Abstract
We consider an optimal control problem for a diffuse interface model of tumor growth. The state equations couples a Cahn–Hilliard equation and a reaction-diffusion equation, which models the growth of a tumor in the presence of a nutrient and surrounded by host tissue. The introduction of cytotoxic drugs into the system serves to eliminate the tumor cells and in this setting the concentration of the cytotoxic drugs will act as the control variable. Furthermore, we allow the objective functional to depend on a free time variable, which represents the unknown treatment time to be optimized. As a result, we obtain first order necessary optimality conditions for both the cytotoxic concentration and the treatment time.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
There has been a recent surge in the development of phase field models for tumor growth. These models aim to describe the evolution of a tumor colony surrounded by healthy tissues which experience biological mechanisms such as proliferation via nutrient consumption, apoptosis, chemotaxis and active transport of specific chemical species. For the case of a young tumor, before the development of quiescent cells, the phase field models often consist of a Cahn–Hilliard equation coupled with a reaction-diffusion equation for the nutrient [13, 22, 26, 27, 37]. One may also treat the tumor cells and the healthy cells as inertia-less fluids, and include the effects of fluid flow into the evolution of the tumor, leading to the development of a Cahn–Hilliard–Darcy system [14, 15, 22, 45].
Current treatments for cancer include surgery, immunotherapy (strengthening the immune system), radiotherapy (using radiation to kill cancer cells) and chemotherapy (using drugs to kill cancer cells). The latter three treatments are typically conducted in cycles. A cycle is a period of treatment followed by a (longer) period of rest, so that the patient’s body can build new healthy cells. The goal of these therapeutic treatments is to shrink the tumor into a more manageable size for which surgery can be applied. Further therapeutic treatments may be necessary in order to destroy the cancer cells that may remain after the surgery.
In this work, we consider an optimal control problem involving a cancer treatment with cytotoxic drugs. It is well-known that while cytotoxic drugs mainly target and damage rapidly dividing cells such as tumor cells, the drugs can also accumulate in the body and cause adverse side-effects to the immune system and various vital organs such as the kidneys and the liver. In a worst case scenario, too much cytotoxic drugs may allow tumor cells to mutate and become resistant to the treatment. Thus, from the viewpoint of the patient, the shortest treatment time in which the objectives of the chemotherapy are achieved is the most ideal. Therefore, the optimal control problem we study involves finding the optimal drug distribution and the optimal treatment time.
For \(T > 0,\) in a bounded domain \(\Omega \subset \mathbb {R}^{3}\) with \(C^{3}\)-boundary \(\Gamma ,\) we consider the following Cahn–Hilliard model for tumor growth,
Here, \(\alpha \) is a positive constant, \(\varphi \) denotes the difference in volume fractions, where \(\varphi = 1\) represents the tumor phase and \(\varphi = {-}1\) represents the healthy tissue phase. The function \(\mu \) is a chemical potential associated to \(\varphi ,\, \Psi ^{\prime }(\varphi )\) is the derivative of a potential \(\Psi (\varphi )\) with equal minima at \(\varphi = {\pm } 1,\, \sigma \) denotes the concentration of an unspecified chemical species acting as nutrient for the tumor cells, while u denotes the concentration of cytotoxic drugs.
The function \(h(\varphi )\) is an interpolation function such that \(h(-1) = 0\) and \(h(1) = 1,\) and the parameters \(\mathcal {P},\, \mathcal {A},\, \mathcal {C},\) and \(\mathcal {B}\) denote the constant tumor proliferation rate, tumor apoptosis rate, nutrient consumption rate, and nutrient supply rate, respectively. The positive constants A and B are related to the thickness of the interfacial layer and the surface tension, while \(\partial _{\varvec{\nu }}f = \nabla f \cdot \varvec{\nu }\) denotes the normal derivative of f where \(\varvec{\nu }\) is the unit outward normal of \(\Gamma .\)
The term \(h(\varphi ) \mathcal {P}\sigma \) models the proliferation of tumor cells which is proportional to the concentration of the nutrient, the term \(h(\varphi ) \mathcal {A}\) models the apoptosis of tumor cells, and \(\mathcal {C}h(\varphi ) \sigma \) models the consumption of the nutrient only by the tumor cells. The term \(\alpha u h(\varphi )\) models the elimination of the tumor cells by the cytotoxic drugs at a constant rate \(\alpha .\) Meanwhile, \(\sigma _{S}\) denotes the nutrient concentration in a pre-existing vasculature, and \(\mathcal {B}(\sigma _{S} - \sigma )\) models the supply of nutrient from the blood vessels if \(\sigma _{S} > \sigma \) and the transport of nutrient away from the domain \(\Omega \) if \(\sigma _{S} < \sigma .\)
In comparison with the models of [22], we have neglected the effects of chemotaxis and active transport, but the new feature of (1.1) is the inclusion of the effects of cytotoxic drugs via the term \(\alpha u h(\varphi ),\) and in this work the function u will act as our control. For realistic applications the control \(u{\text {:}}\,[0,\,T] \rightarrow [0,\,1]\) should be spatially constant, where \(u = 1\) represents a full dosage and \(u = 0\) represents no dosage. However, in the subsequent analysis, we allow for spatial dependence (see Assumption 2.1 below).
For positive constants \(r,\, \beta _{u}\) and \(\beta _{T},\) and nonnegative constants \(\beta _{Q},\, \beta _{\Omega },\) and \(\beta _{S},\) we consider the objective functional \(J_{r}\) given as
In particular, (1.2) can be seen as the relaxation of the following more natural objective functional
Here, \(\tau \in (0,\,T]\) represents the treatment time, \(\varphi _{Q}\) represents a desired evolution for the tumor cells while \(\varphi _{\Omega }\) represents a desired final distribution. The first two terms of J are of standard tracking type that are often considered in the literature of parabolic optimal control, and the third term of J measures the size of the tumor at the end of the treatment. The fourth term penalizes large concentrations of the cytotoxic drugs, and the fifth term of J penalizes long treatment times.
Let us make the following comments:
-
(1)
A large value of \(\left| {\varphi - \varphi _{Q}}\right| ^{2}\) would mean that the patient suffers from the growth of the tumor, and a large value of \(\left| {u}\right| ^{2}\) would mean that the patient suffers from high toxicity of the drug.
-
(2)
The function \(\varphi _{\Omega }\) can be a stable configuration of the system, so that the tumor does not grow again once the treatment is completed. One can also choose \(\varphi _{\Omega }\) as a configuration which is suitable for surgery.
-
(3)
The variable \(\tau \) can be regarded as the treatment time of one cycle, i.e., the amount of time the drug is applied to the patient before the period of rest, or the treatment time before surgery.
-
(4)
It is possible to replace \(\beta _{T} \tau \) by a more general function \(f(\tau )\) where \(f{\text {:}}\, \mathbb {R}_{\ge 0} \rightarrow \mathbb {R}_{\ge 0}\) is continuously differentiable and increasing.
-
(5)
We consider \(T \in (0,\,\infty )\) as a fixed maximal time in which the patient is allowed to undergo a treatment obtained from this optimal control problem.
For technical reasons highlighted below, we consider an optimal control problem with the relaxed objective functional (1.2) and the state equations (1.1). We denote the space of admissible controls as \(\mathcal {U}_{\mathrm {ad}}\) (see Assumption 2.1 below) and the optimal control problem we study in this work can be expressed as follows,
The optimal control problem (P) is a problem involving a free terminal time, and we say that \((u_{*}, \,\tau _{*})\) is a minimizer of (P) if
where the infimum is taken over triplets \((\phi ,\, w,\, s)\) such that \(w \in \mathcal {U}_{\mathrm {ad}},\, s \in [0,\,T]\) and \(\phi \) solves (1.1) with datum w. In ODE constrained optimal control where the cost functional depends on the free terminal time, the necessary optimality condition can be derived with the help of the corresponding Hamiltonian function, see for instance [35, Chap. 20] and [25, 33, 38]. One may use the notion of Hamiltonian functional to derive the optimality condition for the free terminal time when the state equations are partial differential equations, see in particular [2, 39, 40] for semilinear parabolic state equations.
Below we illustrate with an example the optimality conditions obtained with the Hamiltonian from ODE theory and with the Lagrangian method for PDE-constrained optimization, see for instance [44, §2.10] and [32, §1.6.4]. Suppose the objective functional is of the form
and \(\varphi \) satisfies for example
Let \(u_{*}\) denote an optimal control with corresponding state \(\varphi _{*}.\) The Hamiltonian H is defined as
where p act as the adjoint variable to \(\varphi _{*}.\) From the works of [2, 39, 40] and also from the theory of ODE-constraint optimal control, the optimality condition for the optimal time \(\tau _{*}\) is
Now, let us define the Lagrangian
then one obtains from formally differentiating \(\mathcal {L}\) with respect to \(\tau \) the optimality condition for \(\tau _{*},\) which is
The adjoint equation for p is a terminal time boundary value problem:
Using the terminal condition for p in the expression for \(\frac{\partial \mathcal {L}}{\partial \tau }(\tau _{*},\, \varphi _{*},\, u_{*})\) we see that
and thus \(\frac{\partial \mathcal {L}}{\partial \tau }(\tau _{*},\, \varphi _{*},\, u_{*}) = 0\) is equivalent to (1.4). That is, the optimality conditions for the free terminal time obtain from the Hamiltonian formulation and the Lagrangian formulation coincide.
Let us briefly explain the issues with the objective functional (1.3). Formally differentiating (1.3) with respect to \(\tau ,\) we obtain
and in order for the terms in (1.5) to be well-defined, we need that
Furthermore, to rigorously establish the Fréchet differentiability of J with respect to \(\tau ,\) it turns out that we require
Thus, the main mathematical difficulties arise from establishing high temporal regularity for the state variables. A preliminary analysis shows that it is possible to derive such regularity but only under rather strong assumptions such as \(\varphi _{0} \in H^{5}(\Omega ),\,\sigma _{0} \in H^{3}(\Omega )\) and \(\Vert {\partial _{t}u}\Vert _{L^{2}(0,\,T;\,L^{2}(\Omega ))} \le K\) for some fixed \(K > 0.\) The assumption on the a priori boundedness of \(\partial _{t}u\) is not meaningful as in applications it will be hard to verify this condition. Furthermore using the Lagrangian method, one can compute that the terminal condition for the adjoint variable p to \(\varphi _{*}\) is \(p(\tau _{*}) = \beta _{\Omega }(\varphi _{*}(\tau _{*}) - \varphi _{\Omega }) + \frac{\beta _{S}}{2},\) and so we can write (1.5) more compactly as
But this would mean that we require the weak formulation for the equation of \(\varphi _{*}\) to be satisfied pointwise in \([0,\,T],\) that is,
holds for all \(t \in [0,\,T].\) This in turn implies that we need
These difficulties motivates the current study with the relaxed objective functional \(J_{r}\) (1.2).
There have been many recent contributions regarding the well-posedness and asymptotic behaviour for phase field type tumor models, see for example [7,8,9, 17, 20, 21] for the Cahn–Hilliard variant, and [3, 15, 19, 34, 36] for the Cahn–Hilliard–Darcy variant. From the aspect of optimal control, we mention the works of [5, 6, 11, 12, 29, 47] for the Cahn–Hilliard equation, [41, 46, 48, 49] for the convective Cahn–Hilliard equation and [18, 28, 30, 31] for the Cahn–Hilliard–Navier–Stokes system. In the context of PDE-constraint optimal control for diffuse interface tumor models, we have the recent work of [10], where the objective functional (1.3) with \(\beta _{S} = \beta _{T} = 0\) and no dependence of J on \(\tau \) is studied with state equations given by the model of [27] and the control enters the nutrient equation as a source term, similar to the term \(\mathcal {B}(\sigma _{S} - \sigma )\) in (1.1c). With this work we aim to provide a contribution to the theory of free terminal time optimal control in the context of diffuse interface tumor models.
Let us provide some future directions of research motivated by this study:
-
(1)
An optimal control u that is periodic in time, reflecting the cyclic nature of therapeutic treatments.
-
(2)
A feedback mechanism taking into account the patient’s response to the therapy, and the tumor’s resistance to the drug.
-
(3)
Analysis and identification of stable equilibria for diffuse interface models of tumor growth.
Plan of the paper the rest of this paper is organized as follows. In Sect. 2 the general assumptions are outlined and the main results are stated. The well-posedness of the state equations (1.1) is established in Sect. 3. The existence of a minimizer to (P) is proved in Sect. 4, while the unique solvability of the linearized state equations and the Fréchet differentiability of the control-to-state mapping and of the functional \(J_{r}\) are contained in Sect. 5. In Sect. 6, the unique solvability of the adjoint equations is studied and the first order necessary optimality conditions are derived.
2 General Assumptions and Main Results
Notation for convenience, we will often use the notation \(L^{p} := L^{p}(\Omega )\) and \(W^{k,p} := W^{k,p}(\Omega )\) for any \(p \in [1,\,\infty ],\, k > 0\) to denote the standard Lebesgue and Sobolev spaces equipped with the norms \(\Vert {\cdot }\Vert _{L^{p}}\) and \(\Vert {\cdot }\Vert _{W^{k,p}}.\) In the case \(p = 2\) we use \(H^{k} := W^{k,2}\) and the norm \(\Vert {\cdot }\Vert _{H^{k}}.\) Moreover, the dual space of a Banach space X will be denoted by \(X^{*},\) and the duality pairing between X and \(X^{*}\) is denoted by \(\langle {\cdot } ,{\cdot }\rangle _{X}.\) The space–time cylinder \(\Omega \times (0,\,T)\) will be denoted by Q, and we use the notation \(L^{p}(Q)\) to denote the spaces \(L^{p}(\Omega \times (0,\,T))\) for \(1 \le p \le \infty .\) Using Fubini’s theorem we have the isometric isomorphism \(L^{p}(0,\,T;\,L^{p}) \cong L^{p}(Q)\) for \(p \in [1,\,\infty ).\) We point out that \(L^{\infty }(0,\,T;\,L^{\infty }) \subset L^{\infty }(Q),\) but the converse inclusion is not true in general due to measurability issues (see for instance [42, Ex 1.4.2]).
Useful preliminaries the following Gronwall inequality in integral form will often be used (see [21, Lemma 3.1] for a proof). For \(W,\,X,\, Y,\, Z\) real-valued functions defined on \([0,\,T]\) such that W is integrable, X is nonnegative and continuous, Y is continuous, Z is nonnegative and integrable. If Y and Z satisfy the integral inequality
then it holds that
The following Taylor’s theorem with integral remainder will be used to show the Fréchet differentiability of the control-to-state mapping. For \(f \in C^{2}(\mathbb {R})\) and \(a,\,x \in \mathbb {R},\) it holds that
The Gagliardo–Nirenberg interpolation inequality in dimension d is also useful (see [16, Theorem 10.1, p. 27]): let \(\Omega \) be a bounded domain with \(C^{m}\) boundary, and \(f \in W^{m,r}(\Omega ) \cap L^{q}(\Omega ),\,1 \le q,\,r \le \infty .\) For any integer \(j,\, 0 \le j < m,\) suppose there is an \(\alpha \in \mathbb {R}\) such that
If \(r \in (1,\,\infty )\) and \(m - j - \frac{d}{r}\) is a nonnegative integer, we in addition assume \(\alpha \ne 1.\) Under these assumptions, there exists a positive constant C depending only on \(\Omega ,\, m,\, j,\, q,\, r,\) and \(\alpha \) such that
We consider the following assumptions.
Assumption 2.1
-
(A1)
The initial conditions satisfy \(\varphi _{0} \in H^{3}\) with the compatibility condition \(\partial _{\varvec{\nu }}\varphi _{0} = 0\) on \(\Gamma ,\, \sigma _{0} \in H^{1},\) with \(0 \le \sigma _{0} \le 1\) a.e. in \(\Omega ,\) while the target functions satisfy \(\varphi _{Q},\, \varphi _{\Omega } \in L^{2}(Q).\) The vasculature nutrient concentration \(\sigma _{S}\) satisfies \(0 \le \sigma _{S} \le 1\) a.e. in Q.
-
(A2)
The interpolation function \(h{\text {:}}\, \mathbb {R}\rightarrow [0,\,1]\) is twice continuously differentiable and Lipschitz continuous (with Lipschitz constant \(L_{h}\)). The parameters \(\mathcal {P},\, \mathcal {A},\, \mathcal {C}\) and \(\mathcal {B}\) are nonnegative constants, and \(\alpha \) is a positive constant.
-
(A3)
The space of admissible controls is given as
$$\begin{aligned} \mathcal {U}_{\mathrm {ad}} = \left\{ u \in L^{\infty }\left( 0,\,T;\,L^{\infty }\right) {\text {:}}\,0 \le u \le 1 \text { a.e. in } Q \right\} . \end{aligned}$$ -
(A4)
The potential \(\Psi {\text {:}}\,\mathbb {R}\rightarrow \mathbb {R}_{\ge 0}\) is three times continuously differentiable and satisfies for some positive constants \(\{k_{j}\}_{j=0}^{5},\)
$$\begin{aligned}&\left| {\Psi ^{\prime }(s)}\right| \le k_{0} \Psi (s) + k_{1},\end{aligned}$$(2.4)$$\begin{aligned}&\Psi (s) \ge k_{2} \left| {s}\right| - k_{3}, \end{aligned}$$(2.5)$$\begin{aligned}&\left| {\Psi ^{\prime \prime }(s)}\right| \le k_{4}\left( 1 + \left| {s}\right| ^{2}\right) , \end{aligned}$$(2.6)$$\begin{aligned}&\left| {\Psi ^{\prime }(s) - \Psi ^{\prime }(t)}\right| \le k_{5}\left( 1 + \left| {s}\right| ^{2} + \left| {t}\right| ^{2}\right) \left| {s - t}\right| , \end{aligned}$$(2.7)for all \(s,\, t \in \mathbb {R}.\)
We point out that as \(\Omega \) is a bounded domain, there exists an open set \(\mathcal {U} \subset L^{2}(Q)\) such that \(\mathcal {U}_{\mathrm {ad}} \subset \mathcal {U}.\) In this work, we consider quartic potentials \(\Psi \) for the state equations, for which the classical double-well potential \(\Psi (s) = \frac{1}{4}(1-s^{2})^{2}\) is one example. The well-posedness of the state equations with higher polynomial growth for \(\Psi \) is also possible, see for instance the procedure in [17, Proof of Theorem 1], but we restrict our current analysis to that of quartic potentials to simplify the computations.
Theorem 2.1
(Well-posedness) For every \(T \in (0,\,\infty )\) and given data \((\varphi _{0}, \,\sigma _{0},\, u),\) under Assumption 2.1 there exists a unique triplet of solutions \((\varphi ,\, \mu , \,\sigma )\) with
such that \(\varphi (0) = \varphi _{0},\, \sigma (0) = \sigma _{0},\) and for a.e. \(t \in (0,\,T)\) and for all \(\zeta \in H^{1},\)
Furthermore, it holds that
for some positive constant \(\overline{C}\) not depending on \((\varphi ,\, \mu ,\, \sigma ,\, u).\) Let \((\varphi _{i}, \,\mu _{i},\, \sigma _{i})_{i = 1,2}\) denote two weak solutions to (1.1) satisfying (2.8) corresponding to \(\{u_{i}\}_{i=1,2}\) with the same initial data \(\varphi _{0}\) and \(\sigma _{0}.\) Then, there exists a positive constant \(C_{\mathrm {cts}}\) depending only on \(\Vert {\varphi _{i}}\Vert _{L^{\infty }(0,\,T;\,L^{\infty })},\, A,\, B,\, \mathcal {P},\, \mathcal {A},\, \mathcal {C},\, \alpha ,\, k_{5},\, T\) and the Lipschitz constant \(L_{h},\) such that for all \(s \in (0,\,T],\)
The existence of solutions to the state equations (1.1) is proved via a fixed point argument, see also [15] for a similar argument applied to a multispecies tumor model. One may also use a Galerkin approximation, which has been applied to similar systems in [7, 17, 19,20,21, 34, 36]. The key difference here are that we establish boundedness of the nutrient concentration \(\sigma ,\) which comes from the application of a weak comparison principle. Here we also point out that the gradient \(\nabla \varphi \) is continuous on the boundary up to initial time by the embedding \(\varphi \in L^{\infty }(0,\,T;\,H^{2}) \cap H^{1}(0,\,T;\,L^{2}) \subset \subset C^{0}([0,\,T];\,H^{\beta })\) for \(\beta < 2\) and the trace theorem. Hence, the initial condition \(\varphi _{0}\) needs to fulfill the boundary conditions.
The unique solvability of the state equations (1.1) allows us to define a solution operator \(\mathcal {S}\) given as
where the triplet \((\varphi ,\, \mu , \, \sigma )\) is the unique weak solution to (1.1) with data \((\varphi _{0}, \,\sigma _{0},\, u)\) over the time interval \([0,\,T].\) We use the notation \(\varphi = \mathcal {S}_{1}(u)\) for the first component of \(\mathcal {S}(u).\) Then, we deduce the existence of a minimizer to (P).
Theorem 2.2
(Existence of minimizer) Under Assumption 2.1, there exists at least one minimizer \((\varphi _{*},\, u_{*},\, \tau _{*})\) to (P). That is, \(\varphi _{*} = \mathcal {S}_{1}(u_{*})\) with
Note that we cannot exclude the trivial cases where \(\tau _{*} = 0\) or T. To establish the Fréchet differentiability of the solution operator with respect to the control u, we first investigate the linearized state equations. For arbitrary but fixed \(\overline{u} \in \mathcal {U}_{\mathrm {ad}},\) let \((\overline{\varphi },\, \overline{\mu },\, \overline{\sigma }) = \mathcal {S}(\overline{u})\) denote the unique solution triplet to (1.1) from Theorem 2.1. For \(w \in L^{2}(Q),\) we consider the following linearized state equations,
The unique solvability of (2.10) is obtained via a Galerkin procedure.
Theorem 2.3
(Unique solvability of the linearized state equations) For any \(w \in L^{2}(Q),\) there exists a unique triplet \((\Phi ,\, \Xi ,\, \Sigma )\) with
such that for a.e. \(t \in (0,\,T),\) and for all \(\zeta \in H^{1},\)
Furthermore, there exists a constant \(C > 0\) not depending \((\Phi ,\, \Xi ,\, \Sigma , \,w)\) such that
The expectation is as follows. Let \(\overline{u},\, \hat{u} \in \mathcal {U}_{\mathrm {ad}} \subset \mathcal {U}\) be arbitrary, with \((\overline{\varphi },\, \overline{\mu }, \,\overline{\sigma }) = \mathcal {S}(\overline{u})\) and \((\hat{\varphi },\, \hat{\mu },\, \hat{\sigma }) = \mathcal {S}(\hat{u})\) denoting the unique solution triplets to (1.1) corresponding to \(\overline{u}\) and \(\hat{u},\) respectively. Denote by \(w:= \hat{u} - \overline{u} \in L^{2}(Q)\) and let \((\Phi ^{w},\, \Xi ^{w},\, \Sigma ^{w})\) denote the unique solution to the linearized state equations (2.10) associated to w. We define the remainders to be
If, for a suitable Banach space \(\mathcal {Y}\) yet to be identified, we have
then, it holds that the solution operator \(\mathcal {S}\) is Fréchet differentiable at \(\overline{u},\) the Fréchet derivative with respect to the control u, denoted as \(\mathrm{D}_{u} \mathcal {S},\) belongs to \(\mathcal {L}(L^{2}(Q),\, \mathcal {Y}),\) and satisfies
With the unique solvability of the linearized state equations, we have the following result on the Fréchet differentiability of the solution operator.
Theorem 2.4
(Fréchet differentiability with respect to the control) Under Assumption 2.1, the solution operator \(\mathcal {S}\) is Fréchet differentiable in \(\mathcal {U}\) as a mapping from \(L^{2}(Q)\) to the product Banach space
That is, for any \(\hat{u},\, \overline{u} \in \mathcal {U}_{\mathrm {ad}} \subset \mathcal {U}\) with \(w = \hat{u} - \overline{u} \in L^{2}(Q),\) there exists a positive constant \(C_{\mathrm {diff},u}\) not depending on \(\hat{u},\,\overline{u}\) and w such that
where \((\theta ^{w},\, \rho ^{w},\, \xi ^{w})\) are defined as in (2.12).
We now define a reduced functional
For any \(u \in \mathcal {U}_{\mathrm {ad}} \subset \mathcal {U},\) set \(w = u - u_{*} \in L^{2}(Q)\) and let \((\Phi ^{w},\, \Xi ^{w},\, \Sigma ^{w})\) be the unique solution to (2.10) corresponding to w, the optimal control \(u_{*}\) and the corresponding state variables \((\varphi _{*},\, \mu _{*},\, \sigma _{*}).\) By Theorem 2.4, \(\mathcal {J}\) is Fréchet differentiable with respect to the control with
Next, we make use of the following adjoint equation to eliminate the presence of the linearized state variable \(\Phi ^{w}\) in (2.14),
Note that the adjoint system is supplemented with terminal conditions at the optimal treatment time \(\tau _{*}.\) We now state the unique solvability result.
Theorem 2.5
(Unique solvability of the adjoint equations) Under Assumption 2.1, for any \(u \in \mathcal {U}_{\mathrm {ad}}\) there exists a unique triplet \((p,\,q,\,r)\) associated to \(\mathcal {S}(u) = (\varphi ,\, \mu , \,\sigma )\) with
satisfying
for a.e. \(t \in (0,\,\tau _{*})\) and for all \(\eta \in H^{1}\) and \(\zeta \in H^{2}.\)
The first order necessary optimality conditions for the minimizer \((u_{*},\, \tau _{*})\) of Theorem 2.2 also requires the Fréchet derivative of \(\mathcal {J}\) with respect to \(\tau ,\) and for this we make the additional assumption on the target functions \(\varphi _{Q}\) and \(\varphi _{\Omega }.\)
Assumption 2.2
We now assume that \(\varphi _{Q} \in H^{1}(0,\,T;\,L^{2})\) and \(\varphi _{\Omega } \in H^{1}(-r,\,T;\,L^{2}).\)
Furthermore, we extend \(\varphi \) to negative times using the initial condition, i.e., \(\varphi (t) = \varphi _{0}\) for \(t < 0.\)
Theorem 2.6
(Fréchet differentiability of the reduced functional with respect to time) Under Assumptions 2.1 and 2.2, let \(u \in \mathcal {U}_{\mathrm {ad}}\) be arbitrary with corresponding state variables \(\mathcal {S}(u) = (\varphi ,\, \mu ,\, \sigma ).\) The reduced functional \(\mathcal {J}(u, \,\tau )\) is Fréchet differentiable with respect to \(\tau \) with
The first order necessary optimality conditions to (P) for the minimizer \((u_{*},\, \tau _{*})\) are given as follows.
Theorem 2.7
(First order necessary optimality conditions) Under Assumptions 2.1 and 2.2, let \((u_{*},\, \tau _{*}) \in \mathcal {U}_{\mathrm {ad}} \times [0,\,T]\) denote a minimizer to (P) with corresponding state variables \(\mathcal {S}(u_{*}) = (\varphi _{*},\, \mu _{*},\, \sigma _{*})\) and associated adjoint variables \((p,\,q,\,r).\) Then, it holds that
and
Remark 2.1
If we extend p by zero to \((\tau _{*},\, T],\) then we can express (2.17) as
which allows for the interpretation that the optimal control \(u_{*}\) is the \(L^{2}(Q)\)-projection of \(\beta _{u}^{-1} h(\varphi _{*}) \alpha p\) onto \(\mathcal {U}_{\mathrm {ad}}.\)
3 Results on the State Equations
We show the existence of strong solutions to the state equations (1.1) by means of a fixed point argument. The idea is to consider the following two auxiliary problems. Let \(\phi \) be given, we define the solution mapping \(\mathcal {M}_{1}\) by \(\sigma = \mathcal {M}_{1}(\phi ),\) where \(\sigma \) is the unique solution to
with homogeneous Neumann boundary condition and initial condition \(\sigma _{0}.\) Then, we define the solution mapping \(\mathcal {M}\) by \(\varphi = \mathcal {M}(\phi ),\) where \(\varphi \) is the unique solution to
with homogeneous Neumann boundary conditions and initial condition \(\varphi _{0}.\) If \(\tilde{\varphi }\) is a fixed point for \(\mathcal {M},\) with \(\tilde{\sigma } = \mathcal {M}_{1}(\tilde{\varphi })\) and \(\tilde{\mu } = A \Psi ^{\prime }(\tilde{\varphi }) - B \Delta \tilde{\varphi },\) then the triplet \((\tilde{\varphi },\, \tilde{\mu },\, \tilde{\sigma })\) is a solution to (1.1).
3.1 Auxiliary Problems
Lemma 3.1
Let \(\phi \in L^{2}(Q)\) be given. Under Assumption 2.1, there exists a unique solution
to (AP1) such that \(\sigma (0) = \sigma _{0}\) and \(0 \le \sigma \le 1\) a.e. in Q. Furthermore there exists a positive constant \(C_{\mathrm {AP}1}\) not depending on \(\phi \) such that
Proof
As (AP1) is a linear parabolic equation in \(\sigma \), the existence of weak solutions can be shown using a Galerkin approximation, and we will only present the derivation of a priori estimates here. The weak formulation of (AP1) is
for a.e. \(t \in (0,\,T)\) and for all \(\zeta \in H^{1}.\)
First estimate substituting \(\zeta = \sigma \) in (3.2) yields
Neglecting the nonnegative term \(\mathcal {C}h(\phi )\left| {\sigma }\right| ^{2} + \mathcal {B}\left| {\sigma }\right| ^{2},\) and the application of the Gronwall inequality leads to
Second estimate substituting \(\zeta = \partial _{t} \sigma \) in (3.2) yields
where we used the boundedness of h. Integrating in time and using that \(\sigma _{0} \in H^{1},\, \sigma _{S} \in L^{2}(Q)\) we have
Third estimate note that (3.2) can be seen as a weak formulation for the following elliptic problem,
As the right-hand side of (3.5) belongs to \(L^{2}\) for a.e. \(t \in (0,\,T),\) elliptic regularity theory [24, Theorem 2.4.2.7] yields that \(\sigma (t) \in H^{2}\) for a.e. \(t \in (0,\,T),\) with the estimate
where C is a positive constant not depending on \(\sigma \) and \(\phi .\) Integrating in time gives
The a priori estimates (3.3), (3.4) and (3.6) are sufficient to deduce the existence of a strong solution \(\sigma \) satisfying (3.2). The initial condition is attained by the use of the continuous embedding \(L^{2}(0,\,T;\,H^{1}) \cap H^{1}(0,\,T;\,L^{2}) \subset C^{0}([0,\,T];\,L^{2}),\) and using weak/weak-* lower semicontinuity of the norms, we obtain (3.1). We now establish the boundedness property and continuous dependence on the data \(\phi .\)
Boundedness substituting \(\zeta = \sigma ^{-} := \max (-\sigma ,\, 0)\) in (3.2) leads to
As the integrand is nonnegative, we neglect the second term on the left-hand side and upon integrating yields
where we used that \(\sigma _{0} \ge 0\) a.e. in \(\Omega \), and so \(\sigma ^{-}(0) = 0\) a.e. in \(\Omega .\) Thus \(\sigma \ge 0\) a.e. in Q. On the other hand, consider \(\zeta = (\sigma - 1)^{+} = \max (\sigma - 1,\, 0)\) in (3.2), which yields
Using that h is nonnegative and \(\sigma _{S} \le 1\) a.e. in Q, so that \((1-\sigma _{S})(\sigma - 1)^{+}\) is nonnegative, we find that the integrand is nonnegative. Thus, after integrating from 0 to t, we obtain
where we used that \(\sigma _{0} \le 1\) a.e. in \(\Omega .\) This implies that \(\sigma \le 1\) a.e. in Q.
Continuous dependence let \(\{\sigma _{i}\}_{i = 1,2}\) denote two functions satisfying (3.2) corresponding to \(\{\phi _{i}\}_{i = 1,2} \subset L^{2}(Q),\) respectively, and with the same initial condition \(\sigma _{0}\) and nutrient supply \(\sigma _{S}.\) Then the difference \(\sigma := \sigma _{1} - \sigma _{2}\) satisfies
for a.e. \(t \in (0,\,T)\) and for all \(\zeta \in H^{1}.\) Substituting \(\zeta = \sigma \) in (3.9), neglecting the nonnegative term \(\mathcal {C}h(\phi _{2}) \left| {\sigma }\right| ^{2} + \mathcal {B}\left| {\sigma }\right| ^{2},\) and integrate over \([0,\,s]\) for \(s \in (0,\,T],\) we obtain
where we have used the boundedness of \(\sigma _{1}\) and the Lipschitz property of h. Applying Gronwall’s inequality (2.1) yields
where we used that
Next, substituting \(\zeta = \partial _{t} \sigma \) in (3.9) and integrate over \([0,\,s]\) leads to
From (3.10) we have
and so this yields
\(\square \)
Remark 3.1
The main reason we do not employ a Galerkin approximation for the state equation (1.1) is that the computations in the weak comparison principle seems not to apply to the Galerkin solutions, in particular we cannot show that the Galerkin solutions to \(\sigma \) is nonnegative and bounded above by 1. Indeed, for Galerkin solutions \(\varphi _{n}\) and \(\sigma _{n}\) belong to some finite dimensional subspace \(W_{n}\) of \(H^{1}\) satisfying
for all \(v \in W_{n},\) if we test with \(v = \Pi _{n}(\sigma _{n}^{-}) \in W_{n}\) where \(\Pi _{n}\) denotes the orthogonal projection to \(W_{n},\) we have
but we cannot deduce if
is nonpositive. There is also a similar issue with the nonnegativity of
as it is not guaranteed that the projection of a nonnegative function is nonnegative.
Due to the estimate (3.11), we can define a continuous mapping
Lemma 3.2
Let \(\phi \in L^{2}(Q)\) be given. Under Assumption 2.1, there exists a unique solution pair
to (AP2) such that \(\varphi (0) = \varphi _{0}\) and satisfy
for a.e. \(t \in (0,\,T)\) and for all \(\zeta \in H^{1}.\) Furthermore, there exists a positive constant \(C_{\mathrm {AP2}}\) depending only on \(T,\, \Omega ,\, k_{0},\, k_{1},\, k_{2},\, k_{3},\, k_{4},\, A,\, B,\, \mathcal {P},\, \mathcal {A},\, \alpha ,\, \Vert {\varphi _{0}}\Vert _{H^{3}}\) and \(C_{\mathrm {AP}1},\) such that
That is, \(C_{\mathrm {AP2}}\) does not depend on \(\phi .\)
Proof
Let \(\{w_{i}\}_{i \in \mathbb {N}}\) denote the eigenfunctions of the Neumann–Laplacian with corresponding eigenvalues \(\{\lambda _{i}\}_{i \in \mathbb {N}}{\text {:}}\)
Then, it is well-known that \(\{w_{i}\}_{i \in \mathbb {N}}\) forms an orthonormal basis of \(L^{2}\) and an orthogonal basis of \(H^{1}.\) As constant functions are eigenfunctions, we take \(w_{1} = 1\) with \(\lambda _{1} = 0.\) Let \(n \in N\) be fixed and we define \(W_{n} := \mathrm {span}\{w_{1}, \ldots , w_{n}\}\) as the finite dimensional space spanned by the first n eigenfunctions, with the corresponding projection operator \(\Pi _{n}.\) We consider sequences \(\{\phi _{n}\}_{n \in \mathbb {N}},\, \{u_{n}\}_{n \in \mathbb {N}} \subset C^{0}([0,T];L^{2})\). such that \(\phi _{n} \rightarrow \phi \) and \(u_{n} \rightarrow u\) strongly in \(L^{2}(0,\,T;\,L^{2})\) and look for functions of the form
where the coefficients \(\varvec{a}_{n} := \{a_{n,i}\}_{i=1}^{n}\) and \(\varvec{b}_{n} := \{b_{n,i}\}_{i=1}^{n}\) satisfy the following initial-value problem
with prime denoting the time derivative and for \(1 \le i,\,j \le n,\)
Without loss of generality, we assume that \(0 \le u_{n} \le 1\) a.e. in Q for all \(n \in \mathbb {N}\) and from Lemma 3.1 it holds that \(0 \le \mathcal {M}_{1}(\phi _{n}) \le 1\) a.e. in Q and \(\mathcal {M}_{1}(\phi _{n}) \in C^{0}([0,\,T];\,L^{2})\) for all \(n \in \mathbb {N}.\) Substituting (3.15b) into (3.15a) leads to a system of ODEs in \(\varvec{a}_{n}\) with right-hand side depending continuously on t and \(\varvec{a}_{n}.\) By the Cauchy–Peano theorem [4, Chap. 1, Theorem 1.2], there exists a \(t_{n} \in (0,\,T]\) such that (3.15) has a local solution \(\varvec{a}_{n}\) on \([0,\,t_{n})\) with \(\varvec{a}_{n} \in C^{1}([0,\,t_{n});\,\mathbb {R}^{n}).\) Then, \(\varvec{b}_{n}\) can be defined by the relation (3.15b), and we obtain functions \(\varphi _{n},\, \mu _{n} \in C^{1}([0,\,t_{n});\,W_{n})\) satisfying
In the following we will derive a series of a priori estimates leading to the uniform boundedness (in n) of \((\varphi _{n},\, \mu _{n})\) in the following Bochner spaces:
-
(1)
\(\Psi (\varphi _{n}) \in L^{\infty }(0,\,T;\,L^{1}),\, \varphi _{n} \in L^{\infty }(0,\,T;\,H^{1}),\, \mu _{n} \in L^{2}(0,\,T;\,H^{1}),\)
-
(2)
\(\varphi _{n} \in L^{2}(0,\,T;\,H^{3}),\)
-
(3)
\(\mu _{n} \in L^{\infty }(0,\,T;\,L^{2}) \cap L^{2}(0,\,T;\,H^{2}),\, \varphi _{n} \in L^{\infty }(0,\,T;\,H^{2}),\, \partial _{t}\varphi _{n} \in L^{2}(0,\,T;\,L^{2}).\)
In particular for the third estimate, we have to differentiate (3.17b) in time to obtain a system of ODEs involving \(\partial _{t}\mu _{n}.\) Thus, we prescribe additional initial conditions, namely we set
Note that by Assumption 2.1, there exists a positive constant \(C_{\mathrm {ini}},\) not depending on \(\phi \) and n, such that
Furthermore, to approximate \(\varphi _{0}\) by a linear combination of eigenfunctions of the Neumann–Laplacian in \(H^{2},\) we require that \(\varphi _{0}\) satisfies zero Neumann boundary conditions.
First estimate multiplying (3.17a) with \(\mu _{n}\) and (3.17b) with \(\partial _{t}\varphi _{n},\) integrate over \(\Omega \) and integrate by parts, upon adding and using the boundedness of \(h,\, \mathcal {M}_{1}(\phi _{n})\) and \(u_{n},\) we obtain
Let \(C_{u} := \mathcal {P}+ \mathcal {A}+ \alpha ,\) then by the Poincaré inequality in \(L^{1}\) (with constant \(C_{p} > 0\) depending only on \(\Omega \)), Hölder’s inequality and Young’s inequality, the right-hand side of (3.18) can be estimated as follows,
From integrating (3.17b) over \(\Omega ,\) and (2.4), we find that
Hence, we obtain the following differential inequality
By the Sobolev embedding \(H^{1} \subset L^{6}\) and the growth assumption (2.6), it holds that \(\Vert {\Psi (\varphi _{0})}\Vert _{L^{1}} \le C(1 + \Vert {\varphi _{0}}\Vert _{L^{4}}^{4}) \le C(1 + \Vert {\varphi _{0}}\Vert _{H^{1}}^{4}).\) Thus \(c_{0} := A \Vert {\Psi (\varphi _{0})}\Vert _{L^{1}} + \frac{B}{2} \Vert {\nabla \varphi _{0}}\Vert _{L^{2}}^{2}\) is bounded. Integrating over \([0,\,s]\) for \(s \in (0,\,T]\) yields
Applying the Gronwall inequality (2.1) gives
for all \(s \in (0,\,T].\) Taking supremum in s leads to
where the constant C depends only on \(T,\, C_{u},\, C_{p},\, k_{0},\, k_{1},\, A,\, B,\, \left| {\Omega }\right| ,\) and \(\Vert {\varphi _{0}}\Vert _{H^{1}}.\) From (3.20) and (3.22), the mean of \(\mu _{n}\) is bounded in \(L^{\infty }(0,\,T),\) and the Poincaré inequality gives that \(\mu _{n}\) is bounded in \(L^{2}(0,\,T;\,L^{2}).\) Meanwhile, by (2.5) we see that
and thus by (3.22), the mean of \(\varphi _{n}\) is bounded in \(L^{\infty }(0,\,T),\) and by the Poincaré inequality we obtain that \(\varphi _{n}\) is also bounded in \(L^{\infty }(0,\,T;\,L^{2}).\) Thus, there exists a positive constant C, not depending on \(\phi _{n}\) and n such that
and as a result, this guarantees that the Galerkin solutions \((\varphi _{n}, \,\mu _{n})\) can be extended to the interval \([0,\,T],\) and thus \(t_{n} = T\) for each \(n \in \mathbb {N}.\)
Second estimate from (2.6) and the Sobolev embedding \(H^{1} \subset L^{6},\) we have that
where C is a positive constant depending only on \(k_{4}\) and \(\Omega .\) Since \(\Vert {\Pi _{n}(\Psi ^{\prime }(\varphi _{n}))}\Vert _{L^{2}} \le \Vert {\Psi ^{\prime }(\varphi _{n})}\Vert _{L^{2}},\) applying elliptic regularity to (3.17b) yields that \(\varphi _{n}(t) \in H^{2}\) for a.e. \(t \in (0,\,T)\) and satisfies
with a positive constant C depending only on \(\Omega ,\, A\) and B. Then, by the Gagliardo–Nirenburg inequality (2.3) with \(d = 3,\, p = 10,\, j = 0,\, r = 2,\, m = 2,\, q = 6\) and \(\alpha = \frac{1}{5},\) we have
and with \(d = 3,\, p = \frac{10}{3},\, j = 0,\, r = 2,\, m = 1,\, q = 2,\) and \(\alpha = \frac{3}{5},\) we have
Then, by (2.6) we have
and so \(\Psi ^{\prime }(\varphi _{n}) \in L^{2}(0,\,T;\,H^{1}).\) Application of elliptic regularity yields that \(\varphi _{n}(t) \in H^{3}\) for a.e. \(t \in (0,\,T)\) and
for a positive constant C not depending on \(\phi _{n}\) and n.
Third estimate differentiate (3.17b) in time and we obtain
Multiplying (3.24) with \(\mu _{n}\) and (3.17a) with \(B \partial _{t}\varphi _{n},\) integrating over \(\Omega \) and we obtain upon summing
From (2.6), we find that
and so \(\Psi ^{\prime \prime }(\varphi _{n})\) is bounded in \(L^{\infty }(0,\,T;\,L^{3}).\) Applying Hölder’s inequality on the right-hand side of (3.25) yields
where we recall \(C_{u} = \mathcal {P}+ \mathcal {A}+ \alpha \) and \(C_{\mathrm {Sob}}\) is the positive constant from the Sobolev embedding \(H^{1} \subset L^{6}\) depending only on \(\Omega .\) Then, integrating in time and using that \(\mu _{n}\) is bounded in \(L^{2}(0,\,T;\,H^{1}),\) and \(\Vert {\mu _{n}(0)}\Vert _{L^{2}}^{2} \le C_{\mathrm {ini}}\Vert {\varphi _{0}}\Vert _{H^{3}}^{2},\) we have
where the positive constant C depends only on \(\Omega ,\, C_{u},\, A,\, B,\, \Vert {\varphi _{n}}\Vert _{L^{\infty }(0,T;H^{1})},\, \Vert {\mu _{n}}\Vert _{L^{2}(0,T;H^{1})},\, k_{4},\) and \(\Vert {\varphi _{0}}\Vert _{H^{3}}.\) Furthermore, by (2.6) we have that
Together with the improved regularity \(\mu _{n} \in L^{\infty }(0,\,T;\,L^{2}),\) when we revisit the elliptic equation (3.17b) we find that
with a positive constant C not depending on \(\phi _{n}\) and n. Similarly, viewing (3.17a) as an elliptic problem for \(\mu _{n},\) and as \(\partial _{t}\varphi _{n} \in L^{2}(0,\,T;\,L^{2}),\) we have by elliptic regularity
where the positive constant C does not depend on \(\phi _{n}\) or n.
Compactness from the above a priori estimates, we obtain for a non-relabelled subsequence,
and thanks to the compact embedding [1, Theorem 6.3 part III]
where d is the space dimension, we find that \(H^{2}(\Omega )\) is compactly embedded into \(C^{0}(\overline{\Omega }).\) Hence, by [43, §8, Corollary 4] we have the following strong convergences
for any \(1 \le r < 6.\) The initial condition \(\varphi _{0}\) is attained from that fact that \(\varphi \in C^{0}([0,\,T];\,H^{1}).\) It follows from standard arguments that the pair \((\varphi ,\, \mu )\) satisfies (3.13), see for instance [20, 21]. Furthermore, by weak/weak* lower semicontinuity of the norms, we obtain (3.14).
Continuous dependence let \(\{(\varphi _{i},\,\mu _{i})\}_{i=1,2}\) denote two solution pairs satisfying (3.13) with the same initial condition \(\varphi _{0}\) and corresponding data \(\{(\phi _{i},\, u_{i})\}_{i=1,2},\) respectively. Then, it holds that the difference \(\varphi := \varphi _{1} - \varphi _{2}\) and \(\mu := \mu _{1} - \mu _{2}\) satisfy
where
Substituting \(\zeta = B \varphi \) in (3.27a) and \(\zeta = \mu \) in (3.27b), integrating over \([0,\,s]\) for \(s \in (0,\,T]\) and upon adding we obtain
By the boundedness of \(\mathcal {M}_{1}(\phi _{1})\) and \(u_{1},\) the Lipschitz continuity of h, we obtain
while by Hölder’s inequality and Young’s inequality, and the boundedness of h, we have
Using (2.7) and the fact that \(\varphi _{i} \in C^{0}(\overline{Q}),\) we find that
Then, substituting the above three estimates into (3.28) we obtain for \(s \in (0,\,T],\)
where
is a positive constant. Applying (2.1) yields for any \(s \in (0,\,T],\)
where we used that \(W(t) := \Vert {\overline{u}}\Vert _{L^{2}(0,t;L^{2})}^{2} + \Vert {\overline{\mathcal {M}_{1}}}\Vert _{L^{2}(0,t;L^{2})}^{2}\) is a nondecreasing function of t, and thus
\(\square \)
We point out that although the source term in (1.1a) closely resembles that of [21], we obtain a priori estimates for potentials \(\Psi \) with quartic growth (see (2.6)), which is in contrast to the quadratic potentials considered in [21]. The main difference is that here we have the boundedness of the nutrient, and thus we only require a bound on the mean of \(\mu \) (see (3.19)). But in [21], the presence of the active transport mechanism [modeled by the term \(\, {\mathrm{div}} \, (n(\varphi )\chi \nabla \varphi )\) in the nutrient equation] prevents us from applying a weak comparison principle to deduce the boundedness of the nutrient. Without the boundedness of the nutrient, we have to control the square of the mean of \(\mu \) in order to estimate the source term \(h(\varphi _{n})(\mathcal {P}\sigma _{n} - \mathcal {A}- \alpha u_{n}) \mu _{n}.\)
3.2 Existence by Schauder’s Fixed Point Theorem
Note that if \(\{\phi _{n}\}_{n \in \mathbb {N}}\) is a bounded sequence in \(L^{2}(Q),\) by Lemma 3.1 the corresponding sequence \(\{\sigma _{n} := \mathcal {M}_{1}(\phi _{n})\}_{n \in \mathbb {N}}\) satisfies \(0 \le \sigma _{n} \le 1\) a.e. in Q, and by Lemma 3.2 we have that the corresponding solution pair \(\{\varphi _{n},\,\mu _{n}\}_{n \in \mathbb {N}}\) is bounded uniformly in
which yields a strongly convergent non-relabelled subsequence \(\{\varphi _{n}\}_{n \in \mathbb {N}}\) in \(L^{2}(Q),\) due to the compact embedding
Thus, the mapping
is compact. To apply Schauder’s fixed point theorem [23, Theorem 11.3] and deduce the existence of a fixed point of the mapping \(\mathcal {M},\) we need to check that if there exists a constant M such that
The problem \(\varphi = \lambda \mathcal {M}(\varphi )\) translates to
By Lemma 3.1 we have that \(0 \le \sigma \le 1\) a.e. in Q for all \(\lambda \in [0,\,1],\) and thus we can choose M to be the constant \(C_{\mathrm {AP2}}\) in (3.14) which does not depend on \(\varphi \) and \(\lambda \in [0,\,1].\) Thus Schauder’s fixed point theorem yields the existence of a strong solution \((\varphi ,\, \mu ,\, \sigma )\) to the state equations (1.1) with \(0 \le \sigma \le 1\) a.e. in Q and
for some positive constant \(\overline{C}\) not depending on \((\varphi ,\, \mu ,\, \sigma ,\, u).\)
3.3 Continuous Dependence
We now establish continuous dependence on the control u. For this purpose, let \(u_{1},\, u_{2} \in \mathcal {U}_{\mathrm {ad}}\) be given, along with the corresponding solution triplet \((\varphi _{1},\, \mu _{1},\, \sigma _{1})\) and \((\varphi _{2},\, \mu _{2},\, \sigma _{2})\) satisfying the same initial data \(\varphi _{0}\) and \(\sigma _{0}.\) Let \(\varphi = \varphi _{1} - \varphi _{2},\, \mu = \mu _{1} - \mu _{2}\) and \(\sigma = \sigma _{1} - \sigma _{2},\) then from (3.11) we obtain
Substituting this into (3.30) leads to
Setting
we obtain from (2.1) that
Combining with (3.10) and (3.12), we find that there exists a positive constant \(C_{1},\) depending only on \(B,\, \mathcal {Q},\, \mathcal {C},\, L_{h}, \,T\) such that
for \(s \in (0,\,T].\) Next, we find using (2.7) and the fact that \(\varphi _{i} \in C^{0}(\overline{Q})\) for \(i = 1,\,2,\)
and so viewing (3.27b) as an elliptic problem for \(\varphi ,\) we obtain by elliptic regularity
where \(C_{2}\) is a positive constant depending only on \(\Omega ,\, A,\, k_{5},\, \Vert {\varphi _{i}}\Vert _{L^{\infty }(Q)},\, T\) and \(C_{1}.\)
4 Existence of a Minimizer
From (3.31) it holds that
where \(\overline{C}\) is a positive constant independent of \((\varphi ,\, \mu ,\, \sigma ,\, u).\) Hence, we obtain that
As \(J_{r}\) is bounded from below, we can consider a minimizing sequence \((u_{n},\, \tau _{n})_{n \in \mathbb {N}}\) with \(u_{n} \in \mathcal {U}_{\mathrm {ad}},\, \tau _{n} \in (0,\,T)\) and corresponding solutions \((\varphi _{n},\, \mu _{n}, \,\sigma _{n})_{n \in \mathbb {N}}\) on the interval \([0,\,T]\) with \(\varphi _{n}(0) = \varphi _{0}\) and \(\sigma _{n}(0) = \sigma _{0}\) for all \(n \in \mathbb {N},\) such that
In particular, \(u_{n} \in \mathcal {U}_{\mathrm {ad}}\) implies that \(0 \le u_{n} \le 1\) a.e. in Q for all \(n \in \mathbb {N}.\) As \(\{\tau _{n}\}_{n \in \mathbb {N}}\) is a bounded sequence, there exists a non-relabelled subsequence such that
and
where \((\varphi _{*},\, \mu _{*},\, \sigma _{*},\, u_{*})\) satisfy (2.8) with \(0 \le u_{*},\, \sigma _{*} \le 1\) a.e. in Q. Note that by the dominating convergence theorem, for all \(p \in [1,\,\infty ),\)
Then, by the strong convergence of \(\varphi _{n} - \varphi _{Q}\) to \(\varphi _{*} - \varphi _{Q}\) in \(L^{2}(Q)\) and the strong convergence \(\chi _{[0,\tau _{n}]}(t)\) to \(\chi _{[0,\tau _{*}]}(t)\) also in \(L^{2}(Q),\) we have
A similar argument yields
Then, by passing to the limit \(n \rightarrow \infty \) in \(J_{r}(\varphi _{n},\, u_{n},\, \tau _{n})\) and using (4.1) and (4.2), we have
which implies that \((u_{*},\,\tau _{*})\) is a minimizer of (P).
5 Fréchet Differentiability of the Solution Operator
5.1 Unique Solvability of the Linearized State Equations
Recalling the set \(\{w_{i}\}_{i \in \mathbb {N}}\) of eigenfunctions of the Neumann–Laplacian from the proof of Lemma 3.2, we look for functions of the form
satisfying
for all \(v \in W_{n}.\) Substituting \(v = w_{j}\) leads to
where the matrix \(\varvec{S}\) has been defined in (3.16), and for \(1 \le i,\,j \le n,\)
Taking an approximating sequence in \(C^{0}([0,\,T];\,L^{2})\) for \(\overline{u},\) which we will abuse notation and reuse the variable \(\overline{u},\) and then supplementing (5.2) with the initial conditions \(\varvec{\gamma }_{n}(0) = \varvec{0}\) and \(\varvec{\eta }_{n}(0) = \varvec{0}\) leads to a system of ODEs with right-hand sides depending continuously on \((t,\, \varvec{\gamma }_{n},\, \varvec{\eta }_{n}).\) Thus, by the Cauchy–Peano theorem, there exists \(t_{n} \in (0,\,T]\) such that (5.2) has a local solution \((\varvec{\gamma }_{n},\, \varvec{\delta }_{n},\, \varvec{\eta }_{n})\) on \([0,\,t_{n}]\) with \(\varvec{\gamma }_{n},\, \varvec{\delta }_{n},\, \varvec{\eta }_{n} \in C^{1}([0,\,t_{n});\,\mathbb {R}^{n}).\) Then, we obtain functions \(\Phi _{n}, \,\Xi _{n}, \,\Sigma _{n} \in C^{1}([0,\,t_{n});\,W_{n})\) satisfying (5.1).
First estimate substituting \(v = \Phi _{n}\) in (5.1a), \(v = \Delta \Phi _{n}\) in (5.1b) and \(v = \Sigma _{n}\) in (5.1c), integrating over \([0,\,t]\) for \(t \in (0,\,T],\) and integrating by parts, we obtain after summation
where we used that \(\Sigma _{n}(0) = \Phi _{n}(0) = 0\) and have neglected the nonnegative term \((\mathcal {B}+ \mathcal {C}h(\overline{\varphi })) \left| {\Sigma _{n}}\right| ^{2}.\) From Theorem 2.1, we have \(\overline{\varphi } \in C^{0}(\overline{Q}),\) and as \(\Psi ^{\prime \prime },\, \Psi ^{\prime \prime \prime },\,h^{\prime }\) and \(h^{\prime \prime }\) are continuous with respect to their arguments, it holds that there exists a constant \(C_{*} > 0\) such that
Then, applying Hölder’s inequality and Young’s inequality we obtain
Using the estimates for \(I_{1}, I_{2},\, I_{3}\) and \(I_{4},\) we obtain
where \(C_{5},\, C_{6} > 0\) are positive constants depending only on \(C_{*},\, A,\, B,\, \mathcal {P},\, \mathcal {A},\, \mathcal {C},\) and \(\alpha .\) Applying the integral form of Gronwall’s inequality we obtain that
for some constant \(D_{1}\) not depending on n, which in turn implies that
Second estimate substituting \(v = \partial _{t} \Sigma _{n}\) in (5.1c), we obtain
Applying Gronwall’s inequality yields that
where \(D_{2}\) is a positive constant not depending on n. Hence,
Furthermore, since \(\mathcal {C}(h(\overline{\varphi }) \Sigma _{n} + h^{\prime }(\overline{\varphi }) \Phi _{n} \overline{\sigma }) - \partial _{t} \Sigma _{n} \in L^{2}\) for a.e. \(t \in (0,\,T),\) we obtain from elliptic regularity theory that
where C and \(D_{3}\) are positive constants not depending on n. Thus,
Third estimate substituting \(v = 1\) in (5.1b) yields
Then, by the Poincaré inequality we find that
and thus
Substituting \(v = \Xi _{n}\) in (5.1a) and \(v = {-} \partial _{t} \Phi _{n}\) in (5.1b), and upon summing and integrating over \([0,\,t]\) for \(t \in (0,\,T],\) we obtain
Applying Hölder’s inequality and Young’s inequality and (5.7), we observe that
where \(C_{7},\, C_{8} > 0\) are positive constants depending only on \(C_{*},\, C_{p},\, \mathcal {P},\, \mathcal {A},\) and \(\alpha .\) To estimate \(J_{1}\) we first obtain an estimate for \(\Vert {\partial _{t}\Phi _{n}}\Vert _{L^{2}(0,t;(H^{1})^{*})}\) by considering \(v \in L^{2}(0,\,T;\,H^{1})\) in (5.1a) and integrating over \([0,\,t].\) Then, we obtain that
Thus, for \(J_{1}\) we have
where \(C_{9} > 0\) is a positive constant depending only in \(A,\, \mathcal {P},\, \alpha ,\, C_{*},\) and \(\mathcal {A}.\) Returning to (5.8) we have
Applying the integral form of Gronwall’s inequality and recalling (5.7) and (5.9), we find that
where \(D_{4}\) is a positive constant not depending on n, and so
Furthermore, as \(\Xi _{n} - A \Psi ^{\prime \prime }(\overline{\varphi }) \Phi _{n} \in H^{1}\) for a.e. \(t \in (0,\,T),\) applying elliptic regularity to (5.1b) yields that
where C and \(D_{5}\) are positive constants not depending on n. This implies that
The a priori estimates (5.4)–(5.6), (5.10) and (5.11) imply that \((\Phi _{n}, \,\Xi _{n},\, \Sigma _{n})\) can be extended to the interval \([0,\,T],\) and thus \(t_{n} = T\) for each \(n \in \mathbb {N}.\) Furthermore, there exists a non-relabelled subsequence such that
and a standard argument shows that the limit functions \((\Phi ,\, \Xi ,\, \Sigma )\) satisfy (2.11).
Uniqueness let \((\Phi _{i},\, \Xi _{i},\, \Sigma _{i})_{i =1,2}\) denote two weak solution triplets to (2.10) with the same data \(w \in L^{2}(Q).\) Then, as (2.10) is linear in \((\Phi ,\, \Xi ,\, \Sigma ),\) the differences \(\Phi := \Phi _{1} - \Phi _{2},\, \Xi := \Xi _{1} - \Xi _{2}\) and \(\Sigma := \Sigma _{1} - \Sigma _{2}\) satisfy (2.10) with \(w = 0.\) Due to the regularity of the solutions, the derivation of (5.4)–(5.6) remain valid, which implies that
and so \(\Phi = \Sigma = 0.\) Substituting \(\Phi = 0\) in (2.11b) yields that \(\Xi = 0.\)
5.2 Fréchet Differentiability with Respect to the Control
In this section we use the notation \(\varphi ^{w} = \hat{\varphi },\, \mu ^{w} = \hat{\mu },\, \sigma ^{w} = \hat{\sigma }.\) The remainders \((\theta ^{w},\, \rho ^{w},\, \xi ^{w})\) from (2.12) satisfy
for a.e. \(t \in (0,\,T)\) and for all \(\zeta \in H^{1}\) with
Using the Taylor’s theorem with integral remainder (2.2) we see that
and so for \(\varphi ^{w} - \overline{\varphi } = \Phi ^{w} + \theta ^{w},\) we have
where
Thanks to the fact that \(\overline{\varphi },\, \varphi ^{w} \in C^{0}(\overline{Q})\) and the continuity of \(\Psi ^{\prime \prime \prime }\) and \(h^{\prime \prime },\) we see that there exists a constant \(C_{**} > 0\) such that
Furthermore, we can express
Let \(X^{w} := \mathcal {P}\sigma ^{w} - \mathcal {A}- \alpha (\overline{u} + w)\) and \(\overline{X} := \mathcal {P}\overline{\sigma } - \mathcal {A}- \alpha \overline{u}.\) Then, it holds similarly that
and thus, we see that \((\theta ^{w},\, \rho ^{w},\, \xi ^{w})\) satisfy
for a.e. \(t \in (0,\,T)\) and for all \(\zeta \in H^{1}.\)
First estimate let us first compute the following preliminary estimates, using the continuous dependence estimate (2.9), the Lipschitz continuity of h, Hölder’s inequality, Young’s inequality and the embedding \(L^{2}(0,\,T;\,H^{2}) \subset L^{2}(0,\,T;\,L^{\infty }),\) we have that
where \(C_{10}\) is a positive constant depending only on \(C_{\mathrm {cts}},\, C_{\mathrm {Sob}}\) and \(L_{h}.\) Meanwhile, using the boundedness of \(\overline{\sigma },\,h^{\prime }(\overline{\varphi })\) and \(R_{2}^{w}\) in Q, we see that
Thus, when we substitute \(\zeta = \xi ^{w}\) in (5.15c), integrating over \([0,\,s]\) for \(s \in (0,\,T],\) and neglecting the nonnegative term \(\mathcal {B}\left| {\xi ^{w}}\right| ^{2} + \mathcal {C}h(\overline{\varphi }) \left| {\xi ^{w}}\right| ^{2},\) we obtain
Next, substituting \(\zeta = \theta ^{w}\) in (5.15a), \(\zeta = \theta ^{w}\) in (5.15b) and \(\zeta = \frac{1}{B} \rho ^{w}\) in (5.15b), integrating by parts and integrating over \([0,\,s]\) for \(s \in (0,\,T],\) and upon adding leads to
Using (2.9), Hölder’s inequality, Young’s inequality, the boundedness of \(\Psi ^{\prime \prime }(\overline{\varphi })\) and \(R_{1}^{w}\) in Q, we have
where \(C_{11}\) is a positive constant depending only on \(C_{*},\, C_{**},\, C_{\mathrm {cts}},\, A\) and B. Meanwhile, by the Lipschitz continuity of h, and the fact that
we see that
for some positive constant \(C_{12}\) depending only on \(B,\, L_{h},\, \mathcal {P}\) and \(\alpha .\) Furthermore, using the boundedness of \(\overline{X},\,h^{\prime }(\overline{\varphi }),\, R_{2}^{w}\) and \(h(\overline{\varphi })\) in Q, we have
where we recall \(C_{u} = (\mathcal {P}+ \mathcal {A}+ \alpha ).\) Substituting the above estimates into (5.17) we obtain
Then, adding (5.16) and (5.18) we have for \(s \in (0,\,T],\)
where the positive constants \(C_{13},\, C_{14}\) depend only on \(\mathcal {C},\, C_{\mathrm {cts}},\, C_{*},\, C_{**},\, C_{10}, C_{11},\, C_{12},\, \mathcal {P},\, \mathcal {A},\, \alpha ,\,A\) and B. Applying Gronwall’s inequality to (5.19) we have that
for some positive constant \(C_{15}\) depending only on \(C_{13}\) and \(C_{14}.\)
Second estimate substituting \(\zeta = \partial _{t} \xi ^{w}\) in (5.15c), integrating over \([0,\,s]\) for \(s \in (0,\,T]\) leads to
Using the Lipschitz continuity of h, the boundedness of \(\overline{\sigma },\, h(\overline{\varphi }),\, h^{\prime }(\overline{\varphi }),\) and \(R_{2}^{w}\) in Q, Hölder’s inequality, Young’s inequality and (5.20), we obtain
for some positive constant \(C_{16}\) depending only on \(T,\, L_{h},\, \mathcal {C},\, C_{\mathrm {cts}},\, C_{*},\, C_{**}\) and \(C_{15}.\)
Third estimate viewing (5.15b) as the weak formulation of an elliptic problem for \(\theta ^{w},\) by elliptic regularity we obtain
for some positive constant \(C_{17}\) not depending on \(\theta ^{w},\, \rho ^{w}\) and w. Applying (5.20), the boundedness of \(\Psi ^{\prime \prime }(\overline{\varphi })\) and \(R_{1}^{w}\) in Q, we have
for some positive constant \(C_{18}\) depending only on \(C_{15},\, C_{17},\, C_{*},\, C_{**},\, C_{\mathrm {cts}},\, A\) and B. Then, upon integrating (5.15a) over \([0,\,s]\) for \(s \in (0,\,T],\) integrating by parts then yields
By Hölder’s inequality, the boundedness of \(\overline{X} = \mathcal {P}\overline{\sigma } - \mathcal {A}- \alpha \overline{u},\, h^{\prime }(\overline{\varphi }),\, R_{2}^{w}\), and \(h(\overline{\varphi })\) in Q, (2.9), and (5.20) we have that
for some positive constant \(C_{19}\) depending only on \(\mathcal {P},\, \mathcal {A},\, \alpha ,\, C_{*},\, C_{**},\, C_{\mathrm {cts}},\, T\) and \(C_{15}.\) Meanwhile,
where \(C_{20}\) is a positive constant depending only on \(C_{15},\, L_{h},\, \mathcal {P},\, \alpha ,\, C_{\mathrm {cts}}\) and \(\Omega \) (via the Sobolev embedding \(H^{1} \subset L^{6}\)). Hence, we see that
By the continuous embedding \(L^{2}(0,\,T;\,H^{2}) \cap H^{1}(0,\,T;\,(H^{2})^{*}) \subset C^{0}([0,\,T];\,L^{2}),\) we find that there exists a positive constant \(C_{21}\) depending only on \(C_{18},\, C_{19}\) and \(C_{20}\) such that
Combining this with (5.20) and (5.21) yields (2.13).
5.3 Fréchet Differentiability of the Objective Functional with Respect to Time
In this section, we assume that Assumption 2.2 holds. Using the relation
for \(f \in L^{1}(-r,\,T;\,L^{1})\) and \(\tau \in (0,\,T),\) we can define
and upon setting \(\varphi (t) = \varphi _{0}\) for \(t \le 0,\) we can express (1.2) as
Note that only the last two terms on the right-hand side are dependent on \(\tau ,\) and thus the first three terms on the right-hand side will vanish when we compute the Fréchet derivative of \(J_{r}\) with respect to \(\tau \). We now compute for any \(f \in H^{1}(0,\,T) \subset L^{\infty }(0,\,T),\) and \(\tau \in (0,\,T),\) \(h > 0\) such that \(\tau + h \in (0,\,T),\)
This shows that
and a similar argument also yields
Using the fact that \(\varphi _{Q} \in H^{1}(0,\,T;\,L^{2}),\, \varphi _{*},\, \varphi _{\Omega } \in H^{1}(-r,\,T;\,L^{2}),\) we obtain that the optimal control \((u_{*},\, \tau _{*})\) satisfies
where
We can simplify (5.23) with the following argument. If \(\tau _{*}\in (0,\,T),\) choose \(s = \tau _{*}\pm h\) for \(h > 0\) to deduce that \(\mathrm{D}_{\tau } \mathcal {J}(u_{*},\,\tau _{*}) = 0.\) If \(\tau _{*}= 0,\) then from (5.23) we obtain \(\mathrm{D}_{\tau } \mathcal {J}(u_{*},\,\tau _{*}) \ge 0.\) Meanwhile, if \(\tau _{*}= T,\) then \(s - \tau _{*}\le 0\) for any \(s \in [0,\,T],\) and thus \(\mathrm{D}_{\tau } \mathcal {J}(u_{*},\,\tau _{*}) \le 0.\)
6 First Order Necessary Optimality Conditions
6.1 Unique Solvability of the Adjoint System
We apply a Galerkin approximation and consider a basis \(\{w_{i}\}_{i \in \mathbb {N}}\) of \(H^{2}\) that is orthonormal in \(L^{2},\) and we look for functions of the form
which satisfy
for all \(v \in W_{n} := \mathrm {span}\{w_{1}, \ldots , w_{n}\}.\) Substituting \(v = w_{j}\) leads to
where \(\varvec{S}\) is defined in (3.16), and
and we supplement the above backward-in-time system of ODEs with the conditions
Once again, we consider approximating sequences in \(C^{0}([0,\,T];\,L^{2})\) for \(u,\, \varphi _{Q}\) and \(\varphi _{\Omega }\) and use the same variables to denote the approximating functions. Note that the right-hand side of (6.2) depends continuously on \((\varvec{P}_{n},\, \varvec{Q}_{n},\, \varvec{R}_{n})\) but due to the term \(\chi _{(\tau _{*}-r,\tau _{*})}(t) \varvec{G}_{n}\) in the equation for \(\varvec{P}_{n}^{\prime },\) we cannot apply the Cauchy–Peano theorem directly. However, we can consider first solving (6.2) on the interval \((\tau _{*}-r,\,\tau _{*}],\) that is, \(\varvec{P}_{n}\) and \(\varvec{R}_{n}\) satisfy
for \(t \in (\tau _{*}-r,\,\tau _{*}],\) which would yield, via the Cauchy–Peano theorem, the existence of \(t_{n} \in [\tau _{*}-r,\,\tau _{*})\) and a local solution pair \((\varvec{P}_{n},\, \varvec{R}_{n}) \in ( C^{1}((t_{n},\, \tau _{*}];\, \mathbb {R}^{n}) )^{2}\) to (6.3). The a priori estimates derived below will allow us to deduce that \((\varvec{P}_{n},\, \varvec{R}_{n})\) can be extended to \(\tau _{*}-r,\) that is, \(t_{n} = \tau _{*}- r\) for all \(n \in \mathbb {N}.\) Then, we then extend the solutions by solving the system
with terminal conditions at time \(\tau _{*}-r.\) Overall, this procedure yields functions \(p_{n},\, q_{n},\, r_{n} \in C^{1}((t_{n},\, \tau _{*}];\,W_{n})\) satisfying (6.3) for some \(t_{n} \in [0,\,\tau _{*}).\) We now derive the a priori estimates.
First estimate substituting \(v = r_{n}\) in (6.1c) and integrating over \([s,\, \tau _{*}]\) for \(s \in (0,\,\tau _{*})\) leads to
where we neglected the nonnegative term \(\mathcal {B}\left| {r_{n}}\right| ^{2} + \mathcal {C}h(\varphi ) \left| {r_{n}}\right| ^{2}\) and used the boundedness of h, and \(r_{n}(\tau _{*}) = 0.\) Then, substituting \(v = p_{n}\) in (6.1a) and \(v = Bq_{n}\) in (6.1b), integrating over \([s,\,\tau _{*}]\) for \(s \in (0,\,\tau _{*})\) and summing leads to
where we used that \(h(\varphi ) \le 1,\, \sigma \le 1,\, u \le 1\) a.e. in Q, and (5.3). Combining (6.5) and (6.6), and applying Young’s inequality and then Gronwall’s inequality, we see that
for some positive constant C depending only on \(\mathcal {C},\, \mathcal {P},\, \mathcal {A},\, \alpha ,\, C_{*},\, A,\, B,\) and T. This implies that \((p_{n},\, q_{n},\, r_{n})\) can be extended to the interval \([0,\,\tau _{*}],\) and thus \(t_{n} = 0\) for each \(n \in \mathbb {N}.\)
Second estimate viewing (6.1b) as the weak formulation of an elliptic problem for \(p_{n},\) and using that \(q_{n}\) is bounded uniformly in \(L^{2}(0,\,\tau _{*};\,L^{2}),\) we have by elliptic regularity that
for some positive constant C not depending on n.
Third estimate substituting \(v = -\partial _{t}r_{n}\) in (6.1c), integrating over \([s,\,\tau _{*}]\) for \(s \in (0,\,\tau _{*})\) leads to
Thus, by (6.7) we have that
for some positive constant C not depending on n. Furthermore, viewing (6.1c) as a weak formulation of an elliptic problem for \(r_{n}\) and elliptic regularity yields that
for some positive constant C not depending on n.
Fourth estimate integrating (6.1a) over \([0,\,\tau _{*}]\) and integrate by parts, by Hölder’s inequality we obtain that
which yields that \(\{\partial _{t}p_{n}\}_{n \in \mathbb {N}}\) is bounded uniformly in \(L^{2}(0,\,\tau _{*};\,(H^{2})^{*}).\)
It follows from the a priori estimates that we obtain a non-relabelled subsequence \((p_{n},\, q_{n},\, r_{n})\) such that
and by standard arguments the triplet \((p,\,q,\,r)\) satisfies (2.16) and is a solution to the adjoint system (2.15).
Uniqueness let \(p := p_{1} - p_{2},\, q := q_{1} - q_{2}\) and \(r := r_{1} - r_{2}\) denote the difference between two solutions to the adjoint system (2.15). Then, it holds that
for a.e. \(t \in (0,\,\tau _{*})\) and for all \(\eta \in H^{1}\) and \(\zeta \in H^{2},\) with \(p(\tau _{*}) = r(\tau _{*}) = 0.\) Substituting \(\zeta = p \in L^{2}(0,\,T;\,H^{2})\) in (6.8a) and integrate by parts, substituting \(\eta = Bq\) in (6.8b) and \(\eta = r\) in (6.8c), summing and then integrate over \([s,\,\tau _{*}]\) for \(s \in (0,\,\tau _{*})\) leads to
where we neglected the nonnegative term \(\mathcal {B}\left| {r}\right| ^{2} + \mathcal {C}h(\varphi ) \left| {r}\right| ^{2}\) and used that \(h(\varphi ) \le 1,\, \sigma \le 1,\, u \le 1\) a.e. in Q, and (5.3). Estimating the right-hand side with Young’s inequality and a Gronwall argument shows that
which yields that \(p = q = r = 0.\)
6.2 Simplification of the First Order Necessary Optimality Condition for the Control
Let \((u_{*},\,\tau _{*})\) denote the minimizer of (P) from Theorem 2.2, with corresponding state variables \((\varphi _{*},\,\mu _{*},\, \sigma _{*}) = \mathcal {S}(u_{*})\) and adjoint variables \((p,\,q,\,r)\) associated to \((\varphi _{*},\, \mu _{*},\, \sigma _{*}).\) For any \(u \in \mathcal {U}_{\mathrm {ad}},\) let \(w := u - u_{*} \in L^{2}(Q)\) and let \((\Phi ,\, \Xi ,\, \Sigma )\) denote the linearized state variables associated to w. Then, from (2.14), the optimal control \(u_{*}\) satisfies the following first order necessary optimality condition,
Substituting \(\zeta = \Phi \) in (2.16a), \(\eta = \Xi \) in (2.16b) and \(\eta = \Sigma \) in (2.16c), integrate over \([0,\,\tau _{*}]\) leads to
Meanwhile, substituting \(\zeta = p\) in (2.11a), \(\zeta = q\) in (2.11b) and \(\zeta = r\) in (2.11c), integrate over \([0,\,\tau _{*}]\) leads to
Using that \(r(\tau _{*}) = 0,\, p(\tau _{*}) = 0,\, \Sigma (0) = 0,\, \Phi (0) = 0,\, \partial _{t} \Phi \in L^{2}(0,\,T;\,(H^{1})^{*})\) and \(p \in L^{2}(0,\,T;\,H^{2}),\) we have that
Substituting (6.12b) into (6.10c) and comparing with (6.11c) leads to
Meanwhile, substituting (6.12a) and (6.11b) into (6.10a), and using (6.10b) and (6.13) leads to
Comparing the above equality with (6.11a) we obtain
References
Adams, R.A., Fournier, J.J.F.: Sobolev Spaces. Pure and Applied Mathematics, 2nd edn. Elsevier, New York (2003)
Arada, N., Raymond, J.P.: Time optimal problems with Dirichlet boundary controls. Discret. Contin. Dyn. Syst. 9, 1549–1570 (2003)
Bosia, S., Conti, M., Grasselli, M.: On the Cahn–Hilliard–Brinkman system. Commun. Math. Sci. 13(6), 1541–1567 (2015)
Coddington, E.A., Levinson, N.: Theory of Ordinary Differential Equations. International Series in Pure and Applied Mathematics. Tata McGraw-Hill, New York (1955)
Colli, P., Farshbaf-Shaker, M.H., Gilardi, G., Sprekels, J.: Optimal boundary control of a viscous Cahn–Hilliard system with dynamic boundary condition and double obstacle potentials. SIAM J. Control Optim. 53(4), 2696–2721 (2015)
Colli, P., Farshbaf-Shaker, M.H., Gilardi, G., Sprekels, J.: Second-order analysis of a boundary control problem for the viscous Cahn–Hilliard equation with dynamic boundary conditions. Ann. Acad. Rom. Sci. Math. Appl. 7, 41–66 (2015)
Colli, P., Gilardi, G., Hilhorst, D.: On a Cahn–Hilliard type phase field model related to tumor growth. Discret. Contin. Dyn. Syst. 35(6), 2423–2442 (2015)
Colli, P., Gilardi, G., Rocca, E., Sprekels, J.: Vanishing viscosities and error estimate for a Cahn–Hilliard type phase field system related to tumor growth. Nonlinear Anal. Real World Appl. 26, 93–108 (2015)
Colli, P., Gilardi, G., Rocca, E., Sprekels, J.: Asymptotic analyses and error estimates for a Cahn–Hilliard type phase field system modelling tumor growth. Discret. Contin. Dyn. Syst. S 10(1), 37–54 (2016)
Colli, P., Gilardi, G., Rocca, E., Sprekels, J.: Optimal distributed control of a diffuse interface model of tumor growth. Preprint. arXiv:1601.04567 (2016)
Colli, P., Gilardi, G., Sprekels, J.: A boundary control problem for the pure Cahn–Hilliard equation with dynamic boundary conditions. Adv. Nonlinear Anal. 4, 311–325 (2015)
Colli, P., Gilardi, G., Sprekels, J.: A boundary control problem for the viscous Cahn–Hilliard equation with dynamic boundary conditions. Appl. Math. Optim. 73(2), 195–225 (2016)
Cristini, V., Li, X., Lowengrub, J.S., Wise, S.M.: Nonlinear simulations of solid tumor growth using a mixture model: invasion and branching. J. Math. Biol. 58, 723–763 (2009)
Cristini, V., Lowengrub, J.: Multiscale Modeling of Cancer: An Integrated Experimental and Mathematical Modeling Approach. Cambridge University Press, Leiden (2010)
Dai, M., Feireisl, E., Rocca, E., Schimperna, G., Schonbek, M.: Analysis of a diffuse interface model for multispecies tumor growth. Nonlinearity 30, 1639–1658 (2017)
Friedman, A.: Partial Differential Equations. Holt, Rinehart and Winston, New York (1969)
Frigeri, S., Grasselli, M., Rocca, E.: On a diffuse interface model of tumor growth. Eur. J. Appl. Math. 26, 215–243 (2015)
Frigeri, S., Rocca, E., Sprekels, J.: Optimal distributed control of a nonlocal Cahn–Hilliard/Navier–Stokes system in two dimensions. SIAM J. Control Optim. 54(1), 221–250 (2016)
Garcke, H., Lam, K.F.: Global weak solutions and asymptotic limits of a Cahn–Hilliard–Darcy system modelling tumour growth. AIMS Math. 1(3), 318–360 (2016)
Garcke, H., Lam, K.F.: Analysis of a Cahn–Hilliard system with non zero Dirichlet conditions modelling tumour growth with chemotaxis. Discret. Contin. Dyn. Syst. (2017). arXiv:1604.00287
Garcke, H., Lam, K.F.: Well-posedness of a Cahn–Hilliard system modelling tumour growth with chemotaxis and active transport. Eur. J. Appl. Math. 28(2), 284–316 (2017)
Garcke, H., Lam, K.F., Sitka, E., Styles, V.: A Cahn–Hilliard–Darcy model for tumour growth with chemotaxis and active transport. Math. Models Methods Appl. Sci. 26(6), 1095–1148 (2016)
Gilbarg, D., Trudinger, N.S.: Elliptic Partial Differential Equations of Second Order, 2nd edn. Springer, Berlin (2001)
Grisvard, P.: Elliptic Problems on Nonsmooth Domains. Monographs and Studies in Mathematics, vol. 24. Pitman, Boston (1985)
Hartl, R.F., Sethi, S.P.: A note on the free terminal time transversality condition. Z. Oper. Res. 27, 203–208 (1983)
Hawkins-Daarud, A., Prudhomme, S., van der Zee, K.G., Oden, J.T.: Bayesian calibration, validation, and uncertainty quantification of diffuse interface models of tumor growth. J. Math. Biol. 67, 1457–1485 (2013)
Hawkins-Daarud, A., van der Zee, K.G., Oden, J.T.: Numerical simulation of a thermodynamically consistent four-species tumor growth model. Int. J. Numer. Methods Biomed. Eng. 28, 3–24 (2012)
Hintermüller, M., Keil, T., Wegner, D.: Optimal control of a semidiscrete Cahn–Hilliard–Navier–Stokes system with non-matched fluid densities. Preprint (2015). arXiv:1506.03591
Hintermüller, M., Wegner, D.: Distributed optimal control of the Cahn–Hilliard system including the case of a double-obstacle homogeneous free energy density. SIAM J. Control Optim. 50(1), 388–418 (2012)
Hintermüller, M., Wegner, D.: Distributed and Boundary Control Problems for the Semidiscrete Cahn–Hilliard/Navier–Stokes System with Nonsmooth Ginzburg–Landau Energies. Isaac Newton Institute Preprint Series No. NI14042-FRB (2014)
Hintermüller, M., Wegner, D.: Optimal control of a semidiscrete Cahn–Hilliard–Navier–Stokes system. SIAM J. Control Optim. 52(1), 747–772 (2014)
Hinze, M., Pinnau, R., Ulbrich, M., Ulbrich, S.: Optimization with PDE Constraints. Mathematical Modelling: Theory and Applications, vol. 23. Springer, Dordrecht (2009)
Jang, T., Kwon, H.D., Lee, J.: Free terminal time optimal control problem of an HIV model based on a conjugate gradient method. Bull. Math. Biol. 73, 2408–2429 (2011)
Jiang, J., Wu, H., Zheng, S.: Well-posedness and long-time behavior of a non-autonomous Cahn–Hilliard–Darcy system with mass source modeling tumor growth. J. Differ. Equ. 259(7), 3032–3077 (2015)
Lenhart, S., Workman, J.T.: Optimal Control Applied to Biological Models. Mathematical and Computational Biology. Chapman and Hall/CRC, London (2007)
Lowengrub, J.S., Titi, E., Zhao, K.: Analysis of a mixture model of tumor growth. Eur. J. Appl. Math. 24, 691–734 (2013)
Oden, J.T., Hawkins, A., Prudhomme, S.: General diffuse-interface theories and an approach to predictive tumor growth modeling. Math. Models Methods Appl. Sci. 58, 723–763 (2010)
Palanki, S., Kravaris, C., Wang, H.Y.: Optimal feedback control of batch reactors with a state inequality constraint and free terminal time. Chem. Eng. Sci. 49(1), 85–97 (1994)
Raymond, J.P., Zidani, H.: Pontryagin’s principle for time-optimal problems. J. Optim. Theory Appl. 101(2), 375–402 (1999)
Raymond, J.P., Zidani, H.: Time optimal problems with boundary controls. Differ. Integral Equ. 13(7–9), 1039–1072 (2000)
Rocca, E., Sprekels, J.: Optimal distributed control of a nonlocal convective Cahn–Hilliard equation by the velocity in three dimensions. SIAM J. Control Optim. 53(3), 1654–1680 (2015)
Roubíček, T.: Nonlinear Partial Differential Equations with Applications. International Series of Numerical Mathematics, vol. 153. Birkhäuser Verlag, Basel (2005)
Simon, J.: Compact sets in space \(L^{p}(0, T;B)\). Ann. Mat. Pura Appl. 146(1), 65–96 (1986)
Tröltzsch, F.: Optimal Control of Partial Differential Equations: Theory, Methods, and Applications. Graduate Studies in Mathematics, vol. 112. AMS, Providence (2010)
Wise, S.M., Lowengrub, J.S., Frieboes, H.B., Cristini, V.: Three-dimensional multispecies nonlinear tumor growth—I: model and numerical method. J. Theor. Biol. 253(3), 524–543 (2008)
Zhao, X., Duan, N.: Optimal control of the sixth-order convective Cahn–Hilliard equation. Bound. Value Probl. 2014, 206–222 (2014)
Zhao, X., Liu, C.: Optimal control problem for viscous Cahn–Hilliard equation. Nonlinear Anal. 74, 6348–6357 (2011)
Zhao, X., Liu, C.: Optimal control of the convective Cahn–Hilliard equation. Appl. Anal. 92(5), 1028–1045 (2013)
Zhao, X., Liu, C.: Optimal control for the convective Cahn–Hilliard equation in 2D case. Appl. Math. Optim. 70, 61–82 (2014)
Acknowledgements
The financial support of the FP7-IDEAS-ERC-StG #256872 (EntroPhase) and of the Project Fondazione Cariplo-Regione Lombardia MEGAsTAR “Matematica d’Eccellenza in biologia ed ingegneria come accelleratore di una nuona strateGia per l’ATtRattività dell’ateneo pavese” is gratefully acknowledged.
Author information
Authors and Affiliations
Corresponding author
Rights and permissions
About this article
Cite this article
Garcke, H., Lam, K.F. & Rocca, E. Optimal Control of Treatment Time in a Diffuse Interface Model of Tumor Growth. Appl Math Optim 78, 495–544 (2018). https://doi.org/10.1007/s00245-017-9414-4
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00245-017-9414-4
Keywords
- Tumor growth
- Cancer treatment
- Free terminal time
- Distributed optimal control
- Cahn–Hilliard equation
- Well-posedness