Abstract
In the present work we are interested in an optimal design problem for a linear elliptic state equation with a homogeneous boundary Dirichlet condition. The control variables correspond to the coefficients of the diffusion term and the open set where the equation is posed. From the application point of view these variables represent the layout of the materials composing the corresponding domain and its shape. We obtain a relaxed formulation of the problem, the optimality conditions, and we provide a numerical algorithm to solve it. Some numerical simulations are also carried out.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
A classical problem in the optimal design of materials consists in finding the optimal arrangement of two materials in the sense that it minimizes a certain cost functional [1, 26, 31]. In this way, assuming that we are dealing with two isotropic conductive materials given by their diffusion constants \(0<\alpha <\beta \), a model problem is given by
where \(\Omega \) is a bounded open set of \(\mathbb R^N\) and f a given source. The control variable corresponds to the measurable set \(\omega \subset \Omega \) where we place the material \(\alpha \) in the mixture. Clearly, we can consider other boundary conditions and more general cost functionals depending for example on the gradient of the solution. It is also usual to add a constraint in the measure of \(\omega \). It means that one of the materials is better than the other one, but it is also more expensive and thus we want to use a limited amount. It is known that this kind of problems has no solution in general [24]. Thus, it is usual to deal with a relaxed formulation, which can be found using the homogenization theory for elliptic linear equations with varying coefficients [1, 23, 25, 26, 28, 30, 31].
Another classical problem in optimal design is when we only have one conductive material, but the open set where the diffusion equation is posed can be freely chosen. Similarly to (1.1), the model problem is given by
This kind of problems has been studied for example in [4, 5]. Analogously to (1.1), the problem has no solution in general and thus, it is usual to work with a relaxed formulation which can be found using the homogenization theory for Dirichlet linear elliptic problems in varying domains [11, 14,15,16].
Our purpose in the present work is to consider the case, where analogously to (1.1), we look for the optimal distribution of two conductive materials and, similarly to (1.2), we can choose the set where the diffusion equations holds. Adding a restriction on the amounts of the materials used in the mixture, the problem reads as
with \(\kappa ^\alpha \), \(\kappa ^\beta \) two positive constants. Observe that assuming \(\kappa ^\alpha \) or \(\kappa ^\beta \) bigger than \(|\Omega |\) is equivalent to assume that there is no restriction on the amount of the corresponding material. As we commented above for problems (1.1) and (1.2), it is better to work with a relaxed formulation.
The paper is organized as follows:
In Sect. 2 we get the relaxed formulation for (1.3) mentioned above. It will be obtained as a consequence of the results proved in [17]. In this paper we carry out the homogenization of Dirichlet linear elliptic problems where both the coefficients and the open sets where the equations are posed vary (see also [8] for nonlinear problems). For completeness, we recall at the beginning of this section the results on homogenization theory that we will need. Our main result proves that a relaxed formulation of (1.3) is given by problem (2.18) below. In the corresponding state equation, the diffusion term \(-\textrm{div}\big ((\alpha \chi _{\omega ^\alpha }+\beta \chi _{\omega ^\beta })\nabla u\big )\) is replaced by \(-\textrm{div} (A\nabla u)\), where the matrix function A corresponds to a homogenized material obtained as a microscopic mixture of \(\alpha \) and \(\beta \) with proportions \(\theta \) and \(1-\theta \). Moreover, in (2.18), instead of a homogeneous Dirichlet problem in the open set \(\omega ^\alpha \cup \omega ^\beta \), we have a homogeneous Dirichlet problem in the whole open set \(\Omega \). This new equation contains a zero order term \(\mu u\), with \(\mu \) a nonnegative Borel measure vanishing on the sets of null capacity. It is the “strange term” in the Cioranescu–Murat terminology, [14]. Equation (1.3) is a particular case of (2.18) corresponding to \(\mu =\infty \chi _{\Omega {\setminus } (\omega ^\alpha \cup \omega ^\beta )}\).
In Sect. 3 we get a system of optimality conditions for the relaxed formulation (2.18) of problem (1.3). As a consequence we get Theorem 3.2 which proves that if F is concave in the second variable then, the optimal measure \(\mu \) can be chosen as \(\infty \chi _E\) for some quasi-closed set \(E\subset \Omega \).
Section 4 is devoted to the numerical resolution of problem (2.18). For this purpose, we introduce the approximation of this problem given by (4.1). Essentially the idea is to replace the nonnegative measure \(\mu \) in (2.18) which can take the value \(+\infty \) in arbitrary subsets of \(\Omega \) by a function, still denoted by \(\mu \), in \(L^\infty (\Omega ;[0,n])\), with n large (see also [6, 7]). In order to get the existence of solution, we also need to conveniently adapt the volume constraints in (2.18). Theorem 4.1 proves the convergence of the solutions of (4.1) to those of (2.18) when n goes to infinity. With respect to the numerical computation of (4.1) we propose an algorithm based on the gradient descent method adapted to our problem. The minimization of the corresponding derivative uses the Uzawa method to deal with the volume constraints combined with Proposition 4.2, which provides an explicit solution of this minimization problem once the corresponding Lagrange multipliers are known. We finish the section with some numerical experiments.
As a first example we consider the maximization of the potential energy, i.e. \(F(x,u)=-u\). It is a classical example in the case where either the coefficients or the set where the equation is posed vary [5, 20, 26]. Taking into account the invariance by rotations of the problem, we take a particular choice of the volume constraints \(\kappa ^\alpha \) and \(\kappa ^\beta \) in (1.3) in such a way that the solution is unique and explicitly known. We show that our algorithm provides a good approximation of such solution which, as expected, is better when n increases. We also consider the case where the amount of both materials is big enough to fulfill the whole of \(\Omega \). Then, the solution must agree with the classical one in which the equation is posed in a fixed domain (see [1, 10, 19, 22]). This is corroborated by the results that we find.
In the second example we consider a case that has been studied in [7] when only a material is available. In the results obtained in [7], the volume constraint is not saturated. The optimal domain obtained by our algorithm agrees with the one obtained in this paper. It also provides the distribution of the materials \(\alpha ,\beta \) in this domain.
To finish this introduction we refer to [1,2,3, 19, 20, 27, 29] for other algorithms used in the resolution of optimal design problems related to the one studied here.
2 Statement of the problem and relaxed formulation
We consider a bounded open set \(\Omega \subset \mathbb R^N\), a distribution \(f\in H^{-1}(\Omega )\) and a function \(F:\Omega \times \mathbb R\rightarrow \mathbb R\) satisfying
For four positive constants \(\alpha ,\beta ,\kappa ^\alpha ,\kappa ^\beta \), with \(\alpha <\beta \), we are interested in the optimal design problem
From the application point of view, it can be interpreted in the following way. The constants \(\alpha ,\beta \) represent the diffusion coefficients of two conductive materials (electric or thermical). Using these materials we consider a mixture of them consisting in taking the material \(\alpha \) in a certain measurable subset \(\omega ^\alpha \subset \Omega \), and the material \(\beta \) in another measurable set \(\omega ^\beta \subset \Omega \). Our problem is to choose these sets in such a way that the temperature or the electric potential u in \(\omega ^\alpha \cup \omega ^\beta \) minimizes the functional
The amount of the materials \(\alpha \) and \(\beta \) are limited by \(\kappa ^\alpha \) and \(\kappa ^\beta \) respectively. Observe that \(\kappa ^\alpha \ge |\Omega |\) or \(\kappa ^\beta \ge |\Omega |\) means that no restriction on the amount of the corresponding material is imposed.
Problem (2.4) has been widely studied in the case where \(\omega ^\alpha \cup \omega ^\beta =\Omega \) or the case \(\alpha =\beta \). In the first case [1, 9, 10, 12, 19, 20, 22, 24, 26] the problem is to distribute two materials in a given set in an optimal way. In the second one [2, 4,5,6,7] we only dispose of one material and the problem is to get an optimal shape \(\omega \) containing such material. The novelty in the present work is to combine these two problems.
From the above references, we know that problem (2.4) has no solution in general. Therefore, it is necessary to work with a relaxed formulation, which is usually obtained using the homogenization theory. We recall the main tools we need of such theory in order to get this formulation. We start with the following theorem
Theorem 2.1
Let \(\Omega \subset \mathbb R^N\) be a bounded open set, and \(A_n\in L^\infty (\Omega )^{N\times N}\) a sequence of symmetric matrix functions such that there exist \(\alpha ,\beta >0\) satisfying
Then, for a subsequence of n, still denoted by n, there exists a symmetric matrix function \(A\in L^\infty (\Omega )^{N\times N}\), which also satisfies (2.5), such that for every \(f\in H^{-1}(\Omega )\), the solution \(u_n\) of
satisfies
with u the solution of
We say that \(A_n\) H-converges to A and we write \(A_n{\mathop {\rightharpoonup }\limits ^{H}}A\).
Remark 2.2
More generally, in the definition of H-convergence it is not necessary to assume \(A_n\) symmetric. The classical result for the H-convergence proved by Murat and Tartar [25, 31] also provides a corrector for \(\nabla u_n\), i.e. an approximation in the strong topology of \(L^2(\Omega )^N\) which we will not need here. The version given in Theorem 2.1 was previously obtained by S. Spagnolo in [28] who used the denomination of G-convergence instead of H-convergence.
Theorem 2.1 describes the asymptotic behavior of a sequence of elliptic equations where the coefficients vary. More generally, we are interested in the case where the open set \(\Omega \) also varies. This is given by Theorem 2.7 below due to Dal Maso and Murat [17]. We also refer to [14, 15] and [16] for the case where \(\Omega \) varies but \(A_n\) is independent of n, and to [8, 11] to some extensions for nonlinear problems. In order to state Theorem 2.7 we need to recall some results about the capacity.
Definition 2.3
For a bounded open set \(\Omega \subset \mathbb R^N\) and \(E\subset \Omega \), we define the capacity of E in \(\Omega \) as
Some of the main properties of the capacity are given in the following remark (see e.g. [18])
Remark 2.4
As a function of E, the capacity is an outer measure over the class of subsets of \(\Omega \). Although the definition of capacity depends on \(\Omega \), for every bounded open sets \(\Omega _1,\Omega _2\subset \mathbb R^N\) and every \(E\subset \Omega _1\cap \Omega _2\), we have
and therefore the sets of null capacity do not depend on \(\Omega \).
Every set \(E\subset \mathbb R^N\) with null capacity has Hausdorff dimension at most \(N-2\) and conversely, every set E with finite \((N-2)\)-Hausdorff measure has null capacity.
We will say that a property holds quasi-everywhere (q.e.) if it holds up to a set of null capacity.
Every function \(u\in H^1(\Omega )\) has a representative which is defined q.e. in \(\Omega \).
Definition 2.5
A set \(U\subset \Omega \) is said to be quasi-open if for every \(\varepsilon >0\), there exists \(G\subset \Omega \) open such that \(\textrm{Cap}(U\Delta G,\Omega )<\varepsilon \). The complementary in \(\Omega \) of a quasi-open set U is said to be quasi-closed.
We define \({{\mathcal {M}}}_0(\Omega )\) as the set of non-negative Borel measures which vanish on the null-capacity sets of \(\Omega \) and satisfy
Remark 2.6
The elements \(\mu \) of \({{\mathcal {M}}}_0(\Omega )\) are not necessarily Radon measures, i.e. it can exist a compact set \(K\subset \Omega \) such that \(\mu (K)=\infty .\) For every measure \(\mu \in {{\mathcal {M}}}_0(\Omega )\), there exists a unique quasi-closed set \({{\mathcal {C}}}_\mu \) such that
where \(\infty _{{{\mathcal {C}}}_\mu }\) is the measure in \({{\mathcal {M}}}_0(\Omega )\) defined as
Recalling that by Remark 2.4, the functions in \(H^1(\Omega )\) are well defined for the elements of \({{\mathcal {M}}}_0(\Omega )\), we get that for every open set \(\omega \subset \Omega \), every \(A\in L^\infty (\omega )^{N\times N}\) uniformly elliptic and every \(f\in H^{-1}(\Omega )\), the problem
is equivalent to
with \(\mu =\infty _{\Omega {\setminus }\omega }\). For a general measure \(\mu \in {{\mathcal {M}}}_0(\Omega )\), we usually write (2.6) as
Remark that the equation is not necessarily satisfied in the distributions sense because the set of test functions in (2.6) does not contain in general the set \(C^\infty _c(\Omega )\).
The following result is due to G. Dal Maso and F. Murat, [17]. Similarly to Theorem 2.1 this paper also contains a corrector result but we will not need it in the following. Moreover it holds for matrices \(A_n\) not necessarily symmetric.
Theorem 2.7
Assume \(\Omega \subset \mathbb R^N\) a bounded open set, \(A_n\in L^\infty (\Omega )^{N\times N}\) symmetric, which satisfies (2.5) and \(\mu _n\in {{\mathcal {M}}}_0(\Omega )\). Then, for a subsequence of n still denoted by n, there exits a symmetric matrix \(A\in L^\infty (\Omega )^{N\times N}\) and a measure \(\mu \in {{\mathcal {M}}}_0(\Omega )\) such that \(A_n\) H-converges to A and for every \(f\in H^{-1}(\Omega )\) the sequence of solutions of
converges weakly in \(H^1_0(\Omega )\) to the unique solution of (2.7). We will write
Remark 2.8
Taking in particular \(\mu _n=\infty _{\Omega _n}\) with \(\Omega _n\) a sequence of open sets contained in \(\Omega \), Theorem 2.7 provides the limit equation of a sequence of elliptic problems with homogeneous Dirichlet conditions with varying coefficients and domains.
In the present work we are interested in the case of a two-phase material. In this case A is obtained as the mixture of two isotropic materials \(A=\alpha \chi _\omega +\beta \chi _{\Omega {\setminus }\omega }\) with \(\omega \) a measurable subset of \(\Omega \) and \(0<\alpha <\beta \), the diffusion constants of each material. It is well known [1, 26, 28, 31] that this set is not closed for the H-convergence. Thus, it is interesting to replace it by its H-closure, which provides the set of materials obtained by mixing \(\alpha \) and \(\beta \) microscopically. To describe this set we introduce
Definition 2.9
For \(p\in [0,1]\), we denote by \(m^-(p)\) and \(m^+(p)\) the harmonic and arithmetic mean values of \(\alpha \) and \(\beta \) with proportions p and \(1-p\) respectively, i.e.
We also define K(p) as the set of symmetric matrices \(M\in \mathbb R^{N\times N}\) such that their eigenvalues \(\lambda _1\le \cdots \le \lambda _N\) satisfy
The following result has been proved in [23] (for \(N=2\)) and [30].
Theorem 2.10
Assume \(\Omega \subset \mathbb R^N\) a bounded open set and \(A\in L^\infty (\Omega )^{N\times N}\). Then, there exist \(\theta \in L^\infty (\Omega ;[0,1])\) and a sequence of measurable sets \(\omega _n\subset \Omega \) such that
if and only if \(A\in K(\theta )\) a.e. in \(\Omega \).
Remark 2.11
In the previous result, the function \(\theta \) represents the proportion of the material \(\alpha \) that we are using in the mixture. For our purpose, it would be sufficient to work with the set
where B(x, r) denotes the ball in \(\mathbb R^N\) of center x and radius \(r>0\).
Having in mind problem (2.4), the following adaptation of Theorem 2.7 is more convenient.
Theorem 2.12
Assume \(\Omega \subset \mathbb R^N\) a bounded open set, \(\mu _n\in \mathcal{M}_0(\Omega )\), \(\theta _n^\alpha ,\theta _n^\beta \in L^\infty (\Omega ;[0,1])\), and \(A_n\in L^\infty (\Omega {\setminus } \mathcal{C}_{\mu _n})^{N\times N}\) such that
Then, there exist a subsequence of n, still denoted by n, \(\mu \in \mathcal{M}_0(\Omega )\), \(\theta ^\alpha ,\theta ^\beta \in L^\infty (\Omega ,[0,1])\), and \(A\in L^\infty (\Omega {\setminus } \mathcal{C}_\mu )^{N\times N}\), satisfying
such that
and such that for every \(f\in H^{-1}(\Omega )\), the sequence of solutions \(u_n\) of (2.8) converges weakly in \(H^1_0(\Omega )\) to the solution u of (2.7).
Proof
Taking a subsequence of n, we can assume that there exist \(\theta ^\alpha \), \(\theta ^\beta \) satisfying (2.14), such that \(\theta ^\alpha +\theta ^\beta \le 1\) a.e. in \(\Omega \).
Let \(w_n\) be the solution of
From [16], there exist a subsequence of n, still denoted by n, \(w\in H^1_0(\Omega )\cap L^\infty (\Omega )\) and \(\hat{\mu }\in \mathcal{M}_0(\Omega )\), such that \(w>0\) q.e. in \(\Omega {\setminus }\mathcal{C}_{\hat{\mu }}\), \(w=0\) q.e. in \( \mathcal{C}_{\hat{\mu }}\), and \(w_n\) converges weakly to w in \(H^1_0(\Omega )\) and weakly-\(*\) in \(L^\infty (\Omega )\). By the Rellich–Kondrachov theorem we can pass to the limit in
in order to obtain
and then that \(\theta ^\alpha +\theta ^\beta =1\) a.e. in \(\Omega {\setminus } \mathcal{C}_{\hat{\mu }}\).
Now, for \(\hat{\theta }_n=1-\theta _n^\beta \) and \({\hat{A}}_n\in L^\infty (\Omega )^{N\times N}\), with \({\hat{A}}_n\in K(\hat{\theta }_n)\) a.e. in \(\Omega \), defined as
with \(e_N\) the last vector of the canonical basis in \(\mathbb R^N\), we apply Theorem 2.7 combined to
to deduce that, up to a subsequence, there exist \(A,\mu \) such that for every \(f\in H^{-1}(\Omega )\) the solution \(u_n\) of (2.8) converges weakly in \(H^1_0(\Omega )\) to the solution u of (2.7). By [11], it is known that there exist \(c,C>0\) such that
and thus \(\mathcal{C}_\mu =\mathcal{C}_{\hat{\mu }}\). Moreover, by Theorem 2.10, we also have that \(A\in K(\hat{\theta })\) a.e. in \(\Omega \), with \(\hat{\theta }\) the limit of \(\hat{\theta }_n\) in the weak-\(*\) topology of \(L^\infty (\Omega )\). On the other hand, taking \(\varphi \in L^1(\Omega )\), we can pass to the limit in
to get
This proves that \(\hat{\theta }=\theta ^\alpha \) a.e. in \(\Omega {\setminus } \mathcal{C}_\mu \), and \(A\in K(\theta ^\alpha )\) a.e. in \(\Omega {\setminus }\mathcal{C}_\mu \). \(\square \)
Taking into account Theorem 2.12 we can now prove
Theorem 2.13
Under the assumptions at the beginning of this section, a relaxed formulation of (2.4) is given by
Proof
Taking into account Theorem 2.12 and Rellich–Kondrachov’s theorem, the direct method of the Calculus of Variations immediately proves the existence of a solution for (2.15). Now, we observe that the controls in (2.4) belong to the set of controls in (2.15) and that the functional in (2.15) only depends on u and it is sequentially continuous for the weak topology in \(H^1_0(\Omega )\). By means of these remarks, in order to finish the proof of the theorem, it is enough to show that for every control \((\theta ^\alpha ,\theta ^\beta ,A,\mu )\) satisfying the constraints in (2.15) there exists two sequences of measurable sets \(\omega ^\alpha _n\), \(\omega ^\beta _n\) satisfying the constraints in (2.4) such that the sequence of solutions \(u_n\) of
converges weakly in \(H^1_0(\Omega )\) to the solution u of
and
First, we consider the case where \(\mu \) belongs to \(L^\infty (\Omega )\) and A is of the form \(A=(\alpha \chi _{\omega ^\alpha }+\beta \chi _{\omega ^\beta })I\) with \(|\omega ^\alpha |\le \kappa ^\alpha \), \(|\omega ^\beta |\le \kappa ^{\beta }\). Then, using the classical construction by D. Cioranescu and F. Murat [14] we can build a sequence of holes \(Q_n\subset \Omega \) such that \(\omega _n^\alpha =\omega ^\alpha {\setminus } Q_n\), \(\omega _n^\beta =\omega ^\beta {\setminus } Q_n\) meets the required conditions.
Second, we consider the case where \(\mu \) is in \(L^\infty (\Omega )\) and A is an arbitrary measurable function with values in \(K(\theta ^\alpha )\). Then, Theorem 2.10 allows us to approximate the solution of (2.17) by the solutions of
where \(\tilde{\omega }^\alpha _n\), \(\tilde{\omega }_n^\beta \) satisfy the restrictions in (2.4). This allows us to use the previous case to finish the proof.
Third, we assume that \(\mu \) is in \(H^{-1}(\Omega )\). In this case, the density of \(L^\infty (\Omega )\) in \(H^{-1}(\Omega )\) and the second case give the result.
Fourth, we consider the case where \(\theta ^\alpha +\theta ^\beta =1\) a.e. in \(\Omega \). Then, we know by [16] that there exist a nonnegative bounded measure \(\nu \in H^{-1}(\Omega )\) and a non-negative \(\mu \)-measurable function g finite \(\mu \)-a.e. in \(\Omega \) such that \(\mu =g\nu +\infty _{\mathcal{C}_\mu }\). Replacing \(\mu \) by \(T_n(g)\nu +n\chi _{\mathcal{C}_\mu }\) where \(T_n\) denotes the usual truncation at the height \(n\in \mathbb N\), taking the limit when n tends to infinity, and using the third case we get the result.
For the general case, \(\theta ^\alpha +\theta ^\beta \le 1\) a.e. in \(\Omega \), we use that the fourth case provides two sequences \( \omega _n^\alpha \) and \( \omega _n^\beta \) satisfying the required conditions for \(\tilde{\theta }^\beta \) replaced by \(1-\theta ^\alpha \). On the other hand, it is well known that there exits a sequence of measurable sets \(\tilde{\omega }_n^\beta \subset \mathcal{C}_\mu \) such that
Replacing \(\omega _n^\beta \) by \((\omega _n^\beta {\setminus } \mathcal{C}_\mu )\cup \tilde{\omega }_n^\beta \) and taking into account that this does not change the solution of (2.16) we finish the proof. \(\square \)
Remark 2.14
Since in (2.15) we only need to know the values of the matrix function A in the direction of \(\nabla u\), we can replace the condition \(A\in K(\theta ^\alpha )\) a.e. in \(\Omega {\setminus } \mathcal{C}_\mu \) by
with \(\textrm{Sp}(A)\) the spectrum of A. Moreover, replacing \(\theta ^\alpha \) and \(\theta ^\beta \) in (2.15) by \(\theta ^\alpha \chi _{\Omega {\setminus } \mathcal{C}_\mu }\) and \(\theta ^\beta \chi _{\Omega {\setminus } \mathcal{C}_\mu }\), which does not change the solution of the state equation, we can assume that \(\theta ^\alpha \) and \(\theta ^\beta \) vanish on \(\mathcal{C}_\mu \), and then that they are only defined in \(\Omega {\setminus } \mathcal{C}_\mu \). In the same way, we can also assume that A is only defined in \(\Omega {\setminus }\mathcal{C}_\mu \).
With these observations, problem (2.15) can be replaced by
3 Optimality conditions
In the present section, assuming further regularity on F, we provide a system of optimality conditions for the solutions of (2.18). Namely, we assume that F satisfies (2.1) and
Observe that (3.1) and (3.2) imply (2.2) and (2.3). We have
Theorem 3.1
Assume \(\Omega \subset \mathbb R^N\) a bounded open set, \(F:\Omega \times \mathbb R\rightarrow \mathbb R\) satisfying (2.1), (3.1) and (3.2), \(\alpha ,\beta ,\kappa ^\alpha ,\kappa ^\beta >0\), \(\alpha <\beta \), and \((\hat{\theta }^\alpha ,{\hat{A}},\hat{\mu })\) a solution of (2.18). Then, if \({\hat{u}}\) is the corresponding state function, and \({\hat{q}}\) is the adjoint state, solution of
we have
Moreover there exists \(\tau \in \mathbb R\) such that defining
we have a.e. in \(\Omega {\setminus } \mathcal{C}_{\hat{\mu }}\)
with
Proof
We take \(z_{\hat{\mu }}\in L^\infty _{\hat{\mu }}(\Omega ),\) \(z_{\hat{\mu }}\ge 0\) \(\hat{\mu }\)-a.e. in \(\Omega \), \(z\in L^\infty (\Omega )\), \(z\ge 0\) a.e. in \(\Omega \), \(\theta ^\alpha \in L^\infty (\Omega {\setminus } \mathcal{C}_{\hat{\mu }};[0,1])\) and \(A\in L^\infty (\Omega {\setminus } \mathcal{C}_{\hat{\mu }})^{N\times N}\) symmetric such that
Then, for \(\varepsilon \in [0,1]\) we define
which satisfy the constraints in (2.18). Thus, since \((\hat{\theta }^\alpha ,{\hat{A}},\hat{\mu })\) is a solution of (2.18), we have
with \(u_\varepsilon \) the solution of
Using that \(H^1_0(\Omega )\cap L^2_{\mu _\varepsilon }(\Omega )=H^1_0(\Omega )\cap L^2_{\hat{\mu }}(\Omega ),\) we can derive in (3.11) to conclude
with \(u'\) the solution of
Thus, taking into account (3.10) and (3.2), we get
which using the definition (3.3) of \({\hat{q}}\) and (3.13) proves
for every \(z_{\hat{\mu }}\in L^\infty _{\hat{\mu }}(\Omega ),\) \(z\in L^\infty (\Omega )\), \(z_{\hat{\mu }},z\ge 0\), \(\theta ^\alpha \in L^\infty (\Omega {\setminus }\mathcal{C}_{\hat{\mu }};[0,1])\) and \(A\in L^\infty (\Omega {\setminus }\mathcal{C}_{\hat{\mu }})^{N\times N}\) symmetric, satisfying (3.9).
Let us now take in (3.15), \(\theta ^\alpha =\hat{\theta }^\alpha \), \(A={\hat{A}}\).
Taking \(z_{\hat{\mu }}=1\) we deduce
On the other hand, taking \(z=z_{\hat{\mu }}=0\), we get
which by (3.16) proves \({\hat{u}}{\hat{q}}=0\) \(\hat{\mu }\)-a.e. in \(\Omega \).
Now, we take in (3.15) \(\theta ^\alpha =\hat{\theta }^\alpha \), \(z_{\hat{\mu }}=1\), \(z=0\), and \(A={\hat{A}}\) a.e. in \((\Omega {\setminus } (\omega \cup \mathcal{C}_{\hat{\mu }}))\) with \(\omega \subset \Omega {\setminus } \mathcal{C}_{\hat{\mu }}\) measurable. This gives
for every \(\omega \subset \Omega {\setminus } \mathcal{C}_{\hat{\mu }}\) measurable, and every \(A\in L^\infty (\omega )^{N\times N}\) with \(\textrm{Sp}(A)\subset [m^-(\hat{\theta }^\alpha ),m^+(\hat{\theta }^\beta )]\), a.e. in \(\omega .\) This proves
and then that (3.5) holds. Finally, taking into account (3.5), we take in (3.15), \(z_{\hat{\mu }}=1\), \(z=0\), and A such that
we deduce that \(\hat{\theta }^\alpha \) is a solution of the convex minimization problem
with \(E^+\) and \(E^-\) defined by (3.6). By Kuhn-Tucker’s theorem this proves the existence of \(\tau ^\alpha ,\tau ^\beta \ge 0\), with
such that \(\hat{\theta }^\alpha \) is also a solution of
Thus, denoting \(\tau =(\tau ^\alpha -\tau ^\beta )/(\beta -\alpha )\), we have that for a.e. \(x\in \Omega {\setminus }\mathcal{C}_{\hat{\mu }}\), \(\hat{\theta }^\alpha (x)\) is a solution of
Using here the definitions (2.10) of \(m^-\) and \(m^+\), we easily conclude that \(\hat{\theta }^\alpha \) satisfies (3.7).
\(\square \)
As an application of the optimality conditions obtained in Theorem 3.1, let us prove that assuming F(x, s) concave in s, we can choose the optimal measure \(\hat{\mu }\) in (2.18) as \(\hat{\mu }=\infty _C,\) for some quasi-closed set \(C\subset \Omega .\) In particular, this applies for F(x, s) linear in s.
Theorem 3.2
Under the assumptions of Theorem 3.1, if \(F=F(x,s)\) is concave in s, then there exists a solution \((\hat{\theta }^\alpha ,{\hat{A}}, \hat{\mu })\) of (2.18), such that \(\hat{\mu }=\infty _C\) for some quasi-closed set \(C\subset \Omega .\)
Proof
We take \((\tilde{\theta }^\alpha ,{\tilde{A}},\tilde{\mu })\) a solution of (2.18), and \({\tilde{u}}\) the solution of the corresponding state equation. Replacing \(\tilde{\mu }\) by \(\tilde{\mu }+\infty _{\{{\tilde{u}}=0\}}\), which does not change the solution \({\tilde{u}}\) of the state equation, we can always assume
Now, we define \({\tilde{q}}\), the adjoint state, as the solution of
We observe that (3.4), \({\tilde{q}}\in L^2_{\tilde{\mu }}(\Omega )\) and (3.19) imply
We define \(\hat{\mu }=\infty _{\{{\tilde{q}}=0\}}\) which thanks to (3.21), satisfies
We also define \(\hat{\theta }^\alpha \), \({\hat{A}}\) as the restrictions of \(\tilde{\theta }^\alpha \) and \({\tilde{A}}\) to \(\Omega {\setminus } \{{\tilde{q}}=0\}\), and \({\hat{u}}\) as the solution of
Since F is concave, we have
Thanks to (3.22) we can take \({\tilde{u}}-{\hat{u}}\) as test function in (3.20). Using also (3.21) we get
On the other hand, we can also take \({\tilde{q}}\) as test function in the equations for \({\tilde{u}}\) and \({\hat{u}}\), which, using the definition of \({\hat{A}}\), provides
Therefore, integrating in \(\Omega \) in (3.23) we conclude
which proves that \((\hat{\theta }^\alpha ,{\hat{A}}, \hat{\mu })\) is a solution of (2.18). \(\square \)
4 A discrete approximation and some numerical simulations
In this section we are interested in the numerical resolution of (2.18). The idea we use is to replace the control measure \(\mu \), which can take the value \(+\infty \) in arbitrary compact subsets of \(\Omega \), by a non-negative function still denoted by \(\mu \) with values in an interval [0, n], n large. This idea has also been used in [6] and [7].
In this sense the set \(\mathcal{C}_\mu \) would be replaced by \(\{\mu =n\}\) and the constraint
would be replaced by
However these conditions are not convex and then, it is not clear that the corresponding optimization problem has a solution. Taking into account that the convex hull of the function
is given by
we propose the approximation
We have
Theorem 4.1
Problem (4.1) has at least one solution for every \(n\in \mathbb N\). Moreover, for every sequence of solutions \((\theta _n,A_n,\mu _n)\) of (4.1), there exist a subsequence, still denoted by n, and a solution \((\hat{\theta }^\alpha ,\hat{\theta }^\beta ,{\hat{A}},\hat{\mu })\) of (2.15) such that denoting by \(u_n\) and \({\hat{u}}\) the solutions of the respective state equations, we have
Moreover
Proof
The existence of solution for (4.1) just follows from Theorem 2.1 combined with the convexity of the functions \(\theta \mapsto (\theta -{\mu \over n})^+\) and \(\theta \mapsto (1-\theta -{\mu \over n})^+\).
Assume \((\theta _n,A_n,\mu _n)\) a solution of (4.1) and \(u_n\) the solution of the corresponding state equation. From Theorem 2.12, with \(\theta _n^\alpha =\theta _n\), \(\theta _n^\beta =1-\theta _n\), there exist a subsequence of n, still denoted by n, \(\hat{\theta }\in L^\infty (\Omega ; [0,1])\), \(\hat{\mu }\in \mathcal{M}_0(\Omega )\) and \({\hat{A}}\in K(\hat{\theta })\) a.e. in \(\Omega {\setminus } \mathcal{C}_{\hat{\mu }}, \) such that
with \({\hat{u}}\) the solution of
On the other hand, we can also assume the existence of \(\hat{\theta }^\alpha ,\hat{\theta }^\beta \in L^\infty (\Omega ;[0,1])\) such that
Clearly
Since \((\theta _n-{\mu _n\over n})^+\le \theta _n\), \((1-\theta _n-{\mu _n\over n})^+\le 1-\theta _n\), we also get
Now, we observe that thanks to Rellick–Kondrachov’s compactness theorem we have
On the other hand, using that the function \(s\rightarrow s^+\) has constant Lipschitz equals one, we have
Combined with the energy equality
which implies that \(|u_n|^2\mu _n\) is bounded in \(L^1(\Omega )\), inequality (4.5) proves
Therefore, (4.4) provides
A similar reasoning also proves
Using
we also get
Thus, \((\hat{\theta }^\alpha ,\hat{\theta }^\beta ,{\hat{A}},\hat{\mu })\) satisfies the constraints in (2.15).
By the Rellich–Kondrachov’s compactness theorem, we also have
with \(\mathcal{I}\) the minimum in (2.15).
In order to finish the proof of Theorem 4.1 we consider \((\tilde{\theta }^\alpha ,{\tilde{A}},\tilde{\mu })\) giving the minimum in (2.18), and \({\tilde{u}} \) the solution of the corresponding state equation, then (see the proof of Theorem 2.13), there exists \(\tilde{\mu }_n\in L^\infty (\Omega ,[0,n])\), such that the solution \({\tilde{u}}_n\) of
converges strongly in \(H^1_0(\Omega )\) to the solution \({\tilde{u}}\) of
Since \(\tilde{\mu }=\infty _{\mathcal{C}_{\tilde{\mu }}}\) in \(\mathcal{C}_{\tilde{\mu }}\), we can always assume \(\tilde{\mu }_n=n\) in \(\mathcal{C}_{\tilde{\mu }}.\) Extending \(\tilde{\theta }^\alpha \) by zero (for example) in \(\mathcal{C}_{\tilde{\mu }}\), we have
Therefore \((\tilde{\theta }^\alpha ,{\tilde{A}},\tilde{\mu })\) satisfies the constraints in (4.1). Thus, since the sequence \((\theta _n,A_n,\mu _n)\) at the beginning of the proof was a solution of (4.1), we have
which combined with (4.6) proves that \((\hat{\theta }^\alpha ,\hat{\theta }^\beta ,{\hat{A}},\hat{\mu })\) is a solution to (2.15). \(\square \)
By Theorem 4.1, in order to solve numerically (2.18), we must provide an algorithm to solve (4.1). Let us look for a descent gradient method. Namely, we assume that F satisfies conditions (3.1) and (3.2), and that we have found an approximation \((\theta _k,A_k,\mu _k)\) of a solution of (4.1), which satisfies the constraints in this problem. Taking into account the convexity of the set of controls, we look for another approximation
which decreases the cost functional. Here \(\hat{\theta }\), \({\hat{A}}\), \(\hat{\mu }\) must also satisfy the constraints in (4.1), while the step \(\varepsilon _k\) belongs to (0, 1]. Denoting by \(u_k\) the solution of the state problem
and assuming \(\varepsilon _k\) small, we have the Taylor approximation
with \(u'_k\) the solution of
Defining the adjoint state \(p_k\) by
one has
Therefore, taking into account (4.9), the idea is to choose \((\hat{\theta },{\hat{A}},\hat{\mu })\) as a solution of
Let us show that this problem can be solved almost explicitly.
Assuming that \((\hat{\mu },\hat{\theta })\) is known, and taking the maximum in A we have
Replacing this expression in (4.13) we get that \((\hat{\theta },\hat{\mu })\) is a solution of
with
Applying Kuhn-Tucker’s theorem to this convex problem, we deduce that solving (4.15) is equivalent to solving the max-min problem
Let us solve this problem using Uzawa’s method (see e.g. [13] Chapter 9.4). It consists in applying the gradient method to the maximization problem. Namely, starting from \((\lambda ^0_{k,1},\lambda ^0_{k,2})\), we construct a sequence \((\lambda ^j_{k,1},\lambda ^j_{k,2})\) as
with \(\rho >0\) small, and \((\theta ^j_k,\mu ^j_k)\) solution of
Then, an approximated solution of (4.15) is given by \((\mu ^j_k,\theta ^j_k)\) with j large enough.
For every j, k problem (4.19) can be solved as follows:
We assume that \(\theta ^j_k\) is known. Then for a.e. \(x\in \Omega \), \(\mu ^j_k(x)\) solves
Since the function to minimize is piecewise linear, we can find a particular solution taking its values in the set of vertices of the constraint set, i.e. \(\{0, n\theta ^j_k, n(1-\theta ^j_k), n\}\). Replacing in (4.19) \(\mu \) by this solution we get an optimization problem for \(\theta ^j_k\). The resolution of this problem is a simple but tedious calculation and thus, we just give the result in the following proposition without proof.
Proposition 4.2
We denote
then, for \(\Psi \) defined by (3.8), we have that a particular solution of (4.19) is given by
-
1.
Assume \(D_k\le 0\). Then
$$\begin{aligned} \left\{ \begin{array}{ll}\displaystyle \hbox {If }E^-_k\le {\alpha \over \beta }E^+_k& \displaystyle \Rightarrow \theta ^j_k=0,\mu ^j_k=n \\ \displaystyle \hbox {If }{\alpha \over \beta }E^+_k\le E^-_k\le {\beta \over \alpha }E^+_k& \displaystyle \Rightarrow \theta ^j_k=\Psi \Big ({\alpha \beta \,E^-_k\over E^+_k}\Big ), \mu ^j_k=n\\ \displaystyle \hbox {If }{\beta \over \alpha }E^+_k\le E^-_k& \Rightarrow \theta ^j_k=1,\mu ^j_k=n.\end{array}\right. \end{aligned}$$ -
2.
Assume \(0\le D_k\le \min \{\tilde{\lambda }_1,\tilde{\lambda }_2\}\). Then
$$\begin{aligned} \left\{ \begin{array}{ll}\displaystyle \hbox {If }D_k\le E^+_k-{\beta \over \alpha }E^-_k& \displaystyle \Rightarrow \theta ^j_k=0,\mu ^j_k=n\\ \displaystyle \hbox {If }E^+_k-{\beta \over \alpha }E^-_k\le D_k\le E^+_k-\gamma E^-_k& \Rightarrow \displaystyle \theta ^j_k= \Psi \Big ({\alpha \beta E^-_k\over E^+_k-D_k}\Big ),\mu ^j_k=n(1-\theta ^j_k)\\ \displaystyle \hbox {If }\Big |E^+_k-\gamma E^-_k\Big |\le D_k& \displaystyle \Rightarrow \theta ^j_k={1\over 2},\mu ^j_k={n\over 2}\\ \displaystyle \hbox {If } {\alpha \over \beta }E^-_k-E^+_k\le D_k\le \gamma E^-_k-E^+_k& \displaystyle \Rightarrow \theta ^j_k=\Psi \Big ({\alpha \beta E^-_k\over E^+_k+D_k}\Big ),\mu ^j_k=n\theta ^j_k\\ \displaystyle \hbox {If }D_k\le {\alpha \over \beta }E^-_k-E^+_k& \displaystyle \Rightarrow \theta ^j_k=1,\mu ^j_k=n\end{array}\right. \end{aligned}$$ -
3.
Assume \(\tilde{\lambda }_1\le D_k\le \tilde{\lambda }_2\). Then
$$\begin{aligned} \left\{ \begin{array}{ll} \displaystyle \hbox {If }D_k\le E^+_k-{\beta \over \alpha }E^-_k& \displaystyle \Rightarrow \theta ^j_k=0,\mu ^j_k=n\\ \displaystyle \hbox {If }E^+_k-{\beta \over \alpha }E^-_k\le D_k\le E^+_k-\gamma E^-_k& \displaystyle \Rightarrow \left\{ \begin{array}{l}\displaystyle \theta ^j_k=\Psi \Big ({\alpha \beta E^-_k\over E^+_k-D_k}\Big )\\ \displaystyle \mu ^j_k=n(1-\theta ^j_k)\end{array}\right. \\ \displaystyle \hbox {If } E^+_k-\gamma E^-_k\le D_k\le E^+_k-\gamma E^-_k+2\tilde{\lambda }_1 & \displaystyle \Rightarrow \theta ^j_k={1\over 2},\mu ^j_k={n\over 2}\\ \displaystyle \hbox {If } E^+_k-\gamma E^-_k+2\tilde{\lambda }_1 \le D_k\le E^+_k-{\alpha \over \beta }E^-_k+2\tilde{\lambda }_1& \displaystyle \Rightarrow \left\{ \begin{array}{l}\displaystyle \theta ^j_k=\Psi \Big ({\alpha \beta E^-_k\over E^+_k-D_k+2\tilde{\lambda }_1}\Big )\\ \displaystyle \mu ^j_k=n(1-\theta ^j_k)\end{array}\right. \\ \displaystyle \hbox {If }E^+_k-{\alpha \over \beta }E^-_k+2\tilde{\lambda }_1\le D_k & \displaystyle \Rightarrow \theta ^j_k=1,\mu ^j_k=0 \end{array}\right. \end{aligned}$$ -
4.
Assume \(\tilde{\lambda }_2\le D_k\le \tilde{\lambda }_1\). Then
$$\begin{aligned} \left\{ \begin{array}{ll}\displaystyle \hbox {If }D_k\ge {\beta \over \alpha }E^-_k-E^+_k+2\tilde{\lambda }_2& \displaystyle \Rightarrow \theta ^j_k=0,\mu ^j_k=0\\ \displaystyle \hbox {If }{\beta \over \alpha }E^-_k-E^+_k+2\tilde{\lambda }_2\ge D_k\ge \gamma E^-_k-E^+_k+2\tilde{\lambda }_2& \displaystyle \Rightarrow \left\{ \begin{array}{l}\displaystyle \theta ^j_k=\Psi \Big ({\alpha \beta E^-_k\over E^+_k+D_k-2\tilde{\lambda }_2}\Big ) \\ \displaystyle \mu ^j_k=n\theta ^j_k\end{array}\right. \\ \displaystyle \hbox {If }\gamma E^-_k-E^+_k+2\tilde{\lambda }_2\ge D_k \ge \gamma E^-_k-E^+_k& \displaystyle \Rightarrow \theta ^j_k={1\over 2},\mu ^j_k={n\over 2}\\ \displaystyle \hbox {If } \gamma E^-_k-E^+_k \ge D_k\ge {\alpha \over \beta }E^-_k-E^+_k& \displaystyle \Rightarrow \theta ^j_k=\Psi \Big ({\alpha \beta E^-_k\over E^+_k+D_k}\Big ),\mu ^j_k=n\theta ^j_k\\ \displaystyle \hbox {If }{\alpha \over \beta }E^-_k-E^+_k\ge D_k& \displaystyle \Rightarrow \theta ^j_k=1,\mu ^j_k=n\end{array}\right. \end{aligned}$$ -
5.
Assume \(\max \{\tilde{\lambda }_1,\tilde{\lambda }_2\}\le D_k\le \tilde{\lambda }_1+\tilde{\lambda }_2\). Then
$$\begin{aligned} \left\{ \begin{array}{ll}\displaystyle \hbox {If }D_k\ge {\beta \over \alpha }E^-_k-E^+_k+2\tilde{\lambda }_2& \displaystyle \Rightarrow \theta ^j_k=0,\mu ^j_k=0\\ \displaystyle \hbox {If }{\beta \over \alpha }E^-_k-E^+_k+2\tilde{\lambda }_2\ge D_k\ge \gamma E^-_k-E^+_k+2\tilde{\lambda }_2 & \displaystyle \Rightarrow \left\{ \begin{array}{l}\displaystyle \theta ^j_k=\Psi \Big ({\alpha \beta E^-_k\over E^+_k+D_k-2\tilde{\lambda }_2}\Big )\\ \displaystyle \mu ^j_k=n\theta ^j_k\end{array}\right. \\ \displaystyle \hbox {If } D_k\le \min \Big \{\gamma E^-_k-E^+_k+2\tilde{\lambda }_2,E^+_k-\gamma E^-_k+2\tilde{\lambda }_1\Big \} & \displaystyle \Rightarrow \theta ^j_k={1\over 2},\mu ^j_k={n\over 2}\\ \displaystyle \hbox {If } E^+_k-\gamma E^-_k+2\tilde{\lambda }_1 \le D_k \le E^+_k-{\alpha \over \beta }E^-_k+2\tilde{\lambda }_1& \displaystyle \Rightarrow \left\{ \begin{array}{l}\displaystyle \theta ^j_k=\Psi \Big ({\alpha \beta E^-_k\over E^+_k-D_k+2\tilde{\lambda }_1}\Big )\\ \displaystyle \mu ^j_k=n(1-\theta ^j_k)\end{array}\right. \\ \displaystyle \hbox {If } D_k\ge E^+_k-{\alpha \over \beta }E^-_k+2\tilde{\lambda }_1& \displaystyle \Rightarrow \theta ^j_k=1,\mu ^j_k=0 \end{array}\right. \end{aligned}$$ -
6.
Assume \(\tilde{\lambda }_1+\tilde{\lambda }_2\le D_k\). Then
$$\begin{aligned} \left\{ \begin{array}{ll}\displaystyle \hbox {If }\tilde{\lambda }_2-\tilde{\lambda }_1\le E^+_k-{\beta \over \alpha }E^-_k& \displaystyle \Rightarrow \theta ^j_k=0,\mu ^j_k=0\\ \displaystyle \hbox {If } E^+_k-{\beta \over \alpha }E^-_k\le \tilde{\lambda }_2-\tilde{\lambda }_1\le E^+_k-{\alpha \over \beta }E^-_k & \displaystyle \Rightarrow \theta ^j_k=\Psi \Big ({\alpha \beta E^-_k\over E^+_k+\tilde{\lambda }_1-\tilde{\lambda }_2}\Big ),\mu ^j_k=0\\ \displaystyle \hbox {If }\tilde{\lambda }_2-\tilde{\lambda }_1\ge E^+_k-{\alpha \over \beta }E^-_k & \displaystyle \Rightarrow \theta ^j_k=1,\mu ^j_k=0.\end{array}\right. \end{aligned}$$
Following the above description, we propose the following algorithm:
Algorithm.
-
1.
Take \(\rho >0\) small, and \({\bar{j}}\in \mathbb N\).
-
2.
Take \(\lambda _{0,1},\lambda _{0,2}\ge 0\), \(\theta _0\in L^\infty (\Omega ;[0,1])\), \(\mu _0\in L^\infty (\Omega ;[0,n])\), \(A_0\in L^\infty (\Omega )^{N\times N}\) symmetric such that
$$\begin{aligned} \left\{ \begin{array}{l}\displaystyle \int _\Omega \Big (\theta _0-{\mu _0\over n}\Big )^+dx\le \kappa ^\alpha ,\quad \int _\Omega \Big (1-\theta _0-{\mu _0\over n}\Big )^+\le \kappa ^\beta \\ \displaystyle \textrm{Sp}(A_0)\subset \big [m^-(\theta _0),m^+(\theta _0)\big ]\hbox { a.e. in }\Omega .\end{array}\right. \end{aligned}$$(4.21) -
3.
Assume we have constructed \(\lambda _{k,1},\lambda _{k,2}\ge 0\), \(\theta _k\in L^\infty (\Omega ;[0,n])\), \(\mu _k\in L^\infty (\Omega ;[0,n])\), \(A_k\in L^\infty (\Omega )^{N\times N}\) symmetric, satisfying the analogous to (4.21).
-
4.
We compute the solutions \(u_k\), \(p_k\) of (4.8) and (4.11) respectively. Then, we define \(E_k^+\), \(E_k^-\) by (4.16).
-
5.
We denote \(\lambda ^0_{k,1}=\lambda _{k,1}\), \(\lambda ^0_{k,2}=\lambda _{k,2},\) then for \(j\le {\bar{j}}-1\), we define \((\lambda _{k,1}^{j+1},\lambda _{k,2}^{j+1})\) by (4.18) where \(\theta ^j_k\), \(\mu ^j_k\) are defined by Proposition 4.2.
-
6.
We take \(\lambda _{k,1}=\lambda _{k,1}^{{\bar{j}}}\), \(\lambda _{k,2}=\lambda _{k,2}^{{\bar{j}}}\), \(\hat{\theta }=\theta ^{{\bar{j}}}_k\), \(\hat{\mu }=\mu ^{{\bar{j}}}_k\) and \({\hat{A}}\) as a symmetric matrix function in \(L^\infty (\Omega )^{N\times N}\) satisfying (4.14) and such that \(\textrm{Sp}({\hat{A}})\subset [m^-(\hat{\theta }),m^+(\hat{\theta })]\), a.e. in \(\Omega \).
-
7.
For a choice of \(\varepsilon _k\in (0,1]\) (see Remark 4.3), we define \(\theta _{k+1}\), \(A_{k+1}\), \(\mu _{k+1}\) by (4.7).
Remark 4.3
In order to implement the previous algorithm it is necessary to provide a rule to choose \(\varepsilon _k\). We propose three strategies for this purpose. For the second and third one, we need to assume
-
1.
Take \(\varepsilon _k=\varepsilon \) a small fixed number in (0, 1]. More generally one can adapt \(\varepsilon _k\) after a certain number of iterations.
-
2.
By (4.22), it is not difficult to show that for every pair of controls \((\theta ,A,\mu )\) and \((\hat{\theta },{\hat{A}}, \hat{\mu })\) satisfying the constraints in (4.1) the function \(\Phi :[0,1]\rightarrow \mathbb R\), defined by
$$\begin{aligned} \Phi (\varepsilon )=\int _\Omega F(x,u_\varepsilon )\,dx, \end{aligned}$$(4.23)with \(u_\varepsilon \) the solution of
$$\begin{aligned} \left\{ \begin{array}{l}\displaystyle -\textrm{div}\big ((A+\varepsilon ({\hat{A}}- A))\nabla u_\varepsilon \big )+\big (\mu +\varepsilon (\hat{\mu }-\mu ))u_\varepsilon =f\ \hbox { in }\Omega \\ u_\varepsilon \in H^1_0(\Omega ),\end{array}\right. \end{aligned}$$(4.24)belongs to \(C^2([0,1])\). Moreover, there exists \(C>0\) which does not depend on \((\theta ,A,\mu )\) and \((\hat{\theta },{\hat{A}}, \hat{\mu })\) (but it depends on n) such that
$$\begin{aligned} |\Phi ''(\varepsilon )|\le C,\quad \forall \,\varepsilon \in [0,1]. \end{aligned}$$(4.25)Therefore, instead of (4.9), we can write
$$\begin{aligned} \int _\Omega F(x,u_{k+1})dx\le \int _\Omega F(x,u_k)dx+\varepsilon _k\int _\Omega \partial _sF(x,u_k)u'_k\,dx+{C\over 2}\varepsilon _k^2. \end{aligned}$$The idea is to choose \(\varepsilon _k\) minimizing the right-hand side. By (4.12), this gives
$$\begin{aligned} \varepsilon _k=\min \left\{ {D_k\over C},1\right\} ,\ \hbox { with }D_k:=\int _\Omega ({\hat{A}}-A_k)\nabla u_k\cdot \nabla p_k\,dx+\int _\Omega (\hat{\mu }-\mu _k)u_kp_k\,dx. \end{aligned}$$Here we recall that \((\hat{\theta },{\hat{A}}, \hat{\mu })\) is chosen as a solution of (4.13) and thus \(D_k\ge 0\). With this choice, we have
$$\begin{aligned} \int _\Omega F(x,u_{k+1})dx\le \int _\Omega F(x,u_k)dx-\left\{ \begin{array}{cl}\displaystyle {D_k^2\over 2C} & \hbox { if }D_k<C\\ \displaystyle D_k-{C\over 2} & \hbox { if }D_k\ge C.\end{array}\right. \end{aligned}$$This implies that \(\sum _{k=0}^\infty D_k^2<\infty \), and therefore that \(D_k\) tends to zero. In this point, it is interesting to observe that the computation done above to get the direction of maximum descent shows that an optimality condition to be satisfied by the solutions \((\hat{\theta },{\hat{A}},\hat{\mu })\) of (4.1) is given by \((\hat{\theta },{\hat{A}}, \hat{\mu })\) a solution of
$$\begin{aligned} \begin{array}{c}\displaystyle \max \int _\Omega \big (A\nabla {\hat{u}}\cdot \nabla {\hat{p}}+ \mu {\hat{u}}{\hat{p}}\big )dx\\ \displaystyle \left\{ \begin{array}{l} \mu \in L^\infty (\Omega ;[0,n]),\quad \theta \in L^\infty (\Omega ;[0,1]),\quad \textrm{Sp}(A)\subset [m^-(\theta ),m^+(\theta )]\ \hbox { a.e. in }\Omega \\ \displaystyle \int _\Omega \Big (\theta -{\mu \over n}\Big )^+dx\le \kappa ^\alpha ,\quad \int _\Omega \Big (1-\theta -{\mu \over n}\Big )^+dx\le \kappa ^\beta .\end{array}\right. \end{array} \end{aligned}$$with \({\hat{u}}\), \({\hat{p}}\) the solutions of
$$\begin{aligned} \left\{ \begin{array}{l} -\textrm{div}({\hat{A}}\nabla {\hat{u}})+\hat{\mu }{\hat{u}}=f\ \hbox { in }\Omega \\ \displaystyle {\hat{u}}\in H^1_0(\Omega ),\end{array}\right. \qquad \left\{ \begin{array}{l} -\textrm{div}({\hat{A}}\nabla {\hat{p}})+\hat{\mu }{\hat{p}}=\partial _sF(x,{\hat{u}})\ \hbox { in }\Omega \\ \displaystyle {\hat{p}}\in H^1_0(\Omega ).\end{array}\right. \end{aligned}$$ -
3.
Defining \(\Phi \) by (4.23) with \((\theta ,A,\mu )=(\theta _k,A_k,\mu _k)\) and \((\hat{\theta },{\hat{A}},\hat{\mu })\) solution of (4.13), we can replace (4.9) by the Taylor expansion of second order
$$\begin{aligned} \int _\Omega F(x,u_k)dx+\varepsilon _k\int _\Omega \partial _sF(x,u_k)u'_kdx+{1\over 2}\Phi ''(0)\varepsilon _k^2. \end{aligned}$$(4.26)Now, a simple calculation shows
$$\begin{aligned} \Phi ''(0)=\int _\Omega \Big (\partial ^2_{ss}F(x,u_k)|u'_k|^2+2 (A-{\hat{A}})\nabla u'_k\cdot \nabla p_k\,dx+2(\mu -\hat{\mu })u'_kp_k\Big )dx, \end{aligned}$$with \(u'_k\) and \(p_k\) the solutions of (4.10) and (4.11) respectively. Therefore, we can choose \(\varepsilon _k\) minimizing (4.26) in [0, 1]. This needs to compute \(u'_k\) in each iteration.
We finish this section showing some numerical experiments for solving (2.15) based on the approximation given by (4.1) and the algorithm described above. The computation has been carried out using FreeFem++ v4.4-3 ([21], http://www.freefem.org).
First example We have chosen a classical example given by \(\Omega =(0,1)^2\), \(F(x,u)=-u\), \(f=1\), \(\alpha =1\), \(\beta =2\) and some particular choices of \(\kappa ^\alpha \) and \(\kappa ^\beta \). This problem appears in the optimal arrangement of two elastic materials in the cross section of a beam in order to minimize the torsion [9, 23, 26]. It is well known that the problem is equivalent to
Using this formulation and Theorem 3.2, we have the existence of a quasi-closed set \(C\subset \Omega \), such that
We have numerically solved the problem, using the above algorithm, in two cases:
First case We take \(\kappa ^\alpha =\kappa ^\beta =\pi /8\). With this choice, it is not difficult to check that the solution of (2.15) is unique, and it is given by
In Figs. 1 and 2, we represent the functions \(\theta \) and \(\mu \) obtained numerically for \(n=10^3\) and \(n=10^5\) respectively. Just as expected the results for \(n=10^5\) are more accurate.
Second case We take \(\kappa ^\alpha =\kappa ^\beta =1/2\). Then \(\mu \) is the null function and the problem is reduced to
The solution of this problem is not analytically known, but it has been numerically solved by several authors (see e.g. [1] Chapter 5, [10] Chapter 3, [19, 22] ). Our results agree with those obtained by them and have been obtained solving directly problem (4.1). In Fig. 3 we show the optimal proportion \(\theta ^\alpha \) obtained solving (4.1) with \(n=10^4\). Indeed, since in this case the optimal measure \(\hat{\mu }\) in (2.15) is the null measure, the choice of n is not relevant.
Second example We take \(\Omega =(0,1)^2\), \(\alpha =1\), \(\beta =2\), \(\kappa ^\alpha =0.2\), \(\kappa ^\beta =0.3\), \(F(x,u)=u\), and \(f=-\chi _K+\chi _{\Omega {\setminus } K}\), with K the non-symmetric cross given by (see Fig. 4 left)
A related problem to this one has been numerically solved in [7]. Namely, it is considered
which corresponds to (2.4) with \(\alpha =\beta =1\). Indeed, in [7], it is considered the approximated formulation
with \(\varepsilon >0\) small and \(n>0\) large (in [7] \(n=10^4\), \(\varepsilon =3\cdot 10^{-4}\)). Observe that in order to minimize (4.29), we want u to be negative, and then we hope \(\omega \) to be a certain neighborhood of the set K, where f is negative. In the numerical computation in [7], the constraint \(|\omega |\le 0.5\) is not saturated. This is because if \(\omega \) is large then u becomes positive in a part of \(\omega \).
In our case, we solve problem (2.4) using formulation (4.1). Such as we hope by Theorem 3.2, the function \(\mu \) we get is essentially a characteristic function, which is given by Fig. 4 right. It is very similar to that obtained in [7] solving (4.30). The function \(\theta ^\alpha \) is given in Fig. 5 left. We recall that the value of \(\theta ^\alpha \) where \(\mu \) is very large is no relevant. Thus, we represent in Fig. 5 right, the function \(\theta ^\alpha \) restricted to the set where \(\mu \) is close to zero. Since the material \(\alpha \) is better than \(\beta \) for this problem, we get that the constraint
is saturated. The optimal solution completes the set obtained in [7] using the material \(\beta \). Thus, the constraint
is not saturated.
Finally, in Fig. 6 we show the cost evolution of the minimization algorithm, in order to check the decreasing process. The evolution in the other examples is similar.
Data availability
Not applicable.
References
Allaire, G.: Shape optimization by the homogenization method. Appl. Math. Sci., vol. 146. Springer, New York (2002)
Bellido, J.C., Buttazzo, B., Velichkov, B.: Worst-cases shape optimization for the Dirichlet energy. Nonlinear Anal. 153, 117–129 (2017)
Bendsøe, M.P., Sigmund, O.: Topology Optimization. Springer, Berlin (2003)
Buttazzo, G., Dal Maso, G.: Shape optimization for Dirichlet problems: relaxed formulation and optimality conditions. Appl. Math. Optim. 23, 17–49 (1991)
Buttazzo, G., Dal Maso, G.: An existence result for a class of shape optimization problems. Arch. Rational Mech. Anal. 122, 183–195 (1993)
Buttazzo, G., Casado-Díaz, J., Maestre, F.: On the existence of optimal potentials on unbounded domains. SIAM J. Math. Anal. 53, 1088–1121 (2021)
Buttazzo, G., Maestre, F., Velichkov, B.: Optimal potentials for problems with changing sign data. J. Optim. Theory Appl. 178, 743–762 (2018)
Calvo-Jurado, C., Casado-Díaz, J.: The limit of Dirichlet systems for variable monotone operators in general perforated domains. J. Math. Pures Appl. 81, 471–493 (2002)
Casado-Díaz, J.: Smoothness properties for the optimal mixture of two isotropic materials: the compliance and eigenvalue problems. SIAM J. Control Optim. 53, 2319–2349 (2015)
Casado-Díaz, J.: Optimal design of multi-phase materials with a cost functional that depends nonlinearly on the gradient, Springer Briefs in Mathematics. Springer, Cham (2022)
Casado-Díaz, J., Garroni, A.: Asymptotic behavior of nonlinear elliptic systems on varying domains. SIAM J. Math. Anal. 31, 581–624 (2000)
Casado-Díaz, J., Castro, C., Luna-Laynez, M., Zuazua, E.: Numerical approximation of a one-dimensional elliptic optimal design problem. Multiscale Model. Simul. 9, 1181–1216 (2011)
Ciarlet, P.G.: Introduction à l’analyse numérique matricielle et à l’optimisation. Masson, Paris (1988)
Cioranescu, D., Murat, F., Un terme ètrange venu d’ailleurs. In Nonlinear Partial Differential Equations and Their Applications. College de France Seminar, Vols. II and III, Res. Not. Math., vols. 60 and 70. Pitman, London, 1982, pp. 98–138 and pp. 154–178. English translation: D. Cioranescu, F. Murat. A strange term coming from nowhere. In: Cherkaev, L., Kohn, R.V. (eds) Topics in the Mathematical Modelling of Composite Materials. Progress in Nonlinear Diff. Equ. and their Appl., vol. 31. Birkaüser, Boston, pp. 45–93 (1998)
Dal Maso, G., Mosco, U.: Wiener-criterion and \(\Gamma \)-convergence. Appl. Math. Optim. 15, 15–63 (1987)
Dal Maso, G., Garroni, A.: New results on the asymptotic behavior of Dirichlet problems in perforated domains. Math. Models Methods Appl. Sci. 4, 373–407 (1994)
Dal Maso, G., Murat, F.: Asymptotic behaviour and correctors for linear Dirichlet problems with simultaneously varying operators and domains. Ann. Inst. H. Poincaré Anal. Non Linéaire 21, 445–486 (2004)
Evans, L.C., Gariepy, R.F.: Measure theory and fine properties of functions, Studies in Advances Mathematics. CRC Press, Boca Raton (1992)
Glowinski, R.: Numerical simulation for some applied problems originating from continuum mechanics. In: Trends in Applications of Pure Mathematics to Mechanics. Symp., Palaiseau/France 1983, Lecture Notes in Physics, vol. 195, pp. 96–145 (1984)
Goodman, J., Kohn, R.V., Reyna, L.: Numerical study of a relaxed variational problem for optimal design. Comput. Methods Appl. Mech. Eng. 57, 107–127 (1986)
Hecht, F.: New developments in FreeFem++. J. Numer. Math. 20, 251–265 (2012)
Kawohl, B., Stara, J., Wittum, G.: Analysis and numerical studies of a problem of shape design. Arc. Rational Mech. Anal. 114, 343–363 (1991)
Lurie, K.A., Cherkaev, A.V.: Exact estimates of the conductivity of a binary mixture of isotropic materials. Proc. R. Soc. Edinb. 104A, 21–38 (1986)
Murat, F.: Théorèmes de non existence pour des problèmes de contrôle dans les coefficients. C.R.A.S Sci. Paris A 274, 395–398 (1972)
Murat, F.: H-convergence. Séminaire d’Analyse Fonctionnelle et Numérique, 1977-78, Université d’Alger, multicopied, 34 pp. English translation: F. Murat, L. Tartar. H-convergence. In: Cherkaev, L., Kohn, R.V. (eds) Topics in the Mathematical Modelling of Composite Materials, Progress in Nonlinear Diff. Equ. and their Appl., vol. 31. Birkaüser, Boston, pp. 21–43 (1998)
Murat, F., Tartar, L.: Calcul des variations et homogénéisation. In Les méthodes de l’homogénéisation: theorie et applications en physique, Eirolles, Paris, 1985, pp. 319–369. English translation: F. Murat, L. Tartar. Calculus of variations and homogenization. In: Cherkaev, L., Kohn, R.V. (eds) Topics in the Mathematical Modelling of Composite Materials, Progress in Nonlinear Diff. Equ. and their Appl., vol. 31. Birkaüser, Boston, pp. 139–174 (1998)
Pironneau, O.: Optimal Shape Design for Elliptic Systems. Springer, New York (1984)
Spagnolo, S.: Sulla convergenza di soluzioni di equazioni paraboliche ed ellittiche. Ann. Scuola Norm. Sup. Pisa Cl. Sci. 22, 571–597 (1968)
Svanberg, K.: The method of moving asymptotes-a new method for structural optimization. Int. J. Numer. Methods Eng. 24, 359–373 (1987)
Tartar, L.: Estimations fines de coefficients homogénéisés. In: Kree, P. (ed.) Ennio de Giorgi colloquium (Paris, 1983). Research Notes in Math., vol. 125. Pitman, London, pp. 168–187 (1985)
Tartar, L.: The General Theory of Homogenization. A Personalized Introduction. Springer, Berlin (2009)
Acknowledgements
This work has been partially supported by the project PID2020-116809GB-I00 of the Ministerio de Ciencia e Innovación of the government of Spain and the project UAL2020-FQM-B2046 of the Consejería de Transformación Económica, Industria, Conocimiento y Universidades de la Junta de Andalucía.
Funding
Funding for open access publishing: Universidad de Sevilla/CBUA.
Author information
Authors and Affiliations
Corresponding author
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
Casado-Díaz, J., Luna-Laynez, M. & Maestre, F. Control problems in the coefficients and the domain for linear elliptic equations. Rev. Real Acad. Cienc. Exactas Fis. Nat. Ser. A-Mat. 118, 160 (2024). https://doi.org/10.1007/s13398-024-01662-4
Received:
Accepted:
Published:
DOI: https://doi.org/10.1007/s13398-024-01662-4
Keywords
- Optimal design
- Shape optimization
- Two-phase materials
- Relaxation
- Homogenization
- Discrete approximation
- Convergence