Abstract
This paper deals with optimal control of systems driven by stochastic differential equations (SDEs), with controlled jumps, where the control variable has two components, the first being absolutely continuous and the second singular. We study the corresponding relaxed-singular problem, in which the first part of the admissible control is a measure-valued process and the state variable is governed by a SDE driven by a relaxed Poisson random measure, whose compensator is a product measure. We establish a stochastic maximum principle for this type of relaxed control problems extending existing results. The proofs are based on Ekeland’s variational principle and stability properties of the state process and adjoint variable with respect to the control variable.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
We consider mixed relaxed-singular stochastic control problems of systems governed by stochastic differential equations (SDEs) with controlled jumps (1.1), where the control variable has two components, the first being a measure-valued process and the second a bounded variation process, called the singular part. More precisely, the system evolves according to the SDE,
on some probability space \((\Omega ,{\mathcal {F}}\),\(({\mathcal {F}}_{t} \mathcal {)}_{t\ge 0},P)\), such that \({\mathcal {F}}_{0}\) contains the \(P-\)null sets. We assume that \(({\mathcal {F}}_{t})_{t\ge 0}\) is generated by a standard Brownian motion B and an independent Poisson random measure \({\widetilde{N}}^{\mu }\), with compensator \(\mu _{t}\otimes \upsilon (da,d\theta ),\) where \(\mu \) is the relaxed control and \(\upsilon \) is the compensator of Poisson measure \(N^{\mu }\). The control variable is \((\mu ,\zeta ),\) where \(\mu \) is a \({\mathcal {P}}(A_{1})-\)valued process, progressively measurable with respect to \(({\mathcal {F}}_{t}\mathcal {)}_{t\ge 0}\) and \(\zeta :\left[ 0;T\right] \times \Omega \longrightarrow A_{2}\) is an adapted process of bounded variation, nondecreasing, left-continuous with right limits such that \(\zeta _{0}=0.\)
The expected cost to be minimized over the class of admissible controls has the form
A control process that solves this problem is called optimal.
Singular stochastic control problems have been extensively studied in the literature. We refer to Botius [9], for a complete survey. This problem was first introduced by Bather and Chernoff [4] in 1967, by considering a simplified model for the control of a spaceship, known as the monotone follower for Brownian motion. The authors have noted that this model of singular control has a connection with some optimal stopping problem. It was proved, in particular, that the value function of the singular control problem is equal to the gradient of value function of the corresponding optimal stopping problem. After this seminal article, this connection has been deeply investigated in different contexts. Singular control problems find many applications in different areas of engineering such as mathematical finance, manufacturing systems and queuing theory [11, 27]. Two approaches were used to handle singular control problems. The first one is based on dynamic programming, which leads to a variational inequality. This approach has been studied by many authors including Benĕs, Sheep, and Witsenhausen [6], Chow, Menaldi, and Robin [13], Karatzas and Shreve [22, 24], Davis and Norman [14], and Haussmann and Suo [18,19,20]. Probabilistic methods have been used to solve the dynamic programming equation, see, e.g., [16, 22] and the references therein. The second approach to solve singular control problems is the so-called stochastic maximum principle. This approach leads to necessary conditions for optimality satisfied by an optimal control and is based on some adjoint process and a variational inequality between hamiltonians. The first version of the stochastic maximum principle that covers singular control problems was obtained by Cadenillas and Haussman [10] for linear systems. General second-order necessary conditions for optimality for nonlinear SDEs with a controlled diffusion matrix were obtained by S. Bahlali and B. Mezerdi [3], extending Peng’s second-order stochastic maximum principle to singular control problems. We refer also to [1] for systems with nonsmooth coefficients. The stochastic maximum principle for relaxed-singular control problem has been studied by S. Bahlali, B. Djehiche, and B. Mezerdi [2]. The relationship between the dynamic programming principle and the maximum principle has been investigated in [12].
Our main goal in this paper is to derive a stochastic maximum principle for mixed relaxed-singular control problems. This means that the first component of the control is a measure valued process and the second component is a process with increasing sample paths. By using the same weak convergence techniques as in [18, 20], it is not difficult to show the existence of an optimal control for the relaxed-singular control problem and that the value functions of the relaxed problem and the strict control problem are the same. The proof of our stochastic maximum principle is divided into three steps. We first establish necessary conditions for optimality satisfied by an optimal strict control. The second step is devoted to the necessary conditions for near optimality satisfied by a sequence of near optimal controls, by using Ekeland’s variational principle. This auxiliary result is in itself one of the novelties of this paper. Indeed, in most practical situations, it is sufficient to characterize and compute nearly optimal controls. In the third step, we prove the relaxed stochastic maximum principle by passing to the limit in the adjoint processes and the variational inequalities. These properties are based on the stability of the state process and the adjoint processes with respect to the control variable. The novelty of our result is that our maximum principle is given for a relaxed optimal control, which exists and the dynamics involves controlled jumps. In particular, our main result extends the stochastic maximum principle proved in [2, 10] to systems of SDE with controlled jumps (1.1). On the other hand, it extends [5] to systems involving a singular component. The idea of the proof is to use spike variation of the absolutely continuous part of the control and a convex perturbation of the singular part. The principal result is given via an adjoint process of first order and two variational inequalities. The main result is obtained by using some stability properties of the state and adjoint processes with respect to the control variable.
The rest of the paper is organized as follows. In section 2, we formulate the control problem and introduce the assumptions of the model. Section 3 is devoted to the proof of the approximation result. In the last section, we state and prove a maximum principle for our relaxed control problem, which is the main result of this paper.
2 Formulation of the Problem
We consider optimal control problems of systems governed by stochastic differential equations, on some filtered probability space \((\Omega ,{\mathcal {F}}\),\(({\mathcal {F}}_{t}\mathcal {)}_{t\ge 0},P)\), such that \({\mathcal {F}}_{0}\) contains the \(P-\)null sets. We assume that \(({\mathcal {F}} _{t})_{t\ge 0}\) is a complete filtration generated by a standard Brownian motion B and an independent Poisson measure N. Assume that the compensator of N has the form \(\upsilon (d\theta )dt\), where the jumps are confined to a compact set \(\Gamma \) and set
Consider the following sets, \(A_{1}\), is a nonempty compact subset of \( {\mathbb {R}} ^{k}\) and \(A_{2}=\left( \left[ 0;\infty \right) \right) ^{m},\) let \(U_{1}\) the class of measurable, adapted processes \(u:\left[ 0;T\right] \times \Omega \longrightarrow A_{1}\), and \(U_{2}\) the class of measurable, adapted processes \(\zeta :\left[ 0;T\right] \times \Omega \longrightarrow A_{2}.\)
Definition 2.1
An admissible strict control is a term \(\alpha =(\Omega ,{\mathcal {F}},{\mathcal {F}}_{t},P,u_{t},\zeta _{t},W_{t},X_{t})\) such that
(1) \((\Omega ,{\mathcal {F}},{\mathcal {F}}_{t},P)\) is a probability space equipped with a filtration \(({\mathcal {F}}_{t})_{t\ge 0}\) satisfying the usual conditions.
(2) \(u_{t}\) is a \(A_{1}\)-valued process, progressively measurable with respect to \(({\mathcal {F}}_{t})\).
(3) \(W_{t}\) is a \(({\mathcal {F}}_{t},P)\)- Brownian motion, \(\widetilde{N}(dt,d\theta )\) is a compensated Poisson random measure independent from \(\left( W_{t}\right) \) and \((W_{t},{\widetilde{N}}(dt,d\theta ),X_{t})\) satisfies (2.1).
4) \(\zeta \) is of bounded variation, nondecreasing left-continuous with right limits and \(\zeta _{0}=0\)
5) \((u,\zeta )\) satisfies
We denote by \({\mathcal {U}}\) the space of strict controls. The controls as defined in the last definition are called weak controls, because of the possible change of the probability space, the Brownian motion and the Poisson random measure with \(u_{t}.\)
For any \((u,\zeta )\in {\mathcal {U}}\), we consider the following stochastic differential equation (SDE)
where
are continuous functions.
The expected cost is given by:
where
are continuous functions.
The strict optimal control problem is to minimize the functional J(., .) over \({\mathcal {U}}\). A control that solves this problem is called optimal.
The following assumptions will be in force throughout this paper
(H\(_{1})\) The maps b, \(\sigma ,\) f and h are continuously differentiable with respect to x. They and their derivatives \(b_{x}\) \(\sigma _{x}\) \(f_{x}\) and \(h_{x}\) are continuous in (x, u).
(H\(_{2})\) The derivatives \(\sigma _{x}\), \(f_{x}\) and \(g_{x}\) are bounded, and \(b_{x}\)and \(h_{x}\) are uniformly bounded in u.
(H\(_{3})\) There exist \(K>0\) such that, b, \(\sigma \) and f are bounded by \(K(1+\left| x\right| +\left| u\right| ),\) and g is bounded by \(K(1+\left| x\right| ).\)
(H\(_{4})\) G and k are continuous and bounded.
Under the above hypothesis, (2.1) has a unique strong solution and the cost functional (2.2) is well defined from \({\mathcal {U}}\) into \( {\mathbb {R}}.\)
2.1 Examples
Singular control problems were first studied in connection with the so-called monotone follower problem and finite fuel problems. This class of problems is highly relevant in many branches of applied science, such as operation research, insurance problems, mathematical finance. In what follows, we give two examples of singular control problems.
2.1.1 The finite Fuel Problem
The finite fuel problem is a well-known control problem, in which the controller tracks a Brownian motion \(x+B_{t}\) starting at x, by an adapted process \(\varsigma (t)=\varsigma ^{+}(t)-\varsigma ^{-}(t)\) with bounded variation. \(\varsigma ^{+}(t)\) and \(\varsigma ^{-}(t)\) are increasing processes. We assume that the total variation of \(\varsigma ^{-}(t)\) is bounded, \(\left| \varsigma \right| _{t}=\varsigma ^{+}(t)+\varsigma ^{-}(t)\le M.\)
The objective is to minimize the discounted cost functional
\(J(\varsigma )=E\left( \int \limits _{0}^{\tau }\exp (-\alpha t)\lambda X^{2}(t)dt+\int _{\left[ 0,\tau \right] }\exp (-\alpha t)d\left| \varsigma \right| _{t}+\exp (-\alpha \tau )\delta X^{2}(\tau )\right) \)
over such bounded variation processes \(\varsigma (t)\) and stopping times \(\tau .\ \)
This problem has been studied extensively by many authors under different assumptions; we can refer to [14] for more details.
2.1.2 The Portfolio Selection Under Transaction Costs
Assume that the investor has two instruments, a bank account \(S_{0}(t)\) paying a fixed interest rate r and a risky asset (stock) \(S_{1}(t)\), whose price evolves according to a geometric Brownian motion. The investor consumes at a rate c(t) from the bank account, under the constraint that the total wealth should remain positive.
The dynamics for \(S_{0}(t)\) and \(S_{1}(t)\) are given by
\(\left\{ \begin{array}{l} dS_{0}(t)=\left( rS_{0}(t)-c(t)\right) dt-\left( 1+\lambda \right) dL_{t}+\left( 1-\mu \right) dU_{t};\quad S_{0}(0)=x\\ dS_{1}(t)=\alpha S_{1}(t)S_{1}(t)dt+\sigma S_{1}(t)dB_{t}+dL_{t}-dU_{t};\quad S_{1}(0)=y \end{array} \right. \)
A policy of investment and consumption is a triple \(\left( c,L,U\right) ,\) where \(L_{t},U_{t}\) are right continuous nondecreasing processes, which represent the cumulative purchases and sales of stock, respectively.
The investor objective is to maximize the utility
\(J_{x,y}(c,L,U)=E_{x,y}\left( \int \limits _{0}^{+\infty }e^{-\delta t}u(c(t))dt\right) \)
over admissible policies \(\left( c,L,U\right) .\) For more details, see [23]
2.2 The Relaxed-Singular Control Problem
It is a well known that even in simple cases, there is no optimal control in the space of strict controls. The idea is to embed the space of strict controls into a wider space with good compactness properties. The idea of relaxed control is to replace the \(A_{1}\)-valued process \((u_{t})\) with a \({\mathcal {P}}(A_{1})\)-valued process \((\mu _{t})\), where \({\mathcal {P}}(A_{1})\) is the space of probability measures equipped with the topology of weak convergence.
Let \({\mathcal {P}}(A_{1})\) be the space of probability measures on the control set \(A_{1}.\) Let \({\mathbb {V}}\) be the space of measurable transformations \(\mu :[0,T]\longrightarrow {\mathcal {P}}(A_{1}),\) then \(\mu \) can be identified as a nonnegative measure on the product \([0,T]\times A_{1},\) by putting for \(C\in {\mathcal {B}}([0,T])\) and \(D\in {\mathcal {B}}(A_{1})\)
\({\overline{\mu }}(C\times D)=\int \nolimits _{C}\mu _{t}(da)dt.\)
\({\overline{\mu }}\) may be extended uniquely to an element of \({\mathbb {M}} _{+}([0,T]\times A_{1})\) the space of Radon measures on \([0,T]\times A_{1},\) equipped with the topology of stable convergence. This topology is the weakest topology such that the mapping
\({\overline{\mu }}\longrightarrow \int \nolimits _{0}^{T}\int \nolimits _{A_{1}} \phi (t,a).{\overline{\mu }}(dt,da)\)
is continuous for all bounded measurable functions \(\phi \) which are continuous in a.
Equipped with this topology, \({\mathbb {M}}_{+}([0,T]\times A_{1})\) is a compact separable metrizable space. Therefore, \({\mathbb {V}}\) as a closed subspace of \({\mathbb {M}}_{+}([0,T]\times A_{1})\) is also compact (see [15]) for more details.
Notice that \({\mathbb {V}}\) can be identified as the space of positive Radon measures on \([0,T]\times A_{1}\), whose projections on [0, T] coincide with Lebesgue measure.
Let us define the Borel \(\sigma -\)field \(\overline{{\mathbb {V}}}\) as the smallest \(\sigma -\)field such that the mappings
are measurable, where \(\phi \) is a bounded measurable function which is continuous in a.
Let us also introduce the filtration \(\left( \overline{{\mathbb {V}}} _{t}\right) \) on \({\mathbb {V}}\), where \(\overline{{\mathbb {V}}}_{t}\) is generated by \(\left\{ 1_{\left[ 0,t\right] }\mu ,\quad \mu \in {\mathbb {V}}\right\} \).
Definition 2.2
A measure-valued control on the filtered probability space \(\left( \Omega ,{\mathcal {F}},{\mathcal {F}}_{t},P\right) \) is a random variable \(\mu \) with values in \({\mathbb {V}}\) such that \(\mu (\omega ,t,da)\) is progressively measurable with respect to \(({\mathcal {F}}_{t})\) and such that for each t, \(1_{(0,t]}.\mu \) is \({\mathcal {F}}_{t}-\)measurable.
The state variable is governed by a counting measure valued process, called the relaxed Poisson measure, as described in the following definition [5, 25, 26].
Definition 2.3
A relaxed Poisson measure \(N^{\mu }\) is a counting measure valued process such that its compensator is the product measure \(\mu \otimes \upsilon \) of the relaxed control \(\mu \) with the compensator \(\upsilon \) of N, such that for any Borel set \(\Gamma _{0}\subset \Gamma \) and \(A_{0}\subset A,\) the processes
are \({\mathcal {F}}_{t}-\)martingales and are orthogonal for disjoint \(\ \Gamma _{0}\times A.\)
Now let us introduce the precise definitions of a relaxed control.
Definition 2.4
A relaxed control is a term \(\alpha =(\Omega ,{\mathcal {F}},{\mathcal {F}}_{t},P,\mu _{t},W_{t},{\widetilde{N}}^{\mu },X_{t})\) such that
-
(1)
\((\Omega ,{\mathcal {F}},{\mathcal {F}}_{t},P)\) is a probability space equipped with a filtration \(({\mathcal {F}}_{t})_{t\ge 0}\) satisfying the usual conditions.
-
(2)
\(\mu \) is a measure-valued control on \(\left( \Omega ,{\mathcal {F}},{\mathcal {F}}_{t},P\right) .\)
-
(3)
\(W_{t}\) is a \(({\mathcal {F}}_{t},P)\)- Brownian motion, \({\widetilde{N}}^{\mu }{\small (dt,d\theta )}\) is a relaxed Poisson random measure, independent from \(\left( W_{t}\right) \) and \((W_{t},{\widetilde{N}}^{\mu }(dt,d\theta ),X_{t})\) satisfies (2.3).
-
(4)
\(\zeta \) is of bounded variation, nondecreasing left-continuous with right limits and \(\zeta _{0}=0\)
-
(5)
\(\zeta \) satisfies \(E\left[ \left| \zeta _{T}\right| ^{2}\right] <\infty .\)
Accordingly, the relaxed cost functional will be given by
Let us denote by \({\mathcal {R}}\) the set of relaxed-singular controls.
For any \((\mu ,\zeta )\in {\mathcal {R}},\) write the stochastic differential equation with controlled jumps in terms of relaxed Poisson measure as follows:
It is well known that an optimal control exists in the class of relaxed-singular controls and that the value functions of the strict and relaxed control problems are equal. This important result is based on the continuity of the state process and the corresponding cost functional with respect to the control variable [18, 20].
3 Approximation of the Relaxed State Process
In order for the relaxed-singular control problem to be truly an extension of the strict control problem, the infimum of the expected cost over relaxed-singular controls must be equal to the infimum over strict controls. This result is based on the approximation of a relaxed control by a sequence of strict controls and the convergence of corresponding state processes.
Let us recall the so-called chattering lemma [8, 15, 17], whose proof is given here for the sake of completeness.
Lemma 3.1
(Chattering lemma) Let \(\mu \) be a relaxed control. Then there exists a sequence of adapted processes \((u^{n})\) with values in \(A_{1}\), such that the sequence of random measures \(\left( \delta _{u_{t}^{n}}(da)\,dt\right) \) converges in \({\mathbb {V}}\) to \(dt . \mu _{t} (da),\) \(P-a.s.,\) that is for any f continuous in \(\left[ 0,T\right] \times A_{1}\), we have:
\(\underset{n\rightarrow +\infty }{\lim }\int _{0}^{T}f(s,u_{s}^{n})ds=\int _{0} ^{T}\int _{A_{1}}f(s,a)\mu _{t}(da)\) uniformly in \(t\in \left[ 0,T\right] ,\) \(P-a.s.\)
Proof
Suppose that \(\mu (t,da)\) has continuous sample paths. Let \(n\ge 1\), let us divide the interval [0, T] into subintervals \(\left( T_{i}\right) \) of the form \([ t_{i},s_{i}[ \)of length not exceeding \(2^{-n}\). Cover \(A_{1}\) by finitely many disjoint sets \(\left( A_{j}\right) \) such that diameter(\(A_{j})\le 2^{-n}\). Choose a point \((t_{i},a_{ij})\) in \(T_{i}\times A_{j}\) for each \(i,j,t_{i}\) being as before. Let \(\lambda _{ij}=\mu (t_{i},A_{j}),\) then \(\sum _{j}\lambda _{ij}=1.\) Subdivide each \(T_{i}\) further into disjoint left-closed, right-open intervals \(T_{ij}\) of length \(\lambda _{ij}\times \) the length of \(T_{i}\). Let \(\varepsilon >0.\) Since f is uniformly continuous, then for n large enough we have
Defined the sequence of predictable process \(\mu _{n}(.)\) by \(\mu _{n}(t,da)=\delta _{a_{ij}}(da)\) for \(t\in T_{ij}.\)Then
By path-continuity of u(.), we may increase n further if necessary to ensure that the above is bounded by
which achieves the proof. Now if \(\mu (t,da)\) does not have continuous sample paths, approximate it by controls which do, e.g., by \(\mu ^{n}(.)\) defined for a continuous f by
where \(k=[t-(t-1/n)\vee 0].\) \(\square \)
The next theorem gives the stability of the stochastic differential equations, with respect to the control variable, and that the two problems have the same infimum of the expected costs.
Theorem 3.2
Let \((\mu ,\zeta )\) be a relaxed-singular control, and let \(x^{\mu }\) be the corresponding trajectory. Then there exists a sequence \((u^{n},\zeta )\) of strict controls, such that
and
where \(x^{n}\) denote the trajectory associated with the strict control \((u^{n},\zeta ).\)
Proof
See Theorem 1 in [5] \(\square \)
4 The Maximum Principle for Relaxed Control Problems
Our main goal in this section is to establish optimality necessary conditions for relaxed-singular control problems, where the system is described by a SDE driven by a relaxed Poisson measure which is a martingale measure, of the form (2.3) and the admissible controls are measure-valued processes which are called relaxed controls. The proof is based on the chattering lemma 3.1, and Ekeland’s variational principle (4.6). We first derive necessary conditions of near optimality satisfied by a sequence of strict controls. Then by using stability properties of the state equations and adjoint processes, we are able to obtain the maximum principle for our relaxed control problem.
4.1 The Maximum Principle for Strict Controls
The purpose of this subsection is to derive optimality necessary conditions, satisfied by an optimal strict control. The proof is based on strong perturbation for the absolutely continuous part, and the convex perturbation for the singular components of the optimal control \((u^{*},\zeta ^{*})\), which is defined by:
for some \((\nu ,\xi )\in {\mathcal {U}}\).
4.1.1 The First Variational Inequality
To obtain the first variational inequality in the stochastic maximum principle, we use the strong perturbations (4.1). The first variational inequality is derived from the fact that
Indeed since \((u^{*},\zeta ^{*})\) is optimal, then \(J(u^{h},\zeta ^{*})\ge J(u^{*},\zeta ^{*})\) and therefore if the derivative exists we get \(\left. \frac{dJ(u^{h},\zeta ^{*})}{dh}\right| _{h=0}\ge 0.\)
Note that the singular part is not affected by the perturbation (4.1). So, it is easy to check by standard arguments that
where
Under our assumptions, one has
where
\(\left\{ \begin{array}{l} d\varsigma _{t}=h_{x}(t,x_{t}^{*},u_{t}^{*})z_{t}dt \qquad t_{0}\le t\le T\\ \varsigma _{t_{0}}=h(t_{0},x_{t_{0}}^{*},\nu )-h(t_{0},x_{t_{0}}^{*},u_{t_{0}}^{*}), \end{array} \right. \)
and the process z is the solution of the linear SDE
From (H\(_{2}\)) the variational Eq. (4.5) has a unique solution. To prove Prop (4.1.1) we need the following estimates.
Proposition 4.1
Under assumptions (H\(_{1}\)) – (H\(_{3}\)), it holds that
Proof
Since \(x_{t}^{(u^{h},\zeta ^{*})}-x_{t}^{*}\) does not depend on the singular part, the proof follows that of Lemma 6 in [5]. \(\square \)
Let us introduce the adjoint process and the first variational inequality from (4.4). We proceed as in [7].
Let \(\varphi (t,\tau )\) be the solution of the linear equation
This equation is linear with bounded coefficients. Hence it admits a unique strong solution. Moreover, the process \(\varphi \) is invertible, with an inverse \(\psi \) satisfying suitable integrability conditions.
From Ito’s formula, we can easily check that \(d(\varphi (t,\tau )\psi (t,\tau ))=0,\) and \(\varphi (\tau ,\tau )\psi (\tau ,\tau )=I_{d},\) where \(\psi \) is the solution of the following equation
If \(\tau =0\) we simply write \(\varphi (t,0)=\varphi _{t}\) and \(\psi (t,0)=\psi _{t}\).
By the uniqueness property, it is easy to check that
then (4.4) will become
Now, if we define the adjoint process by
where
with
It follows that
Defining the Hamiltonian H from \(\left[ 0;T\right] \times {\mathbb {R}} ^{n}\times A\times {\mathbb {R}} ^{n}\) into \( {\mathbb {R}} \) by
we get from the optimality of \(u^{*}\)
By the Ito representation theorem [21], there exist two processes \(Q\in {\mathcal {M}}^{2}\) and \(R\in {\mathcal {L}}^{2}\) satisfying
hence,
Let
The above discussion will allow us to introduce the next inequality which is the first variational inequality.
where the Hamiltonian H is defined by (4.9).
4.1.2 The Second Variational Inequality
To obtain the second variational inequality of the stochastic maximum principle, we use the second perturbation (4.2) of the optimal control. Since \((u^{*},\zeta ^{*})\) is optimal control, then we have
From this inequality, we will be able to derive the second variational inequality.
Lemma 4.2
Let \(x_{t}^{(u^{*},\zeta ^{h})}\) be the trajectory associated with \((u^{*},\zeta ^{h}),\) and \(x_{t}^{*}\) the trajectory associated with \((u^{*},\zeta ^{*}),\) then the following estimate holds:
Proof
From the boundedness and continuity of \(b_{x},\) \(\sigma _{x},\) and \(f_{x}\) and by using the Burkholder–Davis–Gundy inequality for the martingale part, we get
which implies that
Therefore,
Since \({\small \upsilon (\Gamma )<\infty ,}\) by the Gronwall inequality, the result follows immediately by letting h go to zero. \(\square \)
Lemma 4.3
Under assumptions (H\(_{1})-(\)H\(_{4})\), it holds that
where \(z_{t}\) is the solution of the following equation:
Proof
Let
then,
Hence
where
Therefore,
Since \(b_{x},\) \(\sigma _{x},\) and \(f_{x}\) are bounded, then
We conclude by using the boundedness and continuity of \(b_{x},\) \(\sigma _{x},\) and \(f_{x}\), the dominated convergence theorem and \(\lim _{h\rightarrow 0}E\left| \rho _{t}^{h}\right| ^{2}=0.\) Hence, by Gronwall’s lemma, we get \(\lim _{h\rightarrow 0}E\left| y_{t}^{h}\right| ^{2}=0.\) \(\square \)
Lemma 4.4
The following inequality holds:
Proof
From (4.11), we have
From the continuity and boundedness of \(g_{x}\) and \(h_{x},\) by letting h go to zero, we can deduce the result from (4.12) and (4.13).
Now, we are able to derive the second variational inequality from (4.11). If \(\varphi (t,s)\) denotes the solution of (4.6), it is easy to check that \(z_{t}\) is given by
Replacing \(z_{t}\) by its value, we obtain the second variational inequality
where p is the adjoint process defined by (4.8). \(\square \)
Combining the first and second variational inequalities (4.10) and (4.14), we are able to state the maximum principle for strict controls.
Theorem 4.5
(the maximum principle for strict control problem) Let \((u^{*},\zeta ^{*})\) be the optimal strict control minimizing the cost functional J over \({\mathcal {U}}\) and denote by \(x^{*}\) the corresponding optimal trajectory, then the following variational inequalities hold:
where the Hamiltonian H is defined by (4.9).
4.2 The Maximum Principle for Near Optimal Controls
In this section, we establish necessary conditions of near optimality satisfied by a sequence of nearly optimal strict controls; this result is based on Ekeland’s variational principle, which is given by the next lemma.
Lemma 4.6
Let (E, d) be a complete metric space and \(f:E\rightarrow \overline{ {\mathbb {R}} }\) be lower semicontinuous and bounded from below. Given \(\varepsilon >0\), suppose \(u^{\varepsilon }\in E\) satisfies \(f(u^{\varepsilon })\) \(\le \inf (f)+\varepsilon .\) Then for any \(\lambda >0,\) there exists \(\nu \in E\) such that
-
\(f(\nu )\) \(\le f(u^{\varepsilon })\)
-
\(d(u^{\varepsilon },\nu )\le \lambda \)
-
\(f(\nu )\le f(\omega )+\frac{\varepsilon }{\lambda }d(\omega ,\nu )\) for all \(\omega \ne \nu .\)
To apply Ekeland’s variational principle, we have to endow the set \({\mathcal {U}}\) of strict controls with an appropriate metric. For any \((u,\zeta )\) and \((\nu ,\xi )\in {\mathcal {U}},\) we set
where \(P\otimes dt\) is the product measure of P with the Lebesgue measure dt.
Remark 4.7
According to [28, 29], \(({\mathcal {U}},d)\) is a complete metric space and the cost functional J is continuous from \({\mathcal {U}}\) into \( {\mathbb {R}}\).
Let \((\mu ^{*},\zeta ^{*})\in {\mathcal {R}}\) be an optimal relaxed-singular control and denote by \(x^{*}\) the trajectory of the system controlled by \((\mu ^{*},\zeta ^{*}).\) From Lemma (3.1), there exists a sequence \((u^{n})\) of strict controls such that
and
where \(\left( x^{n}\right) \) is the solution of (2.3) corresponding to \(\mu ^{n}.\)
According to the optimality of \(\mu ^{*}\) and (4.6), there exists a sequence \((\varepsilon _{n})\) of positive numbers with \(\lim _{n\rightarrow \infty }\varepsilon _{n}=0\) such that
A suitable version of Lemma (4.6) implies that, given any \(\varepsilon _{n}>0,\) there exists \((u^{n},\zeta ^{*})\in {\mathcal {U}}\) such that
Let us define the perturbations
From (4.15) we have
Using the definition of the distance d, it holds that
Finally, using the definition of \(d_{1}\) and \(d_{2}\), it holds that
where \(C_{i}\) is a positive constant.
Now, we can introduce the next theorem, which is the main result of this subsection.
Theorem 4.8
For each \(\varepsilon _{n}>0,\) there exists a nearly optimal strict control \((u^{n},\zeta )\in {\mathcal {U}}\) such that there exists a unique triple of square integrable adapted processes \((p^{n},q^{n},r^{n})\) solution of the backward SDE
such that for all \(\ (u,\zeta )\in {\mathcal {U}}\)
where \(C_{i}\) is a positive constant.
Proof
From inequality (4.16), we use the same method as in the previous subsection, to obtain the desired result (4.18).\(\square \)
Remark 4.9
The preceding theorem is interesting by itself, in the sense that in many practical situations in engineering characterizing and computing nearly optimals is sufficient.
4.3 The Relaxed Stochastic Maximum Principle
Now, we can introduce the next theorem, which is the main result of this paper.
Theorem 4.10
(The relaxed stochastic maximum principle) Let \((\mu ^{*},\zeta ^{*})\) be an optimal relaxed-singular control minimizing the functional J (., .) over \({\mathcal {R}}\), and let \(x_{t}^{*}\)be the corresponding optimal trajectory. Then there exists a unique triple of square integrable and adapted processes \((p^{*},q^{\ *},r^{*}),\) solution of the backward SDE
such that for all \((u,\zeta )\in {\mathcal {U}}\):
-
i)
\(E {\displaystyle \int \limits _{0}^{T}} \left[ H(t,x_{t}^{*},u_{t},p_{t}^{*},q^{*},r_{t}^{*}(.))- {\displaystyle \int \limits _{A_{1}}} H(t,x_{t}^{*},a,p_{t}^{*},q^{*},r_{t}^{*}(.))\mu _{t}^{*}(da)\right] dt\ge 0,\)
-
ii)
\(E\left[ {\displaystyle \int \limits _{0}^{T}} (k_{t}+G_{t}p_{t}^{*})d(\zeta _{t}-\zeta _{t}^{*})\right] \ge 0.\)
For the proof of the above theorem, we need the following stability lemma.
Lemma 4.11
Let \((p^{\ n},q^{\ n},r^{\ n})\) and \((p^{*},q^{*},r^{*}),\) be the unique solutions of (4.17) and (4.19), respectively. Then we have
Proof
Since the singular term does not affect the adjoint processes, the proof is the same as the proof of Lemma 8 in [5].\(\square \)
Proof of Theorem 4.9
The result follows immediately by letting n go to infinity in inequalities of 4.18 and using Lemma (4.11). \(\square \)
References
Bahlali, K., Chighoub, F., Djehiche, B., Mezerdi, B.: Optimality necessary conditions in singular stochastic control problems with nonsmooth data. J. Math. Anal. Appl. 355(2), 479–494 (2009)
Bahlali, S., Djehiche, B., Mezerdi, B.: The relaxed stochastic maximum principle in singular optimal control of diffusions. SIAM J. Control. Optim. 46(2), 427–444 (2007)
Bahlali, S., Mezerdi, B.: A general stochastic maximum principle for singular control problems. Electron. J. Probab. 10(30), 988–1004 (2005)
Bather, J., Chernoff, H.: Sequential decisions in the control of a spaceship, in Proceedingsof the Fifth Berkeley Symposium on Mathematical Statistics and Probability, Vol. III:Physical Sciences, University of California Press, Berkeley, CA, pp. 181–207 (1967)
Ben Gherbal, H., Mezerdi, B.: The relaxed stochastic maximum principle in optimal control of diffusions with controlled jumps. Afr. Stat. 12(2), 1287–1312 (2017)
Beneš, V.E., Shepp, L.A., Witsenhausen, H.S.: Some solvable stochastic control problems. Stoch.: An Int. J. Probab. Stoch. Processes 4, 39–83 (1980)
Bensoussan, A.: Lectures on stochastic control. In Nonlinear Filtering and Stochastic Control. Springer, Berlin Heidelberg (1983)
Borkar, V. S.: Optimal control of diffusion processes, Pitman Research Notes in Math. Series, 203. Longmann Scientific & Technical, (1989)
Boetius, F.: Singular Stochastic Control and Its Relations to Dynkin Game and Entry-ExitProblems, Ph.D. thesis. University of Konstanz, UK (2001)
Cadenillas, A., Haussmann, U.G.: The stochastic maximum principle for a singular control problem. Stoch.: An Int. J. Probab. Stoch. Processes 49(3–4), 211–237 (1994)
Chiarolla, M.B., Haussmann, U.G.: Optimal control of inflation: central bank problem. SIAM J. Control. Optim. 36, 1099–1132 (1998)
Chighoub, F., Mezerdi, B.: The relationship between the stochastic maximum principle and the dynamic programming in singular control of jump diffusions. Int. J. Stoch. Anal. 201491, 17 (2014)
Chow, P.L., Menaldi, J.L., Robin, M.: Additive control of stochastic linear systems with finite horizon. SIAM J. Control. Optim. 23(6), 858–899 (1985)
Davis, M.H.A., Norman, A.R.: Portfolio selection with transaction costs. Math. Oper. Res. 15(4), 676–713 (1990)
El Karoui, N., Du Huu, N., Jeanblanc-Picqué, M.: Compactification methods in the control of degenerate diffusions: existence of an optimal control. Stochastics 20(3), 169–219 (1987)
El Karoui, N., Karatzas, I.: Probabilistic aspects of finite-fuel, reflected follower problems. Acta Appl. Math. 11, 223–258 (1988)
Fleming, W. H.: Generalized solutions in optimal stochastic control, Differential Games and Control theory II, Proceedings of 2nd Conference, Univ. of Rhode Island, Kingston, RI, 1976, Lect. Notes in Pure and Appl. Math., 30, Marcel Dekker, New York, 147-165 (1977)
Haussmann, U.G., Suo, W.: Singular optimal stochastic controls I: existence. SIAM J. Control. Optim. 33(3), 916–936 (1995)
Haussmann, U.G., Suo, W.: Singular optimal stochastic controls II: dynamic programming. SIAM J. Control. Optim. 33(3), 937–959 (1995)
Haussmann, U.G., Suo, W.: Existence of singular optimal control laws for stochastic differential equations. Stoch.: An Int. J. Probab. Stoch. Processes 48(3–4), 249–272 (1994)
Ikeda, N., Watanabe, S.: Stochastic differential equations and diffusion processes. North Holland, (1989)
Karatzas, I.: Probabilistic aspects of finite-fuel stochastic control. Proc. Nat. Acad. Sci. USA 82, 5579–5581 (1985)
Karatzas, I., Ocone, D., Wang, H., Zervos, M.: Finite-fuel singular control with discretionary stopping. Stoch. Stoch. Rep. 71(1–2), 1–50 (2000)
Karatzas, I., Shreve, S.: E, Connections between optimal stopping and stochastic control I. Monotone follower problems. Adv. Appl. Probab. 16(1), 15–15 (1984)
Kushner, H.J.: Jump-diffusions with controlled jumps: existence and numerical methods. J. Math. Anal. Appl. 249(1), 179–198 (2000)
Kushner, H.J., Dupuis, P.G.: Numerical methods for stochastic control problems in continuous time (Vol. 24). Springer (2001)
Martins, L.F., Kushner, H.J.: Routing and singular control for queueing networks in heavy traffic. SIAM J. Control. Optim. 28, 1209–1233 (1990)
Mezerdi, B.: Necessary conditions for optimality for a diffusion with a non-smooth drift. Stochastics 24(4), 305–326 (1988)
Mezerdi, B., Bahlali, S.: Necessary conditions for optimality in relaxed stochastic control problems. Stoch. Stoch. Rep. 73(3–4), 201–218 (2002)
Acknowledgements
The second author would like to acknowledge the support provided by the Deanship of Scientific Research (DSR) at King Fahd University of Petroleum and Minerals, KSA (KFUPM), for funding this work through Project No: SB201017. The authors would like to thank the anonymous referee for useful and helpful comments, which lead to a substantial improvement of our paper.
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflict of interest
The authors have no competing interests to declare that are relevant to the content of this article.
Additional information
Communicated by Rosihan M. Ali.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.
About this article
Cite this article
Ben-Gherbal, H., Mezerdi, B. The Relaxed Stochastic Maximum Principle in Singular Optimal Control of Jump Diffusions. Bull. Malays. Math. Sci. Soc. 47, 34 (2024). https://doi.org/10.1007/s40840-023-01632-w
Received:
Revised:
Accepted:
Published:
DOI: https://doi.org/10.1007/s40840-023-01632-w
Keywords
- Stochastic differential equation
- Optimal control
- Jump process
- Relaxed control
- Singular control
- Stochastic maximum principle