A posteriori error bounds in the L (L2 )-norm for the wave problem ∞ Emmanuil Georgoulis Department of Mathematics, University of Leicester, UK Joint work with O. Lakkis (Sussex), C. Makridakis (Crete) European Finite Element Fair, Warwick, 20-21 May 2010 1 Model Problem Let Ω ⊂ Rd , d = 2, 3, open polygonal domain and T > 0. Consider the problem ∂tt u − ∇ · (a∇u) = f u = u0 ut = u1 u=0 in Ω × (0, T ] in Ω × {0} in Ω × {0} on ∂Ω × (0, T ] for a = a(x) ∈ C (Ω̄) with 0 < αmin ≤ a ≤ αmax and f ∈ L∞ (0, T ; L2 (Ω)) 2 Model Problem Let Ω ⊂ Rd , d = 2, 3, open polygonal domain and T > 0. Consider the problem ∂tt u − ∇ · (a∇u) = f u = u0 ut = u1 u=0 in Ω × (0, T ] in Ω × {0} in Ω × {0} on ∂Ω × (0, T ] for a = a(x) ∈ C (Ω̄) with 0 < αmin ≤ a ≤ αmax and f ∈ L∞ (0, T ; L2 (Ω)) Problem in variational form: Find u as above such that h∂tt u, v i + ha∇u, ∇v i = hf , v i ∀v ∈ H01 (Ω), 2 Literature review A posteriori error bounds for FEM for wave problems Johnson (’93), Süli (’96,’97) duality methods for wave problem as 1st order system. Akrivis, Makridakis & Nochetto (’06) Adjerid (’02), Bernardi & Süli (’05) time-discrete schemes for 1st order systems. L∞ (H 1 )-error a posteriori estimates. Surprisingly few a posteriori results for FEM for 2nd order hyperbolic problems... 3 Literature review A posteriori error bounds for FEM for wave problems Johnson (’93), Süli (’96,’97) duality methods for wave problem as 1st order system. Akrivis, Makridakis & Nochetto (’06) Adjerid (’02), Bernardi & Süli (’05) time-discrete schemes for 1st order systems. L∞ (H 1 )-error a posteriori estimates. Surprisingly few a posteriori results for FEM for 2nd order hyperbolic problems... We are concerned with a posteriori bounds for the L∞ (L2 )-error. 3 FEM for wave problem Semi-discrete FEM: Find U ∈ Vh : hUtt , V i + a(U, V ) = hf , V i ∀V ∈ Vh , where a(U, V ) = ha∇U, ∇V i. 4 FEM for wave problem Semi-discrete FEM: Find U ∈ Vh : hUtt , V i + a(U, V ) = hf , V i ∀V ∈ Vh , where a(U, V ) = ha∇U, ∇V i. Fully-discrete FEM: for n = 1, . . . , N, find U n ∈ Vhn : h∂ 2 U n , V i + a(U n , V ) = hf n , V i ∀V ∈ Vhn , where f n := f (t n , ·), kn = t n+1 − t n and 4 FEM for wave problem Semi-discrete FEM: Find U ∈ Vh : hUtt , V i + a(U, V ) = hf , V i ∀V ∈ Vh , where a(U, V ) = ha∇U, ∇V i. Fully-discrete FEM: for n = 1, . . . , N, find U n ∈ Vhn : h∂ 2 U n , V i + a(U n , V ) = hf n , V i ∀V ∈ Vhn , where f n := f (t n , ·), kn = t n+1 − t n and n n−1 U − U n n−1 ∂U − ∂U , n 2 n , ∂U := ∂ U := k 0 n 0 kn V := π u1 for n = 1, 2, . . . , N, (1) for n = 0, with U 0 := π 0 u0 , and π 0 : L2 (Ω) → Vh0 suitable projection. 4 Dealing with “lower-than-energy” norms Semi-discrete case We shall make use the ideas of reconstruction, introduced by Makridakis & Nochetto (’03), whereby u−U = u−w | {z } “evolution error” + w −U | {z } “elliptic error” with w ∈ H01 (Ω) the elliptic reconstruction of U. “Dual” concept of the Ritz projection (a.k.a elliptic projection, Wheeler’s Trick) idea in a-priori analysis of evolution problems. 5 Dealing with “lower-than-energy” norms Semi-discrete case We shall make use the ideas of reconstruction, introduced by Makridakis & Nochetto (’03), whereby u−U = u−w | {z } “evolution error” + w −U | {z } “elliptic error” with w ∈ H01 (Ω) the elliptic reconstruction of U. “Dual” concept of the Ritz projection (a.k.a elliptic projection, Wheeler’s Trick) idea in a-priori analysis of evolution problems. A priori error bounds in the L∞ (L2 ) norm can be found in projection along with a special choice of test function. Baker (’76), utilising Ritz 5 Dealing with “lower-than-energy” norms Semi-discrete case We shall make use the ideas of reconstruction, introduced by Makridakis & Nochetto (’03), whereby u−U = u−w | {z } “evolution error” + w −U | {z } “elliptic error” with w ∈ H01 (Ω) the elliptic reconstruction of U. “Dual” concept of the Ritz projection (a.k.a elliptic projection, Wheeler’s Trick) idea in a-priori analysis of evolution problems. A priori error bounds in the L∞ (L2 ) norm can be found in projection along with a special choice of test function. Baker (’76), utilising Ritz Fully-discrete case Suitable space-time reconstruction (cf. (classical) Baker’s techinique Baker (’76). Lakkis & Makridakis (’06)), together with 5 Elliptic Reconstruction Definition We define the elliptic reconstruction w ∈ H01 (Ω) of U to be the solution of the elliptic problem a(w , v ) = hAU − Πf + f , v i ∀v ∈ H01 (Ω), where Π : L2 (Ω) → Vh denotes the orthogonal L2 -projection on the finite element space Vh , and A : Vh → Vh , denotes the (space-)discrete operator defined by for Z ∈ Vh , h−AZ , V i = a(Z , V ) ∀V ∈ Vh . 6 Error equation for ρ We decompose the error: e := U − u = ρ − , where := w − U, and ρ := w − u, 7 Error equation for ρ We decompose the error: e := U − u = ρ − , where := w − U, and ρ := w − u, Lemma We have h∂tt e, v i + a(ρ, v ) = 0 for all v ∈ H01 (Ω) 7 Error equation for ρ We decompose the error: e := U − u = ρ − , where := w − U, and ρ := w − u, Lemma We have h∂tt e, v i + a(ρ, v ) = 0 for all v ∈ H01 (Ω) or h∂tt ρ, v i + a(ρ, v ) = h∂tt , v i 7 Error equation for ρ We decompose the error: e := U − u = ρ − , where := w − U, and ρ := w − u, Lemma We have h∂tt e, v i + a(ρ, v ) = 0 for all v ∈ H01 (Ω) or h∂tt ρ, v i + a(ρ, v ) = h∂tt , v i Z Baker’s technique: set v = v (t, ·) = τ ρ(s, ·)ds t Baker (’76) 7 Abstract a-posteriori bound Theorem The following error bound holds: √ kekL∞ (0,T ;L2 (Ω)) ≤kkL∞ (0,T ;L2 (Ω)) + 2 ku0 − U(0)k + k(0)k Z T +2 kt k + Ca,T ku1 − Ut (0)k, 0 where Ca,T := min{2T , p 2CΩ /αmin }, where CΩ is the PF constant. 8 Abstract a-posteriori bound Theorem The following error bound holds: √ kekL∞ (0,T ;L2 (Ω)) ≤kkL∞ (0,T ;L2 (Ω)) + 2 ku0 − U(0)k + k(0)k Z T +2 kt k + Ca,T ku1 − Ut (0)k, 0 where Ca,T := min{2T , p 2CΩ /αmin }, where CΩ is the PF constant. Controlling , t : use L2 -norm a posteriori bounds for elliptic problems kk≤ EL2 (U, AU − Πf + f , T ). kt k≤ EL2 (Ut , (AU − Πf + f )t , T ). e.g., Verfürth (’96), Ainsworth & Oden (’01)... 8 Time reconstruction Ideally we want time-reconstruction U(t) ∈ C 1 (0, t N ) such that Utt = ∂ 2 U n 9 Time reconstruction Ideally we want time-reconstruction U(t) ∈ C 1 (0, t N ) such that Utt = ∂ 2 U n We cannot have that... 9 Time reconstruction Ideally we want time-reconstruction U(t) ∈ C 1 (0, t N ) such that Utt = ∂ 2 U n We cannot have that... Define U(t) := t − t n−1 n t n − t n−1 (t − t n−1 )(t n − t)2 2 n U + U − ∂ U , kn kn kn for t ∈ (t n−1 , t n ], n = 1, . . . , N. 9 Time reconstruction Ideally we want time-reconstruction U(t) ∈ C 1 (0, t N ) such that Utt = ∂ 2 U n We cannot have that... Define U(t) := t − t n−1 n t n − t n−1 (t − t n−1 )(t n − t)2 2 n U + U − ∂ U , kn kn kn for t ∈ (t n−1 , t n ], n = 1, . . . , N. Note: U(t) ∈ C 1 (0, t N ) and U(t n ) = U n and Ut (t n ) = ∂U n and... 9 Time reconstruction (cont’d) we have Utt = (1 + µn (t))∂ 2 U n 10 Time reconstruction (cont’d) we have Utt = (1 + µn (t))∂ 2 U n with 1 µn (t) := −6kn−1 (t − t n− 2 ), 1 t n− 2 := 1 n (t + t n−1 ) 2 10 Time reconstruction (cont’d) we have Utt = (1 + µn (t))∂ 2 U n with 1 µn (t) := −6kn−1 (t − t n− 2 ), 1 t n− 2 := 1 n (t + t n−1 ) 2 Remark (vanishing moment property) The remainder µn (t) satisfies Z tn µn (t) dt = 0, t n−1 10 Elliptic reconstruction Definition Let U n , n = 0, . . . , N, be the fully-discrete solution computed by the method, Πn : L2 (Ω) → Vhn be the orthogonal L2 -projection, and An : Vhn → Vhn to be the discrete operator defined by for q ∈ Vhn , hAn q, χi = a(q, χ) ∀χ ∈ Vhn . We define the elliptic reconstruction w n ∈ H01 (Ω) such that a(w n , v ) = hg n , v i ∀v ∈ H01 (Ω), with g n := An U n − Πn f n + f¯n , R tn where f¯0 (·) := f (0, ·) and f¯n (·) := kn−1 t n−1 f (t, ·)dt for n = 1, . . . , N. (Special care for n = 1, also) 11 Elliptic reconstruction Definition Let U n , n = 0, . . . , N, be the fully-discrete solution computed by the method, Πn : L2 (Ω) → Vhn be the orthogonal L2 -projection, and An : Vhn → Vhn to be the discrete operator defined by for q ∈ Vhn , hAn q, χi = a(q, χ) ∀χ ∈ Vhn . We define the elliptic reconstruction w n ∈ H01 (Ω) such that a(w n , v ) = hg n , v i ∀v ∈ H01 (Ω), with g n := An U n − Πn f n + f¯n , R tn where f¯0 (·) := f (0, ·) and f¯n (·) := kn−1 t n−1 f (t, ·)dt for n = 1, . . . , N. (Special care for n = 1, also) w (t) := n−1 n t − t n−1 n t n − t n−1 (t − t n−1 )(t n − t)2 2 n w + w − ∂ w , kn kn kn 11 A posteriori bound for the fully-discrete wave problem Theorem √ kekL∞ (0,t N ;L2 (Ω)) ≤kkL∞ (0,t N ;L2 (Ω)) + 2 ku0 − U(0)k + k(0)k 4 Z tN X kt k + +2 ηi (t N ) + Ca,N ku1 − V 0 k, 0 where Ca,N := min{2t N , p (2) i=1 2CΩ /αmin }. 12 A posteriori bound for the fully-discrete wave problem We have made use of the following notation: mesh change indicator η1 (τ ) : = m−1 X Z tj j=1 + m−1 X k(I − Πj )Ut k + t j−1 Z τ k(I − Πm )Ut k t m−1 (τ − t j )k(Πj+1 − Πj )∂U j k + τ k(I − Π0 )V 0 (0)k j=1 evolution error indicator Z τ kGk, η2 (τ ) := 0 where G : (0, T ] → R with G|(t j−1 ,t j ] := G j , j = 1, . . . , N and G j (t) := with γj := γj−1 + (t j − t)2 j (t j − t)4 (t j − t)3 2 j ∂g − − ∂ g − γj , 2 4kj 3 kj2 j 2 ∂g + kj3 2 j 12 ∂ g , j = 1, . . . , N, γ0 = 0; 13 A posteriori bound for the fully-discrete wave problem data error indicator m−1 Z j 1/2 Z τ 1/2 1 X t 3 ¯j 2 3 ¯m + km kf − f k2 ; kj kf − f k η3 (τ ) := 2π t m−1 t j−1 j=1 time reconstruction error indicator Z m−1 Z j 1/2 1 X t 3 j 2 j 2 1/2 τ 3 . η4 (τ ) := kj kµ ∂ U k + km kµm ∂ 2 U m k2 2π t j−1 t m−1 j=1 14 A posteriori bound for the fully-discrete wave problem Finally, we bound the and t terms using Lemma We have kkL∞ (0,t N ;L2 (Ω)) + √ 2k(0)k ≤ δ1 (t N ) + √ 2Cel E 0 , where n 8k 1 δ1 (t N ) := max Cel E(V 0 , ∂g 0 , T 0 ), 27 35 31 o kj −1 ¯j + max max Cel E j + CΩ αmin kf − f j k , 27 27 1≤j≤N kj−1 0≤j≤N with E j := E(U j , Aj U j − Πj f j + f j , T j ), j = 0, 1, . . . , N. 15 A posteriori bound for the fully-discrete wave problem Lemma We have Z tN kt k ≤ δ2 (t N ), 0 where δ2 (t N ) := N 2X −1 (2kj + kj+1 ) Cel E∂j + CΩ αmin k∂f j − ∂ f¯j k , 3 j=0 with E∂j := E(∂U j , ∂(Aj U j ) − ∂(Πj f j ) + ∂f j , T̂ j ), j = 0, 1, . . . , N. 16 Outlook Current work: (G., Lakkis & Makridakis, in preparation) explicit methods for wave problem high order time-stepping for wave problem 17