#### Abstract

Motivated by the so-called shortfall risk minimization problem,
we consider Merton's portfolio optimization problem in a non-Markovian
market driven by a Lévy process, with a bounded state-dependent utility
function. Following the usual *dual variational approach*, we show that the
domain of the *dual problem* enjoys an explicit “parametrization,” built on
a *multiplicative optional decomposition* for nonnegative supermartingales
due to Föllmer and Kramkov (1997). As a key step we prove a *closure property*
for integrals with respect to a fixed Poisson random measure, extending a
result by Mémin (1980). In the case where either the Lévy measure of has finite number of atoms or for a process and a
deterministic function , we characterize explicitly the admissible trading
strategies and show that the dual solution is a risk-neutral local martingale.

#### 1. Introduction

The task of determining good trading strategies is a fundamental problem in mathematical finance. A typical approach to this problem aims at finding the trading strategy that maximizes, for example, the final expected utility, which is defined as a concave and increasing function of the final wealth. There are, however, many applications where a utility function varies with the underlying securities, or even random. For example, if the market is incomplete, it is often more beneficial to allow certain degree of “shortfall” in order to reduce the “super-hedging cost” (see, e.g., [1, 2] for more details). Mathematically, such a shortfall risk is often quantified by the expected loss

where is a convex increasing “loss” function, is a contingent claim, and is the value process that is subject to the constraint , for a given initial endowment .

The above shortfall minimizing problem can be easily recast as a utility maximization problem with a *bounded state-dependent utility* of the form

as it was first pointed out by Föllmer and Leukert [3] (see Definition 2.3 for a formal description of the general *bounded state-dependent utility*). Then, the minimal shortfall risk cost is given by

where the supremum is taken over all wealth processes generated by admissible trading strategies. We should point out here that it is the boundedness and potential nondifferentiability of such utility function that give rise to some technical issues which make the problem interesting.

The existence and essential uniqueness of the solution to the problem (1.3) in its various special forms have been studied by many authors see, for example, Cvitanić [4], Föllmer and Leukert [3], Xu [5], and Karatzas and Žitković [6], to mention a few. However, while the convex duality approach in [3] succeeds in dealing with the non-Markovian nature of the model, it does not seem to shed any light on how to compute, in a feasible manner, the optimal trading strategy, partly due to the generality of the model considered there. In this paper we will consider a specific but popular model driven by a Lévy process. Our goal is to narrow down the domain of dual problem so that the convex duality method holds true. Furthermore, we will try to give an explicit construction of the dual domain that contains the dual optimizer. Although at this point our results are still rather general, and at a theoretical level, we believe that this is a necessary step towards a feasible computational implementation of the convex duality method.

While the utility maximization problem of this kind can be traced back to Merton [7, 8], in this paper we shall follow the convex duality method, suggested by Karatzas et al. [9], and later extended by Kunita [10] to general Lévy market models. However, we note that in [9] the utility function was required to be *unbounded*, strictly increasing and concave, continuously differentiable, and other technical assumptions including the so-called Inada conditions. On the other hand, since one of the key tools in [10] is an exponential representation for *positive* local supermartingales (see, e.g., [11, Lemma 4.2]), it is required that the utility function satisfies the same conditions as in [9] (in particular, unboundedness), plus that the dual domain contains all positive “risk-neutral” local supermartingales. The boundedness and potential nondifferentiability of the utility function in our case thus cause some technical subtleties. For example, the dual optimal process can be with positive probability, thus the representation theorem of Kunita [11, Lemma 4.2] does not apply anymore.

A key element that we use to overcome these technical difficulties is an exponential representation theorem for nonnegative supermartingales by Föllmer and Kramkov [12]. This result leads to an explicit construction of the dual domain, based on those nonnegative supermartingales that can be written as stochastic exponentials , with being an increasing process and belonging to a class of semimartingales that is closed under Émery's topology. To validate this approach we prove a *closure property* for integrals with respect to a fixed compensated Poisson random measures, a result of interest on its own, which extends the analog property for integrals with respect to a fixed semimartingale due to Mémin [13]. Finally, unlike some previous works on the subject (see, e.g., Föllmer and Leukert [3] and Xu [5]), we do not use the so-called *bipolar theorem* of Kramkov and Schachermayer [14] to argue the attainability of the optimal final wealth. Instead, we shall rely on the fundamental characterization of contingent claims that are super replicable [1, 2], reducing the problem of finding the optimal primal solution to a super-eplication problem.

We believe that the dual problem proposed in this paper offers several advantages. For example, since the dual class enjoys a fairly “explicit” description and “parametrization,” our results could be considered as a first step towards a feasible computational implementation of the covex duality method. Furthermore, the specific results we obtained for the Lévy market can be used to characterize the elements of the dual domain and the admissible trading strategies. In particular, if either (i) the jumps of the price process are driven by the superposition of finitely many shot-noise Poisson processes, or (ii) for a process and a deterministic function , we show that the dual solution is a risk-neutral local martingale.

We would like to remark that some of our results are related to those in Xu [5], but there are essential differences. For example, the model in [5] exhibits only finite-jump activity and allows only downward price jumps (in fact, this assumption seems to be important for the approach there), while our model allows for general jump dynamics, and our approach is also valid for general additive processes, including the time-inhomogeneous cases considered in [5] (see (ii) of Section 6).

The rest of the paper is organized as follows. In Section 2 we introduce the financial model, along with some basic terminology that will be used throughout the paper. The convex duality method is revised in Section 3, where a potential optimal final wealth is constructed. An explicit description of a dual class and a characterizations of the dual optimum and admissible trading strategies are presented in Section 4. In Section 5 we show that the potential optimal final wealth is attained by an admissible trading strategy, completing the proof of the existence of optimal portfolio. In Section 6 we give some concluding remarks. Some necessary fundamental theoretical results, such as the exponential representation for nonnegative supermartingales of Föllmer and Kramkov [12] and the closure property for integrals with respect to Poisson random measures, are collected in Appendix A. Finally, Appendix B briefly outlines the proofs of the convex duality results used in the paper.

#### 2. Notation and Problem Formulation

Throughout this paper we assume that all the randomness comes from a complete probability space , on which there is defined a Lévy process with Lévy triplet (see Sato [15] for the terminology). By the Lévy-Itô decomposition, there exist a standard Brownian motion and an independent Poisson random measure on with mean measure , such that

where . Let be the natural filtration generated by and , augmented by all the null sets in so that it satisfies the *usual conditions* (see, e.g., [16]).

##### 2.1. The Market Model

We assume that there are two assets in the market: a risk free bond (or money market account), and a risky asset, say, a stock. The case of multiple stocks, such as the one studied in [10], can be treated in a similar way without substantial difficulties (see Section 6 for more details). As it is customary all the processes are taken to be discounted to the present value so that the value of the risk-free asset can be assumed to be identically equal to . The (discounted) price of the stock follows the stochastic differential equation

where , , , and (see [17] for the terminology). More precisely, , , and are predictable processes such that a.s. (hence, a.s.), and that the processes

are locally integrable with respect to time. Even though we will work with a finite horizon later on, we choose to define our market model on . Finally, we assume that the market is free of arbitrage so that there exists at least one risk-neutral probability measure such that the (discounted) process , , is an -local martingale under . Throughout, will stand for the class of all equivalent risk neutral measures . It is relevant to mention that we do not impose market completeness, and hence, the class is not assumed to be a singleton.

##### 2.2. Admissible Trading Strategies and the Utility Maximization Problem

A trading strategy is determined by a predictable locally bounded process representing the *proportion of total wealth invested in the stock*. Then, the resulting wealth process is governed by the stochastic differential equation

where stands for the initial endowment. For future reference, we give a precise definition of “*admissible strategies.*”

*Definition 2.1. *The process solving (2.4) is called the value process corresponding to the self-financing portfolio with initial endowment and trading strategy . We say that a value process is “admissible” or that the process is “admissible” for if

For a given initial endowment , we denote the set all admissible strategies for by , and the set of all admissible value processes by . In light of the Doléans-Dade stochastic exponential of semimartingales (see, e.g., [17, Section I.4f]), one can easily obtain necessary and sufficient conditions for admissibility.

Proposition 2.2. *A predictable locally bounded process is admissible if and only if
*

To define our utility maximization problem, we begin by introducing the *bounded state-dependent utility function*.

*Definition 2.3. *A random function is called a “bounded and state-dependent utility function” if (1) is nonnegative, nondecreasing, and continuous on ; (2)for each fixed , the mapping is -measurable; (3)there is an -measurable, positive random variable such that for all , is a strictly concave differentiable function on , and it holds that

Notice that the -measurability of the random variable is automatic because is -measurable in light of the above conditions and We remark that while assumption (2.7) is merely technical, assumption (2.6) is motivated by the shortfall risk measure (1.2). Our utility optimization problem is thus defined as

for any . We should note that the above problem is relevant only for those initial wealths that are smaller than , the super-hedging cost of . Indeed, if , then there exists an admissible trading strategy for such that almost surely, and consequently, (see [1, 2] for this *super-hedging* result).

Our main objectives in the rest of the paper are the following: (1) Define the dual problem and identify the relation between the value functions of the primal and the dual problems; (2) By suitably defining the dual domain, prove the attainability of the associated dual problem; (3) Show that the potential optimum final wealth induced by the minimizer of the dual problem can be realized by an admissible portfolio. We shall carry out these tasks in the remaining sections.

#### 3. The Convex Duality Method and the Dual Problems

In this section we introduce the *dual problems* corresponding to the primal problem (1.3) and revise some standard results of convex duality that are needed in the sequel. Throughout, represents the *convex dual function* of , defined as

We observe that the function is closely related to the Legendre-Fenchel transformation of the convex function . It can be easily checked that is convex and differentiable everywhere, for each . Furthermore, if we denote the *generalized inverse* function of by

with the convention that , then it holds that

and the function has the following representation:

*Remark 3.1. *We point out that the random fields defined in (3.1) and (3.2) are -measurable. For instance, in the case of , we can write
and we will only need to check that is jointly measurable for each fixed . This last fact follows because the random field in question is continuous in the spatial variable for each and is -measurable for each In light of (3.3), it transpires that the random field is jointly measurable. Given that the subsequent dual problems and corresponding solutions are given in terms of the fields and (see Definition 3.3 and Theorem 3.5), the measurability of several key random variable below is guaranteed.

Next, we introduce the traditional *dual class* (cf. [14]).

*Definition 3.2. *Let be the class of nonnegative supermartingales such that (i), (ii)for each locally bounded admissible trading strategy , is a supermartingale.

To motivate the construction of the dual problems below we note that if and is the value process of a self-financing admissible portfolio with initial endowment , then , and it follows that

for any . The dual problem is defined as follows.

*Definition 3.3. *Given a subclass , the minimization problem
is called the “dual problem induced by .” The class is referred to as a dual domain (or class) and is called its dual value function.

Notice that, by (3.6) and (3.7), we have the following *weak duality* relation between the primal and dual value functions:

valid for all . The effectiveness of the dual problem depends on the attainability of the lower bound in (3.8) for some (in which case, we say that *strong duality* holds), and the attainability of its corresponding dual problem (3.7). The following well-known properties will be needed for future reference. Their proofs are standard and are outlined in Appendix B for the sake of completeness.

Proposition 3.4. *The dual value function corresponding to a subclass of satisfies the following properties *(1)* is nonincreasing on and *(2)*If
then is uniformly continuous on , and
*(3)*There exists a process such that *(4)*If is a convex set, then (i) is convex, and with (ii) there exists a attaining the minimum . Furthermore, the optimum can be “approximated” by elements of in the sense that there exists a sequence for which , a.s. *

We now give a result that is crucial for proving the strong duality in (3.8). The result follows from arguments quite similar to those in [9, Theorem 9.3]. For the sake of completeness, we outline the proof in Appendix B.

Theorem 3.5. *Suppose that (3.9) is satisfied and is convex, then, for any , there exist and such that *(i)*; *(ii)*, where
*(iii)*. *

We note that Theorem 3.5 provides essentially an upper bound for the optimal final utility of the form , for certain “reduced” contingent claim . By suitably choosing the dual class , we shall prove in the next two sections that this reduced contingent claim is (super-) replicable with an initial endowment .

#### 4. Characterization of the Optimal Dual

We now give a full description of a dual class for which *strong duality*, that is, , holds. Denote to be the class of all real-valued càdlàg, nondecreasing, adapted processes null at zero. We will call such a process “increasing.” In what follows we let be the Doléans-Dade stochastic exponential of the semimartingale (see, e.g., [17] for their properties). Let

and consider the associated class of exponential local supermartingales:

In (4.1), we assume that , , and that , for all . The following result shows not only that the class

is convex, but also that the dual optimum, whose existence is guaranteed from Theorem 3.5, remains in . The proof of this result is based on a powerful representation for nonnegative supermartingales due to Föllmer and Kramkov [12] (see Theorem A.1 in Appendix A), and a technical result about the closedness of the class of integrals with respect to Poisson random measures, under Émery’s topology. We shall defer the presentation of these two fundamental results to Appendix A in order to continue with our discussion of the dual problem.

Theorem 4.1. *The class is convex, and if (3.9) is satisfied, the dual optimum of Theorem 3.5 belongs to , for any .*

*Proof. *Let us check that meets with the conditions in Theorem A.1. Indeed, each in is locally bounded from below since, defining
where . Condition (i) of Theorem A.1 is straightforward, while condition (ii) follows from Theorem A.3. Finally, condition (iii) holds because the processes in are already local martingales with respect to and hence with . By Proposition A.2 we conclude that is convex and closed under Fatou convergence on dense countable sets. On the other hand, is also convex and closed under Fatou convergence, and thus so is the class . To check the second statement, recall that the existence of the dual minimizer in Theorem 3.5 is guaranteed from Proposition 3.4, where it is seen that is the Fatou limit of a sequence in (see the proof of Proposition 3.4). This suffices to conclude that since is closed under under Fatou convergence.

In the rest of this section, we present some properties of the elements in and of the dual optimum . In particular, conditions on the “parameters” so that is in are established. First, we note that without loss of generality, can be assumed predictable.

Lemma 4.2. *Let
**
Then, there exist a predictable process and a process such that .*

*Proof. *Let . Since , there are stopping times such that
Compare with [17, Theorem II.1.33]. Now, define
and . Then,
where we used that . Therefore, is locally integrable, increasing, and thus, its predictable compensator exists. Now, by the representation theorem for local martingales (see [17, Theorem III.4.34]), the local martingale admits the representation
Finally, The conclusion of the proposition follows since is necessarily in .

The following result gives necessary conditions for a process to belong to . Recall that a predictable increasing process can be uniquely decomposed as the sum of three predictable increasing processes,

where is the *absolutely continuous* part, is the *singular continuous* part, and is the jump part (cf. [18, Theorem 19.61]).

Proposition 4.3. *Let , where ,
**
and is an increasing predictable process. Let be the “sinking time” of the supermartingale :
**
Also, let . Then, is a supermartingale if and only if the following two conditions are satisfied. *(i)*There exist stopping times such that
*(ii)*For -a.e. ,
**for almost every , where
*

*Proof. *Recall that and satisfy the SDE’s
Integration by parts and the predictability of yield that
Suppose that is a nonnegative supermartingale. Then, the integral must have locally integrable variation in light of the Doob-Meyer decomposition for supermartingale (see, e.g., [16, Theorem III.13]). Therefore, there exist stopping times such that
Then, (i) is satisfied with , where and . Next, we can write (4.17) as
By the Doob-Meyer representation for supermartingales and the uniqueness of the canonical decomposition for special semimartingales, the last integral must be increasing. Then, for since and for (see [17, Theorem I.4.61]).

We now turn to the sufficiency of conditions (i) and (ii). Since is locally bounded,

is locally integrable. Then, from (4.17), we can write
Condition (ii) implies that is a supermartingale, and by Fatou, will be a supermartingale. This concludes the prove since for , and thus, , for all .

The following result gives sufficient and necessary conditions for to belong to . Its proof is similar to that of Proposition 4.3.

Proposition 4.4. *Under the setting and notation of Proposition 4.3, belongs to if and only if condition (i) in Proposition 4.3 holds and, for any locally bounded admissible trading strategies ,
*

The previous result can actually be made more explicit under additional information on the structure of the jumps exhibited by the stock price process. We consider two cases: when the jumps come from the superposition of shot-noise Poisson processes, and when the random field exhibit a multiplicative structure. Let us first extend Proposition 2.2 in these two cases.

Proposition 4.5. *
(i) Suppose that is atomic with finitely many atoms then, a predictable locally bounded strategy is admissible if and only if -a.e.**
Suppose that for a predictable locally bounded process such that -a.e. and is locally bounded, and a deterministic function such that , then, a predictable locally bounded strategy is admissible if and only if -a.e.**
where and .*

*Proof. * From Proposition 2.2, recall that -a.s.
for a.e. . Then, for any closed set , , and ,
Taking expectation, we get
Since such processes generate the class of predictable processes, we conclude that -a.e.
Let us prove (ii) (the proof of (i) is similar). Notice that
where is the support of . Suppose that . Then, by considering closed sets such that
as , we can prove the necessity. The other two cases (namely, or ) are proved in a similar way. Sufficiency follows since, -a.s.,

*Example 4.6. *It is worth pointing out some consequences (a)In the time homogeneous case, where , the extreme points of the support of (or what accounts to the same, the infimum and supremum of all possible jump sizes) determine completely the admissible strategies. For instance, if the Lévy process can exhibit arbitrarily large or arbitrarily close to jump sizes, then
a constraint that can be interpreted as absence of shortselling and bank borrowing (this fact was already pointed out by Hurd [19]). (b)In the case that , the admissibility condition takes the form If in addition (such that the stock prices exhibit only downward sudden movements), then and , with arbitrary, is admissible. In particular, from Proposition 4.4, if belongs to , then a.s. This means that if and only if condition (i) in Proposition 4.3 holds and a.s. , for a.e. . For a general and still assuming that , it follows that is admissible and satisfy that -a.s.
for a.e. .

We now extend Proposition 4.4 in the two cases introduced in Proposition 4.5. Its proof follows from Propositions 4.4 and 4.5.

Proposition 4.7. * Suppose that either (i) or (ii) in Proposition 4.5 is satisfied, in which case, define
**
then, a process belongs to if and only if condition (i) in Proposition 4.3 holds, and for -a.e. , , for a.e. .*

We remark that the cases and do not lead to any absurd in the definition of above as we are using the convention that . Indeed, for instance, if , it was seeing that , for a.e. , and thus, we set the second term in the definition of to be zero.

Now we can give a more explicit characterization of the dual solution to problem (3.7), whose existence was established in Theorem 4.1. For instance, we will see that is absolutely continuous up to a predictable stopping time. Below, we refer to Proposition 4.3 for the notation.

Proposition 4.8. *Let , and . The followings two statements hold true. *(1)*. Furthermore, if and only if . *(2)*Suppose that either of the two conditions in Proposition 4.7 are satisfied and denote
where is defined accordingly to the assumed case. Then, , and furthermore, the process belongs to if . *

*Proof. *Let denote the increasing predictable processes in decomposition (4.10) of . Since is predictable, there is no common jump times between and . Then,
where we used that , and . Since both processes and enjoy the same absolutely continuous part, and the same sinking time, the second statement in (1) is straightforward from Proposition 4.4. Part (2) follows from Proposition 4.7 since the process is nonnegative, predictable (since is predictable), and locally integrable (since ).

We remark that part (2) in Proposition 4.8 remains true if we take . The following result is similar to Proposition 3.4 in Xu [5] and implies, in particular, that the optimum dual can be taken to be a local martingale.

Proposition 4.9. * Suppose that either (i) or (ii) of Proposition 4.5 is satisfied. Moreover, in the case of condition (ii), assume additionally that
**
for if , and for if . Let . Then, there exists such that and . Furthermore, is a local martingale for all locally bounded admissible trading strategies .*

*Proof. *Let us prove the case when condition (i) in Proposition 4.5 is in force. In light of Proposition 4.8, we assume without loss of generality that with Assume that . Otherwise if, for instance, , then it can be shown that , a.s. (similarly to case (b) in Example 4.6), and the first term of is under our convention that . Notice that, in any case, one can find a predictable process taking values on , such that
Write for an to be determined in the sequel. For it suffices to prove the existence of a field satisfying both conditions below:
(then, is defined as ). Similarly, for to belong to it suffices that
Taking
clearly nonnegative, (b) and (c) hold with equality. Moreover, the fact that inequalities (c) hold with equality implies that is a local martingale for all locally bounded admissible trading strategy (this can be proved using the same arguments as in the sufficiency part of Proposition 4.3). Now suppose that condition (ii) in Proposition 4.5 holds. For simplicity, let us assume that (the other cases can be analyzed following arguments similar to Example 4.6). Notice that (4.37) implies the existence of a Borel (resp., ) such that on (resp., on ) and . Taking
and (c) above will hold with equality.

#### 5. Replicability of the Upper Bound

We now show that the tentative optimum final wealth , suggested by the inequality (iii) in Theorem 3.5, is (super-) replicable. We will combine the dual optimality of with the *super-hedging theorem*, which states that given a contingent claim satisfying , one can find for any fixed an admissible trading strategy (depending on ) such that almost surely (see Kramkov [2], and also Delbaen and Schachermayer [1]). Recall that denotes the class of all equivalent risk neutral probability measures.

Proposition 5.1. *Under the setting and conditions of Theorem 3.5, for any , there is an admissible trading strategy for such that
**
and thus, the optimum of is reached at the strategy . In particular,
**
when .*

*Proof. *For simplicity, we write , , and
Fix an equivalent risk neutral probability measure , and let be its corresponding density processes. Here, (resp., ) is the restriction of the measure (resp., ) to the filtration . Under , is a local martingale, and then, for any locally bounded , is a -local martingale. By [17, Proposition III.3.8.c], is a -local martingale (necessarily nonnegative by admissibility), and thus, is in . On the other hand, belongs to due to the exponential representation for *positive* local martingales in Kunita [11] (alternatively, by invoking [17, Theorems III.8.3, I.4.34c, and III.4.34], even if were just an additive process ). By the convexity of the dual class and the fact that (see Theorem 4.1), belongs to , for any . Moreover, since is convex and ,
The random variable is integrable since by assumption . We can then apply dominated convergence theorem to get
which is nonnegative by condition (i) in Theorem 3.5. Then, using condition (ii) in Theorem 3.5,
Since is arbitrary, By *the super-hedging theorem*, there is an admissible trading strategy for such that
The second statement of the theorem is straightforward since is strictly increasing on .

#### 6. Concluding Remarks

We conclude the paper with the following remarks.

*(i) The dual class *

The dual domain of the dual problem can be taken to be the more familiar class of equivalent risk-neutral probability measures . To be more precise, define

Since is obviously a convex subclass of , Theorem 3.5 implies that, as far as

for each , there exist and (not necessarily belonging to ) such that (i)–(iii) in Theorem 3.5 hold with . Finally, one can slightly modify the proof of Proposition 5.1, to conclude the replicability of

Indeed, in the notation of the proof of the Proposition 5.1, the only step which needs to be justified in more detail is that

for all , where (here, is a fixed element in ). The last inequality follows from the fact that, by Proposition 5.1 (c), can be approximated by elements in in the sense that a.s. Thus, can be approximated by the elements in , for which we know that

Passing to the limit as , we obtain (6.4).

In particular we conclude that condition (6.2) is sufficient for both the existence of the solution to the primal problem and its characterization in terms of the dual solution of the dual problem induced by . We now further know that belongs to the class defined in (4.3), and hence, enjoys an explicit parametrization of the form

for some triple .

*(ii) Market driven by general additive models*

Our analysis can be extended to more general multidimensional models driven by additive processes (i.e., processes with independent, possibly nonstationary increments; cf. Sato [15] and Kallenberg [20]). For instance, let be a complete probability space on which is defined a dimensional additive process with Lévy-Itô decomposition:

where is a standard -dimensional Brownian motion, is an independent Poisson random measure on , and . Consider a market model consisting of securities: one risk free bond with price

and risky assets with prices determined by the following stochastic differential equations with jumps:

where the processes , , , and are predictable satisfying usual integrability conditions (cf. Kunita [10]). We assume that , where is the natural filtration generated by and ; namely, . The crucial property, particular to this market model, that makes our analysis valid, is the representation theorem for local martingales relative to (see [17, Theorem III.4.34]). The definition of the dual class given in Section 4 will remain unchanged, and only very minor details will change in the proof of Theorem A.3. Some of the properties of the results in Section 4 regarding the properties of will also change slightly. We remark that, by taking a real (nonhomogeneous) Poisson process, the model and results of Chapter 3 in Xu [5] will be greatly extended. We do not pursue the details here due to the limitation of the length of this paper.

*(iii) Optimal wealth-consumption problem*

Another classical portfolio optimization in the literature is that of optimal wealth-consumption strategies under a budget constraint. Namely, we allow the agent to spend money outside the market, while maintaining “solvency” throughout . In that case the agent aims to maximize the cost functional that contains a “*running cost*”:

where is the instantaneous rate of consumption. To be more precise, the cumulative consumption at time is given by and the (discounted) wealth at time is given by

Here, is a (state-dependent) utility function and is a utility function for each . The dual problem can now be defined as follows:

over a suitable class of supermartingales . For instance, if the support of is , then can be all supermartingales such that and is a supermartingale. The dual Theorem 3.5 can be extended for this problem. However, the existence of a wealth-consumption strategy pair that attains the potential final wealth induced by the optimal dual solution (as in Section 5) requires further work. We hope to address this problem in a future publication.

#### Appendices

#### A. Convex Classes of Exponential Supermartingales

The goal of this part is to establish the theoretical foundations behind Theorem 4.1. We begin by recalling an important optional decomposition theorem due to Föllmer and Kramkov [12]. Given a family of supermartingales satisfying suitable conditions, the result characterizes the nonnegative exponential local supermartingales where and , in terms of the so-called *upper variation process* for . Concretely, let be the class of probability measures for which there is an increasing predictable process (depending on and ) such that is a local supermartingale under , for all . The *smallest* of such processes is denoted by and is called the upper variation process for corresponding to . For easy reference, we state Föllmer and Kramkov’s result (see [12] for a proof).

Theorem A.1. *Let be a family of semimartingales that are null at zero, and that are locally bounded from below. Assume that , and that the following conditions hold: *(i)* is predictably convex,*(ii)* is closed under the Émery distance,*(iii)*, **then, the following two statements are equivalent for a nonnegative process : *(1)* is of the form , for some and an increasing process ;*(2)* is a supermartingale under for each . *

The next result is a direct consequence of the previous representation. Recall that a sequence of processes is said to be “Fatou convergent on ” to a process if is uniformly bounded from below and it holds that

almost surely for all .

Proposition A.2. *If is a class of semimartingales satisfying the conditions in Theorem A.1, then
**
is convex and closed under Fatou convergence on any fixed dense countable set of ; that is, if is a sequence in that is Fatou convergent on to a process , then .*

*Proof. *The convexity of is a direct consequence of Theorem A.1, since the convex combination of supermartingales remains a supermartingale. Let us prove the closure property. Fix a and denote . Notice that because is increasing and hence, its jumps are nonnegative. Since , is a supermartingale under . Then, for ,
By Fatou’s lemma and the right-continuity of process ,
Finally, using the right continuity of the filtration, we have
where . Since is arbitrary, the characterization of Theorem A.1 implies that .

The most technical condition in Theorem A.1 is the closure property under Émery distance. The following result is useful to deal with this condition. It shows that the class of integrals with respect to a Poisson random measure is closed with respect to Émery distance, thus extending the analog property for integrals with respect to a fixed semimartingale due to Mémin [13].

Theorem A.3. *Let be a closed convex subset of containing the origin. Let be the set of all predictable processes , and , such that , for all , and , for -a.e. . Then, the class
**
is closed under convergence with respect to Émery’s topology.*

*Proof. *Consider a sequence of semimartingales
in the class . Let be a semimartingale such that under Émery topology. To prove the result, we will borrow some results in [13].

For some , we denote to be the Banach space of all -square integrable martingales on , endowed with the norm , and to be the Banach space of all predictable processes on that have -integrable total variations, endowed with the norm . Below, stands for the localized class of increasing process in . By [13, Theorem II.3], one can extract a subsequence from , still denote it by , for which one can construct a probability measure , defined on and equivalent to (the restriction of on ), such that the following assertions hold:

(i) is bounded by a constant; (ii), , for Cauchy sequences and in and , respectively.

Let us extend and to by setting and for all . Also, we extend for by setting , so that (on ). In that case, it can be proved that This follows essentially from [17, Proposition III.3.5] and Doob's Theorem. Now, let denote the density process. Since is bounded, both and are bounded. By [17, Lemma III.3.14 and Theorem III.3.11], the -quadratic covariation has -locally integrable variation and the unique canonical decomposition of relative to is given by

Also, the -quadratic variation of the continuous part of (relative to ), given by is also a version of the -quadratic variation of the continuous part of (relative to ). By the representation theorem for local martingales relative to (see, e.g., [17, Theorem III.4.34] or [11, Theorem 2.1]), has the representation
for predictable and necessarily satisfying that ,
Then,