Abstract
In this manuscript we consider optimal control problems of stochastic differential equations with delays in the state and in the control. First, we prove an equivalent Markovian reformulation on Hilbert spaces of the state equation. Then, using the dynamic programming approach for infinite-dimensional systems, we prove that the value function is the unique viscosity solution of the infinite-dimensional Hamilton-Jacobi-Bellman equation. We apply these results to problems coming from economics: stochastic optimal advertising problems and stochastic optimal investment problems with time-to-build.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
In this manuscript, we consider a class of stochastic optimal control problems with infinite horizon with delays in the state and in the control. Precisely, the state equation is a stochastic delay differential equation (SDDE) in \(\mathbb {R}^n\) of the form
where u(t) is the control process taking values in a bounded subset U of \(\mathbb {R}^p\), \(\eta _0 \in \mathbb {R}^n,\) \(\eta _1 \in L^2([-d,0];\mathbb {R}^n)\) are the initial conditions of y(t), \(\delta \in L^2([-d,0); U)\) is the initial condition of u(t), W(t) is a Brownian motion with values in \(\mathbb {R}^q\). Denoting \(\eta =(\eta _0,\eta _1)\), the goal is to minimize, over all \(u(\cdot )\in \mathcal {U}\), a functional of the form
The presence of delays is the crucial aspect of (SDDE): these appear linearly via the integral terms
where the first represents the delay in the state and the second the one in the control. For a complete picture of optimal control problems with delays we refer the reader to de Feo et al. (2023), while here we will only recall some results. A similar problem with delays only in the state was treated by means of the dynamic programming method via viscosity solutions in de Feo et al. (2023). In this paper we aim to extend some of these results to the case of delays also in the control.
The main difficulty for delay problems is in the lack of Markovianity, which prevents a direct application of the dynamic programming method, e.g. see (de Feo et al. 2023). The approach we follow here, similarly to de Feo et al. (2023), is to lift the state equation to an infinite-dimensional Banach or Hilbert space (depending on the regularity of the data), in order to regain Markovianity.Footnote 1 This is done at a cost of of moving to infinite-dimension.
In de Feo et al. (2023) (where delays are only in the state) the classical approach of rewriting the state equation in the Hilbert space
was used. However when delays in the control appear this procedure has to be extended carefully. One possible way would be to use the extended delay semigroup (including the control in the state-space of the delay semigroup), e.g. see (Bensoussan et al. 2007, Part II, Chapter 4) for deterministic problems. This approach brings the complication of having an unbounded control operator ("boundary control"). However, when the delays appear in a linear way in the state equation, an alternative approach can be used. Such approach was proposed by Vinter and Kwong (1981) for deterministic control problems (see also (Bensoussan et al. 2007, Part II, Chapter 4)) and extended in Gozzi and Marinelli (2006) to the stochastic setting. In Gozzi and Marinelli (2006) a linear state equation with additive noise is considered and an equivalent abstract representation of the state equation in the Hilbert space X is proved. In this paper we generalize this result of Gozzi and Marinelli (2006) to the following nonlinear state equation with multiplicative noise (note that such state equation is more general than (1.1)):
In this case the abstract state equation in X is of the form
for suitable operators A, M and coefficients \(b, \sigma \). See Theorem 3.2 for the equivalent abstract representation of the state equation in X.
Going back to our original problem, (1.1) can be rewritten on X as
for suitable \(\mathcal {A},f, \sigma \) (see (4.1)). The functional \(\mathcal {J}(\eta ,\delta ;u(\cdot ))\) is rewritten as
for a suitable cost function L (see (3.8)). Having lifted the problem in the space X we are in a similar setting to the one of de Feo et al. (2023), hence we wish to proceed in a similar way. Indeed we want to use the theory of viscosity solutions in Hilbert spaces (see (Fabbri et al. 2017, Chapter 3)) in order to treat the optimal control problem. However, \(\mathcal {A},f,\sigma \) in (1.2) have a different structure than the ones in de Feo et al. (2023). Indeed in de Feo et al. (2023) the unbounded operator is the classical delay operator, while \(\mathcal {A}\) here it is (up to a bounded perturbation) its adjoint and the coefficient f here has a non-zero \(L^2-\)component.
At this point, in order to use the theory of viscosity solutions in Hilbert spaces of (Fabbri et al. 2017, Chapter 3), we rewrite the state equation by introducing a maximal dissipative operator \(\tilde{\mathcal {A}}\) in the state equation, see Proposition 4.1. Next we introduce an operator B satisfying the so called weak B-condition (which is crucial in the theory of viscosity solutions in Hilbert spaces), see Proposition 4.3. Hence, we prove that the data of the problem satisfy some regularity conditions with respect to the norm induced by the operator \(B^{1 / 2}\), see Lemma 4.6. This enables us to characterize the value function of the problem as the unique viscosity solution of the following fully non linear second order infinite-dimensional HJB equation
where H is the Hamiltonian. See Theorem 5.4 for such result. To the best of our knowledge, this is the first existence and uniqueness result for fully nonlinear HJB equations in Hilbert spaces related to stochastic optimal control problems with delays in the state and in the control. This extends the corresponding result of de Feo et al. (2023) to the case of delays (also) in the control. Moreover in the present paper the delay kernels \(a_1,p_1\) are such that their rows \(a_1^j,p_1^j \in L^2([-d,0]; \mathbb {R}^n)\) for every \(j \le n\). Instead in de Feo et al. (2023) a higher regularity of \(a_1,a_2\) (where \(a_2\) is the delay kernel associated to the diffusion) is required, i.e. \(a_1^j,a_2^j \in W^{1,2}([-d,0]; \mathbb {R}^n)\) and \(a_1^j(-d)=a_2^j(-d)=0\) for every \(j \le n\) (of course \(p_1=0\), i.e. only delays in the state are present). However we remark that the structure of the state equation in de Feo et al. (2023), with delays only in the state, is more general.
For a complete picture of the literature related to optimal control problems with delays only in the state we refer to de Feo et al. (2023), here we only list some results: for stochastic optimal control problems with delays only in the state e.g. see (Biffis et al. 2020; Biagini et al. 2022; de Feo and Święch 2023; Djehiche et al. 2022; Di Giacinto et al. 2011; Federico 2011; Federico and Tankov 2015; Fuhrman et al. 2010; Masiero and Tessitore 2022; Pang et al. 2019). For deterministic optimal control problems with delays only in the state e.g. see (Carlier and Tahraoui 2010; Federico et al. 2010, 2011).
Stochastic differential equations with delays also in the control are more difficult since in this case the so called structure condition, that is the requirement that the range of the control operator is contained in the range of the noise, does not hold, e.g. see (Fabbri et al. 2017, Subsection 2.6.8). This fact, together with the lack of smoothing of the transition semigroup associated to the linear part of the equation (this is a common feature also in problems with delays only in the state), prevent the use of standard techniques, based on mild solutions and on backward stochastic differential equations. However stochastic differential equations with delays only in the control, linear structure of the state equation and additive noise (the HJB equation is semilinear) were completely solved in Gozzi and Masiero (2017, 2017, 2021) by means of a partial smoothing property for the Ornstein-Uhlenbeck transition semigroup which permitted to apply a variant of the approach via mild solutions in the space of continuous functions. See also (Gozzi and Masiero 2023), [32], where this approach is extended to stochastic optimal control problems with unbounded control operators and applications to problems with delays only in the control (with delay kernel being a Radon measure) are given. Finally we refer to Federico and Tacconi (2014) for a deterministic optimal control problem with delays only in the control and linear structure of the state equation solved by means of viscosity solutions in the space \(\mathbb {R} \times W^{1,2}([-d,0])\) (see also Remark 4.5).
At the end of the manuscript, we provide applications of our results to problems coming from economics. We consider a stochastic optimal advertising problem with delays in the state and in the control and controlled diffusion, generalizing the one from (Gozzi and Marinelli 2006; Gozzi et al. 2009) (see also Nerlove and Arrow 1962 for the original deterministic model). We characterize the value function as the unique viscosity solution of the fully non-linear HJB equation. We recall that, in the stochastic setting with additive noise, in Gozzi and Marinelli (2006), Gozzi et al. (2009) a verification theorem was proved in the context of classical solutions and optimal feedback strategies were derived. Moreover an explicit (classical) solution of the HJB equation was derived in a specific case. We also recall that in de Feo et al. (2023) the case with no delays in the control (i.e. \(p_1=0\)) was treated via viscosity solutions.
Finally, we consider a stochastic optimal investment problem with with time-to-build, inspired by (Fabbri and Federico 2014, p. 36) (see also e.g. (Bambi et al. 2012, 2017) for similar models in the deterministic setting). We characterize the value function as the unique viscosity solution of the fully non-linear HJB equation.
The paper is organized as follows. In Sect. 2 we introduce the problem and state the main assumptions. In Sect. 3 we prove an equivalent infinite-dimensional formulation for a more general state equation and we rewrite the problem in an infinite dimensional setting. In Sect. 4 we prove some preliminary estimates for solutions of the state equation and the value function. In Sect. 5 we introduce the notion of viscosity solution of the HJB equation and state a theorem about the existence and uniqueness of viscosity solutions, and characterize the value function as the unique viscosity solution. In Sect. 6 we provide applications to problems coming from economics: stochastic optimal advertising models and stochastic investment models with time-to-build.
2 Setup and assumptions
We denote by \(M^{m \times n}\) the space of real valued \(m \times n\)-matrices and we denote by \(|\cdot |\) the Euclidean norm in \(\mathbb {R}^{n}\) as well as the norm of elements of \(M^{m \times n}\) seen as linear operators from \(\mathbb {R}^{m}\) to \(\mathbb {R}^{n}\). We will write \(x \cdot y\) for the inner product in \(\mathbb {R}^n\).
Let \(d>0\). We consider the standard Lebesgue space \(L^2:=L^2([-d,0];\mathbb {R}^n)\) of square integrable functions from \([-d,0]\) to \(\mathbb {R}^{n}\), denoting by \(\langle \cdot ,\cdot \rangle _{L^{2}}\) the inner product in \(L^2\) and by \(|\cdot |_{L^{2}}\) the norm. We also consider the standard Sobolev space \(W^{1,2}:=W^{1,2}([-d,0]; \mathbb {R}^n)\) of functions in \(L^{2}\) admitting weak derivative in \(L^{2}\), endowed with the inner product \(\langle f,g\rangle _{W^{1,2}}:= \langle f,g\rangle _{L^{2}}+\langle f',g'\rangle _{L^{2}}\) and norm \(|f|_{W^{1,2}}:=(|f|^2_{L^{2}}+|f'|^2_{L^{2}})^{\frac{1}{2}}\), which render it a Hilbert space. It is well known that the space \(W^{1,2}\) can be identified with the space of absolutely continuous functions from \([-d,0]\) to \(\mathbb {R}^{n}\).
Let \(\tau =\left( \Omega , \mathcal {F},\left( \mathcal {F}_t\right) _{t \ge 0}, \mathbb {P}, W\right) \) be a reference probability space, that is \((\Omega , \mathcal {F}, \mathbb {P})\) is a complete probability space, \(W=(W(t))_{t \ge 0}\) is a standard \(\mathbb {R}^q\)-valued Wiener process, \(W(0)=0\), and \(\left( \mathcal {F}_t\right) _{t \ge 0}\) is the augmented filtration generated by W. We consider the following controlled stochastic delay differential equation (SDDE)
where
-
(i)
given a bounded measurable set \(U\subset \mathbb {R}^{p}\), \(u(\cdot )\) is the control process lying in the set
$$\begin{aligned} \mathcal {U}_\tau= & {} \{u(\cdot ): \Omega \times [0,+\infty ) \rightarrow U: \ u(\cdot ) \ \text{ is } \ (\mathcal {F}_t)\text{-progressively } \text{ measurable } \text{ and } \text{ integrable } \text{ a.s. } \}; \end{aligned}$$ -
(ii)
\(\eta _0 \in \mathbb {R}^n\) and \(\eta _1 \in L^2\) are the initial conditions of the state y;
-
(iii)
\(\delta \in L^2([-d,0); U)\) is the initial condition of the control \(u(\cdot );\)
-
(iv)
\(b_0 :U \rightarrow \mathbb {R}^n\), \(\sigma _0 :U \rightarrow M^{n\times q}\);
-
(v)
\(a_{1}, p_1:[-d,0]\rightarrow M^{n \times n}\) and if \(a_{1}^{j},p_1^j\) are the j-th row of \(a_1(\cdot )\), \(p_1(\cdot )\) respectively for \(j=1,...,n\), then \(a_1^{j}, p_1^{j}\in L^2([-d,0]; \mathbb {R}^n)\).
Remark 2.1
Similarly to de Feo et al. (2023) we cannot treat the case of pointwise delay (e.g. \(a_1,p_1=\delta _{-d}\) the Dirac’s Delta). In de Feo et al. (2023) a higher regularity of \(a_1,a_2\) (where \(a_2\) is the delay kernel associated to the diffusion) is required, i.e. \(a_1^j,a_2^j \in W^{1,2}([-d,0]; \mathbb {R}^n)\) and \(a_1^j(-d)=a_2^j(-d)=0\) for every \(j \le n\) (of course \(p_1=0\), i.e. in de Feo et al. (2023) only delays in the state are present). See also (Federico et al. 2010, 2011) for similar restrictions in deterministic problems. Here, instead, we require less regularity, i.e. \(a_1^{j}, p_1^{j}\in L^2([-d,0]; \mathbb {R}^n)\).
We will assume the following conditions.
Assumption 2.2
\(b_0 :U \rightarrow \mathbb {R}^n\), \(\sigma _0 :U \rightarrow M^{n\times q}\) are continuous and bounded.
Under Assumption 2.2, by (Revuz and Yor 1999, Theorem IX.2.1), for each initial data \(\eta :=(\eta _0,\eta _1)\in \mathbb {R}^{n}\times L^2([-d,0];\mathbb {R}^n)\), \(\delta \in L^2([-d,0);U)\), and each control \(u(\cdot )\in \ \mathcal {U}\), there exists a unique (up to Indistinguishability) strong solution to (3.1) and this solution admits a version with continuous paths that we denote by \(y^{\eta ,\delta ;u}\).
We consider the following infinite horizon optimal control problem. Given \(\eta =(\eta _0,\eta _1)\in \mathbb {R}^{n}\times L^2\), \(\delta \in L^2([-d,0); U)\), we define a cost functional of the form
where \(\rho >0\) is the discount factor, \(l :\mathbb {R}^n \times U \rightarrow \mathbb {R} \) is the running cost. As in (Fabbri et al. 2017, p. 98, Equation (2.8)), we define
where the union is taken over all reference probability spaces \(\tau \). The goal is to minimize \(\mathcal {J}(\eta ,\delta ;u(\cdot ))\) over all \(u(\cdot )\in \mathcal {U}\). This is a standard setup of a stochastic optimal control problem (see Yong and Zhou 1999; Fabbri et al. 2017) used to apply the dynamic programming approach. We remark (see e.g. Fabbri et al. 2017, Section 2.3.2) that
for every reference probability space \(\tau \) so the optimal control problem is in fact independent of the choice of a reference probability space.
Assumption 2.3
-
(i)
\(l :\mathbb {R}^n \times U \rightarrow \mathbb {R} \) is continuous.
-
(ii)
There exist constants \(K,m>0\), such that
$$\begin{aligned} |l(z,u)| \le K(1+|z|^m) \ \ \ \forall y\in \mathbb {R}^n, \ \forall u \in U. \end{aligned}$$(2.3) -
(iii)
There exists a local modulus of continuity for \(l(\cdot ,u)\), uniform in \(u\in U\), i.e. for each \(R>0\), there exists a nondecreasing concave function \(\omega _{R}:\mathbb {R}^{+}\rightarrow \mathbb {R}^{+}\) such that \(\lim _{r\rightarrow 0^{+}} \omega _{R}(r)=0\) and
$$\begin{aligned} |l(z,u)-l(z',u)| \le \omega _R( |z-z'|) \end{aligned}$$(2.4)for every \(z,z' \in \mathbb {R}^n\) such that \(|z|,|z'| \le R\) and every \(u \in U\).
We will show, suitably reformulating the state equation in an infinite dimensional framework, that the cost functional is well defined and finite for a sufficiently large discount factor \(\rho >0\).
Throughout the paper we will write \(C>0,\omega , \omega _R\) to indicate, respectively, a constant, a modulus continuity, and a local modulus of continuity, which may change from place to place if the precise dependence on other data is not relevant. The equality involving random variables will be intended \(\mathbb {P}-\)a.s..
3 Infinite dimensional Markovian representation
The optimal control problem at hand is not Markovian due to the delay. In order to regain Markovianity and approach the problem by Dynamic Programming we reformulate the state equation in an infinite-dimensional space generalizing a well-known procedure, see (Bensoussan et al. 2007, Part II, Chapter 4), Vinter and Kwong (1981) for deterministic delay equations and Gozzi and Marinelli (2006) for the stochastic case with linear state equation and additive noise.
In this section, in place of (2.1), we will consider the following more general state equation:
where, in this case, \(b_0 :\mathbb {R}^n \times U \rightarrow \mathbb {R}^n\), \(\sigma _0 :\mathbb {R}^n \times U \rightarrow M^{n\times q}\), while all the other terms satisfy the same conditions as in (2.1). In this setting we consider the following assumptions.
Assumption 3.1
\(b_0 :\mathbb {R}^n \times U \rightarrow \mathbb {R}^n\), \(\sigma _0 :\mathbb {R}^n \times U \rightarrow M^{n\times q}\) are continuous and there exist constants \(L,C>0\) such that
for every \(y,y' \in \mathbb {R}^n\), \(u \in U\).
Under Assumption 3.1, by (Revuz and Yor 1999, Theorem IX.2.1), for each initial data \(\eta :=(\eta _0,\eta _1)\in \mathbb {R}^{n}\times L^2([-d,0];\mathbb {R}^n)\), \(\delta \in L^2([-d,0);U)\), and each control \(u(\cdot )\in \ \mathcal {U}\), there exists a unique (up to indistinguishability) strong solution to (3.1) and this solution admits a version with continuous paths that we denote by \(y^{\eta ,\delta ;u}\).
We define \( X:= \mathbb {R}^n \times L^2 \). An element \(x\in X\) is a couple \(x= (x_0,x_1)\), where \(x_0 \in \mathbb {R}^n\), \(x_{1}\in L^{2}\); sometimes, we will write \(x=\begin{bmatrix} x_0\\ x_1 \end{bmatrix}.\) The space X is a Hilbert space when endowed with the inner product
The induced norm, denoted by \(|\cdot |_X\), is then
For \(R>0\), we set the following notation for the open balls of radius R in X, \(\mathbb {R}^n,\) and \(L^2\), respectively:
We denote by \(\mathcal {L}(X)\) the space of bounded linear operators from X to X, endowed with the operator norm
An operator \(L \in \mathcal {L}(X)\) can be seen as
where \(L_{00} :\mathbb {R}^n \rightarrow \mathbb {R}^n\), \(L_{01} :L^2 \rightarrow \mathbb {R}^n\), \(L_{10} :\mathbb {R}^n \rightarrow L^2\), \(L_{11} :L^2 \rightarrow L^2\) are bounded linear operators. Moreover, given two separable Hilbert spaces \((H, \langle \cdot , \cdot \rangle _H), (K, \langle \cdot , \cdot \rangle _K)\), we denote by \(\mathcal {L}_1(H,K)\) the space of trace-class operators endowed with the norm
We also denote by \(\mathcal {L}_2(H,K)\) the space of Hilbert-Schmidt operators from H to K endowed with the norm
When \(H=K\) we simply write \(\mathcal {L}_1(H)\), \(\mathcal {L}_2(H)\). We denote by S(H) the space of self-adjoint operators in \(\mathcal {L}(H)\). If \(Y,Z\in S(H)\), we write \(Y\le Z\) if \(\langle Yx,x \rangle \le \langle Zx,x \rangle \) for every \(x \in H\).
Let us define the unbounded linear operator \(A:D(A)\subset X\rightarrow X\) as follows:
The adjoint of A is the operator \(A^*:D(A^*)\subset X\rightarrow X\) (e.g., see (Federico et al. 2010, Proposition 3.4))
Note that \(A^*\) is the generator of the delay semigroup, see, e.g., (Bensoussan et al. 2007, Part II, Chapter 4). For problems with delays (also) in the control appearing in a linear way in the state equation, its adjoint, i.e. A, is used to reformulate the problem in the space X (see, e.g., (Bensoussan et al. 2007, Part II, Chapter 4)). Indeed, A is the generator of a strongly continuous semigroup \(e^{tA} \) on X, whose explicit expression (see, e.g., (Federico and Tacconi 2014, Eq. (73))) is
where \(\Phi (t)\) is the semigroup of truncated right shift in \(L^{2}\):
Now define \(b :X \times U \rightarrow X\) by
and \(\sigma :X \times U \rightarrow \mathcal {L}(\mathbb {R}^q,X)\) by
Consider the infinite dimensional SDE
By (Fabbri et al. 2017, Theorem 1.127), for each \(u(\cdot ) \in \mathcal {U}\), (3.3) admits a unique mild solution; that is, there exists a unique progressively measurable \(X-\)valued process \(Y=(Y_{0},Y_{1})\) such that
Define the linear operator \(M :X \times L^2([-d,0],U) \rightarrow X\) by
where
Theorem 3.2
Let Assumption 3.1 hold. We have the following claims.
-
(i)
Let \(Y^{x,u}\) be the unique mild solution to (3.3) with initial datum \(x \in X\) and control \(u(\cdot )\in \mathcal {U}\). For every \(t \ge d\)
$$\begin{aligned} Y^{x,u}(t) =(Y^{x,u}_0(t),Y^{x,u}_1(t))= M \left( \left( Y^{x,u}_0(t),Y^{x,u}_0(t+\cdot ) \right) ,u(t+\cdot ) \right) .\end{aligned}$$ -
(ii)
Let \(y^{\eta ,\delta ;u}\) be the solution to SDDE (3.1) with initial data \(\eta ,\delta \) and under the control \(u(\cdot )\in \mathcal {U}\), and let \( x = M(\eta _{0},\eta _{1},\delta ). \) Then, for every \(t\ge 0\),
$$\begin{aligned} Y^{x,u}(t) =(Y^{x,u}_0(t),Y^{x,u}_1(t)) = M \left( \left( y^{\eta ,\delta ,u}(t),y^{\eta ,\delta , u}(t+\cdot ) \right) ,u(t+\cdot ) \right) . \end{aligned}$$In particular, for every \(t\ge 0\),
$$\begin{aligned} y^{\eta ,\delta ,u}(t)=Y^{x,u}_0(t). \end{aligned}$$
Proof
-
(i)
Using (3.2), we can rewrite the two components of (3.4) as
$$\begin{aligned}&\begin{bmatrix} Y_0^{x,u}(t)\\ Y_1^{x,u}(t) \end{bmatrix}\\&\quad =\begin{bmatrix} x_0 +\int _{(-t) \vee (-d)}^0 x_1(\xi )d \xi + \int _0^t \left[ b_0(Y_0^{x,u}(s),u(s))\right. \\ \left. + \int _{(-(t-s)) \vee (-d)}^0 a_1(\xi )Y_0^{x,u}(s)+ p_1(\xi )u(s) d \xi \right] ds\\ + \int _0^t \sigma _0(Y_0^{x,u}(s),u(s)) dW(s)\\ \Phi (t) x_{1}+\int _{0}^{t}\Phi (t-s) a_{1}Y_{0}^{x,u}(s)d s+\int _{0}^{t}\Phi (t-s) p_{1} u(s)d s \end{bmatrix}. \end{aligned}$$Then,
$$\begin{aligned} Y_{1}^{x,u}(t)(\xi )&= 1_{[-d,0]}(\xi -t) x_1(\xi -t)+\int _{0}^{t}1_{[-d,0]}(\xi -t+s) a_{1}(\xi -t+s) Y_{0}^{x,u}(s) d s\nonumber \\&\quad +\int _{0}^{t} 1_{[-d,0]}(\xi -t+s) p_{1}(\xi -t+s) u(s) ds \nonumber \\&= 1_{[-d,0]}(\xi -t) x_1(\xi -t)+\int _{(\xi -t)\vee -d}^{\xi } a_{1}(\alpha ) Y_{0}^{x,u}(t+\alpha -\xi ) d \alpha \nonumber \\&\quad +\int _{(\xi -t)\vee -d}^{\xi } p_{1}(\alpha ) u(t+\alpha -\xi ) d\alpha . \end{aligned}$$(3.5)For \(t \ge d\), we have \(\xi -t \le -d\), so that
$$\begin{aligned} Y_{1}^{x,u}(t)(\xi )= & {} \int _{-d}^{\xi } a_{1}(\alpha ) Y_{0}^{x,u}(t+\alpha -\xi ) d \alpha +\int _{-d}^{\xi } p_{1}(\alpha ) u(t+\alpha -\xi ) d \alpha \\= & {} m\left( Y^{x,u}_{0}(t+\xi ), u(t+\xi ) \right) \end{aligned}$$from which we get the first claim.
-
(ii)
Let \(x=(x_0,x_1)=M\left( \eta _{0}, \eta _1, \delta \right) \). For \(\xi -t \in [-r, 0]\), \(\xi \in [-r, 0]\) we have:
$$\begin{aligned} x_1(\xi -t)=Y^{x,u}_{1}(0)(\xi -t){} & {} =\int _{-d}^{\xi -t} a_{1}(\alpha ) \eta (t+\alpha -\xi ) d \alpha \\{} & {} \quad \ +\int _{-d}^{\xi -t} p_{1}(\alpha ) \delta (t+\alpha -\xi ) d \alpha , \end{aligned}$$so that inserting it into (3.5) we have:
$$\begin{aligned} Y_{1}^{x,u}(t)(\xi )=\int _{-d}^{\xi } a_{1}(\alpha ) \tilde{Y}^{x,u}_{0}(t+\alpha -\xi ) d \alpha +\int _{-d}^{\xi } p_{1}(\alpha ) u(t+\alpha -\xi ) d \alpha , \end{aligned}$$(3.6)where we have defined \({\tilde{Y}}^{x,u}_0\) to be the extension of \(Y^{x,u}_0\) to \([-d,0)\) by
$$\begin{aligned} \tilde{Y}^{x,u}_{0}(s)= {\left\{ \begin{array}{ll}\eta _1(s), &{} s \in [-d, 0), \\ Y^{x,u}_{0}(s), &{} s \ge 0.\end{array}\right. } \end{aligned}$$From (3.6) we have:
$$\begin{aligned} Y^{x,u}(t)&=(Y^{x,u}_0(t),Y^{x,u}_1(t))=M(\tilde{Y}_{0}^{x,u}(t),\tilde{Y}_{0}^{x,u}(t+\cdot ),u(t+\cdot ))\nonumber \\&=(\tilde{Y}_{0}^{x,u}(t),m(\tilde{Y}_{0}^{x,u}(t+\cdot ),u(t+\cdot )). \end{aligned}$$(3.7)To conclude the proof, by uniqueness of strong solutions to (3.1), we need to prove that \({\tilde{Y}}_{0}^{x,u}\) satisfies (3.1). On the other hand, by (Gawarecki and Mandrekar 2010, Theorem 3.2), \(Y^{x,u}(t)\) is also a weak solution of (3.3), i.e. it satisfies
$$\begin{aligned} \langle Y^{x,u}(t),h\rangle _X&= \langle x,h\rangle _X +\int _0^t \langle Y^{x,u}(t),A^* h\rangle _X ds +\int _0^t \langle b(Y^{x,u}(t),u(t)), h\rangle _X ds\\&\quad +\int _0^t \langle \sigma (Y^{x,u}(t),u(t)), h \rangle _X dW(s), \ \ \ \ \forall t \ge 0, \ \forall h \in D(A^*). \end{aligned}$$For \(k\in \mathbb {N}\setminus \{0\}\), let \(h^k=(1,h_1^k)\) be defined by
$$\begin{aligned} h_1^k(\xi )= \int _{-d}^\xi k \textbf{1}_{[-1/k,0]}(r)dr, \ \ \ \ \ \xi \in [-d,0]. \end{aligned}$$Then,
$$\begin{aligned} {\left\{ \begin{array}{ll} h^k\in D(A^*)\ \ \ \forall k\in \mathbb {N}\setminus \{0\},\\ h_1^k(\xi )\rightarrow 0 \ \ \text{ as } \ k \rightarrow \infty \ \ \text{ for } \text{ a.e. } \ \xi \in [-d,0],\\ \lim _{k \rightarrow \infty }\int _{-d}^0 \frac{dh_1^k}{d\xi }(\xi ) z(\xi ) d \xi = z(0), \ \ \ \forall z \in {C}([-d,0];\mathbb {R}^{n}). \end{array}\right. } \end{aligned}$$Therefore, for every \(t \ge 0\), we have:
$$\begin{aligned}&Y^{x,u}_0(t)+\left\langle Y^{x,u}_1(t),h_1^k \right\rangle _{L^2}\\&\quad = x_0+ \langle x_1,h_1^k \rangle _{L^2}+ \int _0^t\left\langle Y^{x,u}_1(s),\frac{dh_1^k}{d\xi } \right\rangle _{L^2} ds + \int _0^t b_0(Y^{x,u}_0(s),u(s))ds\\&\qquad +\int _0^t \left\langle a_1 Y^{x,u}_0(s)+p_1 u(s) , h_1^k\right\rangle _{L^2} ds + \int _0^t \sigma _0(Y^{x,u}_0(s),u(s))dW(s). \end{aligned}$$Note that, for every \(t \ge 0\), on a has \(Y_1(t)(\cdot ) \in {C}([-d,0];\mathbb {R}^{n})\), since \(Y_1(t)(\cdot )\) is the sum of the convolutions of \(L^2\)-functions. Then, taking \(k\rightarrow + \infty \) in the equation above, we get
$$\begin{aligned} Y^{x,u}_0(t)&= x_0+ \int _0^t Y^{x,u}_1(s)(0)ds + \int _0^t b_0(Y^{x,u}_0(s),u(s))ds\\&\quad + \int _0^t \sigma _0(Y^{x,u}_0(s),u(s))dW(s). \end{aligned}$$By (3.6), with \(\xi =0\) we have:
$$\begin{aligned} Y^{x,u}_{1}(s)(0) =\int _{-d}^{0} a_{1}(\xi ) {\tilde{Y}}^{x,u}_{0}(s+\xi ) d \xi +\int _{-d}^{0} p_{1}(\xi ) u(s+\xi ) d \xi , \end{aligned}$$so that, for every \(t \ge 0\),
$$\begin{aligned} Y^{x,u}_0(t)&= x_0+ \int _0^t \left[ \int _{-d}^{0} a_{1}(\xi ) {\tilde{Y}}^{x,u}_{0}(s+\xi ) d \xi +\int _{-d}^{0} p_{1}(\xi ) u(s+\xi ) d \xi \right] ds \\&\quad + \int _0^t b_0(Y^{x,u}_0(s),u(s))ds + \int _0^t \sigma _0(Y^{x,u}_0(s),u(s))dW(s). \end{aligned}$$Recalling the definition of \({\tilde{Y}}_0^{x,u(\cdot )}\), this says that \({\tilde{Y}}_{0}^{x,u(\cdot )}\) satisfies (3.1), so we conclude.\(\square \)
3.1 Objective functional
Using Theorem 3.2, we can give a Markovian reformulation on the Hilbert space X of the optimal control problem.
We present such result for an optimal control problem with the more general state equation (3.1). Consider the functional \(\mathcal {J}\) defined by (2.2) with \(y^{\eta ,u(\cdot ),\delta (\cdot )}\) being the solution of (3.1) (in place of (2.1)). Denoting by \(Y^{x,u}\) a mild solution of (3.3) for general initial datum \(x \in X\) and control \(u(\cdot ) \in \mathcal {U}\) and introducing the functional
where
the original functional \(\mathcal {J}\) and J are related through
We then consider the problem of optimizing J under (3.3) and define the value function V for this problem:
For what we said, an optimal control \(u^*(\cdot )\in \mathcal {U}\) for the functional \(J(x;\cdot )\) with \(x=M(\eta ,\delta )\) is also optimal for \(\mathcal {J}(\eta ,\delta ;\cdot )\). Hence, from now on, we focus on the optimization problem (3.9).
4 B-continuity
In order to get B-continuity of the value function V, needed to employ the theory of viscosity solutions in infinite dimension, we consider the simpler state equation (2.1). In this case, the state equation (2.1) can be rewritten in infinite dimension as
where
and
Indeed, since \(\mathcal {A}\) is the sum of A with a bounded linear operator, by (Engel and Nagel 2000, Corollary 1.7), we have that (3.3) with specifications
and (4.1) are equivalent and have the same (unique) mild solution \(Y^{x,u}(t)\).
4.1 Reformulation with a maximal dissipative operator
The aim of this subsection is to rewrite (4.1) with a maximal dissipative operator \(\tilde{\mathcal {A}}\) in place of \(\mathcal {A}\). The need for that is that we want to use the viscosity solutions theory in infinite dimension to treat the HJB equation associated to the optimal control problem, which requires for the comparison theorem the presence of a maximal dissipative operator in the equation (see (Fabbri et al. 2017, Chapter 3)). The operator \(\tilde{\mathcal {A}}\) is constructed by means of a suitable shift of the operator \(\mathcal {A}\).
Proposition 4.1
There exists \(\mu _0 >0\) such that \(\tilde{\mathcal {A}}_\mu :D(\tilde{\mathcal {A}}_\mu ) =D(A), \subset X \rightarrow X\) defined by
is maximal dissipative for every \(\mu \ge \mu _0\).
Proof
Step 1. We prove that \(\tilde{\mathcal {A}}_\mu \) is dissipative. Let \(x \in D(A)=\left\{ x \in X: x_1 \in W^{1,2}([-d,0]; \mathbb {R}^n), \ x_1(-d)=0\right\} ,\) then
where \(\mu _0:=\max \left\{ |a_0|+1,\frac{1}{2} |a_1|^2_{L^2}\right\} \). This implies that \(\tilde{\mathcal {A}}_\mu \) is dissipative for every \(\mu \ge \mu _0\).
Step 2. We prove that \(\tilde{\mathcal {A}}_\mu \) is maximal for \(\mu \ge \mu _0\). For that we show that there exists \(\lambda >0\) such that \(R(\lambda I - \tilde{\mathcal {A}}_\mu )=X\). This means that, for every \(y\in X\), there exists a solution \(x \in D(\tilde{\mathcal {A}}_\mu )\) to the equation
Fix \(\lambda >0\). Let then \(y \in X\). The equation above rewrites as
This system is uniquely solvable. Indeed, from \((\lambda +\mu ) x_0-a_0 x_0 - x_1 (0) =y_0\), we get uniquely
On the other hand,
yields the unique solution
Taking \(\xi =0\) in this equality and equating with (4.4), we have:
Then, for \(\mu \ge \mu _0\)
Therefore, we have proved that, for every \(\mu \ge \mu _0\) and \(y \in X\), there exists a unique solution \(x \in D(\tilde{\mathcal {A}}_\mu )\) to (4.3) given by
The claim follows. \(\square \)
Now, we fix \(\mu > \mu _0\) and denote
We may rewrite SDE (4.1) as
where
Similarly to before, since the operator \(\tilde{\mathcal {A}}\) is the sum of A with a bounded operator, by (Engel and Nagel 2000, Corollary 1.7), we have that (4.6) is equivalent to (4.1) (and so also to (3.3) with the condition (4.2)) and all them have the same (unique) mild solution, that in terms of \(\tilde{\mathcal {A}}\) writes as
4.2 Weak B-condition
In this subsection, we recall the concept of weak B-condition for the maximal dissipative operator \(\tilde{\mathcal {A}}\) and introduce an operator B satisfying it. This concept is fundamental in the theory of viscosity solutions in Hilbert spaces, see (Fabbri et al. 2017, Chapter 3), which will be used in this paper.
Definition 4.2
(Fabbri et al. 2017, Definition 3.9) We say that \(B \in \mathcal {L}(X)\) satisfies the weak B-condition for \(\tilde{\mathcal {A}}\) if the following hold:
-
(i)
B is strictly positive, i.e. \(\langle Bx,x \rangle _{X} >0\) for every \(x \ne 0\);
-
(ii)
B is self-adjoint;
-
(iii)
\(\tilde{C}^* B \in \mathcal {L}(X)\);
-
(iv)
There exists \(c_0 \ge 0\) such that
$$\begin{aligned} \langle \tilde{\mathcal {A}}^* Bx,x \rangle _{X} \le c_0 \langle Bx,x \rangle _{X}, \quad \forall x \in X. \end{aligned}$$
Let \(\tilde{\mathcal {A}}^{-1}\) be the inverse of the operator \(\tilde{\mathcal {A}}\). Its explicit expression can be derived as in the proof of Proposition 4.1:
Notice that \(\tilde{\mathcal {A}}^{-1} \in \mathcal {L}(X)\) and it is compact. Define now the compact operator
We are going to show that B satisfies the weak-B condition for \(\tilde{\mathcal {A}}\).
Proposition 4.3
B defined in (4.9) satisfies the weak-B condition for \(\tilde{\mathcal {A}}\) with \(c_0=0\).
Proof
Clearly, \(B \in \mathcal {L}(X)\), \(\tilde{\mathcal {A}}^* B= \tilde{\mathcal {A}}^{-1} \in \mathcal {L}(X)\), and B is self adjoint. Moreover, B is strictly positive; indeed
and it is easy to check that, whenever \(x \ne 0\), we have that \( |\tilde{\mathcal {A}}^{-1}x|>0\). Finally, by dissipativity of \(\tilde{\mathcal {A}}\), we have:
with \(y= \tilde{\mathcal {A}}^{-1}x\). \(\square \)
Observe that
By strict positivity of B, we have that \(B_{00}\) and \(B_{11}\) are strictly positive. We introduce the \(|\cdot |_{-1}\)-norm on X by
We define
which is a Hilbert space endowed with the inner product
Notice that \(|x|_{-1}\le |\tilde{\mathcal {A}}^{-1}|_\mathcal {L(X)}|x|_X\); in particular, we have \((X,|\cdot |) \hookrightarrow (X_{-1},|\cdot |_{-1})\). Strict positivity of B ensures that the operator \(B^{1 /2}\) can be extended to an isometry
By (4.11) and an application of (Da Prato and Zabczyk 2014, Proposition B.1), we have \(\text{ Range }(B^{1/2})=\text{ Range }((\tilde{\mathcal {A}}^{-1})^*)\). Since \(\text{ Range }((\tilde{\mathcal {A}}^{-1})^*)=D(\tilde{\mathcal {A}}^*)\), we have
By (4.12), the operator \(\tilde{\mathcal {A}}^* B^{1/2}\) is well defined on the whole space X. Moreover, since \(\tilde{\mathcal {A}}^*\) is closed and \(B^{1/2}\in \mathcal {L}(X)\), \(\tilde{\mathcal {A}}^* B^{1/2}\) is a closed operator. Thus, by the closed graph theorem, we have
Remark 4.4
In the infinite dimensional theory of viscosity solutions it is only required that \(\tilde{\mathcal {A}}^*B \in \mathcal {L}(X)\) (condition (iii) of Definition 4.2). Such an operator can be constructed for any maximal dissipative operator \(\tilde{\mathcal {A}}\) (see, e.g., (Fabbri et al. 2017, Theorem 3.11)). Similarly to de Feo et al. (2023) in the case of the present paper, in addition, we also have \(\tilde{\mathcal {A}}^* B^{1/2} \in \mathcal {L}(X)\). Such stronger condition may be helpful in order to get differentiability properties of the value function (see (de Feo et al. 2023, Proof of Theorem 6.5)) or in order to construct optimal feedback laws (see de Feo and Święch 2023).
Since B is a compact, self-adjoint and strictly positive operator on X, by the spectral theorem B admits a set of eigenvalues \(\{\lambda _i \}_{i \in \mathbb {N}} \subset (0, +\infty )\) such that \(\lambda _{i}\rightarrow 0^{+}\) and a corresponding set \(\{f_i \}_{i \in \mathbb {N}} \subset X\) of eigenvectors forming an orthonormal basis of X. By taking \(\{e_i \}_{i \in \mathbb {N}}\) defined by \(e_i:=\frac{1 }{\sqrt{\lambda }_i} f_i\), we then get an orthonormal basis of \(X_{-1}\). We set \(X^N:= \text{ Span } \{f_1,...f_N\}= \text{ Span } \{e_1,...e_N\}\) for \(N\ge 1\), and let \(P_N :X \rightarrow X\) be the orthogonal projection onto \(X_N\) and \(Q_N:= I - P_N\). Since \(\{e_i \}_{i \in \mathbb {N}}\) is an orthogonal basis of \(X_{-1}\), the projections \(P_N,Q_N\) extend to orthogonal projections in \(X_{-1}\) and we will use the same symbols to denote them.
We notice that
Therefore, since \(|BQ_N|_\mathcal {L(X)}=|Q_NB|_\mathcal {L(X)}\) and B is compact, we get
Remark 4.5
-
(i)
We remark that the following inequality
$$\begin{aligned} |x_0| \le C_R |x|_{-1} \quad \forall x=(x_0,x_1) \in B_R \subset X, \end{aligned}$$(4.16)(which was proven in de Feo et al. (2023)) here is false. We first claim that such inequality does not hold on unbounded sets of X. Indeed, we provide the following counter example which is similar to the one in Federico and Tacconi (2014). Let \(n=1\) and consider
$$\begin{aligned} x^{N}=\left( x_{0}^{N}, x_{1}^{N}\right) , \quad x_{0}^{N}=1, \quad x_{1}^{N}=-N \textbf{1}_{[-1 / N, 0]}, \quad N\ge 1 \text{. } \end{aligned}$$For N big enough \(-1 / N>-d\), then we have
$$\begin{aligned} \left| x^{N}\right| _{-1}{} & {} =\left| \tilde{\mathcal {A}}^{-1} x^{N}\right| =0+\int _{-\frac{1}{N}}^{0}\left| \int _{-\frac{1}{N}}^{\xi } n d s\right| ^{2} d \xi =\int _{-\frac{1}{N}}^{0} N^{2}\left( \xi +\frac{1}{N}\right) ^{2} d \xi \\{} & {} =\frac{1}{3 N} \longrightarrow 0. \end{aligned}$$Therefore, we have \(\left| x_{0}^{N}\right| =1\) and \(\left| x^{N}\right| _{-1} \rightarrow 0\).
Now note that if the inequality (4.16) holded on bounded sets \(B_R\) then it would hold on the whole X with \(C_R=C>0\) which is a contradiction. Indeed assume the inequality holds on bounded sets \(B_R\). Then it is true for every \(x=(x_0,x_1)\) with \(|x|\le 1\), i.e. we have
$$\begin{aligned} |x_0|\le C|x|_{-1}, \quad |x|\le 1 \end{aligned}$$(4.17)Now let any \(y=(y_0,y_1) \in X\), \(y \ne 0\), define \(x=y/|y|=(y_0/|y|,y_1/|y|)\), \(|x|=1\) so that (4.17) holds and then by multiplying the inequality by |y| we have
$$\begin{aligned} |y_0|\le C|y|_{-1}, \end{aligned}$$which is a contradiction.
-
(ii)
We point out that the fact that here (4.17) is false is not in contradiction with de Feo et al. (2023), where such inequality was shown to be true. Indeed we recall that the operators \(\tilde{\mathcal {A}}\) here and in de Feo et al. (2023) are not the same: \(\tilde{\mathcal {A}}\) here is the adjoint (up to some bounded perturbation) of the operator \(\tilde{\mathcal {A}}\) in de Feo et al. (2023).
In Federico and Tacconi (2014), where delays appear in the control similarly to here, the inequality is then proved in the smaller space \(\mathbb {R} \times W^{1,2}([-d,0])\), see (Federico and Tacconi 2014, remark 5.4) and this is enough since the optimal control problem is deterministic. This would not be possible here due to the presence of the Brownian motion, whose trajectories are not absolutely continuous. Indeed, it would not be possible to have \(Y_1(t) \in W^{1,2}\) as in Federico and Tacconi (2014).
4.3 B-continuity of the value function
In this subsection, we prove some estimates for solutions of the state equation and on the cost functional in order to prove the B-continuity of the value function.
Lemma 4.6
There exists \(C>0\) and a local modulus of continuity \(\omega _R\) such that
for every \(x,y \in X, u \in U\), \(R>0\).
Moreover
Proof
-
(4.18) follows immdediately by the definition of \({\tilde{b}}(x,u)\).
-
Similarly for (4.19) we have:
$$\begin{aligned} |{\tilde{b}}(x, u)-{\tilde{b}}(y, u)|_{-1}^2&=\langle {\tilde{b}}(x, u)- {\tilde{b}}(y, u), B(x-y)\rangle \\&= \mu \langle x-y, B(x-y)\rangle = \mu |x-y|_{-1}^2. \end{aligned}$$ -
Recalling the definition of \({\tilde{b}}(x,u)\), (4.20) follows the boundedness of \(b_0(u)\) and the boundedness of U.
-
(4.21) follows by the boundedness of \(\sigma _0 :U \rightarrow M^{n \times q}\) since
$$\begin{aligned} |\sigma (u)|_{\mathcal {L}_{2}(\mathbb {R}^q, X)}^2 =\sum _{i=1}^n |\sigma (u) r_i|^2=\sum _{i=1}^n |\sigma _0(u) r_i|^2\le n |\sigma _0(u)|^2 \le C \end{aligned}$$where \(\{r_i \}_{i=1}^q\) is the classical orthonormal basis of \(\mathbb {R}^q\).
-
For (4.22) note that by (2.4) the function \(L(x,u)=l(x_0,u)\) is weakly sequentially continuous uniformly in \(u \in U\) (as \(x_0 \in \mathbb {R}^n\)), i.e.
$$\begin{aligned} \sup _{u \in U} |L(y,u)-L(x,u)|=\sup _{u \in U} |l(y_0,u)-l(x_0,u)|\le \omega _R(|y_0-x_0|)\rightarrow 0 \end{aligned}$$for every \(y \rightharpoonup x\).
Then the inequality follows by application of (Fabbri et al. 2017, Lemma 3.6 (iii)) which can easily be extended to functions weakly sequentially continuous uniformly with respect to the control parameter.
-
We notice that by (Fabbri et al. 2017, Appendix B) and (4.21), we have:
$$\begin{aligned}&|\text {Tr}\left[ \sigma (u) \sigma (u)^{*} B Q_{N}\right] |\\&\quad \le |\sigma (u) \sigma (u)^{*} BQ_N|_{\mathcal {L}_1(\mathbb {R}^q,X)} \le |\sigma (u) \sigma (u)^{*}|_{\mathcal {L}_1(\mathbb {R}^q,X)} |B Q_N|_{\mathcal {L}(X)} \\&\quad = |\sigma (u)|_{\mathcal {L}_2(\mathbb {R}^q,X)}^2 |B Q_N|_{\mathcal {L}(X)} \le C |B Q_N|_{\mathcal {L}(X)}. \end{aligned}$$Now taking the supremum over u and letting \(N \rightarrow \infty \) by (4.15) we have (4.24).\(\square \)
Remark 4.7
We remark that the fact that (4.16) is false (see Remark 4.5) was the reason that lead us to consider the simpler state equation (2.1), in place of the more general state equation (3.1), in order to use the approach via viscosity solutions: regarding \(b_0,\) for instance, if (4.16) were true, we could have considered a more general \(b_0\) of the form \(b_0(x_0,u),\) satisfying
Indeed, using (4.16), we would have had
from which we would have proved
Similarly we could have allowed \(\sigma _0(x_0,u)\) to depend also on \(x_0\). This would have been enough in order to prove the B-continuity of V and the validity of the hypotheses of the comparison theorem (Fabbri et al. 2017, Theorem 3.56) when the state equation is of the form (3.1). This approach was used in de Feo et al. (2023), where this inequality was proved to be true.
We recall (Fabbri et al. 2017, Proposition 3.24). Set
where C is the constant appearing in (4.20) and (4.21), and m is the constant from Assumption 2.3 and (4.23).
Proposition 4.8
((Fabbri et al. 2017, Proposition 3.24)) Let Assumption 2.2 holds and let \(\lambda >\rho _0\). Let Y(t) be the mild solution of (4.6) with initial datum \(x \in X\) and control \(u(\cdot ) \in \mathcal {U}\). Then, there exists \(C_\lambda >0\) such that
We need the following assumption.
Assumption 4.9
\(\rho >\rho _{0}.\)
Proposition 4.10
Let Assumptions 2.2, 2.3, and 4.9 hold. There exists \(C>0\) such that
Hence,
Proof
By (4.23) and Proposition 4.8 applied with \(\lambda =(\rho +\rho _0)/2\), we have
The estimate on V follows from this. \(\square \)
Next, we show continuity properties of V. We recall first the notion of B-continuity (see (Fabbri et al. 2017, Definition 3.4))
Definition 4.11
Let \(B \in \mathcal {L}(X)\) be a strictly positive self-adjoint operator. A function \(u: X \rightarrow \mathbb {R}\) is said to be B-upper semicontinuous (respectively, B-lower semicontinuous) if, for any sequence \(\left\{ x_{n}\right\} _{n \in \mathbb {N}}\subset X\) such that \( x_{n} \rightharpoonup x \in X\) and \(B x_{n} \rightarrow B x\) as \(n \rightarrow \infty \), we have
A function \(u: X \rightarrow \mathbb {R}\) is said to be B-continuous if it is both B-upper semicontinuous and B-lower semicontinuous.
We remark that, since our operator B is compact, in our case B-upper/lower semicontinuity is equivalent to the weak sequential upper/lower semicontinuity, respectively.
Proposition 4.12
Let Assumptions 2.2, 2.3, and 4.9 hold. For every \(R>0,\) there exists a modulus of continuity \(\omega _R\) such that
Hence V is B-continuous and thus weakly sequentially continuous.
Proof
We prove the estimate
as in (Fabbri et al. 2017, Proposition 3.73), since the assumptions of the latter are satisfied due to Lemma 4.6. Then, (4.25) follows. As for the last claim, we observe that by (4.25) and by (Fabbri et al. 2017, Lemma 3.6(iii)), V is B-continuous in X. \(\square \)
We point out that V may not be continuous with respect to the \(|\cdot |_{-1}\) norm in the whole X.
5 The value function as unique viscosity solution to HJB equation
In this section we prove that the value function V is the unique viscosity solution of the infinite dimensional HJB equation.
Given \(v \in C^1(X)\), we denote by Dv(x) its Fréchet derivative at \(x \in X\) and we write
where \(D_{x_0}v(x), D_{x_1}v(x)\) are the partial Fréchet derivatives. For \(v \in C^2(X)\), we denote by \(D^2 v(x)\) its second order Fréchet derivative at \(x \in X\) which we will often write as
We define the Hamiltonian function \(H:X\times X\times S(X)\rightarrow \mathbb {R}\) by
for every \(x,r \in X, Z \in \mathcal {S}(X).\)
By (Fabbri et al. 2017, Theorem 3.75) the Hamiltonian H satisfies the following properties.
Lemma 5.1
Let Assumptions 2.2 and 2.3 hold.
-
(i)
H is uniformly continuous on bounded subsets of \(X \times X \times S(X)\).
-
(ii)
For every \(x,r \in X\) and every \(Y,Z \in S(X)\) such that \(Z \le Y\), we have
$$\begin{aligned} H(x,r,Y)\le H(x,r,Z). \end{aligned}$$(5.1) -
(iii)
For every \(x,r \in X\) and every \(R>0\), we have
$$\begin{aligned} \lim _{N \rightarrow \infty } \sup \Big \{ |H(x,r,Z+\lambda BQ_N)-H(x,r,Z)|: \ |Z_{00}|\le R, \ |\lambda | \le R \Big \}=0. \end{aligned}$$(5.2) -
(iv)
For every \(R>0\) there exists a modulus of continuity \(\omega _R\) such that
$$\begin{aligned} H \left( z,\frac{B(z-y)}{\varepsilon },Z \right) -H \left( y,\frac{B(z-y)}{\varepsilon },Y \right) \ge -\omega _R\left( |z-y|_{-1}\left( 1+\frac{|z-y|_{-1}}{\varepsilon }\right) \right) \end{aligned}$$(5.3)for every \(\varepsilon >0\), \(y,z \in X\) such that \(|y|_X,|z|_X\le R\), \(Y,Z \in \mathcal {S}(X)\) satisfying
$$\begin{aligned} Y=P_N Y P_N \quad Z=P_N Z P_N \end{aligned}$$and
$$\begin{aligned} \frac{3}{\varepsilon }\left( \begin{array}{cc}B P_{N} &{} 0 \\ 0 &{} B P_{N}\end{array}\right) \le \left( \begin{array}{cc}Y &{} 0 \\ 0 &{} -Z\end{array}\right) \le \frac{3}{\varepsilon }\left( \begin{array}{cc}B P_{N} &{} -B P_{N} \\ -B P_{N} &{} B P_{N}\end{array}\right) . \end{aligned}$$ -
(v)
If \(C>0\) is the constant in (4.20) and (4.21), then, for every \(x \in X, p, q \in X, Y,Z \in \mathcal {S}(X)\),
$$\begin{aligned} | H(x, r+q,Y+Z)-H(x, r, Y)| \le C\left( 1+|x|_X\right) |q|_X+\frac{1}{2}C^2\left( 1+|x|_X\right) ^{2}|Z_{00}|. \end{aligned}$$(5.4)
The HJB equation associated with the optimal control problem is the infinite dimensional PDE
We recall the definition of B-continuous viscosity solution from Fabbri et al. (2017).
Definition 5.2
-
(i)
\(\phi :X \rightarrow \mathbb {R}\) is a regular test function if
$$\begin{aligned} \phi \in \Phi&:= \{ \phi \in C^2(X): \phi \textit{ is }B\textit{-lower semicontinuous and }\phi , \\&\qquad \qquad D\phi , D^2\phi , A^* D\phi \textit{ are uniformly continuous on }X\}; \end{aligned}$$ -
(ii)
\(g :X \rightarrow \mathbb {R}\) is a radial test function if
$$\begin{aligned} g \in \mathcal {G}&:= \{ g \in C^2(X): g(x)=g_0(|x|_X) \textit{ for some } g_0 \\&\qquad \qquad \in C^2([0,\infty )) \textit{ non-decreasing}, g_{0}'(0)=0 \}. \end{aligned}$$
Note that, if \(g\in \mathcal {G}\), we have
Definition 5.3
-
(i)
A locally bounded B-upper semicontinuous function \(v:X\rightarrow \mathbb {R}\) is a viscosity subsolution of (5.5) if, whenever \(v-\phi -g\) has a local maximum at \(x \in X\) for \(\phi \in \Phi , g \in \mathcal {G}\), then
$$\begin{aligned} \rho v(x) - \langle x,\tilde{\mathcal {A}}^* D\phi (x)\rangle _{X} + H(x,D\phi (x)+Dg(x) ,D^2\phi (x)+D^2g(x))\le 0. \end{aligned}$$ -
(ii)
A locally bounded B-lower semicontinuous function \(v:X\rightarrow \mathbb {R}\) is a viscosity supersolution of (5.5) if, whenever \(v+\phi +g\) has a local minimum at \(x \in X\) for \(\phi \in \Phi \), \(g \in \mathcal {G}\), then
$$\begin{aligned} \rho v(x) + \langle x,\tilde{\mathcal {A}}^* D\phi (x)\rangle _{X} + H(x,-D\phi (x)-Dg(x) ,-D^2\phi (x)-D^2g(x))\ge 0. \end{aligned}$$ -
(iii)
A viscosity solution of (5.5) is a function \(v:X\rightarrow \mathbb {R}\) which is both a viscosity subsolution and a viscosity supersolution of (5.5).
Define \(\mathcal {S}:=\{u :X \rightarrow \mathbb {R}: \exists k\ge 0 \ \text{ satisfying } (5.7) \text{ and } {\tilde{C}}\ge 0 \,\text{ such } \text{ that }\, |u(x)|\le {\tilde{C}} (1+|x|_X^k)\},\) where
and C is the constant appearing in (4.20) and (4.21).
We can now state the theorem characterizing V as the unique viscosity solution of (5.5) in \(\mathcal {S}\).
Theorem 5.4
Let Assumptions 2.2, 2.3, and 4.9 hold. The value function V is the unique viscosity solution of (5.5) in the set \(\mathcal {S}\).
Proof
Notice that \(V \in \mathcal {S}\) by Proposition 4.10.
The proof of the fact that V is the unique viscosity solution of the HJB equation can be found in (Fabbri et al. 2017, Theorem 3.75) as all assumptions of this theorem are satisfied due to Lemma 5.1. \(\square \)
Remark 5.5
We remark that, similarly to de Feo et al. (2023), Theorem 5.4 also holds in the deterministic case, i.e. when \(\sigma (x,u)=0\). (in which case we may take \(\rho _0=Cm\) and \(k<\rho /C\) in (5.7)). The theory of viscosity solutions handles well degenerate HJB equations, i.e. when the Hamiltonian satisfies
for every \(Y,Z \in S(X)\) such that \(Z\le Y\). Hence viscosity solutions can be used in connection with the dynamic programming method for optimal control of stochastic differential equations in the case of degenerate noise in the state equation, in particular, when it completely vanishes (deterministic case). This is not possible using the mild solutions approach (see de Feo et al. 2023; Fabbri et al. 2017).
Remark 5.6
In this work we could not prove the partial differentiability of V with respect to \(x_0\) as in (de Feo et al. 2023, Theorem 6.5). Indeed in (de Feo et al. 2023, Theorem 6.5) a key assumption was
In de Feo et al. (2023) this condition holds under some standard assumptions, see (de Feo et al. 2023, Example 6.2). In particular the cost \(l(\cdot ,u)\) is assumed to be Lipschitz (uniformly in u).
However in the present paper we could not prove (5.8): indeed due to Remark 4.5 the following inequality (which holds true in de Feo et al. (2023)) is false
This means that even for a Lipschitz \(l(\cdot ,u)\) (uniformly in u) we cannot use (5.9) in the following way
to get the Lipschitzianity of L with respect to \(|\cdot |_{-1}\) (as it was done in de Feo et al. (2023)). The best we could get in this paper is (4.22), i.e.
but this is of course not enough in order to prove (5.8). Hence we could not proceed as in the proof of (de Feo et al. 2023, Theorem 6.5).
We finally remark that the same reason prevented us to apply \(C^{1,1}-\)regularity results from de Feo et al. (2023), where \(L(\cdot ,u)\) was assumed to be Lipschitz with respect to \(|\cdot |_{-1}\), uniformly in u.
6 Applications
In this section we provide applications of our results to problems coming from economics.
6.1 Optimal advertising with delays
The following model is a generalization of the ones in Gozzi and Marinelli (2006), Gozzi et al. (2009) to the case of controlled diffusion. We recall that in de Feo et al. (2023) the case with no delays in the control (i.e. \(p_1=0\)) was treated.
The model for the dynamics of the stock of advertising goodwill y(t) of the product is given by the following controlled SDDE
where \(d>0\), the control process u(t) models the intensity of advertising spending and W is a real-valued Brownian motion. Moreover
-
(i)
\(a_0 \le 0\) is a constant factor of image deterioration in absence of advertising;
-
(ii)
\(b_0 \ge 0\) is a constant advertising effectiveness factor;
-
(iii)
\(a_1 \le 0\) is a given deterministic function satisfying the assumptions used in the previous sections and represent the distribution of the forgetting time;
-
(iv)
\(p_1 \ge 0\) is a given deterministic function satisfying the assumptions used in the previous sections and it is the density function of the time lag between the advertising expenditure and the corresponding effect on the goodwill level;
-
(v)
\(\sigma _0>0\) is a fixed uncertainty level in the market;
-
(vi)
\(\gamma _0> 0\) is a constant uncertainty factor which multiplies the advertising spending;
-
(vii)
\(\eta _0 \in \mathbb {R}\) is the level of goodwill at the beginning of the advertising campaign;
-
(viii)
\(\eta _1 \in L^2([-d,0];\mathbb {R})\) is the history of the goodwill level.
-
(ix)
\(\delta \in L^2([-d,0]; U)\) is the history of the advertising spending.
Again, we use the same setup of the stochastic optimal control problem as the one in Sect. 2 and the control set U is here \(U= [0,{\bar{u}}]\) for some \({\bar{u}}>0\). The optimization problem is
where \(\rho >0\) is a discount factor, \(l(x,u)=h(u)-g(x)\), with a continuous and convex cost function \(h :U \rightarrow \mathbb {R}\) and a continuous and concave utility function \(g :\mathbb {R} \rightarrow \mathbb {R}\) which satisfies Assumption 2.3.
We are then in the setting of Sect. 4. Therefore we can use Theorem 5.4 to characterize the value function V as the unique viscosity solution to (5.5).
6.2 Optimal investment models with time-to-build
The following model is inspired by (Fabbri and Federico 2014, p. 36). See also, e.g., (Bambi et al. 2012, 2017) for similar models in the deterministic setting.
Let us consider a state process y(t), representing the stock capital of a certain enterprise at time t, and a control process \(u(t) \ge 0\), representing the investment undertaken at time t to increase y. We assume that the dynamics of y(t) is given by the following SDDE
where
-
(i)
\(b_0 :U \rightarrow [0,\infty )\) is a continuous bounded function representing an instantaneous effect of the investment on the capital;
-
(ii)
\( p_1 \ge 0\) is a given deterministic function satisfying the assumptions used in the previous sections and representing the density function of the time-to-build between the investment and the corresponding effect on the stock capital;
-
(iii)
\(\sigma _0 :U \rightarrow [0,\infty )\) is a a continuous bounded function representing the uncertainty of achievement of the investment plans.
-
(iv)
\(\eta _0 \in \mathbb {R}\) is the initial level of the capital;
-
(v)
\(\delta \in L^2([-d,0];U)\) is the history of the investment spending.
Again, we use the same setup of the stochastic optimal control problem of Sect. 2 and the control set U here is \(U= [0,{\bar{u}}]\) for some \({\bar{u}}>0\). The goal is to maximize, over all \(u (\cdot ) \in \mathcal {U},\) the expected integral of the discounted future profit flow in the form
where \(F: \mathbb {R} \rightarrow \mathbb {R}\) is a production function and \(C: U \rightarrow \mathbb {R}\) is a cost function. We assume that F, C satisfy Assumption 2.3. The optimization problem is equivalent to minimize, over all \(u (\cdot ) \in \mathcal {U},\)
where \(l(x,u):=C(u)- F(x) \). We are then in the setting of Sect. 2 (with \(a_1=0\)). Therefore we can use Theorem 5.4 to characterize the value function V as the unique viscosity solution to (5.5).
Notes
For the procedure of rewriting deterministic delay differential equations, we refer the readers to (Bensoussan et al. 2007, Part II, Chapter 4). For the stochastic case, one may consult (Chojnowska-Michalik 1978; Da Prato and Zabczyk 1996; Gozzi and Marinelli 2006; Gozzi et al. 2009; Federico and Tankov 2015) for the Hilbert space case and (Mohammed 1984, 1998; Flandoli and Zanco 2016) for the Banach space case. A “mixed” approach is employed in Federico (2011).
References
Bambi, M., Fabbri, G., Gozzi, F.: Optimal policy and consumption smoothing effects in the time-to-build AK model. Econom. Theory 50(3), 635–669 (2012)
Bambi, M., Di Girolami, C., Federico, S., Gozzi, F.: Generically distributed investments on flexible projects and endogenous growth. Econom. Theory 63(2), 521–558 (2017)
Bensoussan, A., Da Prato, G., Delfour, M.C., Mitter, S.K.: Representation and Control of Infinite Dimensional Systems, 2nd edn. Foundations and Applications, Systems and Control (Birkhäuser, Boston, 2007)
Biffis, E., Gozzi, F., Prosdocimi, C.: Optimal portfolio choice with path dependent labor income: the infinite horizon case. SIAM J. Control. Optim. 58(4), 1906–1938 (2020)
Biagini, S., Gozzi, F., Zanella, M.: Robust portfolio choice with sticky wages. SIAM J. Financial Math. 13(3), 1004–1039 (2022)
Carlier, G., Tahraoui, R.: Hamilton-Jacobi-Bellman equations for the optimal control of a state equation with memory. ESAIM Control Optim. Calc. Var. 16(3), 744–763 (2010)
Chojnowska-Michalik, A.: Representation theorem for general stochastic delay equations. Bull. Acad. Polon. Sci. Sér. Sci. Math. Astronom. Phys. 26(7), 635–642 (1978)
Da Prato, G., Zabczyk, J.: Ergodicity for Infinite-Dimensional Systems. London Mathematical Society Lecture Note Series, vol. 229. Cambridge University Press, Cambridge (1996)
Da Prato, G., Zabczyk, J.: Stochastic Equations in Infinite Dimensions. Encyclopedia of Mathematics and its Applications, vol. 152. Cambridge University Press, Cambridge (2014)
de Feo, F., Federico, S., Świȩch, A.: Optimal control of stochastic delay differential equations and applications to path-dependent financial and economic models. SIAM J. Control Optim. 62(3), 1490–1520 (2024)
de Feo, F., Święch, A.: Optimal control of stochastic delay differential equations: Optimal feedback controls. (2023) arXiv preprint arXiv:2309.05029
de Feo, F., Święch, A., Wessels, L.: Stochastic optimal control in Hilbert spaces: \(C^{1,1}\) regularity of the value function and optimal synthesis via viscosity solutions. (2023) arXiv preprint, arXiv:2310.03181
Di Giacinto, M., Federico, S., Gozzi, F.: Pension funds with a minimum guarantee: a stochastic control approach. Finance Stoch. 15(2), 297–342 (2011)
Djehiche, B., Gozzi, F., Zanco, G., Zanella, M.: Optimal portfolio choice with path dependent benchmarked labor income: a mean field model. Stochastic Process. Appl. 145, 48–85 (2022)
Engel, K.J., Nagel, R.: One-Parameter Semigroups for Linear Evolution Equations. Graduate Texts in Mathematics, vol. 194. Springer, Berlin (2000)
Fabbri, G., Federico, S.: On the infinite-dimensional representation of stochastic controlled systems with delayed control in the diffusion term. Math. Econ. Lett. 2(3), 33–43 (2014)
Fabbri, G., Gozzi, F., Świȩch, A.: Stochastic optimal control in infinite dimension. Dynamic programming and HJB equations. With a contribution by Marco Fuhrman and Gianmario Tessitore. Probability Theory and Stochastic Modelling, vol. 82. Springer, Cham (2017)
Federico, S.: A stochastic control problem with delay arising in a pension fund model. Finance Stoch. 15(3), 421–459 (2011)
Federico, S., Goldys, B., Gozzi, F.: HJB equations for the optimal control of differential equations with delays and state constraints, I: regularity of viscosity solutions. SIAM J. Control. Optim. 48(8), 4910–4937 (2010)
Federico, S., Goldys, B., Gozzi, F.: HJB equations for the optimal control of differential equations with delays and state constraints, II: verification and optimal feedbacks. SIAM J. Control. Optim. 49(6), 2378–2414 (2011)
Federico, S., Tacconi, E.: Dynamic programming for optimal control problems with delays in the control variable. SIAM J. Control. Optim. 52(2), 1203–1236 (2014)
Federico, S., Tankov, P.: Finite-dimensional representations for controlled diffusions with delay. Appl. Math. Optim. 71(1), 165–194 (2015)
Flandoli, F., Zanco, G.: An infinite-dimensional approach to path-dependent Kolmogorov’s equations. Ann. Probab. 44(4), 2643–2693 (2016)
Fuhrman, M., Masiero, F., Tessitore, G.: Stochastic equations with delay: optimal control via BSDEs and regular solutions of Hamilton-Jacobi-Bellman equations. SIAM J. Control. Optim. 48(7), 4624–4651 (2010)
Gawarecki, L., Mandrekar, V.: Stochastic Differential Equations in Infinite Dimensions: with Applications to Stochastic Partial Differential Equations. Springer (2010)
Gozzi, F., Marinelli, C.: Stochastic optimal control of delay equations arising in advertising models. Stochastic partial differential equations and applications-VII, 133–148, Lect. Notes Pure Appl. Math., 245, Chapman & Hall/CRC, Boca Raton, FL, (2006)
Gozzi, F., Marinelli, C., Savin, S.: On controlled linear diffusions with delay in a model of optimal advertising under uncertainty with memory effects. J. Optim. Theory Appl. 142(2), 291–321 (2009)
Gozzi, F., Masiero, F.: Stochastic optimal control with delay in the control I: Solving the HJB equation through partial smoothing. SIAM J. Control. Optim. 55(5), 2981–3012 (2017). (93E20 (35K55 47D07 49L20 60H20))
Gozzi, F., Masiero, F.: Stochastic optimal control with delay in the control II: Verification theorem and optimal feedbacks. SIAM J. Control. Optim. 55(5), 3013–3038 (2017)
Gozzi, F., Masiero, F.: Errata: Stochastic optimal control with delay in the control I: solving the HJB equation through partial smoothing, and stochastic optimal control with delay in the control II: verification theorem and optimal feedbacks. SIAM J. Control. Optim. 59(4), 3096–3101 (2021)
Gozzi, F., Masiero, F.: Stochastic control problems with unbounded control operators: solutions through generalized derivatives. SIAM J. Control. Optim. 61(2), 586–619 (2023)
Gozzi, F., Masiero, F.: Lifting partial smoothing to solve HJB equations and stochastic control problems. https://arxiv.org/pdf/2107.04305.pdf
Masiero, F., Tessitore, G.: Partial smoothing of delay transition semigroups acting on special functions. J. Differ. Equ. 316, 599–640 (2022)
Mohammed, S.E.A.: Stochastic Functional Differential Equations. Pitman Research Notes in Mathematics Series, vol. 99. Longman, Boston (1984)
Mohammed, S.E.A.: Stochastic differential systems with memory: theory, examples and applications. Stochastic Analysis and Related Topics VI, ed. by Decreusefond, L., Gjerde, J., Øksendal, B., Üstünel, A.S., Progress in Probability, vol. 42 (Birkhäuser, Boston, 1998), pp. 1-77
Nerlove, M., Arrow, J.K.: Optimal advertising policy under dynamic conditions. Economica 29(114), 129–142 (1962)
Pang, T., Tao, Yong, Y.: A new stochastic model for stock price with delay effects. In: Proceedings of the Conference on Control and its Applications, pp. 110–117. Society for Industrial and Applied Mathematics (2019)
Revuz, D., Yor, M.: Continuous Martingales and Brownian Motion. 3rd edition, Grundlehren der Mathematischen Wissenschaften, vol. 293 (Springer, Berlin, 1999)
Vinter, R.B., Kwong, R.H.: The infinite time quadratic control problem for linear systems with state and control delays: an evolution equation approach. SIAM J. Control. Optim. 19(1), 139–153 (1981)
Yong, J., Zhou, X.Y.: Stochastic Controls, Hamiltonian Systems and HJB Equations. Applications of Mathematics, vol. 43. Springer, New York (1999)
Acknowledgements
The author is very grateful to Andrzej ’Swiȩch for many useful conversations related to the content of the present paper. He also thanks the two anonymous Referees for their careful reading of the manuscript and their useful comments and suggestions.
Funding
Open access funding provided by Luiss University within the CRUI-CARE Agreement. This research was partially financed by the INdAM (Instituto Nazionale di Alta Matematica F. Severi) - GNAMPA (Gruppo Nazionale per l’Analisi Matematica, la Probabilità e le loro Applicazioni) Project “Riduzione del modello in sistemi dinamici stocastici in dimensione infinita a due scale temporali” and by the Italian Ministry of University and Research (MIUR), within PRIN project 2017FKHBA8 001 (The Time-Space Evolution of Economic Activities: Mathematical Models and Empirical Applications).
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflict of interest statement
The author has no Conflict of interest to declare that are relevant to the content of this article.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
de Feo, F. Stochastic optimal control problems with delays in the state and in the control via viscosity solutions and applications to optimal advertising and optimal investment problems. Decisions Econ Finan (2024). https://doi.org/10.1007/s10203-024-00456-y
Received:
Accepted:
Published:
DOI: https://doi.org/10.1007/s10203-024-00456-y