See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/226896109

Explicit Solution of a General

Consumption/Investment Model

Chapter in Mathematics of Operations Research · April 2006

Impact Factor: 1.31 · DOI: 10.1007/BFb0041165

CITATIONS READS

204 218

4 authors, including:

Ioannis Karatzas Suresh Sethi

Columbia University University of Texas at Dallas
152 PUBLICATIONS 7,955 CITATIONS 431 PUBLICATIONS 10,623 CITATIONS

SEE PROFILE SEE PROFILE

Available from: Suresh Sethi

Retrieved on: 18 May 2016
Explicit Solution of a General Consumption/Investment Problem
Author(s): Ioannis Karatzas, John P. Lehoczky, Suresh P. Sethi, Steven E. Shreve
Source: Mathematics of Operations Research, Vol. 11, No. 2 (May, 1986), pp. 261-294
Published by: INFORMS
Stable URL: http://www.jstor.org/stable/3689808
Accessed: 30/09/2009 15:17

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.

Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at
http://www.jstor.org/action/showPublisher?publisherCode=informs.

Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.

JSTOR is a not-for-profit organization founded in 1995 to build trusted digital archives for scholarship. We work with the
scholarly community to preserve their work and the materials they rely upon, and to build a common research platform that
promotes the discovery and use of these resources. 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. II, No. 2, May 1986
Printed in U.S.A.

EXPLICITSOLUTIONOF A GENERAL
CONSUMPTION/INVESTMENTPROBLEM*t
IOANNIS KARATZAS, : JOHN P. LEHOCZKY, ? SURESH P. SETHI, ** AND
STEVEN E. SHREVEtt
This papersolves a generalconsumptionand investmentdecisionproblemin closed form.
An investorseeks to maximizetotal expecteddiscountedutilityof consumption.Thereare N
distinctriskyinvestments,modelledby dependentgeometricBrownianmotionprocesses,and
one riskless(deterministic)investment.The analysisallowsfor a generalutilityfunctionand
generalratesof return.The modeland analysistake into considerationthe inherentnonnega-
tivity of consumptionand considerbankruptcy,so this papergeneralizesmany of the results
of Lehoczky,Sethi, and Shreve[5]. The value function is determinedexplicitly,as are the
optimal consumptionand investmentpolicies. The analysis is extended to consider more
generalriskyinvestments.Undercertainconditions,the valuefunctionsderivedfor geometric
Brownianmotionare shownto provideupperand lowerboundson the valuefunctionsin the
moregeneralcontext.

Contents
?1. Introduction ???? .262
?2. of
Summary Assumptions and Results . . ??? .263
?3. Admissible Policies. ???? .265
?4. The Bellman Equation. ???? .267
?5. Reduction of the Model: Mutual Fund Theorem ??? .268
?6. Solving the Bellman Equation. Part I ????? .270
?7. Candidate Optimal Policies ????? .272
?8. Performance Evaluation of Candidate Optimal Policies . ? .273
?9. Solving the Bellman Equation. Part II ????? .278
?10. Solution when U'(0) = o, U(O) Is Finite, P = U(0)/3, .279
?11. Solution when U(0) = - oo..... ???? .280
?12. Solution in All Other Cases where the Consumption Constraint
Is Inactive ..284
? 13. Solution when the Consumption Constraint Is Active .284
?14. Solutions for HARA Utility Functions .... .290
?15. Dependence of the Optimal Policy on the Prices . .291
? 16. The Model with Nonconstant a, v and r ... . 292
? 17. Tabulated Results ....... . 292

*ReceivedApril23, 1984;revisedMarch 1, 1985.

AMS 1980 subject classification. Primary: 93E20. Secondary: 49B60, 90A16.
IAOR 1973 subject classification. Main: Finance. Cross references: Control.
OR/MS Index 1978 subject classification. Primary: 197 Finance/Investment.
Keywords.Consumption,investment,utilityfunction,Brownianmotion,controlof diffusions,mutualfund
theorem.
tResearchwas supportedin partby the NationalScienceFoundationundergrantsMCS-8202210(Lehoczky
and Shreve)and DMS-84-16734(Karatzas),and by the NSERCof CanadaundergrantA4619and SSHRC
of Canadaundergrant410-83-0888(Sethi).The authorswish to thankthe refereefor a numberof helpful
comments.
*ColumbiaUniversity.
?Carnegie-MellonUniversity.
** University of Toronto.
ttCarnegie-Mellon University.
261
0364-765X/86/1102/0261$01.25
Copyright ? 1986, The Institute of Management Sciences/Operations Research Society of America
262 I. KARATZAS,J. P. LEHOCZKY,S. P. SETHI & S. E. SHREVE

1. Introduction. This paper investigates the optimal consumption and investment

behavior of a single individual who attempts to maximize total expected discounted
utility of consumption. The paper is a continuation of the work initiated by Lehoczky,
Sethi and Shreve [5]. The investor begins with an initial wealth x and makes
consumption and investment decisions in continuous time. There are N + 1 distinct
investments available to the investor. One is riskless (deterministic) with rate of return
r > 0, whereas the others are risky and, following Merton [71-[9], Black and Scholes
[1], are modelled by geometric Brownian motions. More specifically, let Po(t) be the
price of the riskless asset and let P(t) = (Pl(t), . . ., Ps(t)) be the vector of prices at
time t of the N risky assets. We assume

dPo(t)
dP(t dt (1.1)
Po0(t)
dPi(t)P =
aidt + e,DdwT(t), i = 1,2 ..., N (1.2)

where {w(t), t > 0) is an N-dimensional standard Wiener process, ei is the unit row
vector with a one in the ith position, ai is the average rate of return on the ith asset, D
is an N x N matrix with I = DDT positive definite, and superscript T denotes the
transpose operation.
We allow the investor to specify a consumption rate {c(t), t > 0) and an investment
policy (w(t) = (Orl(t), ... , 7N(t)), t > 0}, where vi(t) denotes the fraction of wealth
invested in the ith investment. The remaining fraction of wealth, iO(t)= 1 - S=lri(t),
is invested in the risk-free asset. The 'r(t) vector is unconstrained, which means that
unlimited borrowing and shortselling are allowed; c(t) must be nonnegative. Both
{c(t), t > 0} and {w((t), t > 0} must be chosen to depend on the price vector (P(t),
t > 0) in a nonanticipative way. We give this statement precise meaning in ?3.
For given {c(t), t > 0) and {r(t), t > 0), the investor's wealth process, as discussed
in ?3, obeys the It6 stochastic differential equation

x(O) = x, (1.3)

dx(t) = (a - rl)r T(t)x(t)dt + (rx(t) - c(t))dt + x(t)r(t)DdwT(t), (1.4)

where a = (a,, . . ., aN) and 1 = (1, . . ., 1).
A complete formulation of the model requires some assumption concerning the
options available to the investor if wealth reaches zero, since further consumption
would result in negative wealth. We refer to the state of having wealth zero as
bankruptcy.Many assumptions about bankruptcy are possible. For example, the agent
might be re-endowed and allowed to restart the decision problem. We incorporate all
such models by assigning a value P to bankruptcy, and including P as a parameter of
our model. More specifically, we define

To= inft > 0: x(t) = 0, (1.5)

and, if To < oo, then the agent receives payment P at time T0, and the decision
problem terminates.
We introduce now a strictly increasing, strictly concave utility function U on (0, oo)
whose first three derivatives are continuous, and we extend U to [0, oo) by defining
U(0)= limc0QU(c).The investor wishes to choose (c(t), t > 0} and {(?(t), t > 0) so as
to maximize the expected total reward

Vc(.),,( )(x) Ex[Ter-U(c(t))dt +

Pe-T]. (1.6)
L ^~~~~~~~~~~~~16
 EXPLICITSOLUTION OF GENERAL CONSUMPTION/INVESTMENT PROBLEM 263

The parameter,f > 0 is a discount factor. Let us note here that setting P = U(O)/1f
would be equivalent to continuing the problem indefinitely after bankruptcybut
allowingonly zero consumption.This value of P will be called the naturalpayment and
will play a distinguishedrole in the analysis.
Severalvariantsof this consumption/investmentproblemhave been studiedexten-
sively ([5]-[11]). Much of the literatureis devoted to drawingconclusionsfrom the
Bellman equation about the nature of optimal policies, and most authorshave not
attended to the difficulties created by the consumption constraint c, > 0 or the
possibilityof bankruptcy.Instead,it has been tacitly assumedthat under an optimal
policy, the consumptionconstraintwill not be active and bankruptcywill not occur.
The condition U'(O)= oo has been cited as sufficientto ensurethis.We shall show that
the consumptionconstraintis active only when U'(0) is finite, and then only for a
particularset of values of P. On the other hand, we show that the possibility of
bankruptcyis not relatedto U'(0); rather,bankruptcyoccurswith positiveprobability
if and only if P > U(0)/f,. See the summaryof our resultsin ?2.
Previousworkshave failed to settlethe questionof existenceof an optimalpolicy. In
this paper,we exhibitthe value functionexplicitlyand use it to give preciseconditions
under which an optimal consumption/investmentpolicy exists. Besides, previous
works have assumed without proof that the value function is twice continuously
differentiableand satisfies the Bellman equation;here we establishthese facts. The
readermay wish to consult[5] for furtherdiscussionabout the relationof our model to
existing economics literature.In [5], the model with N = 1 and a, = r was treated
under the assumptionsc > c > 0, 0 < 77 < 1, as well as for unconstrained7T,. The
latter case (referredto in [5] by the device a = oo) with c = 0 is a special case of the
model of this paper.
This paperis organizedas follows. ??2 and 3 serveas a summaryof resultsand as a
listing of all our assumptionsabout the model. In ?4 we introduce the relevant
Bellmanequation,and we employ it in ?5 to show how the model with one risk-free
and N risky investmentscan be reducedto a situationwith one risk-freeinvestment
and only one risky investment.This reductionis known in the literatureas a Mutual
Fund Theorem(see e.g. [8]).The remainderof the paperthus deals with a model with
N = 1. In ?6 we define a mapping X(.;a,B) from consumption to wealth; this
mappingis suggestedby an analysisof the Bellmanequation.In ?7 this mappingis
used to create a class of candidateconsumption/investmentpolicies, from which we
laterselect an optimalpolicy. In ?8 we show how to evaluatethe performanceof these
candidateoptimalpoliciesand use this performanceevaluationto providesolutionsto
the Bellmanequation.??10-13 are devoted to selectingthe optimalpolicy in various
cases which can arise. In ?14 we discuss our resultswhen the utility function U is a
memberof the HARA class. In ?15, we note that the optimalpolicy we have obtained
is drivenby the asset prices.The generalizationof theseresultsto a largerclass of risky
investments is discussed in ?16. We report that the value functions derived for
investmentsmodelledby (1.1), (1.2) with constanta, r and D can serve as upperand
lower bounds for the value functions which arise when a, r and D are random
processes.A convenienttabularsummaryof our resultsis presentedin ?17.
2. Summaryof assumptionsand results. We recallour assumptionsthat wealthis
governedby (1.4), where2 = DDT is positivedefinite,and that the performanceof a
policy is given by (1.6), where U is a real-valuedfunction, strictlyincreasing,strictly
concave, and C3 on (0, oo). We set
U(0) = lim
cO
U(c),
c0 U'(O) = lim U'(c).
Note that U(0) may be - oo and U'(0) may be + oo. We recall also that / > 0 and
264 I. KARATZAS, J. P. LEHOCZKY, S. P. SETHI & S. E. SHREVE

r > 0, but there is no constraint on the vector a = (a,, .., aN). On the utility
function U we impose the sublinearity condition

U(c) =
lim 0, (2.1)
c->oo C
which is equivalent to
lim U'(c) = 0. (2.2)
C-> 00

We define the nonnegative constant

y = 2(a- rl)- '(a- rl,)T (2.3)

and we consider the equation

yX2- (r - -y)X - r = 0. (2.4)

If y > 0, this equation has two solutions:
A_ < -1, A+ >0; (2.5)
whereas if y = 0 and / < r, we define X_ = - r/(r - 3), which is then the (negative)
solution to (2.4). Whenever X_ is defined as above, we assume

f dO < oo Vc > 0, (2.6)

JC ( '( ))

a condition stronger than (2.2). We shall see in ?14 that failure of (2.6), when U is a
HARA function, results in an infinite value function. Condition (2.6) is weaker than
the condition imposed by Merton [7, condition (41)] to ensure a finite value function;
see also (14.4).
Under these assumptions, and under some technical conditions on {c(t), t > 0} and
{rT(t), t > 0} designed primarily to ensure the existence of a solution to (1.4) (see ?3),
Vc(. ,( .(x) is given by (1.6), and we can define the valuefunction
V*(x) = sup V(.),,(.)(x), x > 0, (2.7)
c(-),ITr(-
)
which is finite for every x > 0 (Corollary 10.3).
The model is interesting only when

U(O) <P< lim U(c). (2.8)

If P > 3-'limc~o U(c), one should consume to bankruptcy quickly, and the value
function is then identically equal to P. There is no optimal policy, since instantaneous
bankruptcy cannot be achieved. If P < U(O)//f, one should behave as if P were the
natural payment U(0)/,8 because, as we shall see, this does not result in bankruptcy.
We henceforth assume (2.8).
In this paper we obtain a closedform expression for V*. From this expression we
draw the following conclusions when y > 0 (i.e., a - rl = 0). These conclusions, along
with those for the case a - rl = 0, are tabulated in ?17.
la. If P = U(0)//, and U'(0) = oo, the optimal consumption is never zero but is not
bounded away from zero, and the optimal wealth process does not lead to bankruptcy
(? 10 and Theorem 11.4).
2a. If P > U(0)/l/ and U'(O)= oo, the optimal consumption is bounded below by a
positive constant a (Theorems 11.2 and 12.1), and the optimal wealth process leads to
bankruptcy with positive probability. The probability of bankruptcy is equal to one if
and only if / > r + y (Remark 11.3).
 EXPLICIT SOLUTION OF GENERAL CONSUMPTION/INVESTMENT PROBLEM 265

3a. If U'(O) is finite and

1(O P <P* A UO
U(0)< P < P*
1 -
(O) UV(O)x
dO-
+-
0 r (2.9)
(2.9)
pxv (u'(O))X

then for low levels of wealth the optimal consumption is identically zero. If P = U(O)
/,B, bankruptcy is avoided, but if P > U(O)/fl, bankruptcy occurs with positive
probability, which is equal to one if and only if ,/ > r + y (?13).
4a. If U'(O) is finite and P > P*, the optimal consumption is never zero and is
bounded away from zero if and only if P > P* (Theorem 12.2). There is a positive
probability of bankruptcy, and this probability is equal to one if and only if 3 > r + y
(Remark 12.3).
When y = 0 (i.e., ac = r, i = 1 ,..., N), this model becomes the a = oo model of [5].
We extract from that paper the parallel cases:
lb. If P = U(O)/,l and U'(O)= o, we have the same conclusions as in (la) above.
The agent places no money in the risky assets.
2b. If P > U(O)/l/ and U'(O)= o, the optimal consumption is bounded away from
zero whenever consumption is relevant. If ,/ > r, the optimal investment strategy is to
invest only in the risk-free asset, and bankruptcy occurs. If /3< r, there is a positive
critical level of wealth x, such that, when initial wealth x is in (0, xx), the investor
places arbitrarily large sums of money in the risky assets by borrowing against the
risk-free one. Regardless of consumption, this causes immediate bankruptcy (with
probability (xo - x)/x,) or causes wealth to move immediately to x, (with probabil-
ity x/xO). When his wealth is in [xx, oo), the agent invests only in the risk-free asset
and consumes so as to increase his wealth. If P = r, constant consumption c = rx,
and any investment policy are optimal so long as wealth is in [0,x.). By choice of
investment policy in this region, the investor can cause the probability of bankruptcy
to be any number between (xo - x)/x, and 1, where x is the initial wealth, assumed
to lie in (0, x,). If wealth equals or exceeds x,, the agent invests only in the risk-free
asset and consumes so as to keep his wealth constant.
3b, 4b. If U'(O) is finite and P = r, then the optimal consumption and investment
policies are as in (2b) (but xo = 0 when P = U(O)//l).
If U'(O) is finite and /f < r, then the optimal investment policy is as described in
(2b) (again with xo = 0 when P = U(O)//). To discuss the consumption policy, we
note that when / < r, X_ is defined and so P* can be defined by (2.9). We can then
define a second critical level of wealth x which is strictly greater than x, when
U(O)/l < ? P < P* and equal to xo when P = P*. As a function of wealth, optimal

consumption is irrelevant in (0, xO), zero in [x,O,x], and increasing in (x, oo). When
P > P*, we have x < xX and optimal consumption on [x,, oo) is bounded away from
zero.
If U'(O) is finite and /f > r, the agent invests only in the risk-free asset. If
P = U(O)//f, consumption is positive but not bounded away from zero; if P > U(0)
//3, consumption is bounded away from zero. In both cases, bankruptcy occurs.

3. Admissible policies. In this section, we discuss informally how the asset price
equations (1.1), (1.2) lead to the wealth equation (1.4), and we then impose conditions
on consumption and investment policies so as to ensure a solution to (1.3), (1.4). The
connection between (1.1), (1.2) and (1.4) is fairly complex and is tangential to the
subject of this paper, so we deal with it briefly. The interested reader may wish to
consult Harrison and Kreps [3] and Harrison and Pliska [4] for a fuller account of
security trading.
Let (c,, t > O0 be the consumption rate process, and let n,i(t), t > O0 be a process
266 I. KARATZAS,J. P. LEHOCZKY,S. P. SETHI & S. E. SHREVE

giving the number of shares held of asset i, i = 0, 1, N, whose price Pi(t) is given
by (1.1), (1.2). The agent's wealth at time t is then
N

x(t)= ni(t)Pi(t). (3.1)

i=O

If the trading of shares is done at discrete time points, say t - h, t, and t + h, and
consumption levels are held constant over these intervals, then we have
N N

(t -)= ni(t - h)Pi(t) = 2 ni(t)P(t) + c,h = x(t) + cth, (3.2)

i=O i=O

where x(t - ) is the wealth just before time t. Equation (3.2) states that the trading of
shares at time t is used to finance consumption during the interval [t, t + h). From
(3.1) and (3.2) we can derive
N

x(t)
- x(t - h) = n(t - h)[P(t) - P(t- h)] - cth. (3.3)
i=0

When trading occurs in continuous time, we adopt the continuous analogue of (3.3):
N
dxt = E n(t) dP(t) - ct dt, (3.4)
i=O

which says that the change in wealth is due only to capital gains from price changes in
the assets and to consumption. If we now set ri(t)= ni(t)Pi(t)/x(t), which is the
fraction of wealth invested in asset i, (1.1), (1.2) and (3.4) yield (1.4).
We now take (1.3), (1.4) as the starting point for a careful formulation of the model.
Let {wt, Y,, t > 0} be an N-dimensional standard Wiener process on a probability
space (2, Y, ), where { ,, t > 0} is a nondecreasing, right-continuous family of
a-fields. An admissible consumption process { c, t > 0} is a nonnegative process
adapted to {(t, which satisfies almost surely

fc,ds< oo, t > 0. (3.5)

For each {(t }-adapted, N-dimensional random process {rt, t > 0}, define the
{Y, }-stopping time

T(r) = sup{t > 0 :,'T ds< oo }. (3.6)

For c and r as described above, we can define a solution to (1.3), (1.4) for
0 < t < T(sr). Indeed, a solution is

- -
x ez,x J'cse ds, 0 < t < T(Qr), where (3.7)

z J t((a - rl)rST + r - ds +JvsDdwr, 0 < t < T(t).

s2r) (3.8)

Note that for t < T('n), the Ito integral in (3.8) is defined because fV%sr,sTds < 0o
almost surely.
Of particular interest is the time of bankruptcy To given by (1.5). For each x > 0, we
call admissible at x any adapted pair of consumption and investment processes { ct, ,,
t > 0} for which either T(r) = oo or To < T(Qr)or lim,TT(,r)xt exists and is zero. This
ensures that To is defined and <
To T(r). The supremum (2.7) is taken over all pairs
in
c(.), ir(.) of consumption and investment processes admissible at x for which
 EXPLICIT SOLUTION OF GENERAL CONSUMPTION/INVESTMENT PROBLEM 267

VC(.),,(-(x) is well defined by (1.6), i.e., the integral and the expectation in (1.6) are
defined. We shall show (Remark 10.5) that in fact VC,(,()(x) is well defined
whenever c(.) and 'w(.) are admissible at x. Note that because (1.4) is a linear
stochastic differential equation, albeit with random coefficients, its solution is unique,
and it is given by (3.7) and (3.8), at least up to time To.
Because of (2.8), there is some E > 0 for which U(E)> fP. The pair c, =_'w"V-_0 is
admissible for any x > 0 and yields V,(.),,(.) > P, so V*(x)> P, x > 0. Of course, we
always have V*(O)= P.

4. The Bellman equation.

THEOREM With P a finite number,let us assume that V: (0, oo)-*(P, 00) is a C2
4.1.
function satisfying the Bellman equation
3V(x) = max[(a - rl)TTx V'(x) + (rx - c)V'(x) + l.V.TX2V,,(x) + U C

x > 0. (4.1)
If U(0) is finite, then V(x) > V*(x), x > 0.
PROOF. Fix x > 0 and select admissible c(.) and 'nu(.). Choose 0 K <<x < 2,2
define

Sn = inf{t > 0 :fuslrsTds = n

and use (4.1) and It6's lemma to write

Exf TEA/ TE2A/\Sne-'tU (c,) dt

< Ex /3An
TE ASn tO[ /V(x,) - (a - rl)lrT'x1V'(x,)

(rx, - C,) V'(x,) - f Urj;a TX2VI(xt)] dt

-

=ExfT T,A1'2 AS[ - d(e -tV(xt))+ e Vx,'(x,)rrDuDdWT]

- Exexp(- f(Tc, A T5, A Sn))V(x(TC, A T~2

A Sn)) + V(x). (4.2)
Since U(0) is finite, we may rewrite (4.2) as

- TE, A T2ASne
V(X) -P
U(0)> Exf [ U(c1) - U(0)] dt

+Eexp(-3(T7 A TA Sn))[ V(x(TA TA A Sn)) - (O)

Now let I40, 52 -) oo and n - oo, so T,' A T52 A SnITo almost surely by the assump-
tion on admissible policies. By the monotone convergence theorem and the nonnega-
tivity of U(c,) - U(0), the first expectation converges to

Exf TOe'lt[U(c,) -U(0)ldt.

On the other hand, (2.8) implies

I U(O) > P -
- U(0 >00
V(x(TC, A TIA Sn))- (0
268 I. KARATZAS,J. P. LEHOCZKY,S. P. SETHI & S. E. SHREVE

Therefore, by Fatou's lemma,

8(T, Eexp(- A T A Sn)) V(x(T A A S)) - -

lim T U(0)]
n--> oo

> E e- P- U(O))

and we conclude with

V(x) - 1 U(O) > Ex[, Te -'U(c))dt + Pe -T -

U(O)

-VC(),(.)(X)- A U(O).

Maximization over c(-), 'w(*)gives the desired result. I

REMARK 4.2. If U(O)= - oo, the integrand on the left-hand side of (4.2) may be
unbounded both above and below. However, if we have
+
E,x Toe ~tU (c)dt< oo, (4.3)

where U + = max{0, U) and U- = max{0, - U}, then (4.2) is valid and we may use
the monotone convergence theorem on the two integrals
Tt A T2
^ASne-Ftu+ (ct)dt

to conclude V(x) > V( .)J()(X). Indeed, if V satisfies the assumptions of Theorem

4.1, P is finite, and (4.3) holds for every admissible c( ), vr(), then V(x) > V*(x) for
all x > 0 even if U(0) = -oo.

5. Reduction of the model: mutual fund theorem. Let us assume for the moment
that y defined by (2.3) is strictly positive. Choose any a and a > 0 so that
I1 (a- r)2
2 a2 (5.1)

and consider the "reduced" consumption/investment model with a single risky asset
whose mean rate of return is a and variance is a2. For example, if one constructs a
"mutual fund" which trades continuously using a self-financing strategy [4, p. 237] to
maintain proportions of the risk-free and N risky assets given by the (N + 1)-
dimensional vector (1 - (a - rl)1- 1T,(a - rl)-1 ), then the mutual fund has aver-
age rate of return a = r + 2y and variance a2 = 2y, so (5.1) holds. Moreover, if
(a - rl)S- 11T 0, then the mutual fund consisting only of risky stocks held with
proportions (a - rl)S-/(a - rl)- llT also satisfies (5.1). The other data of the
model remain unchanged. The reduced wealth equation (cf. (1.4)) takes the form
X = x, (5.2)

dx, = (a - r), dt - dtt+ x 47,tyd, (5.3)

{
where W(, t > 0} is a one-dimensional standard Wiener process, and 47,is the fraction
of wealth placed in the risky investment at time t. The reduced Bellman equation is (cf.
(4.1))
V(x) = max [(a
- r)rxV'(x) +
(rx - c)V'(x) + 72a2x2V"(x) +
U(c)],

x > 0. (5.4)
 EXPLICIT SOLUTION OF GENERAL CONSUMPTION/INVESTMENT PROBLEM 269

This reduced model is the subject of ??6-14. We shall exhibit there a C2, strictly
increasing, strictly concave function V: (0, oo) -(P, oo) satisfying (5.4). For each
x > 0, we shall produce an admissible consumption/investment process {(t, t,, t > 0)
such that when (xt, t > 0} is given by (5.2), (5.3), then c = ct, ~ = 7Tt attain the
maximum in (5.4) with x = Xt; in particular,
A
aV'(x1) ax-r (5 5)
=
=- . (5.5)
xtV (Xt)

Furthermore,
V(x) = Ex fe'U(ct)dt + Pe -o] where (5.6)

To= inf(t > 0: Xt = 0}.

From these facts we shall conclude that V is the value function and { , t, t > 0) is an
optimal policy for the reduced model.
It is a straightforward exercise to verify that the strictly concave solution V to (5.4)
also solves (4.1), and the maximizing wrin (4.1) is

V'(x)
A= (a-rrl)S-1.
xV"(x)
Comparison with (5.5) suggests that for x > 0, we should define a consumption/in-
vestment policy for the nonreduced model by

t= , (5.7)

= A02
't -r (a- rl)- 1* (5.8)

Substituting this policy into (1.4) and using (2.3) and (5.1), we obtain
X0= x, (5.9)
dxt = (a - r)7 dt + (r
(xt - dt +
x)dt t dw,, (5.10)
where
wt 'DwT
a -a r (a-rl)-

is a standard Wiener process. Comparing (5.9), (5.10) with (5.2), (5.3), we conclude
that {x, t > 0} and {xt, t > 0} are equivalent in law, so (5.6) implies

V(x) = Ex[f e-U(ct)dt + Pe -T = VC(.),(.X), (5.11)

where
T = inf{t > 0 : x,= 0.

From (5.11), we see that the value function V* for the nonreduced model dominates
V, i.e.,
V*(x) > V(x), x > 0. (5.12)
If P > - o and U(0) > -oo, Theorem 4.1 implies V= V*, so we have solved the
nonreduced problem. If U(0) = - oo and P > - oo, then either U(c) < 0 for c > 0 or
else we can find c, > 0 and construct another utility function U1 such that U,(c)
= U(c) for c > ci, Ul(c)> U+ (c) for c > 0, and U(0) > - oo. With V and V* both
corresponding to U1 as V and V* correspond to U, we have V*'(x) = Vl(x) < oo, so
Remark 4.2 applies, and we again have V = V*.
270 I. KARATZAS, J. P. LEHOCZKY, S. P. SETHI & S. E. SHREVE

If U(0)= -oo and P = -oo, we let {P,)}, be a sequence of negative numbers

with Pn, - o0. With V1 and V,*corresponding to the model with payment P,, V and
V* corresponding to the model with payment P = -oo, we have V, = V* > V*. We
show in the proof of Theorem 11.4 that limn, Vn = V so V > V*. This, together with
(5.12), shows that even in this case the solution to the reduced problem provides a
solution for the nonreduced problem.
We now turn to the case y = 0, i.e., a = rl. Consider the subclass of admissible
investment policies -r(t) = (vT(t), ... , 7T(t)) in the nonreduced problem for which
= * * *= 7N(t) = 0, i.e., investment is allowed only in the risk-free asset
2(t) 7T3(t)
and the first risky asset. The problem of maximization over this subclass of policies
was solved in [5], an optimal value function V was obtained which clearly must satisfy
V < V*, where V* is the result of maximization over all admissible investment policies,
and the multidimensional version of Theorem 5 of [5] can be employed to show
V > V*. When a = rl, the risky assets are used in the model only as a way to
immediately drive wealth to either zero or a threshold xo above which no further risky
investment is made (see items 2b-4b in ?2), and one risky investment is sufficient for
this purpose.

6. Solving the Bellman equation. Part I. Throughout the remainder of the paper,
we assume a single risky investment so the wealth equation is (cf. (5.2), (5.3))
x = x, (6.1)
dx, = (a - r)7Tx, dt + (rx, - c,)dt + x, Tr, dw,. (6.2)

We assume a # r and define y by (5.1). Note that we allow a < r as well as a > r.
(The case of a = r has been dealt with in [5].) The Bellman equation (4.1) reduces to
(5.4).
In this section we proceed somewhat heuristically, with the aid of hindsight, using
the Bellman equation to discover a class of feedback functions from wealth to
consumption. In ?7 we employ these functions to create consumption/investment
policies, and in later sections we show how to choose the various parameters so that
these policies are optimal. In ?13 we shall be forced by a particularly difficult
combination of the model parameters to generalize the approach of this section; we
could have used that more general methodology throughout the entire paper, but chose
instead to retain the more intuitive approach of this section for the bulk of our
analysis.
We attack the Bellman equation (5.4) under the assumption that V" < 0. After we
obtain the value function V* in closed form, we will be able to verify that its second
derivative is negative. When V" < 0, the maximizing r1in (5.4) is given by

(a - r) V'(x) (6.3)
- , (6.3)
- V- a
2xV"(x)
and the Bellman equation may be rewritten as

-
V(x) = V(x) + max [(rx - c)V'(x) + U(c)]. (6.4)
V"1(x) ,c>O
When the constraint c > 0 is not binding, the maximizing c = C(x) must satisfy

V'(x)= U'(C(x)), (6.5)

and differentiating with respect to x, we obtain

V"(x)= U"(C(x))C'(x).
 EXPLICIT SOLUTION OF GENERAL CONSUMPTION/INVESTMENT PROBLEM 271

We may now rewrite (6.3) as

(a - r) U'(C(x))
7T (6.6)
(6.6)
xU"(C(x))C(x)
which removes all reference to the unknown function V. It remains to determine the
function C(.), and we do that by assuming that C(.) has an inverse function X(-) and
rewriting (6.4) as

= - Y( '(c))'X'(c)+
(U(c)2X
/v(x(c)) (rX(c) - c)U'(c) + U(c).
Ul(c)
Differentiation with respect to c yields the second-order, linear, ordinary differential
equation
U"(c) yU"'(c) U"(c)
yX"(c)= (r U- 2y) (c)+ U"(c) X'(c) + U'() (rX(c)- c). (6.7)

We postulate a solution to the homogeneous version of (6.7) of the form (U'(c))x,

and verify that we have indeed such a solution if and only if the constant X satisfies the
equation yX2 - (r - 8/ - y)X - r = 0. There are two roots: _ < - 1, X+ > 0. A partic-
ular solution to (6.7) can be found by variation of parameters. One such solution is

X(c;a)= c- II (U( c)) A+ dO + (U(c))A x dO

r y(X+ -X_) { X+ a (U'(0))+ X_ I u
(U(O9))

where a > 0 is a parameter. Note that X(c;a) is well defined because of (2.6). In
verifying that X(c; a) solves (6.7), one should note the useful relation
A+A_= -r/y. (6.8)
The general solution of (6.7) is

B( U'(c))X++ B( U(c))- + X(c; a),

but experience suggests that we should discard the rapidly growing term (U'(c))X- by
setting B = 0 and define
c
X(c; a, B ) B(U'(c))x^+

Y 1 ( U'(c))X rc dO
y(A+ -Ax_)l A+ ia (U'(o))^+

+ ( ( c)) ( dO)' , c > a. (6.9)

x-
(u,'(o))-
Note that the derivative of X(c; a, B) with respect to c is given as

X'(c; a, B) = B+ (U'(c))x+ -U"(c)

u"(c) JC
(U(c))+ dO
y (X+ -A_ )( C
(U'(())0

o
We
shall
X(; a , soct increasing
We[/'(c))-I
B) choose
is strictly
a,always ds (6.10)
) on (a, oo).
We shall always choose a > O, B < O, so X(.; a, B) is strictly increasing on (a, oo).
272 I. KARATZAS, J. P. LEHOCZKY, S. P. SETHI & S. E. SHREVE

We set X(a; a, B) 4 limcX(c; a, B). (When a = 0 and U'(O) = oo, we cannot simply
substitute c = a into (6.9).) From (6.8) we have

c _ (U (c)) c d
lim > lim
C-o00X(c; a,B) C-^oc y(X+ -A_)A+ a (U,())+

r y(X+-X )dX+

=lim ( 1 a
C->o
00 r y(X+ -A_ )X+ y(X+ -X )X+

= lim
+c a
+
c-->o r( -X ) y(X+ -X )X+
= 00,
(6.11)
so X(.;a,B) maps [a, o) onto [X(a;a,B), oo), and its inverse function C(.;a,B) is
C2, increasing, and maps [X(a; a, B), oo) onto [a, oo).

7. Candidate optimal policies. In this section we fix a > 0, B < 0 and use the
functions X(c)= X(c;a,B), C(x)= C(x;a,B) of ?6 to create a policy {c(t),r(t),
t > 0}. In ?8 we evaluate the performance of this policy.
We will ultimately select a > 0, B < 0 so that X(a) > O. Let = X(a) and assume
( > 0. Given x0 >
5, we define a wealth process {xI, 0 < t < T7} by (5.10), where (cf.
(6.6))
(a - r)U'(c,)
= 7r=- 2( t
C, C(xt), <) T. (7.1)

For t > T~, we assume c, = q(p(x,), tr,= T2(x,), where q1, q2 are chosen so that c(.) and
are admissible but are otherwise arbitrary.
7T( )
We derive a stochastic differential equation for {c,, 0 < t < T}. Note first that
since X(C(x))= x, we have X'(C(x))C'(x)= 1 and

= 0, x > /.
X"(C(x))(C'(x))2+ X'(C(x))C"(x) (7.2)
From (7.1), (7.2) and (6.7) we have

U"(cl) U"(c,) '(c)

U'( dt - -r U (c,)
dc, = - r + y dt
o
dw,. (7.3)
u", ( u a(c,)) U"(c,)
This equation can be solved explicitly. Definingy, = U'(c,), we obtain by a straightfor-
ward application of Ito's rule:

dy, =-(r - /)ytdt- a- r y,dw,, (7.4)

a linear stochastic differential equation, whose solution is given by

U'(c,) =y, = U'(co)exp[ -(r - , + y)t -a- r (7.5)

w].

Since U is strictly concave, U' has an inverse I. Equation (7.3) yields then a candidate
for the optimal consumption policy in the form:

0 < t < T7.

 EXPLICIT SOLUTION OF GENERAL CONSUMPTION/INVESTMENT PROBLEM 273

In terms of this policy, the candidate for the optimal investment policy is

,- r U'(c,) X'(c,)
77t < t<
a2 U"(c,) X(c, T)'
REMARK7.1. Note that

Tr inft t > 0:x =

= inf{t > O : c = a)
= inf{t 0:yt = U'(a)).
If a = 0 and U'(O) = oo, then T, = oo a.s. because Yt does not explode. Under such
conditions, bankruptcy cannot occur. However, if t = 0 and U'(a) is finite, bankruptcy
will occur when y, rises to U'(a). This happens almost surely if 8/ > r + y, and with
positive probability less than one if ,/ < r + y. (See (7.5).)

8. Performance evaluation of candidate optimal policies. In this section we evalu-

ate the expected return associated with the policy {c(t),7T(t), t > 0) of ?7. We let
v = Vc( ),(. )(), and we assume v is finite.
Now choose x0 > t and let co = C(xo). We have, by the Markov property,

Vc(.),(.)(Xo) = H(co) Exo[Te -U(c)+e-)dt +. (8.1)

We derive a closed form expression for H under the assumption that H(c) is well
defined and finite for all c > a. We justify this assumption later (Remark 10.4).
To aid in the analysis, we define

G(yo) = H(I(yo))

= Exo[fTe-ftiU(I(yt))dt + ve -t i, <y < U'(a), (8.2)

where {y,, 0 < t < T7) is given by (7.4), (7.5), with yo = U'(co). Theorem 13.16 of
Dynkin [2] (sometimes called the Feynman-Kac formula) implies that if G is well
defined and finite, then it is C2 on (0, U'(a)), satisfies

fG(y) = -(r ,- )yG'(y) + yy2G"(y) + U(I(y)), 0 <y < U'(a), (8.3)

and, if U'(a) < oo,

lim G(y) = . (8.4)

y U'(a)

It follows immediately from (8.3), (8.4) and the fact H(c) = G(U'(c)) that, if H is well
defined and finite, then it is C2 on (a, oo), satisfies

U'(c) U'(c) U"'(c)

/H(c) r- +Y ( H (c)
-U"() )

'
+ U"(c) )H"(c) + U(c), c> a (8.5)

and, if U'(a)< oo, then

= .
limH(c)
c,a
(8.6)
274 I. KARATZAS,J. P. LEHOCZKY,S. P. SETHI & S. E. SHREVE

The general solution to (8.5) is

J(c; a,A,A) A (U'(c))P+ +A(U'(c))- +Jo(c;a)

where p+ and p_ are the positive and negative roots, respectively, of the quadratic
-
equation yp2 (r - /3 + y)p - j/ = 0, and a particular solution of (8.5) is given by

U(c) _ 1 (U
U(c)) c de
J0(c,a) y(p+ P- ){ P+ Ta (U'(o))A+

+ U'(c))p o d
P- }(
u())
Note that
p+=l+ X , P+P-= - P/y. (8.7)
For some choice of A and A, we have H() = J( ;a,A,A). Our first task is to show
that because of the rapid growth of (U'(c))P as c-> oo, the proper choice for A is zero.
PROPOSITION 8.1. If H(c) is well defined and finite for all c > a, and if J(c; a,A,A)
H(c) for some choice of c > a, A, and A, where J(; a,A,A) is bounded below on
[c, oo), then J(c; a,A,A) > H(c) for all c > .
PROOF. For c > c, define Sc = inf{ t > 0: = cc. From the definition of H and the
strong Markov property, we have

H(co) = Exo fJSe -'U(c,)dt + H(c)e-s , c, > c. (8.8)

Since J( ; a,A, A) satisfies (8.5), and dc, is given by (7.3), we have

-e- a- r U'(ct) J'

d[e-'J(c, ;a,A,A)]= U(c,)dt + a a,( A)dw, , (8.9)
U"(c,)
and integration yields, for c < c0< c,

J(c; a,A,A) = ExtJSC S^ e-'U(c)dt + e-'ls{sS, H(c)

+ e-Sl{S<)J(c; a,AA)].
As coo, we obtain from the monotone and bounded convergence theorems and
Fatou's lemma:

J(co ;a,AA) > Exo e U(c,)dt+ )e H H( co).

LEMMA 8.2. We have

lim (U'(c))P+ Jo(c; a)

(U )P = 0,- lim < 0.
c-,O (U(C,)) c->OO(U,(c))P-
PROOF. The equality is a direct consequence of (2.2). For the inequality, we note
that
Jo(c;a) U(c) 1_ d
(U'(c))"- /(U'(c)) - yp- (p+ -p) ('())
 EXPLICITSOLUTION OF GENERAL CONSUMPTION/INVESTMENT PROBLEM 275

The integral converges to zero as c - oo because of (2.6), so it suffices to prove

U(c)
(
lim < 0. (8.10)
C-4 ( U'(c))

There is nothing to show if limc_o U(c) < oo. If not, let us assume for a moment that,
contrary to (8.10), there were to exist e > 0 and c, > 0 such that

U(c) > (U'(c))- > 0, >c c,.

Then

U'(o) 1
U(o) <(U())A ' >c,

and integrating from c, to c, we would obtain

E1og
U(c) < c dO f00
< d
<
U(CE) Jc(< 9)
u()
U~(c (Uc, u(0))

Therefore, U must be bounded above, contradicting the assumption limc_ooU(c) = oo.

So (8.10) holds in this case as well. i

LEMMA
8.3. We have limc,o,Jo(c; a) > - oo.

PROOF. For a + 1 < 0 < c, we have U'(c) < U'(0), from which follows

Jo(c;a) > U(c) dO + C

i (U(C))
'YP+(p+ -p-)) (U'(0) a+ (9))0d

)U(c)
-'P+ (p+ -p-)

P) ( ) -U(a
+ )
YP+(+-
U( U,( 0) ) .
From (8.7) we see that
1 1 L-+= P+ >0,
YP+(P+-P-) - (P+-p-)
so J0(c; a) is bounded from below as c - oo. I

PROPOSITION 8.4. If H(c) is well defined and finite for all c > a, thenfor some choice
of A, J(c; a,A, ) = H(c) for all c > a.
PROOF. Let A,A be such that J(c; a,A,A) = H(c) for all c > a. If A < 0, then
Lemma 8.2 shows that limc0oJ(c;a,A,A)= -oo. But (8.8) shows that H(co) is
bounded below as c0-> oo, so we must have A > 0.
Now choose c > a and let A be chosen so J(c; a, A, O) = H(c). According to Lemma
8.3, Jo is bounded below on [c, oo), so J(. ;a,A,0) is also. From Proposition 8.1 we
have

J(c;a,A,O) > H(c) = J(c;a,A,A), c > c.

In light of Lemma 8.2, this can hold only if A = 0. I

276 276
I~~1
KARATZAS,J. P. LEHOCZKY,S. P. SETHI & S. E. SHREVE

To simplify notation, we shall usually write J(c; a, A) instead of J(c; a, A, 0), i.e.,

aA U'(c))P +
) -A
J(c; U/3

I {(U'(c4)+ fC dO +
Y(P+ -P- ) P ,( )
(8.11)
THEOREm 8.5. If U'(a) < oc and H(c) given by (8.1) is well defined and finite for all
c > a, then H(c) = J(c; a, A), c > a, where A satisfies

U(a) _ (U'(a))P- 0 dO =
A(U(a))+ /3 Ta =V (8.12)
yp- (P+ -P ) (U,(O))x

PROOF. This follows from Proposition 8.4 and equation (8.6). I

Because U' is real-valued on (0, oc), the case U'(a) = oc can occur only when a = 0.
We show in Theorem 8.8 below that in this case, H ( ) = J( ; 0, 0). We shall need two
auxiliary results, Lemmata 8.6 and 8.7.
LEMMA8.6. Assume U'(0) = oo. Then

lim J(cO,O

PROOF. We must show

liU(c) (8.13)
CI~O (U'(c))~ P+
0

00
-P+ dO =00 (8.14)
c40 U())) Je(U/(O))A=

If U(0) is finite, (8.13) holds. If U(0) = - oo, then

U(C)
im <0
CI0 (U'(c))P+ 0

But because U is concave, we also have, for fixed C'> 0 and 0 < c < C',U(c) > U(s') -
U'c('- c), from which follows

U(c) . (s
lim = 0.
>1c)P , i

To prove (8.14), we observe that for C'> 0,

00
0?<lim (U'(c))0 -P+ dO
CIOJU'O0)

iii;i
?~ (4~
U'(c) X-
dO+ lim(U'(c))p-- r 00 dO
40IO
< lim 1I cO
40O (U1(c)) X+ J '(
 EXPLICITSOLUTION OF GENERAL CONSUMPTION/INVESTMENT PROBLEM 277

LEMMA8.7. Assume U'(O) = oo. Then

U(O)
limJ(c; 0,0)=
c40 f

PROOF. We distinguish two cases.

Case I: U(O) is finite. We will show

C d =0,
lim(U,(c))+
coo (8.15)
(U'(0))V+

dO =0.
lim(
c4o U,())- f (8.16)
(v'(o ))
We have from the concavity of U

0 < lim (U'(c))P+ f dO

cO Jo
(Uv'())X+
= lim cU'(c) < lim [ U(c) - U(O)] = 0,
c$O c1,O

and (8.15) is proved. To prove (8.16), we choose e > 0 and write

0c lim (U'(c))'P- dO
c4O0
v Jc (Uv'(o))-

lim I( U'(c) ) U'(0)dO+ lim (U'(c)) '- (

dOe
c,o U'(O) v,o (v',(o))

< lim rfc '(O)dO= U(E)- U(O).

Now let cE0.

Case II: U(O)= -oo. We have

1 r.
{ u'(c)
P-
lim JS(c;0, 0) < lim u'(o)dO
yp_(+ -p ) c \ U'(O)

_ f U(c) 00 dO
+ (U'(c))oP-
( u, (O))", I

U(c)
c <
lim
ctO
[ / YP_ (+ -p_) U( )ds

= Vi m- I4-+ 1
(c)-
C0o -/ YP (p+ -P-) , () -(p+ -p_ )
Using (8.7) we see that
I /3+ I P- -
>0,
i YP- (P+ -P-) 8(p + -P- )

and it follows that limc,oJ(c; 0, 0) = - oo.

278 I. KARATZAS,J. P. LEHOCZKY,S. P. SETHI & S. E. SHREVE

THEOREM8.8. If a = 0 and U'(O) = o, and if H(c) is well defined and finite for
every c > 0, then H(c) = J(c; O,0), c > 0.
PROOF. We know from Proposition 8.4 that for some value of A, H(-) = J(; 0, A).
From Remark 7.1 we see that c, can never reach zero in this case, so (8.8) holds for
0 < co < c and, consequently, H is bounded from above on (0,^]. Lemma 8.6 shows
that A must be nonpositive.
We now choose 0 < ? < co < . We can select A so that J( ; ,0, A) defined at the
beginning of this section satisfies J(c; 0, 0, A) = H(c). Integrating (8.9), we obtain

J(co; 0,0,A) = Exo[ ^ce -'U(c,) dt + e c H(c)

+ee l{S<s^}J(c; 0,0,A)J. (8.17)

According to Lemma 8.7,

A^~ ~U(O) = U(
= limJ(c;
limJ(c; 0, 0,) 0,
, 0)
cJO CIO

so J(c; 0,0,A) is bounded above for c in a neighborhood of the origin. Also,

U(c,) < U(c) for t E [0, S A Sj]. Since c, can never reach zero, S. A S -> Sc as cJO,
and Fatou's lemma can be applied to (8.17) to obtain

J(c0;0,A0,) < Exo fe AU(ct)dt + e -sH(c)

=H(co) = J(co; O,A,0), 0 < co < c.

This inequality will fail for co close to the origin if A < 0. We conclude that A = 0. I

9. Solving the Bellman equation. Part II. Given a > 0, B < 0, the function X( ; a,
B) given by (6.9) is strictly increasing and maps [a, oo) onto [X(a; a, B), oo). We regard
the inverse function C(; a,B) as a feedback policy for choosing consumption as a
function of wealth. In ??7 and 8, we saw that when this consumption policy and the
portfolio policy described by (7.1) are employed and initial wealth is x > 0, then for
the proper choice of A, the expected return is J(C(x; a,B); a,A), where J(. ;a,A) is
defined by (8.11). In the next theorem we see that a fortuitous value for A would be
A =+ B/p+ .

THEOREM9.1. For a > 0, B < 0, the function

V(x;a,B) J(C(x;a,B);a,- B) x > X(a;a,B) (9.1)

P+

satisfies the Bellman equation (5.4).

PROOF. From (6.10) and (8.11) we have

V'(x;a,B)= J'(C(x;a,B);a, - B) C(x;a,B)

J'(C(x; a,); a, B)

X'(C(x;a,b);a,B)
= U'(C(x;a, B )), x > X(a;a,B), (9.2)
 EXPLICITSOLUTION OF GENERAL CONSUMPTION/INVESTMENT PROBLEM 279

whereprimedenotes differentiationwith respectto the first variable.Consequently,

V"(x;a,B) = U"(C(x;a,B))C'(x;a,B ), x > X(a; a, B), (9.3)

so V( ;a,B) is strictlyconcave, and (5.4) takes the form (6.4). In light of (9.2) (cf.
(6.5)),it sufficesto verify
(V'(x;a,B))2
/3V(x;a,B)= -y + (rx - C(x;a, B ))V'(x;a,B) + U(C(x;a,B)),
(x;aa, B)
V"(x;
x > X(a;a,B),
and this is equivalentto

P+ (U"(c))2X(C; a, B)
+ (rX(c;a,B) +
( ),(pU'(c (cy B)) c)U'(c) U(c),
c > a.
The verificationof this last equationis a straightforward
exerciseinvolving(6.8), (6.9),
(8.7) and (8.11). I

10. Solution when U'(O)= oo, U(0) is finite, P = U(0)/t.

THEOREM10.1. Assume U'(0)= oo, U(O) is finite, and P = U(O)/fl. Set a = 0,
B = 0. An optimalpolicy is given by (7.1), and the value function is given by (9.1).
PROOF. From (6.9), (8.15) and (8.16), we have

X(0;0,0) 4 limX(c;
c40
0,0) = 0,
so V(x; 0, 0) is defined by (9.1) for x > 0. Accordingto Lemma8.7,

V(0; 0, 0) Alim V(x; 0, 0) = limJ(c; 0, 0) = P,

x4 cJ,0

and since V is strictlyincreasing(cf. (9.2)), we have V(x) > P for x > 0. Theorem4.1
impliesthat V majorizesthe value function V*.
Since U is boundedbelow, H definedby (8.1) with = 0 and v = P is well defined,
and since
H(C(x; 0, 0))= V,()(X)< *(x) V(x; 0, 0) < oo, x > 0,

H is finite. Theorem 8.8 implies H(C(x;0,0))= J(C(x;0,0);0,0) = V(x;0,0), and

the theoremfollows. I
REMARK 10.2. According to Remark 7.1, the optimal policy of Theorem 10.1
avoids bankruptcyalmost surely.
COROLLARY10.3. Under the assumptionsof ?2, V*(x) is finite for every x > 0.
PROOF. Given x0 > 0, the policy c = rxo, T =- 0 leads to expected return
U(rxo)/13
> - oo, which provides a lower bound on V*(xo).
For the upperbound, let U and P be a utility functionand paymentsatisfyingthe
assumptionsof ?2, and constructanothersuch utilityfunction U and paymentP such
that U(c) > U(c) for all c > 0, U(c) = U(c) for all sufficiently large c, P > P, and U
and P satisfythe assumptionsof Theorem10.1.The value functioncorrespondingto U
is finite and majorizesthe value functioncorrespondingto U. I
280 I. KARATZAS,J. P. LEHOCZKY,S. P. SETHI & S. E. SHREVE

REMARK10.4. Since H(c) defined by (8.1) is the expected return associated with a
feasible policy, H is well defined and H(c) < oo for all c > 0. If, in addition, U is
bounded below or c, is bounded away from zero, then H(c) > - oo, and the condition
that H is well defined and finite appearing in Theorems 8.5 and 8.8 is satisfied.
REMARK10.5. Given any utility function U satisfying the assumptions of ?2, we
can construct another such utility function U, as in the proof of Corollary 10.3, such
that U(c) > U + (c) for all c > 0. The value function corresponding to U is finite, so
for any feasible policy,
-
Ex e U(ct)dt < o.

From this we conclude that (4.3) holds, so, according to Remark 4.2, the conclusion of
Theorem 4.1 is valid even if U(0) = - oo. From (4.3) we also conclude that the integral
and expectation in (1.6) are well defined for any pair c(-) and 7r(-) which is admissible
at x. This includes the important special case U(c)= logc.
REMARK 10.6. The choice of a = 0 in Theorem 10.1 comes from the following
observation: with F(c) defined by (11.1), a is the unique, nonnegative solution of the
equation F(c) + p+ P = 0. Indeed, limc,ocU'(c) = 0 (0 < cU'(c) < U(c) - U(O), for
c > 0) under the assumptions of this section, so we have from (11.1), (8.16):

F(O) lim F(c) = P- U(O)= -p P.

From (8.15), (8.16), we have

X(c; 0,B)- B(U'(c))x+

c 1 (U'(c))p+
oc dO
r y(X+ -X_)U'(c) X+ Jo +
(U'(w))

+ (U (C)) r(U d
X (U'(9 ))
converges to zero as cO,. Since we require limc,oX(c; 0, B) = O, this mandates B = 0.

11. Solution when U(0)= -oo. For this and the following section we need a
technical lemma.
LEMMA 11.1. Assume P > U(O)/fi, U'(O) and U(O) are not necessarilyfinite, and
define

F = -
(c) U'),)-+P U'(c)),
(U,(C c c >0. (11.1)
Y, _ p_ Jc (U'())- f r

Then F is strictly decreasing, and limc,F(c) + p + P < 0. In particular, the equation

F(c)= -p+P (11.2)

has at most one positive solution, and it has a solution if and only if F(0) + p + P > 0,
where F(O) lim,o F(c). If U'(O) = oo, then F(O) + p+ P > 0.
PROOF. Using (6.8) and (8.7), we can derive

c
F'(c) = U"(c) -
(c(A)c
'
dO +-- + _ 0,
( 2k-c (U'(c d c >0
 EXPLICITSOLUTION OF GENERAL CONSUMPTION/INVESTMENT PROBLEM 281

which is negative. Because of (2.8), there is a c > 0 for which iPP< U(c). Concavity of
U implies

U(c)- U(c) + c^U'(c) > cU'(c), c> ,

so

< - - U()) + ()
( )(U(c)

r r

U(c)
=_lim = oo),
yX_(U(c-
.
())+ r
and
))
/+ C
lU-CA_< (U((A AU
limp+ P-F(c)+ P 0.
p+
C---> 00

> lim
F() - U'(c) r dO

imply
F(0)
= -p+ U(0)//3 >-+ P. <
< -li -
f [U'(c)dO-P+ U(c))
When
U (0)= -o cto ( yA_p_
s ) () y_ p_
/3 , then (11.2) has a unique

=limI pso x+
c = a > . Choose
solution --a,B) = o, i.e.,
X(a;
r - A) (U = 0. (11.3
(I(a)) ( +( J(

If U(O)< oo
U(0) <conclude
x but thatC
U'(0)= ~~~+10.4)
U'(0)= oo, then
to (8.16) cU'(c) < U(c)
inequality cU'(c)<
(8.16) and the inequality U(0)
U(c)- - U(O)
imply F(0) = - p + U(0)/fl > - + P. I

When U(O)= - oo, we also have U'(0)= oo. If P is finite, then (11.2) has a unique
Vc(
solution c = a > 0. Choose J(C(x;a,B1)a,A),
B so =X(a; a,B) = 0, i.e.,? x > 0,

(U'(a))X- d
B(U'(a))X'+ 4a r = 0. (11.3)
3,h,(X+ ,(U ,(0))x-
-X_)
It is apparent from (11.3) that B < 0. We may now implement the policy c(), g(.)
described by (7.1) with = 00. In (8.1) we set U = P and use Theorem 8.5 (and Remark
10.4) to conclude that

Vc(. -)(x) = J(C(x; a, B ); a, A ), x > 0,

where A is given by (8.12). Multiplying (8.12) by p+/ U'(a), multiplying (11.3) by +,
282 I. KARATZAS,J. P. LEHOCZKY,S. P. SETHI & S. E. SHREVE

and subtractingthe resultingequations,we obtain

(X+ B - p+ A)(U'(a)) U(a))))

+ U)
U X(r +
r
(a) /~

+
1 ^ -^ d =o.
A Jf0
(A+-X_) (Up'())
Using (8.7) and the equationF(a) + p+ P = 0, we can reduce this to A = X+B/p .
THEOREM 11.2. Assume U(O) = - oo and P is finite. Let c = a be the uniquepositive
solution to (11.2), and let B be given by (11.3). An optimalpolicy is given by (7.1), and the
valuefunctionis givenby (9.1).
PROOF.The precedingargumentshows that V*(x) majorizesV(x; a,B) given by
(9.1). From Theorems9.1, 4.1, and Remark10.5,we have the reverseinequality. I
REMARK 11.3. Recall from Remark7.1 that the optimalpolicy leads to bankruptcy
almost surely if ,B > r + y, and it leads to bankruptcywith positive probabilityless
than one if /f < r + y.
11.4. Assume U(O) = - oo and P = U(O)/
THEOREM = - oo. Set a = B = O. An
optimalpolicyis givenby (7.1), the valuefunctionis givenby (9.1), and, underthe optimal
policy,bankruptcy does not occur.
PROOF.Let {Pn}i' be a sequence of negative numbers with P,, - oo. Let
(an,})' be the sequence of positive numbers satisfying F(a) + p + P, = 0. Because F
is strictlydecreasingand real-valuedon (0, oo), we have a,O. Let Bnsatisfy(11.3)with
a replacedby a,, i.e.,
oo
A an- (U'(an))X- dO = O. (11.4)
U'(an))X-+ rW
B"( +^
r - ))0
X (x\+A-_)X_-=0.f((Un
yy_(X (11.4)
n

We show that
dO
lim( U(c))X- 0, (11.5)
cIO0 (JU(o))^-
from which we concludethat
lim B, (U'(a))X+ = O, (11.6)

and, in particular,limMn,Bn= 0. To prove(11.5),choose e > 0 and write

0<lim (U'(c))A-f dO

U'(c) dO+ lim (U'(c))X r0 dO

< _m r
cIoJ ue() IO (veo)^-
<lim 'dO< e.

Letting4e0, we conclude(11.5).
Accordingto Theorem11.2, the function

V(x;an,B,n) A C(x;a Bn;a, +B x > 0, (11.7)

is the value function for the model with payment Pn. Choose x > 0 and let c =
 EXPLICITSOLUTION OF GENERAL CONSUMPTION/INVESTMENT PROBLEM 283

C(x; a, B). Then

X(c, ;0, ) - x = X(c, ;0,O) - X(cn; a, B,)

n
-= BU(UC(c))X
an dO
(11.8)
?UCnB X+(X+-X_ ) fo U(O +
But cn > a,, 0 < U'(c,) < U'(an), and so (11.6) implies lim,,.B,(U'(cn))x = 0. Also,

lim a. d 0,
n-oo (Ut(c,))x+ < limfJnd=
(oU(O))X+ J n-+o

so, letting n - oo in (11.8), we conclude

lim ;0,0) = x. (11.9)
n---) oo X(c,

Togetherwith (6.11), (11.9) impliesthat the sequence( cn,)} is bounded,so it has an

accumulationpoint. Any accumulationpoint c must satisfyX(c; 0,0) = x, so the only
accumulationpoint, and thereforethe limit, is C(x; 0,0), i.e.,
lim C(x; a,n B )= C(x; 0, 0). (11.10)
We may now use (11.10)in (11.7) to conclude
lim V(x; an, B, ) = J(C(x; O,0); 0, 0), x > 0.
Let V* be the value functionfor the model with paymentP = -oo. Since Pn > P,
we have V*() (;an, B,) for every n, so
V(
V*(x) < J(C(x;0,0);0,0). (11.11)
It remainsonly to prove the reverseinequality.
Fix xo > 0, let co = C(xo;0,0), and define x") = X(co;an,Bn). Let {c,, t > 0) be
given by (7.5), and set
= inf t > 0 : c, =
S, an) = inf( t > 0 :Yt = Ut(an)).
The process {(c :0 < t < Sn} is optimal for the model with payment Pn and initial
wealthxn) (take C(.) in (7.1) to be C( ; a,, Bn)).In otherwords,

-+ Bn = V(Xo);
0 an, n --+ Bnn )
J(co;an,o
\^+ P+

= Exo[ fSn e
tU(c) dt + Pne -t]". (11.12)

Now {ct, t > 0) is feasiblein the model with paymentP and initial wealthxo. (Take
C(.) in (7.1) to be C(.;0,0).) Under this policy and {(r, t > 0) given by (7.1),
bankruptcyis avoided almost surely,so
Vc(.),(.)(xo) = e
Exo[fo U(ct)dt].

In light of Remark 10.5, we know that (4.3) holds (with To= oo), so we may apply
Fatou's lemma to (11.12) to obtain

J(C(xo;0,0); 0,0) < E,o[o e-gtU(ct)dt]

=
V*(o),f)(XO)
< V*(xo), X > 0 I
284 I. KARATZAS,J. P. LEHOCZKY,S. P. SETHI & S. E. SHREVE

12. Solution in all other cases where the consumption constraint is inactive. The
optimal consumption processes found in ??10 and 11 were of the form ct = C(xt; a,
B), where {xt, t > 0) is the optimal wealth process, a > 0, and B < 0. If U'(O) = oo
and a = B = 0, or if a > 0 and B satisfies (11.3), then X(a;a,B)= O, so ct = C(x; a,
B) > 0 whenever x, > 0. We say then that the consumption constraint ct > 0 is
inactive. In addition to the cases analyzed in ??10 and 11, there are two other
situations in which the consumption constraint is inactive. The proofs of the following
theorems are the same as the proof of Theorem 11.2.

THEOREM 12.1. Assume U'() = oo, U(O) is finite, and P > U(0)/P/. Let c = a be
the uniquepositive solution to (11.2), and let B be given by (11.3). An optimalpolicy is
given by (7.1), and the valuefunction is given by (9.1).
If U(0) and U'(O) are both finite, we define

1 ( U )) I))
((U'(O))"- o0
P*=l U(O)- (U( d
.d (12.1)

Note that, in this case, (11.2) has a positive solution if and only if P > P*. If P = P*,
(11.2) has solution c = 0, and if U(O)/jf < P < P*, (11.2) has no nonnegative solution.
This last case is the subject of ?13.

THEOREM 12.2. Assume that U(O)and U'(O)are bothfinite, and P > P*. If P > P*,
let a be the uniquepositive solution to (11.2); if P = P*, let a = 0; in either case, let B be
given by (11.3). An optimal policy is given by (7.1), and the value function is given by
(9.1).
REMARK 12.3. Since U'(a) is finite in Theorems 12.1 and 12.2, Remark 7.1 (with
/ = 0) shows that there is a positive probability (equal to one if and only if P > r + y)
of bankruptcy.

13. Solution when the consumptionconstraint is active. The only remaining case is
that of finite U(O) and U'(O) and

U(O) P< PP*, (13.1)

where P* is defined in (12.1). In this section we assume these conditions, and we shall
establish the existence of x > 0 and B < 0 such that the optimal consumption is given
by
0, 0<x<X,
x , (13.2)
C(x;O,B),

where C(x; 0, B) = 0 and C(x; 0, B) > 0 for x > x. When P = U(O)/13, we will have
B = 0; when U(O)// < P < P*, it turns out that
)yP
yB - ( U(O)) (13.3)

where y is defined by

yP = -pX_(p-
ft
'
U(0)) J (u(e))
'! Jo (U^e)^-
(13.4)
v
 EXPLICITSOLUTION OF GENERAL CONSUMPTION/INVESTMENT PROBLEM 285

We will have

x (=B(U'+ - (v'(o))X-) o dd
(U())x (13.5)

(cf. (4.2) of [5]). Note that P < P* implies in (13.4) that yP- < (U'(O))P-, so
y> U'(0). (13.6)
When P = U(0)//8, we adopt the convention y = oo, so (13.6) still holds. When B = 0,
it is clear that x > 0; when U(O)/1f < P < P*, we have from (13.3) and (13.6)

B( U'(O))+ >+- ( P* U(0))

(U'(O))X- oo d
yX (X+ -X )o(u.(O))
so again x > 0. Finally, we note that with X(c; 0, B) defined by (6.9), x = X(0, 0, B), so
C(x; 0, B) = 0.
In the preceding sections, we chose consumption as an intermediate variable, and
we related wealth to consumption via X(c;a,B) and value to consumption via
J(c; a, + B/p+). This choice of intermediate variable is now inappropriate, because
the mapping from wealth to consumption described by (13.2) is not invertible. We will
therefore allow y = dV*(x)/dx to play the role of intermediate variable. We will
discover that V* is strictly concave, so the mapping from x to dV*(x)/dx is invertible.
Moreover, when x > x, we will have wealth and optimal consumption related by
dV*(x)/dx = U'(c) (cf. (6.5)), soy = U'(c) when x > x.
Recall the function I: (0, U'(0)] -[0, oo) which is the inverse of U'. We extend I by
setting I = 0 on [U'(0), oo). If V is C2 and strictly concave, the Bellman equation (5.4)
takes the form (6.4), which can be written as

fV(x)y= V"()) + [rx - I(V'(x))] V'(x) + U(I(V'(x))), x >. (13.7)

By analogy with (6.9) and (8.11) with a = 0 and c = I(y), we define for B < 0,
A ?0,

X(y;B)= By+ + r1 I(y)

(X ) f yX r (Y) dO + y f dO
y(X+-_) Jo
A (U.(o))+ A '
(Y)(U.())
y>0, (13.8)

(y;A) = AyP + U(I(y))

y 1 0y () dO yP { dO
y(A+ -A_) P+ Jo (U.(o))A+ +' - J(
)(ul(O))A-

y>O . (13.9)
For c > 0, we have c = I(U'(c)) and

ct( U'(c); B) = X(c; O,B ), ,( U'(c);A) = J(c; 0,A), c > 0. (13.10)

286 286
I~~~.
KARATZAS,J. P. LEHOCZKY,S. P. SETHI & S. E. SHREVE

Since

d9l"(y; B) =BX,yA?l y(A-I) yA+- 'Jo d

+yA--'IO dO y>0

which is negative,,',&(; B) is strictlydecreasing.From (13.10) and (6.11) we have

From (13.8) we have

lrn XIi(y;B) limrB/K+
y-+oo
y-.*oo

Consequently,if B = 0, X( ;B) maps(0, co) onto (0, oo) and has an inverse 3(.;B):
(0, oo) ono 0 oo). If B < 0, (13.3), (13.4) and (13.6) can be used to show that y
defined by (13.4) satisfiesI(Y) = 0 and X,(y; B) = 0. We considerX ( ; B) restricted
to (0, Y],whichhas range[0, oo) and inverse931(. ; B) :.[0, co~) onto) (0, yj.
REMARK. Conditions(13.3), (13.4) are equivalentto the requirements ? (9, B) = 0
and / (Y;X+ B/lp +) = P. Condition(13.5) amountsto X = (U()B.
LEMMA 13.1. Assume (13.1), (13.3) and (13.4) hold. Define

The function V is strictly increasing, strictly concave, satisfies the Bellman equation
(13.7), and
lim V(x) =P.(1.)

PROOF. From (13.8) and (13.9) we have

V'(x)
=x;B ;A+
V'(X)
T(5,(x; );9B (x; B) >0, x >0, (13.12)

V"(x) = 3W'(x7Y(;B=,B <0, x >0. (13.13)

Withy = ?*"(x;B), (13.7) becomes

13f(Y; > B) = yy2X9i'(y; B) + (rX9l(y;B) - I(y))y + U(I(y)), 0 < y <,

and the verificationof this equationis straightforward.U

We now suppressthe parameterB < 0 and take V to be as definedin Lemma 13.1.
Let xo > 0 be a given initialwealth,and let (cf. (6.3), (6.5))

(a -r)V(xt) 0 < t < To,

= I(V'(x,)), ITt (13.14)
 EXPLICITSOLUTION OF GENERAL CONSUMPTION/INVESTMENT PROBLEM 287

where {x,, 0 < t < To) is defined by (6.2) and To is defined by (1.5). Sety, = V'(x,) so
Yo = V'(x).
LEMMA13.2. The process (y,, 0 < t < To) satisfies (cf. (7.4), (7.5))

yt =y0exp[-(r- + y)t- r w,] (13.15)

and
To= inf{t > 0 :y =}. (13.16)

PROOF. From (13.12) and (13.13) we havey, = *(xt),

= '( =
V"(xt) (x)

Therefore, with B suppressed, we have

a
ct = I(y,), Tt = rYt,(y,),

dx, = -2Yyt'(yt) dt + (rx - I(yt)) dt - - r y ,'(y,) dw.

Direct calculation shows

yy2'T"(y) = (r - / - 2y)y'T(y) + r((y) - I(y), y > 0.

By analogy with (7.2), we have

' %'
"(xt)( t'(y,))2+ (x,)t "(yt) = O.
According to It6's rule

dyt = 9'(xt) dx, + yy2 (T(yt))2 "(x) dt

-
=-2yytdt + (rx, - I(yt))Y'(x,) dt - a rydwt

+ yyt2 (" (yt))2* "(Xt) dt

-(r - 3)ytdt- a- r ytdw,.

The solution to this stochastic differential equation is given by (13.15). Finally,

To Ainf{t > 0:xt,= (yt)= 0}

=inf{t>0 :y = (0)=Y}. !
REMARK 13.3. If P = U(0)/,8, then B = 0 and, by convention, y = oo. Also, (13.16)
is valid and says that bankruptcy is avoided almost surely. In this case, xt and y, are
defined for t > 0. If P > U(0)//8, so B < 0, then y is finite and yo < y. Bankruptcy
will occur with positive probability, and this probability is equal to one if and only if
,f > r + y. When To < oo, x, and Yt are defined for 0 < t < To. In particular, 7r(.) is
admissible as defined in ?3.
We now evaluate the performance of the policy (13.14). Define

t(yo) = Exo[Jo ReTU(I(yt))dt + Pe 3T], O <yo <y. (13.17)

Corollary 10.3 and the fact that U is bounded below by U(0) imply that , is well
defined and finite. Recall (13.6). If 0 < yo < U'(0), we may take x to be given by (13.5)
288 I. KARATZAS,J. P. LEHOCZKY,S. P. SETHI & S. E. SHREVE

so x = X(U'(0);B) and T = inf{t > 0:y, = U'(0)}, and we may use the strong
Markov property to obtain

At(Yo) = Exo ,fe -'U(I(y,))dt +

Y( U'(O))e -,
0 <yo < U'(O). (13.18)

Comparison with (8.2) shows that jK(yo) = H(I(yo)), 0 < Yo < U'(0), so, by Theorem
8.5,
a(yo) = J(I(y0); 0,A) = (Yo; A), 0 < yo < U'(0), (13.19)
for proper choice of A.
According to Theorem (13.16) of Dynkin [2], X given by (13.17) is C2 on (0, Y),
satisfies (8.3) on this interval, and, if y if finite,

limi"(y) = P. (13.20)
The general solution to (8.3) is f(y;A)+ Ay--, and (13.19) shows we should set
A = 0 to obtain J. If y is finite, A can be determined from (13.20). Indeed, we have
- dO
Ay =
AP+- - - 1 U()+
U( 0) + _( - )Jof ( u(o))A

and from (13.3), (13.4) and (8.7), we have

p+AyP+= p+( l- P- UO)

x+- (P - U(O))
-X7
-y(X+-
=X+By+,

so

A= B.
P+
In other words, when U(0)//i < P < P*, x0 > 0, and the policy described by (13.14) is
employed, the expected return is

Vc .),( .)(XO) = ((xo;B B,)B xO> 0. (13.21)

THEOREM13.4. Assume U(O) and U'(O) are finite and U(0)//3 < P < P*. Let B < 0
be defined by (13.3) and (13.4). The valuefunction V is given by (13.21), and the policy
defined by (13.14) is optimal.
PROOF. According to Theorem 4.1 and Lemma 13.1, the function on the right-
hand side of (13.21) majorizes V*. Since Vc( ),,(.) < V*, the theorem follows. I
It remains to study the model with P = U(0)/1f. Our candidate optimal policy
avoids bankruptcy in this case, so the boundary condition (13.20) cannot be invoked
to determine A. We will consider this problem as the limit of a sequence of problems
in which the nth problem has payment P, E (U(0)/,, P*) and Pn, U(O)//f. The value
function in the nth problem is f('(x; B,); X Bn/p+ ), where B, is given by (13.3)
and (13.4) with P replaced by P,. One can verify that B,,O.
Let V* be the value function in the problem with payment P. The problem with
 EXPLICITSOLUTION OF GENERAL CONSUMPTION/INVESTMENT PROBLEM 289

payment P,, is more favorable, but the difference between the expected returns in the
two problems under the same policy cannot exceed P,, - U(0)//3. Consequently,
X
*x; ,,
6f
B P"- ( < V*(x) < fr j'(x; Bn ); -
B

x ) 0. (13.22)

LEMMA 13.5. With P,4~U(O)/13 and with Bn,IOas described above, we have

lim ~f(Y(x; Bn);?i B) f(*?(x; 0);0), x >0O.

PROOF. We first show

lim *3/(x; Bn, = li M lX; 0), x > 0. (13.23)
nl->,oo n o

Let x > 0 be fixed andy,, = 3K(x; Bn,), So

X(Yn,+I , Bn, 1) = x = T9(y,, ;Bn,)

= Tf(y, ;0) + BYX

? TK(y, ;0) + Bn+ Y

=
X9(yn, ;Bn + )
Since ?X(.Bn + ) is decreasing, we have yn, < y,, + . We also have x =
y; n
< ?X1(yn,; 0), so y,, < 3 (x; 0) for every n. Consequently, yn4y, where y < 00, and

,Ty)= lim [,T(y, ;0) + Bjy2+]

- lim T6(y ; Bn,)= X.

Equation (13.23) follows. From (13.23) we obtain

lim f( 3((x;B,,)>0) +

THEOREM13.6. Assume U(0) and U'(0) are finite and P = U(0)//3. Then the value
function is V(x) A4 f(*3"(x; 0); 0), x > 0, and the policy defined by (13.14) is optimal.
PROOF. Let n -:> oo in (13.22) and use Lemma 13.5. U
REMARK 13.7. Since *3~(5;B) = U'(0), (13.12) implies that when U(0)/13 < P
K 0*, we have

d V*(x) > U'(0), < <

dx
d
V*(x) < U'(0), x > iT.
From (13.10) and (13.12) we see that the policy defined by (13.14) agrees with that of
(13.2).
290 I. KARATZAS,J. P. LEHOCZKY,S. P. SETHI & S. E. SHREVE

13.8. We have
REMARK
00
(X;O) =(U(O))(X/)l/ ' 0/v<x < - (U(O))A dO
/A/_ yX (X+ -X)Jo (U )
so when P = U(0)/i,

V*(x) = /((x; O);O) = + U(O), 0< x < .

P- U'(O)(x/x)'
In particular,
lim d V*(x) = oo,
x4odx
which is the value assignedtoy for this case. In the other case U(0)//f < P < P*, we
have from (13.12) dV*(O)/dx = y.

14. Solutionsfor HARA utility functions. The utility functions on (0, oo) which
are said to have HyperbolicAbsoluteRisk Aversionare the functionsof the form

U(c)= (c + r) , 0<8<1, 1 >0, (14.1)

U(c)= -(c+ 7)8, <
8<0, >0 (14.2)
and
U(c)= log(c + q7), 7 > 0. (14.3)
The last function is referredto as the HARA function with parametersiq > 0 and
=0.
When 8 < 0, U satisfies the growth condition (2.6) because A_ < - 1. When
0 < 8 < 1, (2.6) is equivalent to

>rS+ 1-8' (14.4)

a conditionweakerthan > 8 [r + y(2 - 8)/(1 - 8)], whichwas imposedby Merton
[7, condition(41)] to guaranteethe finitenessof the value function.
In this section we specializethe resultsof this paper to the HARA functionswith
= 0. We also show that for any 8 < 1 and any q > 0, (14.4) is a necessarycondition
for a finite value function.The fact that (14.4) is sufficientfor a finite value function
follows from Corollary10.3.
Using (6.8) and the fact that + +X_ = (r - f - y)/y, one can show that when
= 0, (6.9) yields for all 8 < 1,
(1-8)c
X(c; O, )= r-y/(- c > 0. (14.5)

When8 < 0, the denominatorin (14.5) is positive,and when 8 > 0, the denominatoris
positive under condition (14.4). From (8.7), (8.11) and the fact that p+ +p_ =(r-
fl + y)/y one can show that whenr = 0, we have
(1- 6)c8
J(c;0,0)= ft-rS8/(l-8)' c>0, 0<8<1,

(1 - 6)c6
J(c;0,0)= - r-y8)' c>0, 8<0,

r- + c >0, 8=0.
J(c;0,0)= logc+
ftP~, 2
 EXPLICITSOLUTION OF GENERAL CONSUMPTION/INVESTMENT PROBLEM 291

We can invert X(c; 0,0) in (14.5) to obtain

,0)= 1^ r- 1-r
C(x; ( )x

and then set VJ(x) = J(C(x; O, ); 0, 0), so

)
Vs(x)~ (x)= - r - y/(l - 6) xs' >
xx>-0, < ,
, 0<8<1, (14.)
(14.6)
1-r'-?/'1-

r, - - yS/(l -)) X8, x >O <0, (14.7)

Vs(X) r-log + r- + x >0, =0. (14.8)

Then from (6.3) we obtain a constant optimal investment policy:

a t > 0, < 1.
,= -r2
(1 -8)a2
THEOREM 14.1. Consider the consumption/investment problem with U given by
(14.1), (14.2) or (14.3) with q1= 0 and with P = 0 if 0 < 8 < 1, or P = - oo if 8 < O. If
0 < 8 < 1, assume (14.4). Then the value function is given by (14.6), (14.7) or (14.8),
respectively.
PROOF. This theorem is a direct consequence of Theorem 10.1 and the fact that
when 0 < 6 < 1, (14.4) implies (2.6). I
COROLLARY 14.2. If 0 < 8 < 1, 0 and (14.4) is violated, then V*(x) = oo for all
> 0,
x > 0 and all P.
PROOF. Since the HARA functions are nondecreasing in q7for fixed c, we may
assume q = 0. Without loss of generality, we may take P = U(O)/8P. Let the value
function V* be indexed by the discount factor /3 > 0. By definition, this function is
nonincreasing in P/. Let /,o = r8 + yS/(l - 8) and assume (14.4) is violated when
f/ = ,/, i.e., assume i0o > fl. From Theorem 14.1 we have for /f > Po,

(f,-r8- 3,/(1 -3 ) X8= V (x) < Vo(x) < V (x), x>, so

= lim (x)l
V < V(x), x > 0.

15. Dependence of the optimal policy on the prices. In ?3, we specified that
consumption and investment processes for the model with N risky assets must be
adapted to the filtration of the driving N-dimensional Wiener process {w,, YS, t > 0),
a filtration which may be larger than the filtration {(-, } generated by this Wiener
process. We point out here that the optimal consumption and investment policies we
have obtained are, in fact, adapted to {-,o }. Because the matrix D in (1.2) is
nonsingular, {,-? } is also the filtration generated by the vector of asset prices {P(t),
t > 0}. Thus, our optimal policy is driven by the asset prices.
To see that the optimal {c,, t > 0} and {r,, t > 0} are adapted to {Y?o}, we note
that for the reduced model with one risky asset, the optimal {c,, t > 0} and {7r,, t > 0)
are given by (13.14), where xt can be determined from the equation

d V*(x,),X
dt- X X t ' (151)
(.1
292 I. KARATZAS,J. P. LEHOCZKY,S. P. SETHI & S. E. SHREVE

and y,; t > 0) is given by (13.15).(Whenthe consumptionconstraintis not active,we

have U'(c,)= dV*(x,)/dx and (13.14), (13.15) follow from (7.1), (7.5).) Because V* is
strictlyconcave, (15.1) can be solved for x, as a function of y,. Since (y,, t > 0) is
,?-adapted, so are {c,, t > 0) and {r,, t > 0). Equation (5.8) with 4 = 7, shows that
when we revertto the originalmodel with N risky assets, we obtain an optimal ({V,,
t > 0) which is also adapted to f ,?)}.

16. The model with nonconstant a, I and r. The model addressed by this paper
has constant coefficients.Let V( ; y, r) denote the value function for such a model,
where y is given by (2.3). Now suppose that a, D and r are replaced by random
processes a,, D, and rt adapted to the right-continuous family of o-fields {(ft, t > 0).
Such a model was proposedby Merton[7, p. 873] as a startingpoint for equilibrium
T 1
analysis.We define S, = DD and assume g- exists for all t > 0. Set
t= (a,- rtl)/, -rtl)
One can show that if 0 < y' < y, < y" and r' < r, < r" holds for ail t > 0 a.s., then
V(x; ',r') < V(x) < V(x; y",r"), x > 0,
where V denotesthe value functionfor the model governedby a,, Dt and r,. If U is of
the HARA class with r = 0 (cf. (14.1)-(14.3)) and the payment P is U(0)/fl, then
V(x; y', r') is a lower bound for V(x) provided only that

i I ( , -y )
+ r-or' ds> 0.

In the interestof brevity,we have omittedthe derivationof these results.

17. Tabulatedresults. The following tables summarizethe results of this paper

and [5]. In both tables, q denotes the probabilityof bankruptcyunder the optimal
policy, y is given by (2.3), and P* is given by (2.9). Table 1 recordsthe findingsof this
paperunderall combinationsof assumptionsabout U'(0) and P. Appropriatesections
and theoremsof this paperare cited. Table 2 recordsthe findingsof [5] for the model
of this paperwhena = rl. In both tables,referencesare made to the summary(la-4b)
of ?2.
TABLE I
a 7 rl
1 (O 1 1
P< U(O) U(O) < P < P* P* < P < U(o) AU() < P
P P P P
c, > O, c, > a > O,
q=O. O<q< if /< r+y, o
8 (This is the q= I if f/ > r + .
II case studied
? by Merton I
L [8].) I x
?2, ?10,
Theorem 11.4, (la) Theorems 11.2, 12.1, Remark 11.3, (2a)
c, = 0 if c, = if x, E(0, ], c, > a > 0,
8 x E (0, x], c,>0 if x,(x,oo). O<q< 1 if f/<r+y, r
V c, > 0 if (x = when P = P*.) q= I if 8/ > r + y. ' E
o x, (X,oo), 0<q<l if P/<r+y, "
L 7= O. q= I if f > r + y. Theorem 12.2, V Z
?13. (3a) ?12, ?13, (3a,4a) Remark 12.3, (3a, 4a)
 EXPLICITSOLUTION OF GENERAL CONSUMPTION/INVESTMENT PROBLEM 293

TABLE 2
a = rl
I
P <- U(O) U(O) < P < P* p*< P < U(oo) U(oo) < P
/3 /3 p P
ct > a >0,
't = 0,
A c, >0,
cQ. q == 1. (2h)
(2b)

c, irrelevant, 7, = oo if x, E [0, xoo),

8 Vrt=0, q = (x - x)/xo if x e [0, x);
V
0 C33 c, > a >0, 7 =0 if x,E[x ,oo),
:
q=O if x [0,x,o). (2b)

q =0. c, = rxo , r,tarbitrary if x, E [0, x,o),

3r. x, - x
11
x <q< I if xE[0, xo);
C.Q.

(lb) Ct= rx, , Tr = 0 if x, E [xoo, oo),

q=O if xe[xOo,oo).
c, = rx,,
z3
C.)
rt = 0, t)
A0.
I..
EL
11 q =0. 14
Q0.. 0
cl
;3n
(3b, 4b) (2b, 3b, 4b) 0
.1
C.Il
1~4
0
0.
u
c, > 0, c, > a > 0, C)
ol
4)
0.
r't= 0, g/ = 0,
a
0
;;
. q= I if q=1. ct
A
Q 1 (A
P = U(O). 0
=t
8
V (3b, 4b) (3b, 4b)
o

c, irrelevant if x, E [0, x ), c, irrelevant if x, E[0, xo0),

c,= 0 if c,=O if xtE[xo,x],

x, E (0, X],
c, >0 if c, > if x,E (x, o), c, > a > 0 if x,Et [xo, oo);
xt, (X, oo), (xoo = x if P= P*);

oo if xt [0, x), oo if xt [0,xo),

V %t =0, Tt
0C. 0 if xt E [xo,oo); t0O if xt E[x, oo);

x-x- x
if xE(O,x0o), if x E(0, xo),
q=0. q_0 XOO q= x,o0
1?
if x E[xO, oo), 0 if x E[xo, oo),

(3b, 4b) (3b, 4b)

(3b,4b) (3b, 4b)
(3b,4b)
294 I. KARATZAS,J. P. LEHOCZKY,S. P. SETHI & S. E. SHREVE

References
[1] Black, F. and Scholes, M. J. (1973). The Pricing of Options and Corporate Liabilities. J. Political
Econom. 81 637-654.
[2] Dynkin, E. B. (1965). Markov Processes. II. Academic Press, New York.
[3] Harrison,J. M. and Kreps,D. M. (1979).Martingalesand Arbitragein MultiperiodSecurityMarkets.
J. Econom. Theory 20 381-408.
[4] and Pliska, S. R. (1981). Martingales, Stochastic Integrals, and Continuous Trading. Stochastic
Process. Appl. 11 215-260.
[5] Lehoczky,J., Sethi,S. and Shreve,S. (1983).OptimalConsumptionand InvestmentPoliciesAllowing
ConsumptionConstraintsand Bankruptcy.Math.Oper.Res. 8 613-636.
[6] , and . (1985).A MartingaleFormulationfor OptimalConsumption/Investment
Decision Making. In Optimal Control Theory and Economic Analysis 2, G. Feichtinger (ed.), North-
Holland,Amsterdam,135-153.
[7] Merton,R. C. (1969). LifetimePortfolioSelectionunder Uncertainty:The Continuous-TimeCase.
Rev. Econom. Statist. 51 247-257.
[8] . (1971). (1973). OptimumConsumptionand PortfolioRules in a ContinuousTime Model.
J. Economic Theory3 373-413. Erratum. J. Economic Theory 6 213-214.
[9] . (1973).An Intertemporal CapitalAsset PricingModel.Econometrica 41 867-887.
[10] Richard,S. F. (1979).A GeneralizedCapitalAsset PricingModel.In PortfolioTheory,25 YearsAfter,
Elton, E. J. and Gruber, M. J. (eds.), TIMS Studies in the Management Sciences 11 215-232.
[111 Samuelson,P. A. (1969). LifetimePortfolioSelection by Dynamic StochasticProgramming.Rev.
Econom. Statist. 51 239-246.

KARATZAS:DEPARTMENTOF STATISTICS,COLUMBIAUNIVERSITY,NEW YORK, NEW

YORK 10027
LEHOCZKY: DEPARTMENT OF STATISTICS, CARNEGIE-MELLON UNIVERSITY, PITTS-
BURGH, PENNSYLVANIA 15213
SETHI: FACULTY OF MANAGEMENT STUDIES, UNIVERSITY OF TORONTO, TORONTO,
ONTARIO,CANADA M5S 1V4
SHREVE: DEPARTMENT OF MATHEMATICS, CARNEGIE-MELLON UNIVERSITY, PITTS-
BURGH, PENNSYLVANIA 15213