Abstract
Finite element approximations of Dirichlet boundary control problems governed by parabolic PDEs on convex polygonal domains are studied in this paper. The existence of a unique solution to optimal control problems is guaranteed based on very weak solution of the state equation and \(L^2(0,T;L^2(\varGamma ))\) as control space. For the numerical discretization of the state equation we use standard piecewise linear and continuous finite elements for the space discretization of the state, while a dG(0) scheme is used for time discretization. The Dirichlet boundary control is realized through a space–time \(L^2\)-projection. We consider both piecewise linear, continuous finite element approximation and variational discretization for the controls and derive a priori \(L^2\)-error bounds for controls and states. We finally present numerical examples to support our theoretical findings.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
In this paper we study the following parabolic optimal control problem:
subject to
where \(\varOmega _T=\varOmega \times (0,T], \,\varSigma =\partial \varOmega \times (0,T]\) with \(\varOmega \) denoting an open bounded domain with boundary \(\varGamma :=\partial \varOmega \), \(U_{ad}\) is the admissible control set which is assumed to be of box type
with \(u_a< u_b\) denoting constants. For the convenience we make the following assumption on the domain \(\varOmega \) and the given data which shall be valid throughout the paper without explicit mentioning:
Assumption 1
We assume that \(\varOmega \) is an open bounded, convex polygonal domain in \(\mathbb {R}^2\). \(\alpha >0, \,f\in L^2(0,T;L^2(\varOmega )), \,y_0\in L^2(\varOmega ), \,y_d\in L^2(0,T;L^2(\varOmega ))\) and \(T>0\) are fixed data.
Dirichlet boundary control is important in many practical applications such as the active boundary control of flows, see e.g. [13, 18, 20]. If one is, e.g. interested in blowing and suction as control on part of the boundary, controls with low regularity should be admissible, which could have jumps and satisfy pointwise bounds. In the mathematical theory one has to use the concept of very weak solutions in this situation, see [4] for a more detailed discussion of this fact.
In the present work we consider a parabolic Dirichlet boundary control problem of tracking type, which may be regarded as prototype problem to study Dirichlet boundary control for time-dependent PDEs. For parabolic optimal boundary control problems of Dirichlet type, only few contributions can be found in the literature [2, 3, 23]. Kunisch and Vexler [23] considered a semi-smooth Newton method for the numerical solution of parabolic Dirichlet boundary control problems. A Robin penalization method using Robin-type boundary conditions applied to parabolic Dirichlet boundary control problems is investigated in [3]. However, to the best of the authors’ knowledge no error analysis is available for the finite element approximation of this kind of problems. With the present paper we intend to fill this gap and derive a priori error estimates for parabolic Dirichlet boundary control problems. Compared to the elliptic case, parabolic Dirichlet boundary control problems are more involved in both the definition of discrete schemes and the a priori error analysis, since the regularity of the involved state variable is low.
Finite element approximations of optimal control problems are important for the numerical treatment of optimal control problems related to practical applications, see e.g. [22, Ch. 4]. An overview on the numerical a priori and a posteriori analysis for elliptic control problems can be found in [22, Ch. 3] and [28], respectively. To the best of the authors’ knowledge the first contribution to parabolic optimal control problems is given in [36]. The state of the art in the numerical a priori analysis of distributed parabolic optimal control problems can be found in [30, 31]. More recent contributions with higher order in time Galerkin schemes can be found in [1, 32, 35]. Residual-based a posteriori error estimates are presented in [26] and [27]. For boundary control problems with parabolic equations we refer to [15]. There is a long list of contributions to boundary control of elliptic PDEs, see e.g. [7, 8, 10, 11, 14, 16, 21, 29, 34]. Further references can be found in [22, Ch. 3].
In this paper we use the very weak solution concept for the state equation and \(L^2(0,T;L^2(\varGamma ))\) as control space to argue the existence of a unique solution to the optimal control problems (1.1)–(1.2). For the numerical discretization of the optimal control problem we discretize the state using standard piecewise linear and continuous finite elements in space and dG(0) scheme in time. The Dirichlet boundary conditions are approximated based on the space–time \(L^2\)-projection. The control is discretized in space either by piecewise linear finite elements or implicitly through the discretization of the adjoint state, the so-called variational discretization (see [19]). For both cases we derive a priori error bounds for the state and control in the \(L^2\)-norm for problems posed on polygonal domains. As main result we obtain the error bound
under the coupling \(k=O(h^2)\) with both full control discretisation and variational control discretisation for the optimal solution (y, u) of (1.1), where \(Y_{hk}\) and \(U_{hk}\) denote the optimal discrete state and control, see also Corollary 1. We present several numerical examples which support our theoretical findings.
The rest of our paper is organized as follows. In Sect. 2 we present the analytical setting of the parabolic Dirichlet boundary control problem and argue the existence of a unique solution. In Sect. 3 we establish the fully discrete finite element approximation to the state equation and the corresponding stability results. Then we formulate the fully discrete approximation for parabolic Dirichlet boundary control problems. The a priori error analysis for the finite element approximation and the variational discretization of the optimal control problems posed on convex, polygonal domains is studied in Sect. 4. Furthermore, we present some numerical experiments in Sect. 5 to support our theoretical results.
2 Optimal Control Problem
For \(m\ge 0\) and \(1\le s\le \infty \), we adopt the standard notation \(W^{m,s}(\varOmega )\) for Sobolev spaces on \(\varOmega \) with norm \(\Vert \cdot \Vert _{m,s,\varOmega }\) and seminorm \(|\cdot |_{m,s,\varOmega }\), where \(H^m(\varOmega )=W^{m,2}(\varOmega ), \,\Vert \cdot \Vert _{m,\varOmega }=\Vert \cdot \Vert _{m,2,\varOmega }\) and \(|\cdot |_{m,\varOmega }=|\cdot |_{m,2,\varOmega }\) for \(s=2\). Note that \(H^0(\varOmega )=L^2(\varOmega )\) and \(H_0^1(\varOmega )=\{v\in H^1(\varOmega );\ v=0\ \text{ on }\ \partial \varOmega \}\). We denote by \(L^r(0,T;W^{m,s}(\varOmega ))\) the Banach space of all \(L^r\) integrable functions from [0, T] into \(W^{m,s}(\varOmega )\) with norm \(\Vert v\Vert _{L^r(0,T;W^{m,s}(\varOmega ))}= \Big (\int _0^T\Vert v\Vert ^r_{m,s,\varOmega }dt\Big )^{{1\over r}}\) for \(1\le r<\infty \), and with the standard modification for \(r=\infty \). For a Banach space Y, we use the abbreviations \(L^2(Y)=L^2(0,T;Y), \,H^s(Y)=H^s(0,T;Y), \,s=[0,\infty )\), and \(C(Y)=C([0,T];Y)\). We denote the \(L^2\)-inner products on \(L^2(\varOmega ), \,L^2(\varOmega _T)\) and \(L^2(\varGamma )\) by \((\cdot ,\cdot ), \,(\cdot ,\cdot )_{\varOmega _T}\) and \(\langle \cdot ,\cdot \rangle \), respectively. In addition c and C denote generic positive constants.
Let
The standard weak form for the parabolic equation (1.2) is to find \(y\in L^2(H^1(\varOmega ))\cap H^1(H^{-1}(\varOmega ))\) with \(y|_{\varSigma }=u\) and \(y(\cdot ,0)|_\varOmega =y_0(\cdot )\) such that
This setting requires \(u\in L^2(H^{1\over 2}(\varGamma ))\). Motivated by practical considerations (see e.g. the discussion in [4]) we are interested in controls \(u\in U_{ad}\) defined in (1.3). For a proper treatment of the state equation in this case we use the transposition technique introduced by Lions and Magenes (see [24, Ch. 2, Sec. 5.2] and [25, Ch. 2]) to argue the existence of a unique solution to the state equation (1.2) in the present paper. The very weak form of (1.2) that we shall utilize reads: Find \(y\in L^2(L^2(\varOmega ))\) such that
with \(z(\cdot ,T)=0\) holds, where \(\partial _{{n}}v:=\nabla v\cdot n\) with n denoting the unit outward normal to \(\varGamma \). Then the existence and uniqueness of a very weak solution of (2.2), which we denote by \(y=\mathcal {G}(u)\), is shown in the following lemma (see, e.g. [25])
Lemma 1
For each \(u\in L^2(L^2(\varGamma ))\), there exists a unique very weak solution \(y\in L^2(L^2(\varOmega ))\) of (2.2) satisfying
Proof
For \(y_0\in L^2(\varOmega ), \,f\in L^2(L^2(\varOmega ))\) and \(u\equiv 0\) it is straightforward to show that (1.2) admits a unique solution \(y\in L^2(H^1_0(\varOmega ))\cap H^1(H^{-1}(\varOmega ))\) in the sense of (2.1), which also satisfies (2.3). To prove the lemma in the case \(u\ne 0\) it is sufficient to consider the case \(f\equiv 0, \,y_0\equiv 0\), where we follow the constructive approach of [10]. For each \(g\in L^2(L^2(\varOmega ))\) we denote by \(z\in L^2(H^2(\varOmega )\cap H_0^1(\varOmega ))\cap H^1(L^2(\varOmega ))\) the solution of
Then we have \(\partial _{{n}}z\in H^{\frac{1}{4}}(L^2(\varGamma ))\) according to [23, Th. 3.2]. Moreover, from the fact that \(z\in L^2(H^2(\varOmega ))\) and \(z=0\) on \(\varSigma \) we obtain that \(\partial _{{n}}z\in L^2(H^{1\over 2}(\varGamma ))\) according to Lemma A.2 in [6]. We denote by \(T:L^2(L^2(\varOmega ))\rightarrow L^2(L^2(\varGamma ))\) the continuous linear operator which is defined by \(Tg:=-\partial _{{n}}z|_{\varSigma }\) and denote its adjoint by \(T^*\). Then with \(y=T^*u\) we have
which verifies that y satisfies (2.2). The estimate (2.3) follows by observing that
\(\square \)
Now we are ready to formulate the optimal control problem considered in the present paper. It reads
By standard arguments (see, e.g. [24, Ch. 2, Sec. 1.2]), there exists a unique solution (y, u) for problem (2.5). Let \(J(u):=J(y(u),u)\) denote the reduced cost functional, where for each \(u\in L^2(L^2(\varGamma ))\) the state y(u) is the unique very weak solution of (2.2). Then J is infinitely often Fréchet differentiable. Moreover, the first order sufficient and necessary optimality conditions for problem (2.5) are given by
Theorem 1
Assume that \(u\in L^2(L^2(\varGamma ))\) is the unique solution of problem (2.5) and let y be the associated state. Then there exists a unique adjoint state \(z\in L^2(H^1_0(\varOmega ))\cap H^1(H^{-1}(\varOmega ))\) such that
and
We note that (2.7) is equivalent to
or
where for each \(v\in L^2(L^2(\varGamma )), \,y(v)\) is the solution of problem (2.2) with u replaced by v, and \(P_{U_{ad}}:L^2(L^2(\varGamma ))\rightarrow U_{ad}\) denotes the orthogonal projection.
We now turn to the regularity properties of optimal controls u on \(\varSigma \). The proof of the following theorem can be found in, e.g. [23, Th. 3.4].
Theorem 2
Let \((y,u,z)\in L^2(L^2(\varOmega ))\times L^2(L^2(\varGamma ))\times L^2(H^1_0(\varOmega ))\cap H^1(H^{-1}(\varOmega ))\) be the solution of optimal control problem (2.5)–(2.8). Then we have
and
Proof
From \(f\in L^2(L^2(\varOmega )), \,y_0\in L^2(\varOmega )\) and \(u\in L^2(L^2(\varGamma ))\) we conclude that \(y\in L^2(L^2(\varOmega ))\) according to Lemma 1. Thus, \(y_d\in L^2(L^2(\varOmega ))\) implies \(z\in L^2(H^2(\varOmega )\cap H_0^1(\varOmega ))\cap H^1(L^2(\varOmega ))\), which in turn implies \(\partial _n z\in L^2(H^{1\over 2}(\varGamma ))\cap H^{\frac{1}{4}}(L^2(\varGamma ))\) (see [6, 17, 23]). From (2.9) we obtain that \(u\in L^2(H^{1\over 2}(\varGamma ))\cap H^{\frac{1}{4}}(L^2(\varGamma ))\) and thus \(y\in L^2(H^1(\varOmega ))\cap H^{\frac{1}{2}}(L^2(\varOmega ))\) (see [25, Vol. II, p. 78]). This completes the proof. \(\square \)
In our analysis we frequently use results of the following backward in time parabolic problem:
If \(g\in L^2(L^2(\varOmega ))\), then (2.12) has a unique solution \(w\in L^2(H^2(\varOmega )\cap H_0^1(\varOmega ))\cap H^1(L^2(\varOmega ))\) satisfying
3 Finite Element Discretization of the State Equation and Optimal Control Problems
At first let us consider the finite element approximation of the state Eq. (1.2). For the spatial discretization we consider conforming Lagrange triangular elements.
Let \(\mathcal {T}^h\) be a quasi-uniform partitioning of \(\varOmega \) into disjoint regular triangles \(\tau \), so that \(\bar{\varOmega }=\bigcup _{\tau \in \mathcal {T}^h}\bar{\tau }\). Associated with \(\mathcal {T}^h\) is a finite dimensional subspace \(V^h\) of \(C(\bar{\varOmega })\), such that for \(\chi \in V^h\) and \(\tau \in \mathcal {T}^h, \,\chi |_{\tau }\) are piecewise linear polynomials. We set \(V^h_0=V^h\cap H_0^1(\varOmega )\).
Let \(\mathcal {T}_U^h\) be a partitioning of \(\varGamma \) into disjoint regular segments s, so that \(\varGamma =\bigcup _{s\in \mathcal {T}_U^h}\bar{s}\). Associated with \(\mathcal {T}_U^h\) is another finite dimensional subspace \(U^h\) of \(L^2(\varGamma )\), such that for \(\chi \in U^h\) and \(s\in \mathcal {T}_U^h, \,\chi |_s\) are piecewise linear polynomials. Here we suppose that \(\mathcal {T}_U^h\) is the restriction of \(\mathcal {T}^h\) on the boundary \(\varGamma \) and \(U^h=V^h(\varGamma )\), where \(V^h(\varGamma )\) is the restriction of \(V^h\) on the boundary \(\varGamma \).
For the standard Lagrange interpolation operator \(I_h:C(\bar{\varOmega })\rightarrow V^h\), we have the following error estimate (see, e.g. [9, Sec. 3.1])
To define our discrete scheme, we need to introduce some projection operators. Here \(Q_h:L^2(\varGamma )\rightarrow V^h(\varGamma )\) and \(\tilde{Q}_h:L^2(\varOmega )\rightarrow V^h_0\) denote the orthogonal projection operators. Furthermore, \(R_h: H^1_0(\varOmega )\rightarrow V^h_0\) denotes the Ritz projection operator defined as
It is well known that the Ritz projection satisfies (see, e.g. [9, Sec. 3.1])
For the \(L^2(\varGamma )\) projection operator \(Q_h\) we also have (see [9] and [12, pp. 85–86, Eq. (25) and (28)])
and
In our following analysis we need estimates for discrete harmonic functions.
Lemma 2
Let \(v_h\in V^h(\varGamma )\), and suppose that \(w\in H^1(\varOmega )\) is the solution of
and \(w_h\in V^h\) is the solution of
Then
Proof
The proof of (3.8) and (3.9) can be found in [5, Lm. 3.2], [7, Th. 5.4] and [11, Lm. 1]. Here we provide a proof of (3.10). For each \(g\in H^{-{1\over 2}}(\varOmega )\) let \(\psi _g\in H^{3\over 2}(\varOmega )\cap H^1_0(\varOmega )\) be the solution of
Then we have \(\Vert \psi _g\Vert _{{3\over 2},\varOmega }\le C\Vert g\Vert _{-{1\over 2},\varOmega }\). Note that from (3.6) and (3.7) we have
where \(I_h\psi _g\) is the linear Lagrange interpolation of \(\psi _g\) [9]. Then standard error estimates lead to
where we have used the estimate (3.8). This implies
which proves (3.10).\(\square \)
The semi-discrete finite element approximation of (1.2) reads: Find \(y_h(u)\in L^2(V^h)\) such that
with \(v_h(\cdot , T)=0, \,y_0^h=\tilde{Q}_hy_0\in V^h\) an approximation of \(y_0\) using the \(L^2\)-projection, and \(Q_h\) the projection operator from \(L^2(\varGamma )\) to \(V^h(\varGamma )\). Note that the above semi-discrete scheme is well-defined and admits a unique solution \(y_h(u)\in L^2(H^1(\varOmega ))\), which we denote by \(y_h(u)=\mathcal {G}_h(u)\), since \(Q_h(u)\in L^2(H^{1\over 2}(\varGamma ))\), thus we use a standard bilinear form \(a(\cdot ,\cdot )\) compared to the very weak form (2.2).
The semi-discrete finite element approximation of (1.1)–(1.2) reads as follows:
where \(y_0^h\in V^h\) is an approximation of \(y_0\), and \(U_{ad}^{h}\) is an appropriate approximation to \(U_{ad}\) depending on the discretization scheme for the control.
It follows that the control problem (3.14) has a unique solution \((y_h,u_h)\) and that a pair \((y_h,u_h)\) is the solution of the problem (3.14) if and only if there is a co-state \(z_h\in L^2(V^h_0)\) such that the triplet \((y_h,z_h,u_h)\) satisfies (3.13) and the following optimality conditions:
where \(y_h(v_h)\in L^2(V^h)\) is the solution of state equation (3.13) with Dirichlet boundary condition \(Q_h(v_h)\).
We next consider the fully discrete approximation for above semi-discrete problem by using the dG(0) scheme in time [33]. We note that the dG(0) scheme is equivalent to the backward Euler method with the right hand side approximated by the averaged integral.
Let \(0=t_0<t_1<\cdots <t_{N-1}<t_N=T\) be a time domain partitioning with \(k_n=t_n-t_{n-1}, \,n=1,2,\ldots ,N\) and \(k=\max \limits _{1\le n\le N}k_n\). We assume that the time partitioning is quasi-uniform, i.e., there exist positive constants \(c_1\) and \(c_2\) such that \(c_1k_n\le k\le c_2k_n\) holds for each \(n=1,2,\ldots ,N\). We also set \(I_n:=(t_{n-1},t_n]\). For \(n=1,2,\ldots ,N\), we construct the finite element spaces \(V^h\subset H^1(\varOmega )\) with the mesh \(\mathcal {T}^h\). Similarly, we construct the finite element spaces \(U^h\subset L^2(\varGamma )\) with the mesh \(\mathcal {T}^h_U\). In our case we have \(U^h=V^h(\varGamma )\). Then we denote by \(V^h\) and \(U^h\) the finite element spaces defined on \(\mathcal {T}^h\) and \(\mathcal {T}^h_U\) on each time step.
Let \(V_k\) denote the space of piecewise constant functions on the time partition. We define the \(L^2\) projection operator \(P_k:L^2(0,T)\rightarrow V_k\) on \(I_n\) through
Then we have the following estimate
where H denotes some separable Hilbert space .
We consider a dG(0) scheme for the time discretization and set
i.e. \(\phi \in V_{hk}\) is a piecewise constant polynomial w.r.t. time. We also set \(V_{hk}(\varGamma )\) as the restriction of \(V_{hk}\) on \(L^2(L^2(\varGamma ))\). We set \(Q=Q_hP_k=P_kQ_h\). Thus, we have \(Q:L^2(L^2(\varGamma ))\rightarrow V_{hk}(\varGamma )\). For \(Y,\varPhi \in V_{hk}\) we set
where \(\varPhi ^n:=\varPhi ^n_-=\lim _{s\rightarrow 0^+}\varPhi (t_n-s), \,\varPhi ^{n+1}:=\varPhi ^n_{+}=\lim _{s\rightarrow 0^+}\varPhi (t_n+s)\).
For each \(u\in L^2(L^2(\varGamma ))\) the fully discrete dG(0)-cG(1) finite element approximation of (3.13) now reads: Find \(Y_{hk}\in V_{hk}\) such that
where \(V_{hk}^0\) denotes the subspace of \(V_{hk}\) with functions vanishing on the boundary \(\varGamma \).
It is easy to see that on each time interval \(I_n, \,Y_{hk}^n\in V^h\) solves the following problem:
Here we use the \(L^2\)-projection to approximate the non-smooth Dirichlet boundary condition in (3.18).
In the following we need to investigate the stability behavior of the fully discrete scheme (3.18) with respect to the initial value \(y_0\), the right hand side f and the Dirichlet boundary conditions u.
Lemma 3
There exists a constant C independent of h, k and the data \((f,y_0)\) such that
and
hold in case \(f\equiv 0, \,y_0\equiv 0\). In the case \(u\equiv 0\) the estimate
is valid.
Proof
Let us first assume that \(f\equiv 0, \,y_0\equiv 0\). The proof follows the idea of [12] completely and here we give a sketch for the case with variable time steps. To begin with we introduce the following problem: Find \(y_u\in V_{hk}\) with
For arbitrary \(y_h\in V^h\) we have the splitting
where \(\tilde{Q}_hy_h\in V_0^h\) and \(R_hy_h\in V_0^h\) are the \(L^2\)-projection and Ritz-projection of \(y_h\), respectively. Then we have \(y_2|_{\varGamma }=y_h, \,y_1|_{\varGamma }=y_h\) and
Let \(y^n_u=y_2^n+\tilde{Q}_hy_u^n\). Then (3.23) delivers
Similar to the proof of Proposition 1 in [12, P. 88] we conclude from (3.24) that
For \(y^n_u\in V^h\) we also have the splitting \(y^n_u=y_1^n+R_hy_u^n\). It follows from the proof of Lemma 3 in [12, P. 87] that
Similarly, we also have
We note that \(y_1^n|_{\varGamma }=y^n_u=Q_h(P_k^nu)\in H^{1\over 2}(\varGamma )\) and
Let \(w^n\in H^1(\varOmega )\) be the solution of (3.6) with \(v_h\) substituted by \(Q_h(P_k^nu)\). Then
and \(y_1^n\) is the finite element approximation to \(w^n\). So we deduce from Lemma 2 that
and
which in turn give
Inverse estimates also yield
With the help of above estimates and norm interpolation we are led to
for all \(0\le s\le 1\) and \(n=1,2,\ldots ,N\). Thus, from the quasi-uniformality of time partioning we have
This gives
Similarly, we can derive from (3.26) and the \(W^{s,q}(\varGamma )\) (\(0\le s\le 1\) and \(1\le q\le \infty \)) stability of \(L^2\)-projection operator \(Q_h\) (see [7, P. 1601]) that
Combining (3.28) and (3.29) we prove the case of \(f\equiv 0\) and \(y_0\equiv 0\) with \(Y_{hk}^n=y_u^n\).
For the case \(u\equiv 0\), let \(y_f\in L^2(H_0^1(\varOmega ))\cap H^{1}(H^{-1}(\varOmega ))\) be the solution of following problem
Then we have
Let \(y_f^n\in V^h, \,n=1,2,\ldots , N\) be the solutions of following problems:
Then \( y_f^n\) is the standard fully discrete approximation of \(y_f\). Let \(w_h=k_n(y^n_f- y_f^{n-1})\) in (3.32) we get
thus we have
Summing the above equations over n from 1 to N we obtain
where we used the estimate \(\Vert \tilde{Q}_hy_0\Vert ^2_{1,\varOmega }\le Ch^{-2}\Vert \tilde{Q}_hy_0\Vert ^2_{0,\varOmega }\le Ch^{-2}\Vert y_0\Vert ^2_{0,\varOmega }\). This proves the case \(u\equiv 0\) with \(Y_{hk}^n=y_f^n\). \(\square \)
We next consider the fully discrete approximation for above semi-discrete optimal control problems by using the dG(0) scheme in time. The fully discrete approximation scheme of (3.14) is to find \((Y_{hk},U_{hk})\in V_{hk}\times U_{ad}^{hk}\), such that
subject to
Here \(U_{ad}^{hk}\) is an appropriate approximation to \(U_{ad}\). We set \(U_{ad}^{hk}= V_{hk}(\varGamma )\cap U_{ad}\) for the full discretization of the control problem (1.1)–(1.2) and \(U_{ad}^{hk}\equiv U_{ad}\) for its variational discretization.
It follows from standard arguments (see [24]) that the above control problem has a unique solution \((Y_{hk},U_{hk})\), and that a pair \((Y_{hk},U_{hk})\in V_{hk}\times U_{ad}^{hk}\) is the solution of (3.34)–(3.35) if and only if there is a co-state \(Z_{hk}\in V_{hk}^0\), such that the triplet \((Y_{hk},Z_{hk},U_{hk})\in V_{hk}\times V_{hk}^0\times U_{ad}^{hk}\) satisfies (3.35) and the following optimality conditions:
where \(Y_{hk}(v_{hk})\) is the solution of problem (3.35) with Dirichlet boundary conditions \(Q(v_{hk})\).
To derive an expression for the derivative of \(J_{hk}:L^2(L^2(\varGamma ))\rightarrow \mathbb {R}\) analogous to the one of J given by formula (2.7) we have to define a discrete normal derivative \(\partial ^{hk}_nZ_{hk}\in V_{hk}(\varGamma )\) satisfying
It is easy to verify that the linear form
is well defined on \(V_{hk}(\varGamma )\) and is also continuous. Thus from Riesz representation theorem the equation (3.38) admits a unique solution \(\partial ^{hk}_nZ_{hk}\) in \(V_{hk}(\varGamma )\). For an analogous reconstruction of discrete normal derivatives for elliptic Dirichlet boundary control problems we refer to [7]. With the help of (3.38) it is not difficult to show that
for \(v_{hk}\in U_{ad}^{hk}\), which in turn implies
where \(P_{U_{ad}^{hk}}:L^2(L^2(\varGamma ))\rightarrow U_{ad}^{hk}\) denotes the orthogonal projection in \(L^2(L^2(\varGamma ))\) onto \(U_{ad}^{hk}\).
4 Error Estimates for the Optimal Control Problems
As a preliminary result we first estimate the error introduced by the discretization of the state equation, i.e., the error between the solutions of problems (2.2) and (3.18).
Theorem 3
Let \(y\in L^2(L^2(\varOmega ))\) and \(Y_{hk}(u)\in V_{hk}\) with \(Y_{hk}(u)|_{\varSigma }=Q(u)\) be the solutions of problems (2.2) and (3.18), respectively. Then for \(u\in L^2(L^2(\varGamma ))\) we have
and for \(u\in L^2(H^{1\over 2}(\varGamma ))\cap H^{1\over 4}(L^2(\varGamma ))\) we have
Proof
In view of the linearity of the problem it is sufficient to consider the problems with either \(f\equiv 0, \,y_0\equiv 0\) or \(u\equiv 0\).
Let us first assume that \(f\equiv 0, \,y_0\equiv 0\) and \(u\in L^2(L^2(\varGamma ))\). We first note that according to [12] \(y\in L^2(H^{1\over 2}(\varOmega ))\) holds. Let \(w\in L^2(H^2(\varOmega )\cap H_0^1(\varOmega ))\cap H^1(L^2(\varOmega ))\) be the solution of problem (2.12) with right hand side \(g=y- Y_{hk}(u)\). Since \(w(T)=0\), we from (2.2) and (3.18) deduce that
We treat \(E_1\) by exploiting the properties of \(P_k\) and \(Q_h\):
From the Young’s inequality, the trace inequality and a norm interpolation inequality we derive (see, e.g. [12])
Setting \(\epsilon = k^{-\frac{1}{2}}\) and using the approximation property (3.17) of \(P_k\) gives
We also have
Using the Cauchy-Schwarz inequality and stability results for \(Q_h\) and \(P_k\) we estimate
Next we estimate \(E_2\). Considering (3.18) and \(w^N=w(\cdot ,T)=0\) we calculate
By the Cauchy-Schwarz inequality we have
where
and
In view of the stability result (3.20) of Lemma 3 we have
It remains to estimate \(F_2\). To begin with we note that
and
It is straightforward to show that
and
Using the stability estimate (3.20) of Lemma 3 we conclude
From the estimates (4.3)–(4.12) we conclude the desired result in the case \(f\equiv 0, \,y_0\equiv 0\) and \(u\in L^2(L^2(\varGamma ))\) that
If \(u\in L^2(H^{1\over 2}(\varGamma ))\cap H^{1\over 4}(L^2(\varGamma ))\) we can estimate \(E_1\) as
Combining the estimate (4.11) of \(F_2\) and the stability result (3.21) in Lemma 3 we are led to
This combining with (4.3) and (4.14) gives
If \(u\equiv 0, \,f\in L^2(L^2(\varOmega ))\) and \(y_0\in L^2(\varOmega )\) we have \(y\in L^2(H_0^1(\varOmega ))\cap H^1(H^{-1}(\varOmega ))\) (see, e.g. [25]). Then similar to the above error estimate and using the stability estimate (3.22) of Lemma 3, it is straightforward to prove that (see also [12])
Actually, by using the duality argument it follows from (4.3) and (4.5) that
where we used the fact that the second term in (4.5) vanishes because of \(Y_{hk}^n(u)\in V_0^h\). Note that
It follows from (3.22), (4.7), (4.9), (4.10) that
this together with (4.18), (4.19) gives (4.17). Combining both cases we complete the proof. \(\square \)
Now we are in a position to derive our main result of this section: the a priori error estimates for optimal control problems. At first we consider the fully discrete case, i.e., \(U_{ad}^{hk}=U_{ad}\cap V_{hk}(\varGamma )\).
Theorem 4
Let \((y,u,z)\in {L^2(L^2(\varOmega ))}\times {L^2(L^2(\varGamma ))}\times {L^2(H^2(\varOmega ))}\cap H^1(L^2(\varOmega ))\) and \((Y_{hk},U_{hk},Z_{hk})\in V_{hk}\times U_{ad}^{hk}\times V_{hk}^0\) be the solutions of problem (2.5)–(2.8) and (3.34)–(3.35) with \(U_{ad}^{hk}=U_{ad}\cap V_{hk}(\varGamma )\), respectively. Then we have the a priori error estimate
with a constant \(C>0\) independent of h and k.
Proof
Let us recall the continuous and discrete optimality conditions
and
Setting \(v=U_{hk}\in U_{ad}\) and \(v_{hk}=Q(u)\in U_{ad}^{hk}\) we have
where \(y(U_{hk})\in L^2(L^2(\varOmega ))\) with \(y(U_{hk})|_{\varSigma }= U_{hk}\) solves
with \(v(\cdot ,T)=0\), and \(Y_{hk}(Qu)\in V_{hk}\) solves
With Young’s inequality we deduce
Taking \(\sigma >0\) small enough, we from (4.23)–(4.26) obtain
Note that from the standard error estimates for the \(L^2\)-projection and the regularity of u we have
Thus we are led to
Since \(Y_{hk}(Qu)\) is the fully discrete finite element approximation of y, the error estimate (4.2) of Theorem 3 gives
Then it remains to estimate \(I_3\). From (2.2), (2.6), (4.24) and (4.25) we have
Note that
It is straightforward to estimate
Define \(E_{hk}:=Y_{hk}- Y_{hk}(Qu)\). Using the proof technique of Theorem 3 we from (3.35), (4.25) obtain
With the help of projection error estimate and proceeding as in the estimate of (4.5) we have
From Lemma 3 we conclude
Since the projection operator \(P_{U_{ad}}\) is continuous on \(L^2(H^{\frac{1}{2}}(\varGamma ))\) and \(H^{\frac{1}{4}}(L^2(\varGamma ))\) (see, e.g. [23, Lm. 3.3]), we have from Theorem 2 that
From standard regularity result for parabolic equation [25] and Lemma 1 we have
Combining above results and using the Cauchy-Schwarz inequality and Young’s inequality completes the proof. \(\square \)
If we use variational discretization concept introduced in [19], i.e., \(U_{ad}^{hk}=U_{ad}\), we can prove the following error estimates in a similar way.
Theorem 5
Let \((y,u,z)\in {L^2(L^2(\varOmega ))}\times {L^2(L^2(\varGamma ))}\times {L^2(H^2(\varOmega ))}\cap H^1(L^2(\varOmega ))\) and \((Y_{hk},U_{hk},Z_{hk})\in V_{hk}\times U_{ad}\times V_{hk}^0\) be the solutions of problem (2.5)–(2.8), and (3.34)–(3.35) with \(U_{ad}^{hk}=U_{ad}\), respectively. Then we have following a priori error estimate
with a constant \(C>0\) independent of h and k.
Proof
In the proof of Theorem 4 it suffices to set \(v=U_{hk}\) in (4.21) and \(v_{hk}=u\) in (4.22) and add the corresponding inequalities. This directly gives
Thus
The rest of proof is along the lines of the estimation of the terms \(I_2\) and \(I_3\) in the proof of Theorem 4. \(\square \)
As the final result, we can conclude from Theorems 4 and 5 the explicit convergence rate with respect to h and k for the fully discrete finite element approximation of the optimal control problems under the assumption \(k=O(h^2)\).
Corollary 1
Assume that the spatial mesh size h and time step k satisfy the coupling \(k=O(h^2)\). Then we have the following a priori error estimate
for both full control discretization and variational control discretization with a constant \(C>0\) independent of h and k.
Remark 1
The error estimates we obtained in Theorems 4, 5 and Corollary 1 reflect the worst cases we can expect for parabolic Dirichlet boundary control problems defined on convex polygonal domains.
Since the regularity of parabolic equations depends on the maximum interior angle of the domain, the state admits the improved regularity \(y\in L^2(W^{1,p}(\varOmega ))\) for \(2\le p < p_{*}\) with \(p_{*}=\frac{2\omega }{2\omega -\pi }\) depending on the maximum interior angle \(\omega \) of the polygonal domain \(\varOmega \) and also the data (see [29] for more details). Moreover, for problems defined on curved domains with smooth boundary, we have higher regularity for the optimal control problems as stated in Theorem 3.4 of [23]. Improved space regularity leads to better approximation properties of the state and thus to better convergence rates for space discretization, as is reported in our numerical results. For the elliptic case with polygonal boundaries we refer to [7] where an approximation order for the controls of \(O(h^{1-\frac{1}{p}})\) is derived for some \(2<p\le p_{*}\) with \(p_{*}\) depending on the data and the maximum interior angle of the domain. For the error estimates of elliptic Dirichlet boundary control problems defined on curved domains we refer to [8], [10] and [16] for more details.
Note that \(y\in L^2(H^1(\varOmega ))\cap H^{1\over 2}(L^2(\varOmega ))\), if in addition \(y_d\in L^2(H^1(\varOmega ))\cap H^{1\over 2}(L^2(\varOmega ))\) we may derive that \(z\in L^2(H^3(\varOmega ))\cap H^{3\over 2}(L^2(\varOmega ))\) and \(\partial _nz\in H^{3\over 4}(L^2(\varGamma ))\) under appropriate assumption on the domain \(\varOmega \), and thus \(u\in H^{1\over 2}(L^2(\varGamma ))\) (see the proof of Theorem 2 and [23]). This improved time regularity may deliver higher order convergence for time discretization, compared to the estimates (4.29) and (4.32). This may explain the higher order convergence for the time discretization observed in the numerical experiments.
5 Numerical Experiments
In this section we will carry out some numerical experiments to support our theoretical findings. We consider the optimal control problems (1.1)–(1.2) of tracking type with control set \(U_{ad}\) defined as follows
In the numerical experiments we illustrate the convergence orders with respect to the spatial and time discretizations separately by setting h and k small enough respectively, although we derived a priori error estimate with coupling \(k=O(h^2)\). The numerical tests indicate that such a coupling of k and h seems not to be needed. We expect that an according analysis is possible with adapting the techniques of [30] and [31] to the present setting.
Although we do not consider problems defined on curved domains in our numerical analysis, we include some numerical examples on both polygonal and curved domains using full discretization and variational discretisation of the control. For the numerical approximations of Dirichlet boundary control problems defined on curved domains we refer to [8, 10] and [16]. We use \(\Vert \cdot \Vert _{L^2}\) to denote the \(L^2(L^2(\varGamma ))\)-norm error for the optimal control u and the \(L^2(L^2(\varOmega ))\)-norm errors for the state y and adjoint state z.
Example 1
The first example is a unconstrained problem defined on the unit square \(\varOmega =[0,1]\times [0,1], \,T=1\). The data is chosen as
with \(\alpha =1\), so that the optimal solution is given by
At first we consider the error with respect to spatial mesh size h. We fix the time step to \(k= \frac{1}{8192}\) and present the errors of optimal control u, state y and adjoint state z in Table 1. Then we consider the convergence order of error with respect to time step size k. We fixed the space mesh with \(DOF= 22785\) and present the errors of optimal control u, state y and adjoint state z in Table 2. We observe an order of convergence \(\frac{3}{2}\) for the control and order 2 for the state and adjoint state for spatial discretization, and order 1 convergence for both of them for the time discretization. This is the best result we can expect for linear finite elements and dG(0) approximations.
Example 2
The second example is an unconstrained problem defined in a polygonal domain with maxminum interior angle \(\omega _{max}={5\over 6}\pi \)(see [29]), so that the optimal solution may have only reduced regularity. The data is chosen as
with \(f=1, \,g=\frac{1}{(x_1^2+x_2^2)^{\frac{1}{3}}}\).
There is no exact solution for this example. We take the solution with \(k=\frac{1}{4096}\) and \(Dof=158561\) in the spatial discretization as reference solution. Similarly as in the previous example, we investigate the convergence order with respect to the spatial and time discretization separately. Although the assumption \(k=O(h^2)\) is not satisfied in this example, the analysis and numerical results in [30] and [31] suggest \(O(h^{1\over 2})\) convergence for spatial discretization and \(O(k^{1\over 4})\) convergence for time discretization in our case. We can observe in Table 3 nearly \(O(h^{1\over 2})\)-order convergence for the spatial discretization of the control. The convergence order for the time discretization reported in Table 4 is higher than \(O(k^{1\over 4})\) which might be caused by a higher regularity of the control w.r.t time, compare Remark 1 and [23, Th. 3.4]. Caused by \(y_d\) we may expect a regularity loss w.r.t. time at \(t=0.5\), which in our opinion might only have a mild influence on the convergence order of the numerical computations.
Example 3
This example is a control constrained problem defined in a smooth domain (see [10]). The domain is the unit circle \(\varOmega =B(0,1)\) with center at zero and radius 1, \(T=1\). The data is presented in polar coordinates. We set
so that the optimal solution is given by
We set \(\alpha =1\).
First we consider the error with respect to spatial mesh size h. We fix the time step \(k= \frac{1}{4096}\) and present the error of the optimal control u, the state y and the adjoint state z in Table 5 with full discretisation, and in Table 6 with variational discretisation. We as expected observe that both approaches deliver similar results. Then we consider the convergence order of the time error. We fix the space mesh with \(DOF= 16641\) and present the error of the optimal control u, the state y and the adjoint state z in Table 7. We observe higher order convergence w.r.t. the spatial discretization for both the control u and the state y.
References
Apel, T., Flaig, T.G.: Crank–Nicolson schemes for optimal control problems with evolution equations. SIAM J. Numer. Anal. 50, 1484–1512 (2012)
Arada, N., Raymond, J.P.: Dirichlet boundary control of semilinear parabolic equations: part I. Appl. Math. Optim. 45, 125–143 (2002)
Belgacem, F.B., Bernardi, C., Fekih, H.E.: Dirichlet boundary control for a parabolic equation with a final observation I: a space–time mixed formulation and penalization. Asymptot Anal. 71, 101–121 (2011)
Berggren, M.: Approximation of very weak solutions to boundary value problems. SIAM J. Numer. Anal. 42, 860–877 (2004)
Bramble, J.H., Pasciak, J.E., Schatz, A.H.: The construction of preconditioners for elliptic problems by substructing I. Math. Comput. 47, 103–134 (1986)
Casas, E., Mateos, M., Raymond, J.P.: Penalization of Dirichlet optimal control problems. ESAIM Control Optim. Calc. Var. 15, 782–809 (2009)
Casas, E., Raymond, J.P.: Error estimates for the numerical approximation of Dirichlet boundary control for semilinear elliptic equations. SIAM J. Control Optim. 45, 1586–1611 (2006)
Casas, E., Sokolowski, J.: Approximation of boundary control problems on curved domains. SIAM J. Control Optim. 48, 3746–3780 (2010)
Ciarlet, P.G.: The Finite Element Methods for Elliptic Problems. Elsevier, North-Holland (1978)
Deckelnick, K., Günther, A., Hinze, M.: Finite element approximation of Dirichlet boundary control for elliptic PDEs on two- and three-dimensional curved domains. SIAM J. Control Optim. 48, 2798–2819 (2009)
French, D.A., King, J.T.: Approximation of an elliptic control problem by the finite element method. Numer. Funct. Anal. Optim. 12, 299–314 (1991)
French, D.A., King, J.T.: Analysis of a robust finite element approximation for a parabolic equation with rough boundary data. Math. Comput. 60, 79–104 (1993)
Fursikov, A.V., Gunzburger, M.D., Hou, L.S.: Boundary value problems and optimal boundary control for the Navier-Stokes systems: the two-dimensional case. SIAM J. Control Optim. 36, 852–894 (1998)
Geveci, T.: On the approximation of the solution of an optimal control problem governed by an elliptic equation. RAIRO Anal. Numer. 13, 313–328 (1979)
Gong, W., Yan, N.N.: A posteriori error estimate for boundary control problems governed by the parabolic partial differential equations. J. Comput. Math. 27, 68–88 (2009)
Gong, W., Yan, N.N.: Mixed finite element method for Dirichlet boundary control problem governed by elliptic PDEs. SIAM J. Control Optim. 49, 984–1014 (2011)
Grisvard, P.: Singularities in Boundary Value Problems. Springer, Berlin (1992)
Gunzburger, M.D., Hou, L.S.: Treating inhomogeneous essential boundary conditions in finite element methods and the calculation of boundary stresses. SIAM J. Numer. Anal. 29, 390–424 (1992)
Hinze, M.: A variational discretization concept in control constrained optimization: the linear-quadratic case. Comput. Optim. Appl. 30, 45–63 (2005)
Hinze, M., Kunisch, K.: Second order methods for boundary control of the instationary Navier–Stokes system. ZAMM Z. Angew. Math. Mech. 84, 171–187 (2004)
Hinze, M., Matthes, U.: A note on variational discretization of Neumann boundary control problems. Control Cybern. 38, 577–591 (2009)
Hinze, M., Pinnau, R., Ulbrich, M., Ulbrich, S.: Optimization with PDE Constraints, Mathematical Modelling: Theory and Applications, vol. 23. Springer, Berlin (2009)
Kunisch, K., Vexler, B.: Constrained Dirichlet boundary control in \(L^2\) for a class of evolution equations. SIAM J. Control Optim. 46, 1726–1753 (2007)
Lions, J.L.: Optimal Control of Systems Governed by Partial Differential Equations. Springer, Berlin (1971)
Lions, J.L., Magenes, E.: Non-homogeneous Boundary Value Problems and Applications, vol. I, II. Springer, Berlin (1972)
Liu, W.B., Ma, H.P., Tang, T., Yan, N.N.: A posteriori error estimates for discontinuous Galerkin time-stepping method for optimal control problems governed by parabolic equations. SIAM J. Numer. Anal. 42, 1032–1061 (2004)
Liu, W.B., Yan, N.N.: A posteriori error estimates for optimal control problems governed by parabolic equations. Numer. Math. 93, 497–521 (2003)
Liu, W.B., Yan, N.N.: Adaptive Finite Element Methods for Optimal Control Governed by PDEs. Science press, Beijing (2008)
May, S., Rannacher, R., Vexler, B.: Error analysis fo a finite element approximation of elliptic Dirichlet boundary control problems. SIAM J. Control Optim. 51(3), 2585–2611 (2013)
Meidner, D., Vexler, B.: A priori error estimates for space–time finite element discretization of parabolic optimal control problems. Part I: problems without control constraints. SIAM J. Control Optim. 47(3), 1150–1177 (2008)
Meidner, D., Vexler, B.: A priori error estimates for space–time finite element discretization of parabolic optimal control problems. Part II: problems with control constraints. SIAM J. Control Optim. 47(3), 1301–1329 (2008)
Meidner, D., Vexler, B.: A priori error analysis of the Petrov-Galerkin Crank–Nicolson scheme for parabolic optimal control problems. SIAM J. Control Optim. 49(5), 2183–2211 (2011)
Thomée, V.: Galerkin Finite Element Methods for Parabolic Problems. Springer, Berlin (2006)
Vexler, B.: Finite element approximation of elliptic Dirichlet optimal control problems. Numer. Funct. Anal. Optim. 28, 957–973 (2007)
von Daniels, N., Hinze, M., Vierling, M.: Crank-Nicolson time stepping and variational discretization of control-constrained parabolic optimal control problems. SIAM J. Control Optim. 53(3), 1182–1198 (2015)
Winther, R.: Error estimates for a Galerkin approximation of a parabolic control problem. Ann. Math. Pura App. 117(4), 173–206 (1978)
Acknowledgments
The first author would like to thank the Alexander von Humboldt Foundation for the support during the stay in University of Hamburg, Germany where this work was initialized. This work was supported by the National Basic Research Program of China under Grant 2012CB821204, the National Natural Science Foundation of China under Grant 11201464 and 91330115, and the scientific Research Foundation for the Returned Overseas Chinese Scholars, State Education Ministry. The second author gratefully acknowledges the support of the DFG Priority Program 1253 entitled “Optimization with Partial Differential Equations”. The third author was supported by National Natural Science Foundation of China under Grant 11301311. The authors also would like to thank two anonymous referees for their valuable suggestions which lead to an improved paper.
Author information
Authors and Affiliations
Corresponding author
Rights and permissions
About this article
Cite this article
Gong, W., Hinze, M. & Zhou, Z. Finite Element Method and A Priori Error Estimates for Dirichlet Boundary Control Problems Governed by Parabolic PDEs. J Sci Comput 66, 941–967 (2016). https://doi.org/10.1007/s10915-015-0051-2
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10915-015-0051-2
Keywords
- Optimal control problem
- Parabolic equation
- Finite element method
- A priori error estimate
- Dirichlet boundary control