Experimental Study of the HUM Control Operator for Linear Waves Gilles Lebeau1 1 Université Maëlle Nodet2 de Nice, Institut Universitaire de France 2 Université de Grenoble, INRIA Orléans, juin 2009 Experimental study of HUM for waves Outline 1 HUM control operator 2 Numerical method 3 Experimental study of HUM operator properties Convergence rate Frequency localization Space localization Condition number Non controlling domains Lebeau, Nodet 2 / 36 Experimental study of HUM for waves HUM control operator Outline 1 HUM control operator 2 Numerical method 3 Experimental study of HUM operator properties Convergence rate Frequency localization Space localization Condition number Non controlling domains Lebeau, Nodet 3 / 36 Experimental study of HUM for waves HUM control operator Lebeau, Nodet Controllability of linear waves For a given f = (u0 , u1 ) ∈ H01 (Ω) × L2 (Ω), the problem is to find a source v (t, x)∈ L2 (0, T ; L2 (Ω)) such that the solution u = S(v ) of the linear wave equation 2 (∂tt − ∆)u = χv in ]0, +∞[×Ω u = 0, t > 0 (1) |∂Ω (u|t=0 , ∂u|t=0 ) = (0, 0) reaches the state f = (u0 , u1 ) = (u(T , .), ∂t u(T , .)) at time T , where: Ω is a bounded open subset of Rd , the “control domain” U is a non empty open subset of Ω, χ(t, x) = ψ(t)χ0 (x) where χ0 is a real L∞ function on Ω, such that support(χ0 ) = U and χ0 (x) is continuous and positive for x ∈ U, ψ ∈ C ∞ ([0, T ]) and ψ(t) > 0 on ]0, T [. 4 / 36 Experimental study of HUM for waves HUM control operator Lebeau, Nodet Controllability of linear waves The reachable set at time T is the subspace of H = H01 (Ω) × L2 (Ω): RT = {f = (u0 , u1 ) ∈ H, ∃v , (S(v )(T , .), ∂t S(v )(T , .)) = (u0 , u1 )} Approximate controllability: RT is dense in H. Exact controllability: RT = H. 5 / 36 Experimental study of HUM for waves HUM control operator Lebeau, Nodet The HUM method The Hilbert Uniqueness Method (HUM) [Lions 1988] consists in choosing the function v with L2 -minimal norm. Then v is necessarily of the form χ∂t w where w is a solution of the dual control problem: 2 (∂tt − ∆)w = 0 in ]0, +∞[×Ω w = 0, t > 0 |∂Ω (w |t=T , ∂t w |t=T ) = (w0 , w1 ) = h ∈ H = H01 (Ω) × L2 (Ω) The HUM control operator is then defined by Λ: H → H data f = (u0 , u1 ) 7→ control h = (w0 , w1 ) 6 / 36 Experimental study of HUM for waves HUM control operator Lebeau, Nodet HUM optimal control operator Let A = A∗ be the operator on H = H01 (Ω) × L2 (Ω) defined by 0 Id iA = ∆ 0 √ Let λ = −∆D . Then (1) becomes (∂t − iA)u = B(t)g with 0 0 0 λ−1 χ(t, .) ∗ B(t) = , B (t) = χ(t, .)λ 0 0 0 and g = (λ−1 v , 0). Then we have exact controlability iff Z T ∗ ∃C > 0, MT = eitA B(T − t)B ∗ (T − t)e −itA dt ≥ C Id 0 And in that case we have Λ = MT−1 7 / 36 Experimental study of HUM for waves HUM control operator Lebeau, Nodet Geometric control condition We recall that the source v in (1) is multiplied by χ(t, x) = ψ(t)χ0 (x), where χ0 ∈ L∞ (Ω), such that support(χ0 ) = U and χ0 (x) is continuous and positive for x ∈ U, ψ ∈ C ∞ ([0, T ]) and ψ(t) > 0 on ]0, T [. We also assume that there is no contact of infinite order between ∂Ω and the optical rays of the wave operator in the free space. Let us recall the Geometric Control Condition of C. Bardos, G. Lebeau and J. Rauch: GCC Every geodesic ray of Ω traveling with speed 1 and starting at t = 0 enters the open set U = {x ∈ Ω, χ0 (x) 6= 0} in time t < T . Theorem If χ and T are such that the GCC condition holds true, then MT is an isomorphism, i.e. one has exact controlability (RT = H). 8 / 36 Experimental study of HUM for waves Numerical method Outline 1 HUM control operator 2 Numerical method 3 Experimental study of HUM operator properties Convergence rate Frequency localization Space localization Condition number Non controlling domains Lebeau, Nodet 9 / 36 Experimental study of HUM for waves Numerical method Lebeau, Nodet Previous numerical approach Idea: [R. Glowinski et al.] first discretize the continuous wave equation, then compute the control of the discrete system. Major drawback: As observed by R. Glowinski et al. and precisely studied by E. Zuazua, the discrete model is not uniformly exactly controllable when the mesh size goes to zero, and the interaction of waves with the numerical mesh produces spurious high frequency oscillations. In other words, the processes of numerical discretization and control do not commute for the wave equation. 10 / 36 Experimental study of HUM for waves Numerical method Lebeau, Nodet Spectral Galerkin method Let (ωj2 ) be the sequence of −∆D eigenvalues, and (ej ) the associated orthonormal basis of L2 (Ω): −∆ej = ωj2 ej , ej |∂Ω = 0 For a given cutoff frequency ω, we define L2ω = Span{ej , ωj ≤ ω} and we denote by Πω the orthogonal projector on L2ω , which obviously acts also on H = H01 × L2 : Hω = Πω (H01 × L2 ) and we define the matrix MT ,ω : MT ,ω = Πω MT Πω MT ,ω,n,m = (MT φn |φm )H where (φn ) is an orthonormal basis of Hω . 11 / 36 Experimental study of HUM for waves Numerical method Lebeau, Nodet Spectral Galerkin method Let us recall that Λ = MT−1 and MT ,ω = Πω MT Πω . Then we have: MT ,ω is invertible on Hω (always true) kMT−1,ω k is bounded uniformly in ω (because of exact controlability) Lemma Assume GCC. Then there exists c > 0 such that for all f ∈ H: kΛ(f ) − MT−1,ω (fω )kH ≤ ckf − fω kH + kΛ(fω ) − MT−1,ω (fω )kH with fω = Πω f and limω→∞ kΛ(fω ) − MT−1,ω (fω )kH = 0. In other words, the processes of Galerkin approximation and inversion “almost” commute for MT . 12 / 36 Experimental study of HUM for waves Numerical method Lebeau, Nodet Comparison between ((MT )−1 )ω and ((MT )ω )−1 log abs (trunc(inv JMJ) − inv(trunc JMJ)) −5 100 −10 200 −15 300 −20 400 −25 500 −30 600 −35 700 −40 800 −45 900 −50 1000 100 200 300 400 500 600 700 800 900 1000 13 / 36 Experimental study of HUM for waves Numerical method Lebeau, Nodet Computation of MT ,ω Z MT = T ∗ eitA B(T − t)B ∗ (T − t)e −itA dt 0 Computation of the eigenvectors (ej ) of −∆ in Ω Computation of MT ,ω,m,n : MT ,ω,m,n = (MT φn |φm )H = Dn,m (T , ψ, Ω)Gn,m (χ0 , Ω) Dn,m is a time integral Gn,m is a space integral: Z Gn,m = χ0 (x)2 φn (x)φm (x) dx Ω Inversion of MT ,ω OFFLINE: for each Ω and χ0 (x): eigenvectors, G matrix ONLINE: for various T and data f = (u0 , u1 ), computation of Λ(f ) 14 / 36 Experimental study of HUM for waves Numerical method Lebeau, Nodet Geometries, control domains and smoothing Three geometries: square, disc and general domain For each geometry, one “standard” control domain satisfying GCC Two cases: non-smooth control: χ(t, x) = 1[0,T ] 1U smooth control: χ(t, x) = ψ(t)χ0 (x) with −t) ψ(t) = 4t(T 1[0,T ] and χ0 (x) similarly smoothed T2 control domain smooth control domain control domain smooth control domain control domain smooth control domain 15 / 36 Experimental study of HUM for waves Numerical method Lebeau, Nodet 16 / 36 Impact of smoothing on MT ,ω Smooth control Non-smooth control log J*M*inv(J) log J*M*inv(J) 0 −5 500 −5 500 −10 1000 −10 1000 −15 1500 −15 1500 −20 −20 2000 2000 −25 −25 2500 2500 −30 −30 3000 −35 −40 3500 4000 −45 500 1000 1500 2000 2500 3000 (same color scaling) 3500 4000 3000 −35 −40 3500 −45 4000 500 1000 1500 2000 2500 3000 3500 4000 Experimental study of HUM for waves Experimental study of HUM operator properties Outline 1 HUM control operator 2 Numerical method 3 Experimental study of HUM operator properties Convergence rate Frequency localization Space localization Condition number Non controlling domains Lebeau, Nodet 17 / 36 Experimental study of HUM for waves Experimental study of HUM operator properties Convergence rate Outline 1 HUM control operator 2 Numerical method 3 Experimental study of HUM operator properties Convergence rate Frequency localization Space localization Condition number Non controlling domains Lebeau, Nodet 18 / 36 Experimental study of HUM for waves Experimental study of HUM operator properties Convergence rate Lebeau, Nodet Theoretical results Convergence lemma: with ω0 ≤ ω and fω0 ,ω = (Πω − Πω0 )f we have kΛ(fω ) − MT−1,ω (fω )kH ≤ (1 + ckMT k) (rΛ (ω, ω0 )kf kH + ckfω0 ,ω kH ) where the rate of convergence rΛ (ω, ω0 ) defined by: rΛ (ω, ω0 ) = kΠ⊥ ω ΛΠω0 kH satisfies the following lemma: Lemma Assume GCC holds true and χ is smooth. Then there exists g with rapid decay such that: ω rΛ (ω, ω0 ) ≤ g ω0 19 / 36 Experimental study of HUM for waves Experimental study of HUM operator properties Convergence rate Lebeau, Nodet Convergence rate: reconstruction errors FINITE DIFF: Reconstruction error as a function of the cutoff frequency 0 square ; no smoothing square ; smoothing trapezoid ; no smoothing trapezoid ; smoothing −1 Reconstruction error in % (log10 scale) −2 −3 −4 −5 −6 −7 −0.5 0 0.5 1 Cutoff frequency (log10 scale) 1.5 2 20 / 36 Experimental study of HUM for waves Experimental study of HUM operator properties Frequency localization Outline 1 HUM control operator 2 Numerical method 3 Experimental study of HUM operator properties Convergence rate Frequency localization Space localization Condition number Non controlling domains Lebeau, Nodet 21 / 36 Experimental study of HUM for waves Experimental study of HUM operator properties Frequency localization Lebeau, Nodet Theoretical results Let ψk (D), k ∈ N, be the spectral localization operators associated to the Littlewood-Paley decomposition: X X ψk (D)( aj ej ) = ψk (ωj )aj ej , j j Sk (D) = k X ψj (D), k ≥ 0 j=0 Theorem (Dehman-Lebeau) Assume that the geometric control condition GCC holds true, and that the control function χ(t, x) is smooth. There exists C > 0 such that for every k ∈ N, the following inequality holds true kψk (D)Λ − Λψk (D)kH ≤ C 2−k kSk (D)Λ − ΛSk (D)kH ≤ C 2−k 22 / 36 Experimental study of HUM for waves Experimental study of HUM operator properties Frequency localization Lebeau, Nodet 23 / 36 Frequency localization in the square Data Non smooth control Spectral coeffs of the data 20 15 10 5 Control, with smoothing 140 70 120 60 absolute value of the spectral coefficient absolute value of the spectral coefficient absolute value of the spectral coefficient 25 0 100 80 60 40 20 0 20 40 frequency Smooth control Control, no smoothing 30 60 0 50 40 30 20 10 0 20 40 frequency 60 0 0 20 40 frequency 60 Experimental study of HUM for waves Experimental study of HUM operator properties Space localization Outline 1 HUM control operator 2 Numerical method 3 Experimental study of HUM operator properties Convergence rate Frequency localization Space localization Condition number Non controlling domains Lebeau, Nodet 24 / 36 Experimental study of HUM for waves Experimental study of HUM operator properties Space localization Space localization in the square Lebeau, Nodet 25 / 36 Experimental study of HUM for waves Experimental study of HUM operator properties Space localization Space localization in the disc Lebeau, Nodet 26 / 36 Experimental study of HUM for waves Experimental study of HUM operator properties Space localization Lebeau, Nodet Box experiment in the square, no smoothing 27 / 36 Experimental study of HUM for waves Experimental study of HUM operator properties Space localization Lebeau, Nodet Box experiment in the square, with smoothing 28 / 36 Experimental study of HUM for waves Experimental study of HUM operator properties Space localization Lebeau, Nodet Theoretical results Our experiments suggest that Λ preserves both space and frequency localization, so that the following conjecture may hold true: Conjecture Assume that the geometric control condition GCC holds true, that the control function χ(t, x) is smooth, and that the optical rays have no infinite order of contact with the boundary. Then Λ is a microlocal operator. 29 / 36 Experimental study of HUM for waves Experimental study of HUM operator properties Condition number Outline 1 HUM control operator 2 Numerical method 3 Experimental study of HUM operator properties Convergence rate Frequency localization Space localization Condition number Non controlling domains Lebeau, Nodet 30 / 36 Experimental study of HUM for waves Experimental study of HUM operator properties Condition number Lebeau, Nodet Condition number of MT ,ω as a function of ω Condition number of HUM matrix as a function of the cutoff frequency 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 40 Condition number of MT (log scale) 35 30 25 20 15 10 5 60 80 100 120 140 Cutoff frequency 160 180 200 220 31 / 36 Experimental study of HUM for waves Experimental study of HUM operator properties Condition number Lebeau, Nodet Theoretical results Our experiments lead us to think that the following conjecture may be true for a “generic” choice of the control function χ(t, x): Conjecture There exists C (T , U) such that lim ω→∞ log kMT−1,ω kH ω with C (T , U) = 0 for T ≥ TGCC C (T , U) > 0 for T < TGCC = C (T , U) 32 / 36 Experimental study of HUM for waves Experimental study of HUM operator properties Non controlling domains Outline 1 HUM control operator 2 Numerical method 3 Experimental study of HUM operator properties Convergence rate Frequency localization Space localization Condition number Non controlling domains Lebeau, Nodet 33 / 36 Experimental study of HUM for waves Experimental study of HUM operator properties Non controlling domains Lebeau, Nodet 34 / 36 Non-controlling domains and special eigenvectors 500 500 500 500 450 450 450 450 400 400 400 400 350 350 350 350 300 300 300 300 250 250 250 250 200 200 200 200 150 150 150 150 100 100 100 100 50 50 50 100 200 300 400 500 100 200 300 400 50 500 100 200 300 Eigenvector 53 400 500 100 200 300 400 Eigenvector 60 500 2.5 500 4 450 2 450 3 400 1.5 400 350 1 350 0.5 300 0 250 −0.5 200 −1 150 2 1 300 0 250 −1 200 150 −2 100 −3 −1.5 100 −2 50 50 −4 −2.5 100 200 300 400 500 100 200 300 400 500 500 Experimental study of HUM for waves Experimental study of HUM operator properties Non controlling domains Lebeau, Nodet Influence of the choice of the data Experiments with the first non-controlling domain U (radius truncated around the boundary), and two different data to be reached. On the left, the 53rd eigenmode is localized around the boundary of the circle, as is the control domain. On the right, the 60th eigenmode is localized around the center of the circle. smooth? no yes Condition number 4960 6042 53rd Eigenmode Error (%) kw k 0.63 15 0.00065 24 60th Eigenmode Error (%) kw k 26 4583 0.11 8872 35 / 36 Experimental study of HUM for waves Experimental study of HUM operator properties Non controlling domains Lebeau, Nodet 36 / 36 Influence of the choice of the domain Experiments with the data u53 localized at the boundary of the disc for two different non-controlling domains U, the first one being localized around the boundary, the second one around the center. smooth? no yes Cond. 4960 6042 First domain Error (%) kw k 0.63 15 0.00065 24 Second domain Cond. Error (%) kw k 6 3.6 10 68 1.9 104 5 3.3 10 0.0094 6.5 105 Experimental study of HUM for waves Movie Lebeau, Nodet Time evolution of the reconstructed solution 37 / 37 Experimental study of HUM for waves Errors Lebeau, Nodet 38 / 42 Frequency loc: reconstruction error Non smooth control −3 2.5 Smooth control −7 Spectral coeffs relative error U0 x 10 9 Spectral coeffs relative error U0 x 10 8 7 relative error relative error 2 1.5 1 6 5 4 3 2 0.5 1 0 0 20 40 60 80 100 120 140 160 0 180 0 20 40 60 80 frequency −3 2.5 −6 Spectral coeffs relative error U1 x 10 100 120 140 160 180 120 140 160 180 frequency 4.5 Spectral coeffs relative error U1 x 10 4 3.5 relative error relative error 2 1.5 1 3 2.5 2 1.5 1 0.5 0.5 0 0 20 40 60 80 100 frequency 120 140 160 180 0 0 20 40 60 80 100 frequency Experimental study of HUM for waves Errors Space loc: error in the square Lebeau, Nodet 39 / 42 Experimental study of HUM for waves Errors Space loc: error in the disc Lebeau, Nodet 40 / 42 Experimental study of HUM for waves Errors Error for the box experiment Lebeau, Nodet 41 / 42 Experimental study of HUM for waves Errors Error for the rotated box experiment Lebeau, Nodet 42 / 42