Institut für Numerische Mathematik Boundary and Finite Element Methods for Dirichlet Boundary Control Problems G. Of, T. X. Phan, O. Steinbach Institut für Numerische Mathematik, Technische Universität Graz SFB: Mathematical Optimization and Applications in Biomedical Sciences Subproject: Fast Finite and Boundary Element Methods for Optimality Systems (FEMBEM) T. X. Phan Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 1 / 29 Institut für Numerische Mathematik Outline Dirichlet boundary control problem Boundary integral equations Bi-Laplace boundary integral operators Boundary element methods Error estimates Finite element methods Numerical results T. X. Phan Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 2 / 29 Institut für Numerische Mathematik Dirichlet boundary control problem Dirichlet boundary control problem Z 1 J(u, z) = [u(x) − u(x)]2 dx + 2 to minimize 1 αkzk2A 2 for (u, z) ∈ H 1 (Ω) × H 1/2 (Γ) Ω subject to −∆u(x) = f (x) for x ∈ Ω, u(x) = z(x) for x ∈ Γ = ∂Ω where the operator A A : H 1/2 (Γ) −→ H −1/2 (Γ), kzk2A ≃ hAz, ziΓ ≃ kzk2H 1/2(Γ) e modified hypersingular operator Example: A = D e v iΓ := hDu, v iΓ + hu, 1iΓ hv , 1iΓ hDu, T. X. Phan Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 3 / 29 Institut für Numerische Mathematik Dirichlet boundary control problem Primal Dirichlet boundary value problem −∆u(x) = f (x) for x ∈ Ω, u(x) = z(x) for x ∈ Γ Adjoint Dirichlet boundary value problem −∆p(x) = u(x) − u(x) for x ∈ Ω, p(x) = 0 for x ∈ Γ Optimality condition αAz = ∂ p(x) ∂n for x ∈ Γ Laplace fundamental solution 1 − log |x − y | ∗ 2π U (x, y ) = 1 1 4π |x − y | T. X. Phan for d = 2 for d = 3 Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 4 / 29 Institut für Numerische Mathematik Boundary integral equation for primal problem Primal Dirichlet boundary value problem −∆u(x) = f (x) for x ∈ Ω, Boundary integral equation 1 I + K z − N0 f , Vt = 2 where u(x) = z(x) for x ∈ Γ t= ∂ u ∈ H −1/2 (Γ) ∂n (Vt)(x) := Z U ∗ (x, y )t(y ) dsy for x ∈ Γ, (Kz)(x) := Z ∂ ∗ U (x, y )z(y ) dsy ∂ny for x ∈ Γ, Z U ∗ (x, y )f (y ) dy for x ∈ Γ. Γ Γ (N0 f )(x) := Ω T. X. Phan Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 5 / 29 Institut für Numerische Mathematik Boundary integral equation for adjoint problem Adjoint Dirichlet boundary value problem −∆p(x) = u(x) − u(x) for x ∈ Ω, p(x) = 0 for x ∈ Γ Representation formula for x ∈ Ω Z Z ∂ ∗ p(x) = U (x, y ) p(y ) dsy + U ∗ (x, y )[u(y ) − u(y )] dy ∂n Γ Ω Boundary integral equation Vq = 1 I +K 2 where q= T. X. Phan p − N0 (u − u) ∂ p ∈ H −1/2 (Γ) ∂n Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 6 / 29 Institut für Numerische Mathematik Boundary integral equation for adjoint problem Due to p = 0 on Γ, we obtain Z Z U ∗ (x, y )q(y ) dsy = − U ∗ (x, y )[u(y ) − u(y )] dy Γ for x ∈ Γ Ω U ∗ (x, y ) = ∆x V ∗ (x, y ) = ∆y V ∗ (x, y ) Bi-Laplace fundamental solution V ∗ (x, y ) − 1 |x − y |2 (log |x − y | − 1) ∗ 8π V (x, y ) = 1 |x − y | 8π T. X. Phan for d = 2 for d = 3 Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 7 / 29 Institut für Numerische Mathematik Boundary integral equation for adjoint problem Newton potential Z Ω T. X. Phan ∗ U (x, y )u(y ) dy = Z ∆y V ∗ (x, y )u(y ) dy Ω Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 8 / 29 Institut für Numerische Mathematik Boundary integral equation for adjoint problem Newton potential Z ∗ U (x, y )u(y ) dy = Ω = Z Γ T. X. Phan Z ∆y V ∗ (x, y )u(y ) dy Ω ∂ ∗ V (x, y )u(y ) dsy − ∂ny Z Γ V ∗ (x, y ) ∂ u(y ) dsy + ∂n Z V ∗ (x, y )∆u(y ) dy Ω Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 8 / 29 Institut für Numerische Mathematik Boundary integral equation for adjoint problem Newton potential Z ∗ U (x, y )u(y ) dy = Ω = ∆y V ∗ (x, y )u(y ) dy Ω Z ∂ ∗ V (x, y )u(y ) dsy − ∂ny = Z Z V ∗ (x, y ) ∂ u(y ) dsy + ∂n Γ Γ Γ T. X. Phan Z ∂ ∗ V (x, y )z(y ) dsy − ∂ny Z V ∗ (x, y )∆u(y ) dy Ω Z Γ V ∗ (x, y )t(y ) dsy − Z V ∗ (x, y )f (y ) dy Ω Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 8 / 29 Institut für Numerische Mathematik Boundary integral equation for adjoint problem Representation formula for x ∈ Ω Z Z ∗ p(x) = U (x, y )q(y ) dsy + Γ − Z Γ ∗ ∂ ∗ V (x, y )z(y ) dsy ∂ny Γ V (x, y )t(y ) dsy − Z ∗ V (x, y )f (y ) dy − Ω Z U ∗ (x, y )u(y ) dy Ω Boundary integral equation for x ∈ Γ Vq = V1 t − K1 z + N0 u + M0 f T. X. Phan Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 9 / 29 Institut für Numerische Mathematik Boundary integral equation for adjoint problem Representation formula for x ∈ Ω Z Z ∗ p(x) = U (x, y )q(y ) dsy + Γ − Z Γ ∗ ∂ ∗ V (x, y )z(y ) dsy ∂ny Γ V (x, y )t(y ) dsy − Z ∗ V (x, y )f (y ) dy − Ω Z U ∗ (x, y )u(y ) dy Ω Boundary integral equation for x ∈ Γ Vq = V1 t − K1 z + N0 u + M0 f Normal derivative for x ∈ Γ 1 ′ q= I + K q − D1 z − K1′ t − N1 u − M1 f = αAz 2 T. X. Phan Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 9 / 29 Institut für Numerische Mathematik Boundary integral equations We end up with a system (Nonsymmetric formulation) 1 Vt = 2 I + K z − N0 f Vq = V1 t − K1 z + N0 u + M0 f αAz = q where t, q ∈ H −1/2 (Γ), z ∈ H 1/2 (Γ). The Schur complement system Tα z = V −1 N0 u + V −1 M0 f − V −1 V1 V −1 N0 f where the operator Tα is defined by 1 Tα := V −1 K1 − V −1 V1 V −1 ( I + K ) + αA 2 T. X. Phan Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 10 / 29 Institut für Numerische Mathematik Bi-Laplace boundary integral operators Consider Bi-Laplace equation −∆2 u(x) = 0 for x ∈Ω Representation formula for xe ∈ Ω Z Z ∂ ∗ ∂ U (e x , y )u(y ) dsy u(e x) = U ∗ (e x , y ) u(y ) dsy − ∂n ∂ny Γ Γ Z Z ∂ ∂ ∗ ∗ x , y ) ∆u(y ) dsy − x , y )∆u(y ) dsy V (e + V (e ∂n ∂ny Γ T. X. Phan Γ Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 11 / 29 Institut für Numerische Mathematik Boundary integral equations Taking the limit Ω ∋ e x → x ∈ Γ and using the following notations v (x) = γ0int u(x), r (x) = γ0int ∆u(x), w (x) = γ1int u(x), s(x) = γ1int ∆u(x) we obtain a system, including the so-called Calderón projection C , 1 V −K1 V1 v v 2I − K 1 ′ ′ w D I + K D K w 1 1 2 = 1 r r 0 0 I − K V 2 1 ′ s s 0 0 D 2I + K T. X. Phan Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 12 / 29 Institut für Numerische Mathematik Mapping properties Lemma The Calderón operator C is a projection, i.e., C = C 2 . Lemma For all boundary integral operators there hold the relations VK ′ = KV , K ′ D = DK , VD = K1 V − V1 K ′ = VK1′ − KV1 , DV1 + D1 V + K ′ K1′ + K1′ K ′ = 0, T. X. Phan 1 I − K 2, 4 DV = 1 I − K ′2 , 4 K ′ D1 − D1 K = DK1 − K1′ D, VD1 + V1 D + KK1 + K1 K = 0. Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 13 / 29 Institut für Numerische Mathematik Mapping properties Lemma For any q ∈ H −1/2 (Γ) there holds the equality e qk2 2 = hK1 Vq, qiΓ − hV1 ( 1 + K ′ )q, qiΓ . kV L (Ω) 2 Theorem The boundary integral operator 1 Tα := V −1 K1 − V −1 V1 V −1 ( I + K ) + αA : H 1/2 (Γ) → H −1/2 (Γ) 2 is self-adjoint, bounded and H 1/2 (Γ)-elliptic. The Schur complement system Tα z = V −1 N0 u + V −1 M0 f − V −1 V1 V −1 N0 f admits a unique solution z ∈ H 1/2 (Γ). T. X. Phan Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 14 / 29 Institut für Numerische Mathematik Boundary element method We have to find t, q ∈ H −1/2 (Γ), z ∈ H 1/2 (Γ) such that −V1 t + Vq + K1 z = N0 u + M0 f Vt − 21 I + K z = −N0 f q − αAz = 0 Let −1/2 Sh0 (Γ) = span{ϕ0k }N (Γ), k=1 ⊂ H 1/2 SH1 (Γ) = span{ϕ1k }M (Γ) k=1 ⊂ H be some boundary element spaces of piecewise constant and piecewise linear basis functions. T. X. Phan Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 15 / 29 Institut für Numerische Mathematik Galerkin variational formulation The Galerkin boundary element formulation then reads to find (th , qh , zH ) ∈ Sh0 (Γ) × Sh0 (Γ) × SH1 (Γ) such that i + hK1 zH , wh iΓ = hN0 u + M0 f , wh iΓ , −hV1 th , wh iΓ + hVqh , w h Γ 1 hVth , τh iΓ − h 2 I + K zH , τh iΓ = −hN0 f , τh iΓ , −hqh , vH iΓ + αhAzH , vH iΓ = 0 is satisfied for all (wh , τh , vH ) ∈ Sh0 (Γ) × Sh0 (Γ) × SH1 (Γ). Linear system of algebraic equations, t f1 −V1,h Vh K1,h Vh − 21 Mh − Kh q = f 2 −MhT αAH 0 z T. X. Phan Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 16 / 29 Institut für Numerische Mathematik Unique solvability The Galerkin matrix Vh is positive definite and therefore invertible. Hence we obtain the Schur complement system 1 αAH − MhT Vh−1 V1,h Vh−1 ( Mh + Kh ) + MhT Vh−1 K1,h z 2 = MhT Vh−1 f 1 + V1,h Vh−1 f 2 where eα,H := αAH − M T V −1 V1,h V −1 ( 1 Mh + Kh ) + M T V −1 K1,h T h h h h h 2 defines Galerkin boundary element approximation of the Schur complement boundary integral operator Tα . T. X. Phan Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 17 / 29 Institut für Numerische Mathematik Unique solvability Theorem eα,H is positive definite, i.e. The approximate Schur complement T eα,H z, z) > (T 1 T c kzH k2H 1/2 (Γ) 2 1 for all z ∈ RM ↔ zH ∈ SH1 (Γ), if h 6 c0 H for some constant c0 > 0. T. X. Phan Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 18 / 29 Institut für Numerische Mathematik Error estimates Theorem Let t, z be the unique solution of system of boundary integral equations minimizing the reduced cost functional J(u, z). Let th , zH be the unique solution of the Galerkin variational problem. When assuming z ∈ H 2 (Γ) there hold the error estimates kz − zH kH 1/2 (Γ) 6 C H 3/2 |z|H 2 (Γ) + h3/2 kzkH 2(Γ) + h3/2 , kt − th kH −1/2 (Γ) 6 C H 3/2 |z|H 2 (Γ) + h3/2 kzkH 2(Γ) + h3/2 . Remark: In L2 (Γ) norm kz − zH kL2 (Γ) = O(h2 + H 2 ), kt − th kL2 (Γ) = O(h + H). T. X. Phan Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 19 / 29 Institut für Numerische Mathematik Symmetric formulation Boundary integral equations −V1 V K1 t N0 u + M 0 f V 0 − 21 I − K q = −N0 f 1 ′ ′ z N1 u + M 1 f −K1 2 I + K −(αA + D1 ) The Schur complement system h 1 1 K1′ V −1 ( I + K ) + ( I + K ′ )V −1 K1 2 2 i 1 1 ′ −1 − ( I + K )V V1 V −1 ( I + K ) + (αA + D1 ) z 2 2 1 ′ −1 ′ =K1 V N0 f − ( I + K )V −1 V1 V −1 N0 f 2 1 1 ′ −1 + ( I + K )V N0 u + ( I + K ′ )V −1 M0 f − N1 u − M1 f 2 2 T. X. Phan Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 20 / 29 Institut für Numerische Mathematik Unique solvability for symmetric formulation Theorem The boundary integral operator 1 1 1 Tα := K1′ V −1 ( I + K ) − ( I + K ′ )V −1 V1 V −1 ( I + K ) 2 2 2 1 ′ −1 1/2 +( I + K )V K1 + (αA + D1 ) : H (Γ) → H −1/2 (Γ) 2 is self-adjoint, bounded and H 1/2 (Γ)-elliptic. Linear system of algebraic equations, −V1,h Vh K1,h t a Vh − 12 Mh − Kh q = b T K1,h − 21 MhT − KhT αAH + D1,H c z T. X. Phan Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 21 / 29 Institut für Numerische Mathematik Finite Element Methods Primal Dirichlet boundary value problem −∆u(x) = f (x) for x ∈ Ω, u(x) = z(x) for x ∈ Γ u = u0 + Ez, u0 ∈ H01 (Ω): Z Z ∇[u0 (x) + Ez(x)]∇v (x) dx = f (x)v (x) dx Ω for all v ∈ H01 (Ω) Ω Adjoint Dirichlet boundary value problem −∆p(x) = u(x) − u(x) for x ∈ Ω, p(x) = 0 for x ∈ Γ p ∈ H01 (Ω) is the weak solution of the variational problem Z Z ∇p(x)∇v (x) dx = [u(x) − u(x)]v (x) dx for all v ∈ H01 (Ω) Ω T. X. Phan Ω Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 22 / 29 Institut für Numerische Mathematik Finite Element Methods Optimality condition α(Az)(x) = q(x) = ∂ p(x) for x ∈ Γ ∂n w ∈ H 1/2 (Γ) αhAz, w iΓ = Z = Z ∂ p(x)w (x) dsx ∂nx Γ ∇p(x)∇Ew (x) dx − Ω Finite element discretization −MII KII KII MIC −KIC T. X. Phan Z [u(x) − u(x)]Ew (x) dx Ω uI −f 1 −MCI p = f 2 KCI f3 (MCC + αAh ) z Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 23 / 29 Institut für Numerische Mathematik Numerical results (nonsymmetric formulation) In our numerical example, we consider the square Ω = (0, 0.5)2 , h = H, and 1 8 u(x) = − 4 + [x1 (1 − 2x1 ) + x2 (1 − 2x2 )] , f (x) = − , α = 0.01 α α e: A=D e v iΓ := hDu, v iΓ + hu, 1iΓ hv , 1iΓ hDu, with non zero control z(x), (see paper of S. May, R. Rannacher, B. Vexler). Mesh 16 32 64 128 256 512 1024 expected T. X. Phan kth − th9 kL2 (Γ) Error Order 33.5093 24.0337 0.48 14.7956 0.70 8.4434 0.81 4.5915 0.88 2.4033 0.93 1.1718 1.04 1.00 kzh − zH9 kL2 (Γ) Error Order 0.45219 0.12773 1.82 0.03536 1.85 9.026e-03 1.97 2.175e-03 2.05 5.158e-04 2.08 1.389e-04 1.89 2.00 Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 24 / 29 Institut für Numerische Mathematik Numerical results (symmetric formulation) We consider the square Ω = (0, 0.5)2 , h = H, and 1 8 u(x) = − 4 + [x1 (1 − 2x1 ) + x2 (1 − 2x2 )] , f (x) = − , α α e: A=D α = 0.01 e v iΓ := hDu, v iΓ + hu, 1iΓ hv , 1iΓ hDu, with non zero control z(x). Mesh 16 32 64 128 256 512 1024 expected T. X. Phan kth − th9 kL2 (Γ) Error Order 39.8736 28.6702 0.48 16.2201 0.82 8.8797 0.87 4.7272 0.91 2.4459 0.95 1.1853 1.05 1.00 kzh − zH9 kL2 (Γ) Error Order 2.24551 0.46550 2.27 0.08836 2.39 0.01631 2.44 3.019e-03 2.43 5.731e-04 2.40 1.243e-04 2.20 2.00 Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 25 / 29 Institut für Numerische Mathematik Compare numerical results: BEM and FEM Consider the square Ω = (0, 0.5)2 , h = H and 1 [x1 (1 − 2x1 ) + x2 (1 − 2x2 )] , u(x) = − 4 + α e: A=D 8 f (x) = − , α α = 0.01 e v iΓ := hDu, v iΓ + hu, 1iΓ hv , 1iΓ hDu, with non zero control z(x). Level 2 3 4 5 6 7 8 T. X. Phan Symmetric BEM FEM kL2 (Γ) kzH − zHBEM 9 order kL2 (Γ) kzH − zHFEM 9 order kL2 (Γ) kzH − zHBEM 9 2.24551 0.46550 0.08836 0.01630 3.019e-03 5.731e-04 1.243e-04 2.27 2.39 2.44 2.43 2.40 2.20 0.38934 0.10745 0.02809 7.281e-03 1.872e-03 4.691e-04 1.060e-04 1.86 1.93 1.95 1.96 2.00 2.15 0.38936 0.10747 0.02811 7.298e-03 1.889e-03 4.883e-04 1.337e-04 Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 26 / 29 Institut für Numerische Mathematik Linear systems: BEM and FEM Finite element system −MII KII KII MIC −KIC uI −MCI −f 1 p = f 2 KCI (MCC + αAh ) f3 z Boundary element system −V1,h Vh Vh T K1,h − 21 MhT − KhT K1,h t a − 12 Mh − Kh q = b αAH + D1,H c z • Preconditioned iterative system solvers (saddle point structure) • Construction of preconditioners (BEM and FEM) T. X. Phan Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 27 / 29 Institut für Numerische Mathematik Outlook/Plan • Stability analysis for symmetric formulation: h = H • Mixed boundary control problems • Box boundary control problems • Parabolic control problems • BIE for time dependent problems T. X. Phan Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 28 / 29 Institut für Numerische Mathematik Thanks for your attention! T. X. Phan Boundary and Finite Element Methods for Dirichlet Boundary Control Problems SFB Status Seminar, 07/11/2008 29 / 29