Matrix square root and interpolation spaces Mario Arioli and Daniel Loghin m.arioli@rl.ac.uk STFC-Rutherford Appleton Laboratory, and University of Birmingham Householder, Berlin, 2008 – p.1/23 Outline Norms and duality in finite dimensional Hilbert spaces Discrete Interpolation Norms The continuous case and finite-element approximation An example: the biharmonic operator Summary and open problems Householder, Berlin, 2008 – p.2/23 TRUTH may be misleading In a finite dimensional space all norms are equivalent i.e. Householder, Berlin, 2008 – p.3/23 TRUTH may be misleading In a finite dimensional space all norms are equivalent i.e. c(N )||v||1 ≤ ||v||2 ≤ C(N )||v||1 Householder, Berlin, 2008 – p.3/23 TRUTH may be misleading In a finite dimensional space all norms are equivalent i.e. c(N )||v||1 ≤ ||v||2 ≤ C(N )||v||1 Identify the norms for which we have c||v||1 ≤ ||v||2 ≤ C||v||1 Householder, Berlin, 2008 – p.3/23 Finite dimensional Hilbert spaces and IRN (·, ·) : H p × H → IR scalar product and kukH = (u, u) ∀u ∈ H norm. Householder, Berlin, 2008 – p.4/23 Finite dimensional Hilbert spaces and IRN (·, ·) : H p × H → IR scalar product and kukH = (u, u) ∀u ∈ H norm. ∃{ψi }i=1,...,N a basis for H PN ∀u ∈ H u = i=1 ui ψi ui ∈ IR i = 1, . . . , N Householder, Berlin, 2008 – p.4/23 Finite dimensional Hilbert spaces and IRN (·, ·) : H p × H → IR scalar product and kukH = (u, u) ∀u ∈ H norm. ∃{ψi }i=1,...,N a basis for H PN ∀u ∈ H u = i=1 ui ψi ui ∈ IR i = 1, . . . , N Representation of scalar product in IRN . PN PN Let u = i=1 ui ψi and v = i=1 vi ψi . Then N X N X (u, v) = ui vj (ψi , ψj ) = vT Hu i=1 j=1 where Hij = Hji = (ψi , ψj ) and u, v ∈ IRN . Moreover, uT Hu > 0 iff u 6= 0 and, thus H SPD. Householder, Berlin, 2008 – p.4/23 Dual space H′ f ∈ H′ : H → IR (functional); f (αu + βv) = αf (u) + βf (v) ∀u, v ∈ H H′ is the space of the linear functionals on H kf kH′ f (u) = sup u6=0 kukH Householder, Berlin, 2008 – p.5/23 Dual space H′ f ∈ H′ : H → IR (functional); f (αu + βv) = αf (u) + βf (v) ∀u, v ∈ H H′ is the space of the linear functionals on H f (u) kf kH′ = sup u6=0 kukH PN If H finite dimensional and u = i=1 ui ψi , then PN f (u) = i=1 ui f (ψi ) = f T u Householder, Berlin, 2008 – p.5/23 Dual space H′ f ∈ H′ : H → IR (functional); f (αu + βv) = αf (u) + βf (v) ∀u, v ∈ H H′ is the space of the linear functionals on H f (u) kf kH′ = sup u6=0 kukH PN If H finite dimensional and u = i=1 ui ψi , then PN f (u) = i=1 ui f (ψi ) = f T u Dual vector Let u ∈ H, u 6= 0, then ∃fu ∈ H′ such that fu (u) = kukH (Hahn-Banach). Householder, Berlin, 2008 – p.5/23 Dual space H′ Let H be a Hilbert finite dimensional space and H the real N × N matrix identifying the scalar product. Householder, Berlin, 2008 – p.6/23 Dual space H′ Let H be a Hilbert finite dimensional space and H the real N × N matrix identifying the scalar product. fu (u) = f T u = (uT Hu)1/2 The dual vector of u has the following representation: Householder, Berlin, 2008 – p.6/23 Dual space H′ Let H be a Hilbert finite dimensional space and H the real N × N matrix identifying the scalar product. fu (u) = f T u = (uT Hu)1/2 The dual vector of u has the following representation: Hu f= kukH and kfu k2H′ = uT Hu = f T H−1 f Householder, Berlin, 2008 – p.6/23 Linear operator A : H → V H and V finite dimensional Hilbert spaces. Householder, Berlin, 2008 – p.7/23 Linear operator A : H → V H and V finite dimensional Hilbert spaces. kAkH,V = max u6=0 kAukV = kV1/2 AH−1/2 k2 kukH Householder, Berlin, 2008 – p.7/23 Linear operator A : H → V H and V finite dimensional Hilbert spaces. kAkH,V = max u6=0 kAukV = kV1/2 AH−1/2 k2 kukH The result follows from the generalized eigenvalue problem in IRN AT VAu = λHu Householder, Berlin, 2008 – p.7/23 Linear operator A : H → V H and V finite dimensional Hilbert spaces. kAkH,V = max u6=0 kAukV = kV1/2 AH−1/2 k2 kukH The result follows from the generalized eigenvalue problem in IRN AT VAu = λHu κH (M ) = kM kH,H −1 kM −1 kH −1 ,H . Householder, Berlin, 2008 – p.7/23 Linear operator A : H → V H and V finite dimensional Hilbert spaces. kAkH,V = max u6=0 kAukV = kV1/2 AH−1/2 k2 kukH The result follows from the generalized eigenvalue problem in IRN AT VAu = λHu κH (M ) = kM kH,H −1 kM −1 kH −1 ,H . The interesting case is κH (M ) independent of N Householder, Berlin, 2008 – p.7/23 Interpolation spaces H = (IRN , (u, v)H = uT Hv) M = (IRN , (u, v)M = uT M v) (u, v)H = (u, Sv)M = (Su, v)M S = M −1 H S self-adjoint in the good scalar product! n o Sx = µx ⇔ Hx = µM x ⇒ µ = δ 2 > 0 ∃W s.t. M = W T W, H = W T ∆2 W, ∆ diagonal Householder, Berlin, 2008 – p.8/23 Interpolation spaces H = (IRN , (u, v)H = uT Hv) M = (IRN , (u, v)M = uT M v) (u, v)H = (u, Sv)M = (Su, v)M S = M −1 H S self-adjoint in the good scalar product! n o Sx = µx ⇔ Hx = µM x ⇒ µ = δ 2 > 0 ∃W s.t. M = W T W, H = W T ∆2 W, ∆ diagonal ∆ ≥ 0 Householder, Berlin, 2008 – p.8/23 Interpolation spaces Λ = W −1 ∆W Λ1/2 = W −1 ∆1/2 W S = M −1 H = W −1 W −T W T ∆2 W = W −1 ∆W W −1 ∆W = Λ2 M Λ = W T W W −1 ∆W −T W T W = ΛT M ⇒ (u, Λv)M = (Λu, v)M (Λ1/2 u, Λ1/2 u)M = (u, Λu)M Householder, Berlin, 2008 – p.9/23 Interpolation spaces h 1/2 o i n H, M = u ∈ IRN ; (u, u)M + (u, S 1−ϑ u)M h ϑ i H, M 1/2 n = u ∈ IRN ; (u, u)M + (u, Λu)M 1/2 o Householder, Berlin, 2008 – p.10/23 Interpolation spaces h 1/2 o i n H, M = u ∈ IRN ; (u, u)M + (u, S 1−ϑ u)M h ϑ i H, M 1/2 n = u ∈ IRN ; (u, u)M + (u, Λu)M 1/2 o ||v||2ϑ = ||v||2Hϑ = v T M + M S 1−ϑ v Hϑ = M I + S 1−ϑ = W T I + ∆2(1−ϑ) W Householder, Berlin, 2008 – p.10/23 Interpolation spaces (duality) M′ and H′ dual spaces of M and H h i′ h i H, M = M′ , H′ ϑ 1−ϑ S ′ = M H −1 = W T ∆−2 W −T Householder, Berlin, 2008 – p.11/23 Interpolation spaces (duality) M′ and H′ dual spaces of M and H h i′ h i H, M = M′ , H′ ϑ 1−ϑ S ′ = M H −1 = W T ∆−2 W −T −1 ′ = H1−ϑ = W −1 ∆−2ϑ W −T H1−ϑ Householder, Berlin, 2008 – p.11/23 Interpolation spaces (∞ dimension case) X, Y two Hilbert spaces with X ⊂ Y , X dense and continuously embedded in Y . h·, ·iX , h·, ·iY and k · kX , k · kY the respective norms. (Riesz representation theory) ∃J : X → Y positive and self-adjoint with respect to h·, ·iY such that hu, viX = hu, J viY . E = J 1/2 : X → Y , 2 2 1/2 . X = D(E) with kukX ∼ kukE := kukY + kEukY 1/2 2 1−θ 2 kukθ := kukY + kE ukY . 1−θ The interpolation space of index θ [X, Y ]θ := D(E ), 0 ≤ θ ≤ 1, with the inner-product hu, viθ = hu, viY + u, E 1−θ v Y is a Hilbert space (Lions Magenes 1968). [X, Y ]0 = X and [X, Y ]1 = Y . If 0 < θ1 < θ2 < 1 then X ⊂ [X, Y ]θ1 ⊂ [X, Y ]θ2 ⊂ Y. ∀θ ∈ (0, 1) π ∈ L(X; X ) ∩ L(Y ; Y) =⇒ π ∈ L([X, Y ]θ ; [X , Y]θ ). Householder, Berlin, 2008 – p.12/23 Interpolation spaces (∞ dimension case) Ω ⊂ Rn open bounded with smooth boundary Γ and let α denote a multi-index of order m where m is a positive integer m α 2 H (Ω) = u : D u ∈ L (Ω), |α| ≤ m (H 0 (Ω) = L2 (Ω)) H s (Ω) := [H m (Ω), H 0 (Ω)]1−s/m H0s (Ω) completion of C0∞ (Ω) in H m (Ω), where s > 0. For 0 ≤ s2 < s1 , ( (1−θ)s1 +θs2 (Ω) if (1 − θ)s1 + θs2 6= k + 1/2 (k [H0s1 (Ω), H0s2 (Ω)]θ = H0 k+1/2 [H0s1 (Ω), H0s2 (Ω)]θ = H00 k+1/2 (Ω) ⊂ H0 if (1 − θ)s1 + θs2 = k + 1/2 (k H −s (Ω) = (H0s (Ω))′ s > 0 If (1 − θ)s1 + θs2 = 1/2 ′ 1/2 [H −s1 (Ω), H −s2 (Ω)]θ = H00 (Ω) . Householder, Berlin, 2008 – p.13/23 Finite-element example 1/2 H00 (Ω) = [H01 (Ω), L2 (Ω)]1/2 . Let Xh ⊂ H01 (Ω), Yh ⊂ L2 (Ω). Let {φi }1≤i≤n ∈ Xh be a spanning set for Yh and let Lk ∈ Rn×n denote the Grammian matrices corresponding to the h·, ·iH0k (Ω) -inner product (H 0 (Ω) = L2 (Ω)): (Lk )ij = hφi , φj iH k (Ω) . 0 H = L1 , M = L0 and H1/2,h Moreover, we have 1/2 = L0 I + (L−1 L ) 1 0 H1/2,h ∼ H1/2 = L0 L−1 0 L1 1/2 Householder, Berlin, 2008 – p.14/23 Finite-element example ′ 1/2 H00 (Ω) = [H −(1) 0 (Ω), H (Ω)]1/2 = ′ 1 0 [H0 (Ω), H0 (Ω)]1/2 . Let Xh , Yh be defined as above. Let Yh′ ⊂ Xh′ = span {ψi }1≤i≤n where ψi are basis functions dual to φi , i.e.,hψi , φj iH −1 (Ω)×H01 (Ω) = δij . If Yh′ ∋ z = n X zi ψi = i=1 δij = hψi , φj iH −1 (Ω)×H01 (Ω) = n X wi φi , φl = i=1 n X n X Kli ψi i=1 Kil−1 hφl , φj iH 1 (Ω) = 0 i=1 n X Kil−1 (L1 )lj i=1 ′ = kwk , kzk and the so that z = L0 w and kzkHY′ = kwkL−1 H L0 L−1 X 1 L0 0 ′ matrix representation of HY′ , HX are respectively L0 and L0 L−1 1 L0 . ′ −1/2 H1/2 = L0 (L−1 L ) = H−1/2 . 1 0 Householder, Berlin, 2008 – p.15/23 Evaluation of Hθ z Generalised Lanczos HX Vk = HY Vk Tk + βk+1 Hy vk+1 eTk , VkT HY Vk = Ik (Tk tridiagonal). Householder, Berlin, 2008 – p.16/23 Evaluation of Hθ z Generalised Lanczos HX Vk = HY Vk Tk + βk+1 Hy vk+1 eTk , VkT HY Vk = Ik (Tk tridiagonal). v0 = z Hθ z ≈ HY Vk Tk1−θ e1 kzkHY and Hθ,h z ≈ HY Vk (Ik + Tk1−θ )e1 kzkHY . Householder, Berlin, 2008 – p.16/23 Evaluation of Hθ z Generalised Lanczos HX Vk = HY Vk Tk + βk+1 Hy vk+1 eTk , VkT HY Vk = Ik (Tk tridiagonal). v0 = z Hθ z ≈ HY Vk Tk1−θ e1 kzkHY and Hθ,h z ≈ HY Vk (Ik + Tk1−θ )e1 kzkHY . v0 = HY−1 z Hθ−1 z ≈ Vk Tkθ−1 e1 kzkHY−1 and −1 Hθ,h z ≈ Vk (Ik + Tk1−θ )−1 e1 kzkHY−1 . Householder, Berlin, 2008 – p.16/23 Evaluation of Hθ z Generalised Lanczos HX Vk = HY Vk Tk + βk+1 Hy vk+1 eTk , VkT HY Vk = Ik (Tk tridiagonal). v0 = z Hθ z ≈ HY Vk Tk1−θ e1 kzkHY and Hθ,h z ≈ HY Vk (Ik + Tk1−θ )e1 kzkHY . v0 = HY−1 z Hθ−1 z ≈ Vk Tkθ−1 e1 kzkHY−1 and −1 Hθ,h z ≈ Vk (Ik + Tk1−θ )−1 e1 kzkHY−1 . Alternative: N. Hale, and N. J. Higham and L. N. Trefethen, SIAM J. Numer. Anal. Householder, Berlin, 2008 – p.16/23 An example: biharmonic operator Consider the biharmonic problem in a polygonal convex open domain Ω ⊂ R2 with boundary Γ = ∪K i=1 Γi . ( ∆2 u = f in Ω, u = ∂u/∂n = 0 on Γ. −∆u = f v + ∆u = 0 u = ∂u/∂n = 0 in Ω, in Ω, on Γ. Householder, Berlin, 2008 – p.17/23 An example: biharmonic operator (Pironneau-Glowinski) (i) ( Sλ = ∂u0 /∂ν (ii) (iii) −∆v0 = f v0 = 0 ( −∆v1 = 0 v1 = λ in Ω, on Γ, ( −∆u0 = v0 u0 = 0 in Ω, on Γ, ( −∆u1 = v1 u1 = 0 in Ω, on Γ, on Γ, in Ω, on Γ, Householder, Berlin, 2008 – p.18/23 An example: biharmonic operator S is a boundary operator which is defined on H −1/2 (Γ) and which induces a bilinear form s(·, ·) : H −1/2 (Γ) × H −1/2 (Γ) s(·, ·) is symmetric, positive-definite and H −1/2 (Γ)−elliptic, i.e., c1 kλk2H −1/2 (Γ) ≤ s(λ, λ) ≤ c2 kλk2H −1/2 (Γ) . Householder, Berlin, 2008 – p.19/23 An example: biharmonic operator L ZT Z −MBB ! x vB ! = ! g 0 The Schur complement associated with L in the matrix is S = −MBB − Z T L−1 Z. Let Xh ⊂ H 1 (Γ) denote the space spanned by the restriction of the basis functions of VIh to the boundary Γ. If λh ∈ Xh has a vector of coefficients λ , then λ. s(λh , λh ) = λ T Sλ Householder, Berlin, 2008 – p.20/23 An example: biharmonic operator A discrete H −1/2 -norm on Xh can be defined as a sum of norms corresponding to each open segment of the polygonal boundary Γ: !1/2 K X kλh kH −1/2 (Γ) := kλh k2H −1/2 (Γi ) . i=1 1/2 In particular, H −1/2 (Γi ) = (H00 (Γi ))′ . λk2H{−1/2,h} kλh k2H −1/2 (Γ) = kλ for λh ∈ Xh where H{−1/2,h} = K M λk2H{−1/2} or = kλ ′{i} H1/2,h i=1 K M {i} H{−1/2} = H−1/2 , i=1 {i} H−1/2 −1/2 L ) = L0,i (L−1 1,i 0,i Householder, Berlin, 2008 – p.21/23 An example: biharmonic operator P = L Z PS ! Householder, Berlin, 2008 – p.22/23 An example: biharmonic operator P = L Z PS ! n m I H{−1/2,h} H{−1/2} b {−1/2} H 84,610 640 26 10 12 12 337,154 1,280 30 9 11 11 1,346,050 2,560 36 9 11 11 FGMRES iterations for model problem . Householder, Berlin, 2008 – p.22/23 Open problems Domain decomposition: preliminary results show total independence from mesh size Householder, Berlin, 2008 – p.23/23 Open problems Domain decomposition: preliminary results show total independence from mesh size Strange domains: 1D simplex (wirebasket) and QUANTUM GRAPH Householder, Berlin, 2008 – p.23/23 Open problems Domain decomposition: preliminary results show total independence from mesh size Strange domains: 1D simplex (wirebasket) and QUANTUM GRAPH 3D PDEs Householder, Berlin, 2008 – p.23/23 Open problems Domain decomposition: preliminary results show total independence from mesh size Strange domains: 1D simplex (wirebasket) and QUANTUM GRAPH 3D PDEs Utilization in Image Denoising Householder, Berlin, 2008 – p.23/23 Open problems Domain decomposition: preliminary results show total independence from mesh size Strange domains: 1D simplex (wirebasket) and QUANTUM GRAPH 3D PDEs Utilization in Image Denoising Hs s > 1 Householder, Berlin, 2008 – p.23/23 Open problems Domain decomposition: preliminary results show total independence from mesh size Strange domains: 1D simplex (wirebasket) and QUANTUM GRAPH 3D PDEs Utilization in Image Denoising Hs s > 1 Generalization to Banach spaces (K-method Peetre) Householder, Berlin, 2008 – p.23/23 Open problems Domain decomposition: preliminary results show total independence from mesh size Strange domains: 1D simplex (wirebasket) and QUANTUM GRAPH 3D PDEs Utilization in Image Denoising Hs s > 1 Generalization to Banach spaces (K-method Peetre) For which class of matrices the Schur complement is spectrally equivalent to an algebraic interpolation space matrix Householder, Berlin, 2008 – p.23/23