An energy space finite element approach for elliptic Dirichlet boundary control problems

In this paper we present a finite element analysis for a Dirichlet boundary control problem where the Dirichlet control is considered in a convex closed subspace of the energy space H1/2(Γ)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H^{1/2}(\Gamma )$$\end{document}. As an equivalent norm in H1/2(Γ)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H^{1/2}(\Gamma )$$\end{document} we use the energy norm induced by the so-called Steklov–Poincaré operator which realizes the Dirichlet to Neumann map, and which can be implemented by using standard finite element methods. The presented stability and error analysis of the discretization of the resulting variational inequality is based on the mapping properties of the solution operators related to the primal and adjoint boundary value problems, and their finite element approximations. Some numerical results are given, which confirm on one hand the theoretical estimates, but on the other hand indicate the differences when modelling the control in L2(Γ)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L_2(\Gamma )$$\end{document}.


Introduction
In this paper, the focus is on the a priori error analysis of the finite element approximation to minimise the cost functional where the state u is the unique solution of a second order elliptic partial differential equation in a polygonal bounded domain , satisfying a Dirichlet boundary condition u = z on = ∂ . The non-negative cost functional F(u) describes, e.g., some physical quantity to be minimised, while z 2 V reflects either the costs of the control z ∈ U ⊂ V, or represents some regularization when considering, for example, an inverse source problem. Note that V is an appropriate Hilbert space to be specified, and U ⊂ V is a convex closed subset which describes additional constraints on the control.
For a general analysis on optimal control problems governed by partial differential equations we refer to, e.g., [28,41], for numerical methods to solve related variational inequalities, see, e.g., [13], and for a more recent overview on the theory and numerics of optimal control problems, see, for example [16].
Optimal control problems (1.1) with a Dirichlet boundary control play an important role, for example, in the context of computational fluid mechanics., see, e.g., [12,15], and the references given therein. In [15], the cost functional J (u, z) = F(u) is the domain integral over the strain tensor of the velocity field u satisfying the steady state Navier-Stokes equations with the Dirichlet boundary condition u = z ∈ U ⊂ H 1/2 ( ). A similar minimisation problem is considered in [12], where the cost functional J (u, z) is either written as boundary integral to describe the work needed to overcome the drag exerted on a given body, or equivalently as a domain integral over the strain of the velocity, and some additional boundary integral. In both cases, the cost functional J (u, z) describes an energy in H 1 ( ), or equivalently, in the Sobolev trace space H 1/2 ( ). Hence, when considering the minimisation problem (1.1) subject to a Dirichlet boundary value problem, V = H 1/2 ( ) appears as a natural choice [3]. To obtain smoother optimal solutions one may even consider more regular cost functionals: In [21], V = H 2 ( ) is considered where the norm · H 2 ( ) is realized by using the Laplace-Beltrami operator on the boundary . Note that such an approach requires sufficient regularity of the domain which is assumed to be of the class C 2,1 . In [28], several Neumann and Dirichlet boundary control problems with observations in the domain and on the boundary are considered from an analytic point of view. To describe the involved Sobolev norms on the boundary , in particular H −1/2 ( ) and H 1/2 ( ), certain fractional powers of the Laplace-Beltrami operator are used which seems to be complicated from a numerical point of view. The situation simplifies when U = L 2 ( ) is used as control space. But in this case, the associated partial differential equation has to be considered within an ultraweak variational formulation, see, for example [28], and [4] for an appropriate finite element approximation using standard piecewise linear basis functions. The use of the ultra-weak variational formulation of the primal Dirichlet boundary value problem in the context of an optimal control problem requires the adjoint variable p to be sufficiently regular, i.e., p ∈ H 2 ( ) ∩ H 1 0 ( ). Since the adjoint variable p itself is the unique solution of the adjoint partial differential equation with homogeneous Dirichlet boundary conditions, either a smooth boundary , or a polygonal or polyhedral but convex domain has to be assumed. For related finite element approximations, see, e.g., [8,10,14,26,30], or [44] in the case of a finite dimensional Dirichlet control. Several variational formulations of Dirichlet control problems are discussed in [25]. To include a Dirichlet boundary condition u = z ∈ L 2 ( ) in a standard variational formulation, alternatively one may consider a penalty approximation of the Dirichlet boundary condition by using a Robin boundary condition, see, e.g., [3,17,19,20]. Again, sufficient smoothness of the boundary has to be assumed.
The aim of this paper is to present a numerical analysis of an energy space finite element approach when the control z is considered as an element of the boundary trace space H 1/2 ( ), where an equivalent norm is induced by the so-called Steklov-Poincaré operator which realizes the Dirichlet to Neumann map. Note that in this case the costs represent the energy of the harmonic extension of the Dirichlet control z. The main difference to the more common approach when using L 2 ( ) as control space appears in the optimality condition. In particular, the Steklov-Poincaré operator links the Dirichlet control to the normal derivative of the adjoint variable. The related optimality condition results then in a higher regularity of the control, and requires less assumptions on the smoothness of the adjoint variable. Indeed, the use of H 1/2 ( ) as control space reflects the proper mapping properties of the Dirichlet to Neumann map which appears within the optimality condition. As a consequence, we also obtain higher order convergence results for the approximate finite element solution. Instead, when using L 2 ( ) as control space, Dirichlet and Neumann boundary data are identified with each other. In particular for polygonal and polyhedral domains the control z will then be zero in all corner points which seems to be not motivated from the application in mind.
The paper is organised as follows: In Sect. 2, we describe the considered Dirichlet boundary control problem, the primal boundary value problem, and the reduced cost functional as well as the related adjoint boundary value problem. The minimiser of the reduced cost functional is characterised as the unique solution of a variational inequality of the first kind. The finite element discretization of the variational inequality is described in Sect. 3, where also finite element approximations of both the primal and adjoint boundary value problems and related error estimates are given. These approximations are used to describe and to analyse the finite element discretization of the perturbed variational inequality. The main results of this paper are the error estimates as given in Corollary 3.8. Some numerical results are finally given in Sect. 4, where we also give a comparison with the more common approach when considering the control in L 2 ( ).
For the ease of presentation we restrict our considerations to the case of a convex and polygonal bounded domain. However, the presented approach applies to general two-and three-dimensional Lipschitz domains when taking into account related results on the regularity of the solutions and on appropriate finite element approximations.
For an overview on the used Sobolev spaces in the domain and on the boundary, see, for example, [1,31,37,40].

Dirichlet control problems
Let ⊂ R 2 be a convex polygonal bounded domain with boundary = ∂ . As a model problem, we consider the Dirichlet boundary control problem to minimise the cost functional and where the control satisfies the box constraints We assume f ∈ L 2 ( ), u ∈ L 2 ( ), ∈ R + , g a , g b ∈ H 1/2 ( ), and g a < g b on .
Note that higher regularity will be required later to ensure appropriate finite element error estimates. The choice V = H 1/2 ( ) ensures the well-posedness of the standard variational formulation of the Dirichlet boundary value problem (2.2). There are several possibilities to describe a (semi-)norm in H 1/2 ( ), e.g., by using the Sobolev-Slobodeckii semi-norm other realizations rely on the use of Fourier or multilevel representations which imply equivalent (semi-)norms in H 1/2 ( ). However, since H 1/2 ( ) is the boundary trace of H 1 ( ), we may also consider the equivalent Dirichlet trace norm In particular, for any z ∈ H 1/2 ( ) there exists the harmonic extension u z ∈ H 1 ( ) as the unique solution of the homogeneous Dirichlet boundary value problem Using Green's first formula, this motivates the use of the semi-norm which describes the energy of the harmonic extension u z ∈ H 1 ( ) of the Dirichlet control z ∈ H 1/2 ( ). By introducing the Steklov-Poincaré operator S : where ·, · denotes the duality pairing of H −1/2 ( ) and H 1/2 ( ).

Primal boundary value problem
To rewrite the Dirichlet boundary control problem (2.1)-(2.3) by using a reduced cost functional, we introduce a linear solution operator describing the application of the constraint (2.2). Therefore we consider the homogeneous partial differential equation (2.4) with the control as Dirichlet boundary condition, and an inhomogeneous partial differential equation with zero Dirichlet boundary conditions to describe a particular solution. Let u f be the weak solution of the Dirichlet boundary value problem i.e., u f ∈ H 1 0 ( ) is the unique solution of the variational problem The solution of the Dirichlet boundary value problem (2.2) is given by u = u z + u f , where u z ∈ H 1 ( ) is the weak solution of the Dirichlet boundary value problem (2.4). By applying the inverse trace theorem, see, e.g., [22,31,40] where H * : where the adjoint variable p is the weak solution of the Dirichlet boundary value problem Finally, by using Green's first formula we can rewrite (2.10) for ϕ ∈ H 1/2 ( ) as

Optimality condition
To characterise the minimiser of the reduced cost functional (2.9) we introduce the self-adjoint and bounded operator satisfying and define Hence we can rewrite (2.9) as Proof For z ∈ H 1/2 ( ), let u z = Hz ∈ H 1 ( ) be the harmonic extension satisfying the Dirichlet boundary value problem (2.4). Then, by (2.5) and the assertion follows from the trace theorem.
Since U ⊂ H 1/2 ( ) is convex and closed, and since T is self-adjoint and H 1/2 ( )elliptic, the minimisation of (2.15) is equivalent to solving a variational inequality to find z ∈ U such that Since (2.16) is an elliptic variational inequality of the first kind, we can use standard arguments as given, for example in [6,13,28,29], to establish unique solvability of the variational inequality (2.16).
In what follows we will give another characterization of the solution z ∈ U of the variational inequality (2.16) as a Dirichlet trace of a solution of a bilateral Signorini boundary value problem. We introduce where p is the unique solution of the adjoint Dirichlet boundary value problem and where u z is the harmonic extension of z. Then we can rewrite the variational inequality (2.16) as Since the solution z ∈ U is uniquely determined, we introduce the active zones and the inactive zone Let us first investigate the behavior of the solution z in the vicinity of the upper bound g b . For an arbitrary test function w ∈ U with w ≤ z we define a non-negative function Recall that for a subset 0 ⊂ Hence we have for all appropriate test functions ϕ. By using a scaling argument we conclude λ ≤ 0 in the sense of H −1/2 ( \ a ). In the same way we find λ ≥ 0 in the sense of H −1/2 ( \ b ) when considering the lower bound, and we conclude i.e., λ = 0 in the sense of H −1/2 ( n ). Hence we have obtained the complementary conditions which have to be understood as above.
Finally we have that u z ∈ H 1 ( ) is the unique solution of the Signorini boundary value problem − u z = 0 in (2.18) with the bilateral constraints on  Proof For the solution of the adjoint boundary value problem (2.17) we first have Let z ∈ H 1/2 ( ) be the unique solution of the variational inequality (2.16), and let u z ∈ H 1 ( ) be the harmonic extension, which satisfies the bilateral Signorini boundary value problem (2.18) Since the given boundary data of the mixed boundary value problem (2.22) are sufficient regular, it remains to consider boundary points where the boundary condition changes from an inactive zone to an active one. In particular we investigate the existence of singularity functions in the vicinity of those points. In contrast to boundary value problems with mixed boundary conditions of Dirichlet and Neumann type, no singularity functions appear in the case of Signorini boundary conditions. Without loss of generality let x 0 ∈ a ∩ n . By introducing local polar coordinates, see Fig. 1 x the solution u z can be written in the vicinity of x 0 as To determine the coefficients A and B of the singular part we consider the Signorini type boundary conditions on a and n , respectively.
For x ∈ n , i.e. φ = 0, and n x = (0, −1) , the normal derivative of u z reads and for the Neumann boundary condition on n we obtain Since the regular functions p, u reg z ∈ H 2 ( ) do not contain any singularity, we conclude B = 0. On n we also have the complementary condition g a = u a < u z , which gives for φ = 0 a (0, 0). A first order Taylor expansion of the regular part then gives and therefore, Since the first order derivative of the regular part is bounded for all r ∈ (0, r ), we conclude A ≥ 0 when considering r → 0. For x ∈ a , i.e. φ = φ 0 , we have from which we conclude For x ∈ a the normal vector is n x = (− sin φ 0 , cos φ 0 ) . Therefore, the normal derivative of u z on a is given as The complementary Neumann condition on a then reads from which we conclude A ≤ 0. Hence, together with A ≥ 0 this gives A = 0, i.e. no singularity function occurs, in particular we have u z ∈ H 2 ( ).
Remark 2.1 Although we have given a proof of Theorem 2.2 for the particular case of a convex and polygonal bounded domain only, we may comment on more general situations as follows: 1. As explicitly shown in the proof of Theorem 2.2 the solutions of Signorini boundary value problem do not involve any singularity function. This remains true for smooth boundaries. For a more general discussion of bilateral Signorini boundary conditions, see, e.g., [2,43]. 2. In the case of additional Dirichlet or Neumann boundary conditions one may have only a reduced regularity, see, e.g., the example as given in [24, p. 617f.]. 3. Depending on the regularity of the given data one may even have higher regularity results than proven in Theorem 2.2, but even for smooth data one cannot expect more than u z ∈ H 5/2−ε ( ), i.e. z ∈ H 2−ε ( ). 4. In the case of a non-convex polygonal bounded domain we can not ensure p ∈ H 2 ( ) which would also reduce the regularity of the solution of the bilateral Signorini problem.

Finite element approximations
For the ease of presentation we consider the case where is a convex two-dimensional polygonal bounded domain. Let be some conforming finite element spaces of, e.g., piecewise linear and continuous basis functions ϕ k , which are defined with respect to some admissible domain triangulation h of mesh size h. Let be the boundary finite element space of, e.g., piecewise linear and continuous basis functions φ i which are the boundary traces of the domain basis functions ϕ M +i as given in (3.1). For continuous functions g a , g b ∈ H 1/2 ( ) ∩ C( ) we define the discrete convex set Then the Galerkin discretization of the variational inequality (2.16) is to find Remark 3.1 For a more regular solution z ∈ H 2−ε pw ( ) we obtain an almost quadratic order of convergence when considering the error in L 2 ( ), see [38].
Although the error estimates (3.4) and (3.5) seem to be optimal, the operator T as considered in the variational inequality (3.3) does not allow a practical implementation, since this would require the discretizations of the composed solution operator H * H and of the Steklov-Poincaré operator S, which are not possible in general. Hence, instead of (3.3) we need to consider a perturbed variational inequality to find z h ∈ U h such that where T and g are appropriate approximations of T and g, respectively. The following theorem presents an abstract consistency result, which will be used to analyse the finite element approximations of the Steklov-Poincaré operator S, and of the primal and adjoint boundary value problems.
where z h ∈ U h is the unique solution of the discrete variational inequality (3.3).
Proof The unique solvability of the discrete variational inequality (3.6) follows from the S 1 h ( )-ellipticity and boundedness of T . We further obtain The assertion now follows from the triangle inequality, and by using the boundedness of T and T , respectively.
It remains to define a suitable approximation T := H * H + S which is based on finite element approximations H and H * of the primal and adjoint boundary value problem as well as on the approximation S of the Steklov-Poincaré operator.

Primal boundary value problem
The application of the solution operator u z = Hz = u 0 + E z, u 0 ∈ H 1 0 ( ), is given as the unique solution of the variational problem (2.8). The finite element approximation of (2.8) is to find u 0,h ∈ S 1 h,0 ( ) such that By using Cea's lemma, see, e.g., [5] or [37,Theorem 8.1], and by using the approximation property of S 1 h,0 ( ), see, e.g., [37,Theorem 9.10], we conclude the unique solvability of the Galerkin formulation (3.8) as well as the error estimate (3.9) By applying the Aubin-Nitsche trick, see, e.g., [5] or [37,Theorem 11.1], we also obtain an error estimate in L 2 ( ), However, instead of the variational problem (3.8), we will consider a perturbed variational problem to find an approximate solution u 0,h ∈ S 1 h,0 ( ) such that h ( ) denotes a quasi-interpolation operator [35]. Since the perturbed variational problem (3.11) admits a unique solution u 0,h ∈ S 1 h,0 ( ), a finite element approximation Hz of the solution operator Hz = u 0 + E z can be defined by  [35]. The error estimate (3.13) follows from the application of the Strang lemma, see, e.g., [5] or [37,Theorem 8.2], the Aubin-Nitsche trick, see, e.g., [37,Lemma 11.3], and by using the L 2 error estimate of the linear quasi-interpolation operator Q h , see [35].
In the same way as above we may define a finite element approximation of the particular solution u f , i.e., u f,h ∈ S 1 h,0 ( ) is the unique solution of the Galerkin variational problem (3.14) in the case of a convex polygonal bounded domain .

Adjoint boundary value problem
Next we consider a finite element approximation of the adjoint solution operator τ = H * ψ as defined in (2.10), i.e., of where p ∈ H 1 0 ( ) is the unique solution of the variational problem (2.12). The finite element approximation of (2.12) is to find Again, we conclude the unique solvability of the Galerkin formulation (3.17) by means of Cea's lemma, as well as the quasi-optimal error estimate, in the case of a convex polygonal bounded domain, Now we are able to define an approximation τ = H * ψ of τ = H * ψ for ψ ∈ L 2 ( ) by
Then, H * : L 2 ( ) → H −1/2 ( ) is bounded, and there holds the error estimate Proof Both the boundedness of H * : L 2 ( ) → H −1/2 ( ) and the error estimate (3.20) follow from the finite element error estimate (3.18) when using duality arguments and the inverse trace theorem.

Approximation error estimates
By using the finite element approximations H, H * , and S as defined in (3.12), (3.19) and (3.15), respectively, we can introduce the finite element approximations to be used within the perturbed variational inequality (3.6). For the application of Theorem 3.2 we need to estimate the related approximation error.
Proof By the triangle inequality and by using the boundedness of H * : The assertion now follows from the error estimates (3.13), (3.16), and (3.20).

Lemma 3.7 Let u f,h ∈ S 1
h,0 ( ) be the finite element approximation of the particular solution u f ∈ H 2 ( ). Let p u , p u f ∈ H 2 ( ) be the weak solutions of the adjoint boundary value problem (2.11) with ψ = u and ψ = u f , respectively. Then there holds the error estimate (3.23) Proof From the triangle inequality we first have The assertion now follows from the boundedness of H * : L 2 ( ) → H −1/2 ( ), and the finite element error estimates (3.14) and (3.20). For the application of Lemma 3.5 we need to assume p u , p u f ∈ H 2 ( ) where p u and p u f are the weak solutions of the adjoint boundary value problem (2.11) with ψ = u and ψ = u f , respectively. But this assumption is satisfied in the case of a convex polygonal domain.
It remains to prove the S 1 h ( )-ellipticity of the finite element approximation T . This will be a consequence of the matrix representation of T as discussed in the next subsection. Now we are in the position to present the final error estimate for the approximate solution z h of the variational inequality (3.6).
Proof The error estimate (3.24) follows from Theorem 3.2 by using the approximation error estimates (3.22) and (3.23). By using the Aubin-Nitsche trick for the perturbed variational inequality (3.6) we obtain the error estimate (3.25).
From the error estimate (3.25) we also conclude, by applying the Aubin-Nitsche trick for the primal boundary value problem (2.2), the error estimate (3.26) Moreover, by using the Aubin-Nitsche trick for the adjoint boundary value problem (2.11) with ψ = u − u we also obtain the error estimate (3.27)

Approximate variational inequality
Now we are in a position to describe the finite element approximation of the perturbed variational inequality (3.6) to find z h ∈ U h such that

the unique solution of the Galerkin variational formulation
Let S h be the Galerkin matrix of the approximate Steklov-Poincaré operator S, i.e., for j, = 1, . . . , M . The matrix representation of the variational inequality (3.6) is then given by the discrete variational inequality Since the mass matrix is positive definite, the assertion follows.
In the particular case of a non-constrained minimisation problem, instead of the discrete variational inequality (3.30) we have to solve the linear system which is equivalent to the system Note that the coupled linear system (3.32) also results from a direct finite element approximation of the adjoint boundary value problem, of the primal boundary value problem, and of the optimality condition.

Numerical results and concluding remarks
As numerical example we consider as in [8], see also [30], the Dirichlet boundary control problem (2.1)-(2.3) for the domain = (0, and without box constraints. For the finite element discretization we introduce a uniform triangulation of on several refinement levels L with the mesh size h L = 2 −(L+2) , where for L = 0 the coarsest level is given by 4 uniform triangles. Since the minimiser of (2.1) is not known for this example, we use the finite element solutions (z h 9 , u h 9 ) on the 9th level as reference solutions.  In Table 1 we give the finite element errors for the control variable z in the L 2 ( )norm, for the primal variable u and for the adjoint state p in the L 2 ( )-norm. We observe an estimated order of convergence (eoc) close to 2 as predicted by the error estimates (3.26) and (3.27) for u and p. For the control z the observed order of convergence is larger than 3/2 which is predicted by the error estimate (3.25). So far we are not able to prove this higher order of convergence for a finite element discretization but for the boundary element discretization presented in [32] when assuming z ∈ H 2 pw ( ). In this case the second order of convergence reflects the optimal approximation property when using piecewise linear basis functions on the boundary.
For comparison we present in Table 2 the numerical results when considering the Dirichlet control problem in L 2 ( ), see also [8,30]. While we obtain a quadratic convergence behaviour for the adjoint state p with slightly larger errors than for H 1/2 ( ) control, the convergence rates for the control z and for the primal variable u change significantly. The total errors are worse by orders of magnitude than the ones in Table 1. The main reason for this different convergence behaviour lies in the optimality condition which reads for the L 2 ( ) control where P U denotes the pointwise projection onto U.  Indeed, for the given u we obtain z ∈ H 2/3 ( ) only, instead of z ∈ H 3/2 pw ( ) when considering the control in H 1/2 ( ). Moreover, from the optimality conditions (4.1) we conclude, that z is zero in all corner points due to the zero Dirichlet boundary condition of the adjoint state p. Hence we obtain a zero control z in all corner points independent of the target function u. Instead, when considering the control in H 1/2 ( ), the application of S reflects the proper mapping properties of the Dirichlet to Neumann map. This results in a more feasible Dirichlet control in corner points. For illustration we give the plots of the states u in Fig. 2 for the H 1/2 ( ) setting, and in Fig. 3 for the L 2 ( ) setting. Note that for this example we used a H 1/2 norm, which is defined by the semi-norm induced by the Steklov-Poincaré operator S plus a L 2 ( ) equivalent term, and different values of to ensure comparable values for the tracking functional u − u L 2 ( ) .
In Table 3, we present the computational times in seconds related to Table 1 and  Table 2. N denotes the number of triangles of the triangulation. The slightly higher effort for the approach of the H 1/2 ( ) semi-norm is due to a few more iteration  steps needed to solve the Schur complement system of (3.32) iteratively. Note that advanced preconditioners are not applied here. Taking into account the higher order of convergence, the additional effort is justifiable.
Recall that there are no box constraints considered in this example, i.e., we have to solve the coupled linear system (3.32), or the related Schur complement system. Since the Schur complement T ,h defines an equivalent discrete norm in H 1/2 ( ), appropriate preconditioners can be constructed, e.g. by using multilevel methods. For the solution of the discrete variational inequality (3.30) one may consider either multigrid methods, see, e.g., [18,34], or semi-smooth Newton methods, see, e.g., [23,42].
As an example with active constraints and a H 1/2 ( ) control, we consider the first example where = 1 and choose g b = 2.2 and g a inactive. Again we use the semi-norm induced by the Steklov-Poincaré operator. Figures 4 and 5 clearly show the difference of the solutions of the unconstrained and the constrained case. The constrained problem is solved by a semi-smooth Newton method, see e.g. [23]. Figure 6 shows the control of the constrained and unconstrained problem along the line x 2 = 0.
In this paper we have presented a rigorous stability and error analysis for a finite element approximation of a Dirichlet boundary control problem where the control is considered in the energy space H 1/2 ( ). For the ease of presentation we have restricted our considerations to the case of a convex two-dimensional polygonal bounded domain. However, using standard finite element approximations for partial differential equations which are formulated in domains with smooth curved boundaries, our approach can be extended to the case of two-and three-dimensional domains with smooth boundary. Moreover, the finite element analysis can be transferred also to general polygonal or polyhedral bounded domains, but then related regularity results for Signorini boundary value problems are required.  Note that for the approximation of the operator T : H 1/2 ( ) → H −1/2 ( ) as defined in (2.14) we may also consider boundary element methods which require an appropriate handling of the inhomogeneous adjoint partial differential equation in (2.11). The stability and error analysis of the resulting boundary element approach for the solution of Dirichlet control problems is discussed in [32], see also [33].
Instead of the Steklov-Poincaré operator S : H 1/2 ( ) → H −1/2 ( ) one may use any other bounded and semi-elliptic operator to realize an equivalent norm in H 1/2 ( ). A possible choice is to consider as in [32] the so-called hypersingular boundary integral operator D which does not require an inversion as in the definition of S h . But the use of the hypersingular integral operator may result in a lower regularity of the control z, and therefore in a lower order of convergence of the finite element approximation, see [32,33].