and others (e.g. [4], [9, [[11] and [15]) to characterizethe
optimal value functionas the solutionto the versionof the quasi-variational inequali-
ties that is appropriateto this problem.The optimaltimes of interventionare then the
times of entrance to the "action set", determinedby the value function, and the
optimal portfolios are maximizingargumentsof a function dependenton the utility
and the optimal value. In fact, Bensoussanand Lions [2] suggestan impulseportfolio
control model without explicitlylisting constraintsor giving solutions. Their model
differsfrom ours in some ways,chieflythe following.The idea that we are maximizing
an expected total utility of consumptionwith an interventioncost raises the question
of divergenceof the interventiontimes, which is usually easily dismissedin impulse
control problems in which one simply minimizesa total runningcost plus impulse
costs. Basicallythe issue can be resolvedby forbiddingthe explosionof total wealth.
In ?1 we formulate the problem rigorously.?2 introduces the quasi-variational
inequalitiesand the associatedimpulse control. Some attentionis paid to regularity
conditions of the control.In particularit is shown that underhypothesesof finiteness
and quasi-left-continuity(see [5]) of the price process,the interventiontimes cannot
accumulatebefore the investmentperiod ends. The proof of the optimalityof the
Q.V.I. control makes up ?3; it mostly follows the well-knownmethods.Thus a very
generalimpulse portfolioproblemis reducedto solvingthe quasi-variational inequali-
ties, which can, of course,be difficult.In ?4 we solve the problemin the specialcase of
linear coefficients,no jump structure,one risky asset, proportionatevariabletransac-
tions costs, and risk-neutralutility. Numericaltechniquesfor more generalsituations
are the object of futurestudy.

1. Problem formulation. Let P = (P1, P2,..., P,) be a vector-valuedstochastic

process describingthe pricesof n risky assets. Supposethat P satisfiesthe stochastic
differentialequationwithjumps:

(1.1) dP(r) = p(P(r), r) dr + o(P(r), r) dZ(r) + f y(P(r), r,w)u(dr, dw)

with initial condition P(t) = p, t e [0, T], and with 0 as an absorbingboundaryfor

each P,. Here, a is n x d-dimensional,Z is a d-dimensionalstandardWienerprocess,
and

v((O, s], A) = v((O, s], A) - s q(A),

where v is a Poissonrandommeasurewith meanmeasuredt q(dw) (see Gikhmanand

Skorokod [7]). A unique solution to (1.1) exists if the coefficients,L, a, and y are
linearlyboundedand uniformlyLipschitz([7, Theorem3.3]).We adjointo the systema
risk-freeasset that satisfies

(1.2) dPO(r) = ,O(Po(r), r)dr,

where the initial condition is Po(t) = Po > 0. Also, I0 > 0, so that POis positive and
increasing.The vector price process P is now (n + 1) - dimensional,and its initial
component Po has a0 = yo = 0.
590 JEROMEF. EASTHAM& KEVIN J. HASTINGS

During It, T) an investor holds a portfolio of shares in each asset represented by a

vector s whose ith component is the number of shares of asset i. He is permitted to
intervene at various instants, in order to rebalance his portfolio and consume wealth.
For i = 1, 2,..., n, let Bi be the available number of shares in risky asset i, and denote

B= [0, o) x I[0, B].

i-1

This indicates that there is no inherent bound on the available nonrisky shares. Define
the feasible set K(p, s) corresponding to price vector p and portfolio vector s to be

(1.3) K(p,s) = (s' E B: C(p,s,s') p (s - s') - c - I(p,s,s') > 0}.

In (1.3), C(p,s,s') is the amount the investor consumes if the price vector is p, the
portfolio is changed from s to s', and a fixed cost c plus a nonnegative variable cost '
is paid for transacting. Thus, (1.3) implies that money cannot be borrowed. There are
states for which the feasible set is empty. Denote K = {(p,s): K(p,s) # 0 }. We
assume from now on that ' (and hence C) is continuous, and that

(1.4) K(p, s) $ 0 * 0 is feasible for (p, s).

Under (1.4), K is closed, for if (Pk,Sk) is a sequence in K converging to (p, s), then
the continuity of C shows that 0 must be feasible for (p,s), hence (p,s) E K. We
remark that (1.4) holds under some very sensible conditions. For instance, if

t(P, S, S) = E i(Pi, Si, Si),

i=1

where, for each i and each fixed Pi, si, the function p,(si - s') - *i(pi, si, s[) has its
maximum at s[ = 0, then C(p,s,0) > C(p,s,s') for all feasible s', and (1.4) results.
Once such I is
n

(P,,s,s') = Eci Pi(Si - S,

i=1

where 0 < ci < 1 for each i.

Let G c R"+1 x R"'1 be an open set. The admissibility conditions below give the
interpretation that the investor is required to intervene when the joint price-share
process enters GC. An impulse control v = ((k, Sk)} is a sequence of pairs in which
(0k) is a nondecreasing sequence of random times, 01 > t, and (Sk) is a sequence of
R"+1-valued random variables. The controlled share process S with initial portfolio s is

if t < r <
0< ;
(1.5)s
(1.5) if Oi< r < 0k where k = inf{ m: 0, > '
S(r) = \Sk_1 }

The extra complication in the second half of the definition is forced on us because we
are allowing 0k = k+ 1. The optimal control that we will construct may demand
immediate rebalancing for certain utility functions.
OPTIMAL IMPULSE CONTROL OF PORTFOLIOS 591

Let X be the joint process(P, S). The impulsecontrol v is called admissibleif

(1.6-a) Ok < 0k+1 k,

(1.6-b) Ok is a stoppingtime of X, Vk,

(1.6-c) P lim Ok < T = 0,

k --oo

(1.6-d) Ok+l < inf{r > k :X(r) E Gc},

(1.6-e) Sk E K(P(k), Sk-1),

(1.6-f) whereHr = o (X(s); s < r).

Sk is HOk-measurable,

Let U = U(z, r) be a continuous,real-valuedutility function on R X R+. If v =

{((0,S,)} is an admissibleimpulse control, then denote C, = C(P(0i),S,_i,Si) and
CT= C(P(T), S(T), 0). Define the value of v by
00

(1.7) J(p,s, t, v) - Ep t E (C,, Oi)I[t,T)(0i) + U(CT, T)

i-=1

Note that the only admissibleaction at the terminaltime T is to sell off all holdings,
despite the fact that CT may be negative.For this reason we assume that U(z, t) is
positive or negativeaccordingto whetherz is positive or negative.The problemis to
find an admissible v* such that J(p,s, t, v*) > J(p,s, t, v) for each p ER-++', s E B,
t e [0, T] and each admissiblev.

2. The Q.V.I. control. Let u = u(p,s, r) be a continuous,real-valuedfunctionon

Rn++'x R++' x R+ with continuoustime partial,and continuous1st and 2nd order
partialsin p. Define the operator

(2.1) Lu(p,s, r) = u,(p,s, r) + Vu(p,s, *

r) -p(p, r)

n n
+ E u,j(p,s, r)A j(p, r)
i=1 j=l

+ [u(p + y(p, r,w),s, r) - u(p,s, r)

-Vu(p,s, r) * y(p, r,w)]q(dw),

whereV indicatesthe (n + 1)-dimensionalgradientin p, u,i is the 2nd orderpartialin

pi and pi, and A = aa'. Also define

(2.2) Mu(p,s,r)= sup {U(C(p,s,s'),r) + u(p,s',r)}.

s' K(p, s)

We say that a function u with the above smoothnesspropertiesis a solution of the

592 JEROMEF. EASTHAM& KEVIN J. HASTINGS

quasi-variational inequalities if:

(2.3-a) u > Mu and u = Mu on G' X [0, T),

(2.3-b) Lu < 0,

(2.3-c) (u - Mu) Lu = 0,

(2.3-d) u(p, s, T) = U(C(p,s, ), T).

Define the continuation set C = (u > Mu) and the action set A = CC.Note that by
(2.3-a), GC x [0, T) c A. Now it is clear that for (p,s) E K, the feasible set K(p,s) is
nonempty and compact. We will show in Lemma (2.7) below that K is upper-semicon-
tinuous (u.s.c.) as a multifunction. Since the function to be maximized in (2.2) is
continuous, and B is locally compact, a measurable selection theorem of Schail [14,
Corollary 4] implies that there is a Borel measurable function f: K x [0, T) -* B such
that

(2.4) <((p,s,r) K(p,s) and

= Mu(p,s, r)
U(C(p,s, (p,s, r)), r) + u(p, 0 (p,s, r), r)

for all (p, s, r) E K X [0, T). Note that if K(p, s) is empty, then Mu(p, s, r)= - c
and (p, s, r) E C.
This permits us to define the Q.V.I. impulse control. Again, let the initial data be
(p, s, t) and write 80* = t and So* = s for convenience. Define
(2.5-a) 0k = inf{ r > 0* : (P(r)S, r) E A}),

(2.5-b) Sk = 0(P(' )Sk-1, k*).

In (2.5-a) we adopt the convention that, if the process never enters A before T, then
0* = oo. In Lemma (2.8) we will show that Mu is u.s.c., which, together with the
closure of K, implies that A is closed. The right continuity of P yields that
(P(k ), Sk*1,I ) e A and hence (P(Ok*),Sk_1) K, so that S is well defined. It
may happen that (P(k ), Sk , Ok) is again in A, in which case k+ 1 = 09*.Despite this,
the times ,"*are stopping times of (X,), since q4is a measurable function. In fact, one
can do an easy induction argument to prove that for all k,

(2.6) { * > r} E (P(s); s < r) Fr and S, is Feo-measurable.

The remainder of this section checks that v* = {(*, ,S)} is admissible. The
construction of v* and the remarks above imply that (1.6-a, b, d, e, f) hold. It remains
to prove the two semicontinuity lemmas on which the remarks were based, and the
divergence of the intervention times (1.6-c).

(2.7) LEMMA. The multifunctionK: K - B is u.s.c.

PROOF. It suffices to show that

lim sup K(pk, sk) c K(p, s),

k- oo

where (pk, Sk) (p,s) in K. (Note that when the latter is true, U lK(pk, sk) is
bounded.) Given any subsequence (ki) and s' e K(pk,,Sk,) converging to some s' as
OPTIMAL IMPULSE CONTROL OF PORTFOLIOS 593

i -o oo, it suffices to show that s' E K(p, s). Since B is closed, it is clear that s' E B.
Moreover, since C(pk,,ski, s) > 0 for all i, C is continuous, and (Pk, Ski,s ) - (p, s, s'),
E
we have that C(p, s, s') > 0. Hence s' K(p, s), which completes the proof.

(2.8) LEMMA. Mu is u.s.c. on K X [0, T).

PROOF. Let (pk,Sk, rk) be a sequence in K x [0, T) converging to (p,s, r) in

K x [0, T). It is not hard to see that since K is u.s.c., for each 8 > 0 there exists
N = N(8) such that for all k > N,

(2.9) K(pk,s) C
N(limsupK(pj, sj); 8) c N(K(p,s); 8)
j-- oo

where N(B; 8) denotes the closed 8-neighborhood of a set B. Then any s$ E K(pk, sk)
has a counterpart s' E K(p, s) within a distance of 8, as long as k > N. Denote

(2.10) f(p,s,s', r) = U(C(p,s,s'), r) + u(p,s', r).

It is a straightforward matter to use (2.9) to show that given E> 0,

(2.11) f(pk,sk,sk, rk) < sup f(p,s,s, r) +

s' K(p, s)

for large enough k and all s$ E K(pk, sk). Let E - 0, and conclude that
limsupk,o Mu(pk,sk, rk) < Mu(p, s, r), from which the lemma follows.

(2.12) LEMMA. If P is almost surelyfinite and quasi-left-continuous,then

P lim < =0.

k-I oo 9k T]

(REMARK.Note that the proof does not depend on the specific nature of P, only
these two properties. Our P has the jump times of a Poisson process, and hence is q.l.c.,
and it is also a.s. finite in the case of linearly bounded coefficients.)

PROOF. Let us adopt the notation

D = lim
0m < T},
k- -oo

S* = ith component of Sk,

M= max{Bi; i= 1,..., n},

n
'
Wk(t) = , Pi(t) Sk,i, t> Ok*,
i=O

and also drop the superscript "*" from the impulse control notation. The finiteness of
P and that of Sk,0 (shown below) guarantee that except for outcomes w in a set of
measure zero, Wk < oo. We assert first that

C
(2.13) D c lim sup Wk(Ok+l) - Wk(k) > almost surely.
k-o oo
594 JEROMEF. EASTHAM& KEVIN J. HASTINGS

By contradiction,suppose that w E D and that there is I = I(o) such that for all
k > I, Wk(Gk+l) - Wk(Ok) < c/2. Now in order for an intervention to occur at a time
0i, admissibility requires that Wi(0i) - Wi_l(0i) < -c. If j > I, then

Wj_,() = [W-(wJ) - W-l(0-l)] + [w-J(-1) - Wj-2(j-1)]

+ -- +[w,(e,+1)- w,(,)] + W,(01)

< -(j I- -1) + W,(O,).

Except for w in a set of measurezero, Wj_l(0j(w))can be made negativefor j large

enough, a contradiction.Assertion(2.13) is thereforeproven.
We now wish to bound Sk, from above (recall that there is no upperlimit to the
available shares of the risk-free asset). Using the admissibility condition Wk_-(Ok) -
Wk(Mk) > c, and solving for Sk o gives
n
Pi(0)k)
4 ^ - c
-
Sk,o < Sk-1,0+
po() (Sk-l,i Sk,i) )

Repeating the procedure to bound Sk_ ,, Sk_2, etc. yields

k n k
pi (O) - c
Sk,o < ,o + E po() (S-.- 1 s,,) (,)
_=1 P,= (0)
j=1 j=1 P0(o)

Interchangethe orderof summationto obtain

Pnp(4l0) Pi(Ok)
(2.14) Sk,o<So + Si S
P(1)s
P(01)
O pi(
PO(Ok) k

k
k-'pi(0+) i,(Oi) c
+ E
Po(Oj+x)- Po() S.i - Ec
j=l , .JP0(0) j=- Po(i)

M 'k-1
J np( -
kc
< so + si + E IP,(j +) - P,(0j)[-
nPo(T)

by the monotonicityof PO.

Thus, by assertion(2.13) and (2.14),

D c limsup Wk(Ok+l) Wk(Ok) > } n D

k c/

S
c limsup [Po(0k+) -
)] + E PO Si
k ^ . i=l Po

Mn 'k-1 kc
+ [PO(k+l) - - -
PO(O
k) Po i=l r,j=l Pi(j Pi(0 kl nPO(T)

n
+ E [Pi(0k+l) - P,(Ok)] 'Sk,i > n D
i=1 +?[^^i)- (M -^ ic}
OPTIMAL IMPULSE CONTROL OF PORTFOLIOS 595

In the set above, we have the sum of 2n + 1 termsexceedingc/2 for infinitelymanyk.

By the pigeonholeprinciple,at least one of these termsmust exceed c/2(2n + 1) for
infinitelymany k. That is,

(2.15) D c limsup [PO(Ok+) -P(Ok)] So+ ? Pi()

k PO. 1 Po

cn
2(2n+ 1) } D

U U limsup ([PO(k+1)
[P PO(k)]
i=l1M k-1

i(+) -p
PPo
j=1

kc c
> 2(2n + 1) n D
nPO(T)
n
U Ulimsup1 [P(k+) - + .
P,()]S, >i 2(2n 1) n

The first lim sup is containedin

lim SUP IPO (ok +l) -- P0(6k) > cp?

.
limsupk Po(k+l) Po(Ok) >
2(2n + l)(sop0 + 2P,(,1)s,)

sopo + -Pi(O1)si > O} n D

which in turn is clearlycontainedin the set where PObecomesinfinite at some time

t < T, hence it has probability0. Also, the boundednessof Sk i shows that for each
i= 1,...,n,

limsup [Pi(0k+l) - Pi(k)k, > n

2(2n + 1)

c limsup Pi(Ok+i) - n D.
Pi(Ok) > 2(2n + 1) M

This set is clearlycontainedin the set of outcomessuch that eitherquasi-left-continuity

or finiteness of P is violated, and thus it also has no probabilityweight. Using
monotonicityof POagain,we see that each set in the middle of the union in (2.15) is
containedin
k-1
limsup( k IP,(O+1)-Pi(- )
k j=-1

Poc Poc nD
> MnP (T) +kM 2(2n + 1)(Po(T) - po)

C -
{(Pi(Oj+l) Pi(j)l 0 as j -, o) n D.
596 JEROMEF. EASTHAM& KEVIN J. HASTINGS

Again this last set is containedin a set of measurezero, by hypothesis.The proof is

complete.
It may be worthwhileto summarizethe main assumptions.If (1) there are finitely
many availablesharesof each riskyasset; (2) the transactioncosts are as in (1.3); (3)
model condition(1.4) holds; and (4) the priceprocessis q.l.c. and almostsurelyfinite,
then an admissibleimpulsecontrolcan be constructedfrom a smooth solution u (if
one exists) to the quasi-variationalinequalities.This control demandsa transaction
when (P, S, r) enters A = {u = Mu). The transactionmay not immediatelyreplace
the processinto C = { u > Mu but, almostsurely,afterfinitelymanytransactionsthe
processwill be replacedinto C. It is interestingto ask what extrahypothesisabout the
model ensures that replacementis made into C, i.e. 0+l 1 > 90. Such a hypothesis
might be called a "lazy investor"hypothesis.
Suppose that

(2.16-a) Vp,s,s',s", 4(p,s,s") <4I(p,s,s') + 4(p,s',s");

U = U(z, r) is monotoneincreasingin z, and givenCo > 0 thereis 8 > 0 such that for
all z1, z2 > 0 with zx + z2 < C - c,

(2.16-b) U(z1 + z2 + c, r) > U(z1, r) + U(z2, r) + 8.

Fix p, s, and r; since the time r does not changethroughoutthe followingwe suppress
it in the notation. Denote 4) = 4(p,s). The strict inequality0k*+ > 90 for all k will
follow if we can show that thereis 8 = 8(p,s, r) > 0 such that for all s' E K(p, 4),

(2.17) u(p, 4) > U(C(p, .,s')) + u(p,s') + 6.

It is easy to check that (2.16-a)implies

(2.18) C(p,s,s') > C(p,s, +)+ C(p, ,s') + c,

which also shows that s' is feasiblefor s. Since 4) is the maximizerfor (p,s) we have

(2.19) u(p, 4) + U(C(p,s, 4)) > u(p,s') + U(C(p,s,s')).

In light of (2.19), to show (2.17) it sufficesto show that, for all s' e K(p, 4),

(2.20) U(C(p,s,s')) > U(C(p,s, 4)) + U(C(p, , s')) + 6.

But (2.20) follows from (2.18) and (2.16-b), so that we have proved that under
assumption(2.16), 0k*+l > 06, for everyk.
3. Optimalityof the Q.V.I.control. As in the previoussection,let u be a smooth
solutionof the quasi-variational inequalities.We have seen that undersome reasonable
model conditions, u gives rise to an admissibleimpulsecontrol v*. It is clearfrom the
descriptionof P that undersome otherreasonableconditionson the growthof u, and
on the coefficientsju,a, and y, we can use Ito's Lemmato conclude:

(3.1) E[u(P(eI),Sk-1, k) - u(P(O-1), Sk,_ _-)]

uk'

E f Lu(P(r),Sk1_, r) dr,
k- 1
OPTIMAL IMPULSE CONTROL OF PORTFOLIOS 597

where v = {( ok Sk)} is any admissible impulse control and 8X = 0k A T. We will delay

until after the proof of the following theorem our remarks on sufficient conditions for
equation (3.1).

(3.2) THEOREM. Suppose that u is a smooth solution of the quasi-variationalinequali-

ties (2.3) that gives rise to an admissiblecontrolv = ((, S' )}. Also suppose that (3.1)
holds. Then for all (p, s, t) e G x [0, T) and all admissible controls v = ((ok Sk)},
U(P,s, t) = J(p, s, t, v*) > J(p, s, t, v), that is, v* is an optimal impulse control.

PROOF. The notation will be complicated by the fact that successive intervention
times could be equal. Given w, let m0 = 0, ml = inf{k: Ok(W) > t},..., m,
inf{ k: Ok(O) > 0,,,l(w)). Again denote O"= Ok A T. It can easily be checkedthat

(3.3) u((,)SOu,,~ - (P(O",~),(, ,)e,0

=uP(POI( O,,M i,OMI ) - )(P(o )Sm.i,O,,Mi)

?I . ui)S)-
u( -
o )

M,
+1 --1f
(P(o,), Smi , - u(P(i-1) , Smi i' 8i)

=
U(P(om,
+Ii< ),Smikmi-ti-
T) Om, F) - u(P(Om,i %Smi
[U(P(k),SkIOm) ) - U(P(O)mi,Sk-Om)I
1,O, 1)

Mi1-
+
I(Gi<T} E - U(PCk k, Ok) UP(0Sk1OJ

k=mi

by (2.3-a). The rest of the proof is now standard, and will therefore be done rapidly.
Sum inequality(3.3) from i = 1 to j, take expectation,use Ito's Lemmaand (2.3-b)on
the first difference, and we have

-
Ep, S, I uP(o), S(Oj),
O u(p, S,t)I

M+1-1
d Ep, s,IEU(Ck,Ok)IJOk<T)
k=1

Rearrange the inequality and send j to oo to obtain:

00

u(p,s, t) > Ep,1

S U(Ck, Ok)ifk<T) + U(CT, T)

= J(p,s, t, V)

by the boundary condition (2.3-d). For the Q.V.I. control v", the inequalities are
replaced by equalities, which proves the theorem.
598 JEROMEF. EASTHAM& KEVIN J. HASTINGS

Equation (3.1) follows from Ito's lemma if several expectations are finite:
0

(El) E | lur(P(r),Sk_, r) + Vu- (P(r),Sk_, r)

Ok'-1

+Aijuij(P(r),Sk-_, r)| dr < oo;

Ok'

(E2) E / IVu a(P(r),Sk_1, r)12 dr < C;

O-_

k,

(E3) E /
6k-1 W
flu(P(r) + Y,Sk_, r)- u(P(r),Skl, r)

-v y(P(r),S,_l, r,w)lq(dw) dr < oo;

k'

(E4) E f lu(P(r)+ y,Sk_,r) - u(P(r),Sk,_,r)2q(dw) dr < oo.

One way to guarantee this is to put growth conditions on u and its partials, linear
growth on the coefficients p, o, and y, and finiteness up to a certain order of the
moment functions mk(r) = E[IP(r)lk]. Recalling inequality (2.4) for Sk,0, and the fact
that the available shares in risky assets are bounded by M, it can be shown that the
following conditions together are sufficient for the finiteness of the expectations
(E1)-(E4):
u and its partials ur u,, uj satisfy

(3.4) If(p,s, r)l < Ko(1 + Iplm+ Isolm),

r E [0, T], si E [0, M], i = 1,..., n;

(3.5) IP(P,r)12 + IO(p,r)12+ jly(p,r,w)12q(dw) < Kl(1 + lpl)2;

(3.6) fly(p, r, w)l'q(dw) < K2(1 + IPl) for I up to order2m + 4;

(3.7) mk(r) is finite for k up to order 2m + 4.

Of course in the absence of any jump structure on P, (E3)-(E4) do not arise, and (3.6)
is not needed. In this case the well-known bounds (e.g. Gikhman and Skorohod [7,
Lemma 3.8]) on the moments of a diffusion process satisfying (3.5) show that condition
(3.7) is automatic, and (3.4) and (3.5) alone are sufficient.

4. Model problem. We now illustrate the application of the optimal impulse

control construction for a portfolio containing two types of assets, shares of a risk-free
OPTIMAL IMPULSE CONTROL OF PORTFOLIOS 599

investmentwhich appreciatedeterministically:

(4.1) dP,(t) = loPo(t) dt,

and sharesof a riskyasset whose priceis governedby a continuousSDE:

(4.2) dP,(t) = /lPl(t) dt + aP1(t) dZ(t),

where Z is a one-dimensional standard Weiner process, and ,LO, ,U, a are positive
constants. Thus the partial differential operator associated with the optimal value
function is:

(4.3) Lu u + ioPoUp, +
Plpoup + 2pUplpl.

For simplicity we take the laissez-faire domain G to be R3+x [0, B1] and choose the
risk-neutral utility function U(x) - x. If the fixed cost is c > 0 and the variable costs
are proportional to the value of risky assets traded, then the feasible set will be

(4.4) K(p, s) (s' j(p, s') E G, p * (s - s') > clp, s, - sl + c}

where 0 < c < 1. For fixed (p, s), K is a closed, bounded subset of R x [0, B1]. Hence
the sup-operator M is well defined.
Previously we showed that the optimal expected utility of consumption is character-
ized as a sufficiently regular solution of the time-dependent implicit obstacle problem:

Lu < 0,

(4.5) u > Mu,

(Lu)(u - Mu) = 0

for all t < T, subject to the terminal condition:

(4.6) Ult-T = Po0s + (1 - Cl)P11 - C.

The usual technique for solving implicit obstacle problems such as (4.5)-(4.6)
involves an 'iteration upon the obstacle' (8). Beginning with the choice of suitable u0,
one defines ui, i = 1 2, 3,..., as the solutions of the problems:

Lui < 0,

(4.7) Ui> 1,

(Lu,)(ui - Mu,_i) = 0

for all t < T, subject to terminal condition (4.6).

Suppose that uo is chosen to be the solution of the partial differential equation:

(4.8) Luo =0

satisfying the terminal condition (4.6). This u0 can be interpreted as the expected
utility of consumption if no transactions are permitted before t = T. By induction one
can interpret ui as the optimal expected utility of consumption if at most i transac-
600 JEROMEF. EASTHAM& KEVIN J. HASTINGS

tions are permittedpriorto t = T. Accordingto this interpretation:

(4.9) < u < u2 <

uO< .* < u (pointwise).

It turnsout that, by an applicationof the earlierresulton finitenessof the numberof

transactions,the u, may be shownto convergeto u. Thus the solutionof the implicit
obstacleproblem(4.5)-(4.6) maybe obtainedby solvinga sequenceof explicitobstacle
problems(4.7) with the same terminalcondition(4.6).
Actually,for the case at hand we can show that u = ul. It is easy to verify that:

(4.10) uo = poeo(T-e) + (1 - c1)plel(T-t)s - c.

If t,1 = Po, then uo > Muo everywhere, thus u = uo. There is no benefit to transac-
tions prior to t = T, hence in that case the policy of noninterventionis optimal.
Now considerthe case , < g.o.It is helpfulto considerthe function:

(4.11) i = (poSo + (1 - c1)plsl)e#0(T-')- c.

It is easy to verify that i satisfies

Lu< 0,
u > Mui,

u(T) = poSo+ (1 - c1)ps, - c.

This function can be interpretedas the value of a policy for a slightly different
problem.If the fixed cost c is paid only at time T, then by selling all risky assets at
time t < T and investingthe proceedsin risk-freeassets, one has an expected total
consumption(4.11).
Now u > u since the firstpartof the proof of (3.2) appliesto uias well as to u. Then
ui(p,s, t) > J(p, s, t, v) for any admissiblev. In particular,

(4.12) u >~u,,

since the controlwhichgivesrise to ul is admissible.The sup-operatorM is monotone,

so combining(4.9) and (4.12):

(4.13) Mu > Mu1 > Mu0.

calculations(illustratedshortly)show that
Straightforward

(4.14) Mui = Mu

so that Mu =- Muo. ReplacingMuo by Mu, in (4.7) showsthat ul solves the implicit

obstacleproblem(4.5)-(4.6), i.e. u = u1.
If /Io < j1, then a similartechniqueshows u = ul. The main differenceis that we
need a more complicateddefinitionof u. Denote:

+l
To= T - ( -
Ao) -ln( 1- C,
c
TIMAL IMPULSE CONTROL OF PORTFOLIOS 601

SI

B
8

A
0 sO

FIGURE1. A Typical Region K of Feasible Transactions.

and define iu = uo upon [To, T]. For t < To:

- c
[poSo - (1 + cl)pl(Bl - sl)]e?(T-) + (1 - cl)plBlel(

)j} ~ if (1 + cl)pl(B1 - s) < (poo - c)

(1 - + cpl)-Pos +pisi] el(T - c

)[(l
if (1 + cl)pl(B1 - sl) > (poso - c).

The analysisis complicatedby thejump discontinuitiesin the derivativesof ui.Despite

this it may be shown that u < u and Mui= Muo. Details will be given in a
forthcomingpaper.
We returnto the case ,u < ,o by illustratingthe computationof Muo. We seek to
maximize: C(p,s, s') + uO(p,s', t) over the region K shown in Figure 1. (If poSo< c
then the picture changes slightly but the conclusions remain the same.) Since this is a
piecewise linear continuous function on a bounded convex set, it is enough to compare
values at the comer points:

A = (So + pao((1 - c1)p1s- c),O),

B = (So - poC, S1),

D = (, s, + (1 + c p)-p
-(poSO- c)),
E= (0, s),

0= (0,0).

Since to > O, it can be shown that s' = A gives a larger value than s' = 0 and that
s'= B is better than s' - E. Hence we can reduceto consideringA, B, and D. These
are such that C(p,s,s') = 0, so we actuallyare to comparethe values of uo at these
602 JEROMEF. EASTHAM& KEVIN J. HASTINGS

points:

UOls-A - UOs-B =
=(eo(- t) _ el(T-t))pl(1 - Cl)s,

eeo(T1-' ) - -
UOls-B
-
Uols-D ( i )e
-e
t(p oSo c).

Clearly /o > 1l implies that Muo = Uols.A.

The price po of the risk-freeasset behaves deterministically,so one would expect
that dependenceof the solution u on po could be eliminated.This can be accom-
=
plished by restricting u to the characteristic surface poe'?(T-t) constant, where the
constantis the value of a shareof risk-freeasset at time T. Since thereis no restriction
on the number s0 of risk-freeshares,the terminalvalue of Po can be set equal to 1
without loss of generality.
One can further simplify matters by considering w = ul - u0, which satisfies the
modifiedproblem:
= L(ul -
Low Uo)lpo=e-(T -) < 0,

(4.15) w > M[w + Uo] - U Mow,

(Low)(w - Mow) = 0,

subject to the homogeneousterminalcondition:

(4.16) Wlt-T = 0.

Some work and the relation Mul = Muo are needed to obtain the following
simplifiedexpressions:

Low - w + 'LplWp
t +4- 2ap ,pl
(4.17)
_
Mow (eMo(T-t)_ eli(T-t))(l C)P1S1 -Ce-o(T-t)

After taking the difference Mow = M[w + Uo] - Uo, the variable s0 no longer appears
in the problem,whichimpliesthat the optimalstrategywhen ,ul < to does not require
knowledge of s0. The form of the differentialoperator Lo is unchangedby the
substitution:

(4.18) x = (1 - cl)p1s.

Thus we can rewritethe problemas:

Low wt + iy1xwx + 1a2X2Wx < 0,

(4.19)
W > (eto(T-t) - - celo(T-t) Mow,
ePl(T-t))x

(Low)(w - Mow) = 0,
(4.20)
w(x, T) = 0,
where all direct referencesto p, s are eliminated.Note that x is the effectivewealthin
risky assets.
OPTIMAL IMPULSE CONTROL OF PORTFOLIOS 603

This time-dependentexplicit obstacle problem can be approximatelysolved by

formulatinga sequenceof problemsinvolvingan ordinarydifferentialoperator.Let
the time-step k = At > 0 be given, and discretizew in time by defining w()(x) =
w(x, T - jk). Replace the time derivativein Lo by a forwarddifferenceapproxima-
tion:

(4.21) wt,(x t) w(x, t + At) - w(x, t)

At

We then have the fully-implicitsemi-discreteschemefor (4.19):

_02
AW()x= XX -A1XW(j + k-1lw( > k-1w(-1),

w(i) (e ojk -e-lik)x - ceuoik _= 4ji)

(4.22) (AkW(j))(w() - (j) = 0

at time t = T - jk. For simplicity we take k = T and consider only j = 1, i.e. a single
'backward' time-step from w() = 0 at time t = T to w(l) at time t = 0. Then provided
that ,1T< 1, the solutionis:
r if 0 < x < x ,
w)(x) x)
= X 423/ i x
(4.23) ( ) \otherwise,

where r is the positive root of the initial equationfor AT, and y is determinedby
requiringx * to be a point of osculationbetween w() and (I(). That is:

r= 4[T( (2i - a2)2+ 8a2T-1 + (21i - a2))]

(4.24)
X* [ LOT
x(r - er '
e eiOT_

and y = ()(x,)/x*. The restriction,juT< 1 impliesthat the solutionwill remainin

contact with the obstaclefor large x. The form given for r (rationalizingthe numera-
tor) is preferredfor numericalcomputationbecauseit avoids subtractivecancellation
when jL1> a2/2.

OSC.

A- ptosc.
pts.

t_-.I t .1 ";U ;2 C.
,} U - ,4 -
E I

FIGURE2. One-Step Approximations w(1) for o2 = 0.01, 0.03, 0.05 (resp. x.* 55.15, 58.06, 60.17).
604 JEROMEF. EASTHAM& KEVIN J. HASTINGS

/
/

OSC.
pts.

60' " " 1 n"0 2:

P'?0't""^ " I J :_', ' ,0t I,
FIGURE3. Comparison of One- and Two-Step Approximations (resp. x, = 55.15, 53.53).

In Figure2 we show the effectof the parametera2 on the solution w(). Increasing
the value of a2 increasesboth the value of w(1) and the position of the contact point
x,. In the case illustrated, c = 1, io0= 0.05, /L = 0.03, and a2 = 0.01, 0.03, 0.05.
Recall that an optimal control is to interveneonly when the portfolio holdings
belong to the action set A. By (4.18) all dependenceof the policy on cl, pl, s1 has
been subsumedby x. Now x*, the point of contact between w?() and (P), is the
semidiscretizedapproximationto the boundaryof the action set, dA. The approxima-
tion of x * to dA may be worsethan the approximationof w ) to wIt,.-k. Figure3
shows both a one-step and a two-step approximationto w at time t = 0, with
AO= 0.05, ~t = 0.03, a2 = 0.01, c = 1.
One can treatthe case of discontinuouspriceprocessesby the same techniquesif the
jump sizes have a specialform. In this case a modifiedtrend pt4is obtained,and the
analysisagainproceedsaccordingto whethert0 <.< [ or t0o> p. Details arepresented
in a forthcomingpaper.

Acknowledgement.The authorswouldlike to expressappreciationto C. S. Tapiero,

who gave a talk at the Universityof Delawarein 1983 that motivatedthe problem
studied here. Thanks are also due to the referees whose careful reading of the
manuscriptled to substantialimprovementsin the presentation.

References
[1] Bensoussan, A. and Lions, J. L. (1982). Applications of VariationalInequalities in Stochastic Control.
North Holland, Amsterdam.
[2] and . (1984). Impulse Control and Quasi-Variational Inequalities. Gauthier-Villars,
Paris.
[3] and . (1975). Nouvelles methodes en contrble impulsionnel. Appl. Math. Optim. 1, 4
289-312.
[4] and Tapiero, C. S. (1982). Impulse Control in Management. J. Optim. Theory Appl. 37, 4
419-442.
[5] Blumenthal, R. M. and Getoor, R. K. (1968). Markov Processes and Potential Theory.Academic Press,
N.Y.
[6] Doshi, B. T. Optimal Control of a Diffusion Process with Reflecting Boundaries and Both Continuous
and Lump Costs. Proc. Internat. Conf. Dynamic Programming, Academic Press, Vancouver (to
appear).
[7] Gikhman, I. I. and Skorohod, A. V. (1979). ControlledStochastic Processes. Springer-Verlag,N.Y.
[8] Glowinski, R., Lions, J. L. and Tr6molihres,A. (1981). Numerical Analysis of VariationalInequalities.
North Holland, New York.
[9] Kushner, H. (1977). Approximations and Computational Methods for Optimal Stopping and Stochas-
tic Impulsive Control Problems. Appl. Math. Optim, 3, 2/3 81-99,
OPTIMALIMPULSECONTROLOF PORTFOLIOS 605

[10] Lehoczky,J., Sethi,S. and Shreve,S. (1983).OptimalConsumptionand InvestmentPoliciesAllowing

ConsumptionConstraintsand Bankruptcy.Math.Oper.Res. 8, 4 613-636.
[11] Lepeltier,J. P. and Marshal,B. (1979). TechniquesProbabilistesdans le Contr6leImpulsionnel.
Stochastics.2 243-286.
[12] Magill, M. and Constantinides,G. (1976). Portfolio-Selection
with TransactionsCosts. J. Economic
Theory.13 245-263.
[13] Merton, R. (1971). OptimumConsumptionand PortfolioRules in a Continuous-TimeModel. J.
EconomicTheory.3 373-413.
[14] Schal,M. (1974).A SelectionTheoremfor OptimizationProblems.Arch. Math.25 219-224.
[15] Yavin, Y. and Venter,A. (1981).Computationof ImpulseControlLawsfor a Non-LinearStochastic
Oscillator.Comput.MethodsAppl. Mech. Engrg.28 129-143.

EASTHAM: DEPARTMENTOF MATHEMATICS,UNIVERSITYOF DELAWARE,NEWARK,

DELAWARE19716
HASTINGS: DEPARTMENTOF MATHEMATICSAND COMPUTERSCIENCE,KNOX COL-
LEGE,GALESBURG,ILLINOIS61401