MATH 536 - Numerical Solution of Partial Differential Equations
Transcription
MATH 536 - Numerical Solution of Partial Differential Equations
MATH 536 - Numerical Solution of Partial Differential Equations Peter Minev March 31, 2016 ii Contents 1 Systems of Linear Equations 1.1 Direct Methods – Gaussian Elimination . . . . . 1.1.1 Forward Elimination . . . . . . . . . . . . 1.2 Gaussian Elimination with Pivoting . . . . . . . 1.3 Special Cases . . . . . . . . . . . . . . . . . . . . 1.3.1 Cholesky Decomposition . . . . . . . . . . 1.3.2 Thomas Algorithm . . . . . . . . . . . . . 1.4 Matrix Norms . . . . . . . . . . . . . . . . . . . . 1.5 Error Estimates . . . . . . . . . . . . . . . . . . . 1.6 Iterative Method Preliminaries . . . . . . . . . . 1.7 Iterative Methods . . . . . . . . . . . . . . . . . . 1.7.1 Jacobi’s Method . . . . . . . . . . . . . . 1.7.2 Gauss-Seidel (GS) Iteration . . . . . . . . 1.7.3 Successive Over-relaxation (SOR) Method 1.7.4 Conjugate Gradient Method . . . . . . . . 1.7.5 Steepest Descent Method . . . . . . . . . 1.7.6 Conjugate Gradient Method (CGM) . . . 1.7.7 Preconditioning . . . . . . . . . . . . . . . 2 Solutions to Partial Differential Equations 2.1 Classification of Partial Differential Equations . 2.1.1 First Order linear PDEs . . . . . . . . . 2.1.2 Second Order PDE . . . . . . . . . . . . 2.2 Difference Operators . . . . . . . . . . . . . . . 2.2.1 Poisson Equation . . . . . . . . . . . . . 2.2.2 Neumann Boundary Conditions . . . . . 2.3 Consistency and Convergence . . . . . . . . . . 2.4 Advection Equation . . . . . . . . . . . . . . . 2.5 Von Neumann Stability Analysis . . . . . . . . 2.6 Sufficient Conditions for Convergence . . . . . 2.7 Parabolic PDEs – Heat Equation . . . . . . . . 2.7.1 FC Scheme . . . . . . . . . . . . . . . . 2.7.2 BC scheme . . . . . . . . . . . . . . . . 2.7.3 Crank-Nicolson Scheme . . . . . . . . . 2.7.4 Leapfrog Scheme . . . . . . . . . . . . . 2.7.5 DuFort-Frankel Scheme . . . . . . . . . 2.8 Advection-Diffusion Equation . . . . . . . . . . 2.8.1 FC Scheme . . . . . . . . . . . . . . . . 2.8.2 Upwinding Scheme (FB) . . . . . . . . . iii . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1 1 2 4 4 6 6 8 8 10 10 10 11 11 12 13 14 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 23 23 23 24 25 26 26 29 31 33 34 34 36 36 37 38 38 38 40 iv 3 Introduction to Finite Elements 3.1 Weighted Residual Methods . . . . . . . 3.1.1 Sobolev spaces . . . . . . . . . . 3.1.2 Weighted Residual Formulations 3.1.3 Collocation Methods . . . . . . . 3.2 Weak Methods . . . . . . . . . . . . . . 3.3 Finite Element Method (FEM) . . . . . 3.4 Gaussian Quadrature . . . . . . . . . . . 3.5 Error Estimates . . . . . . . . . . . . . . 3.6 Optimal Error Estimates . . . . . . . . . 3.6.1 Other Boundary Conditions . . . 3.7 Transient Problems . . . . . . . . . . . . 3.8 Finite Elements in Two Dimensions . . 3.8.1 Finite Element Basis Functions . 3.8.2 Discrete Galerkin Formulation . CONTENTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41 41 41 42 43 43 47 48 49 50 52 53 58 59 60 Chapter 1 Systems of Linear Equations 1.1 1.1.1 Direct Methods – Gaussian Elimination Forward Elimination (k-th step): [k−1] [k] aij = aij [k−1] ℓi,k−1 = [k−1] ai,k−1 , [k−1] ak−1,k−1 [k−1] [k] [k−1] [k−1] − ak−1,j ℓi,k−1 , bi = bi In n − 1 steps: [k−1] [k−1] [1] 0 j = k − 1, . . . , n k = 2, . . . , n i = k, . . . , n − bk−1 ℓi,k−1 , a11 0 . . . or i = k, . . . , n, [1] a12 [2] a22 .. . 0 ... ... .. . i = k, . . . , n [1] a1n [2] a2n .. . [n] . . . ann x1 x2 .. . xn = [1] b1 [2] b2 .. . [n] bn [u] [x] = [c] Operations Count n−1 X Where we have used 3 2 (n − j) +2 (n − j) (n − j + 1) = 2n + n − 7n | {z } | {z } 6 2 6 j=1 divisions multiply and sum m X j=1 j= m (m + 1) , 2 m X j2 = j=1 m (m + 1) (2m + 1) 6 Backward Substitution n X 1 [i] [i] aij xj xi = [i] bi − aii j=i+1 1 2 CHAPTER 1. SYSTEMS OF LINEAR EQUATIONS Operation Count n X i=1 2 2 (n − i) + |{z} 1 =n | {z } Multiplication and Subtraction division Matrix Representations Define: 1 0 0 Lk = 0 . .. 0 .. . A[2] = L1 A[1] , 0 1 ... . . . −ℓk+1,k .. . 0 Lemma 1.1.1 ................ ... −ℓnk 0 . . . . . . 0 0 . . . 0 1 . . . 0 .. . 0 ... 1 A[1] = A, b[1] = b b[2] = L1 b[1] Therefore, A[n] = U = Ln−1 Ln−2 · · · L1 A or −1 −1 A = L−1 1 L2 · · · Ln−1 U = LU Proposition 1.1.2 −1 L = L−1 1 · · · Ln−1 is a lower triangular matrix with unit diagonal. Definition A Permutation matrix is a matrix that has exactly one nonzero entry in each row and column, and this entry is equal to 1. The elementary permutation matrix (EPM) is defined as a matrix produced from the identity matrix by interchanging exactly one pair k, m of its columns. We denote it as Pkm = Pk . |{z} k<m For example: P23 1 0 = 0 0 0 0 1 0 0 1 0 0 0 0 0 1 Proposition 1.1.3 If A ∈ Rn×n and Pk,m ∈ Rn×n is EPM, then Pk,m A differs from A by an interchange of rows k and m. 1.2 Gaussian Elimination with Pivoting A[n] = U = Ln−1 Pn−1 . . . L1 P1 A Where U is upper triangular, Lk is defined as before, and Pk = Pkm with m ≥ k. Equivalently, −1 A = P1 L1−1 P2 L−1 2 · · · Pn−1 Ln−1 U {z } | L∗ 1.2. GAUSSIAN ELIMINATION WITH PIVOTING 3 −1 Lemma 1.2.1 If i ≥ j > k and Pj = Pji then Pj L−1 k Pj is produced form Lk by interchanging the j-th and i-th entry in the k-th column. Proof .. . 1 .. . ℓjk .. . .. . 1 .. ℓik .. . . 1 .. . → .. . 1 .. . .. . ℓik .. . 0 ℓjk .. . 1 1 .. . 0 .. . .. → . 1 .. . .. . ℓik .. . 1 .. ℓjk .. . . 1 .. . Theorem 1.2.2 Consider P = Pn−1 . . . P1 . Then P is a permutation matrix and P A = P L∗ U = LU , with L being a lower triangular matrix with unit diagonal. Proof −1 −1 L = P P1 L−1 1 P2 L2 · · · Pn−1 Ln−1 = Pn−1 Pn−2 · · · P1 P1 L−1 P L−1 · · · Pn−1 L−1 n−1 | {z } 1 2 2 I −1 −1 = Pn−1 Pn−2 · · · P2 L−1 1 P2 P3 · · · Pn−1 Pn−1 · · · P3 L2 · · · Pn−1 Ln−1 {z } | = L−1 1∗ Pn−1 · · · P3 L−1 2 P3 I −1 · · · Pn−1 Ln−1 −1 −1 −1 = · · · = L−1 1∗ L2∗ · · · Ln−2∗ Ln−1 [k] Theorem 1.2.3 The pivot entries akk , k = 1, 2, . . . , n − 1 are nonzero if and only if Âk are non-singular for k = 1, . . . , n − 1, where a11 . . . a1k a21 . . . a2k Âk = . . . .. . . .. ak1 . . . akk is called the k th main principle sub-matrix. [k] Proof (i) Suppose first that all akk 6= 0. Then since Âk = L̂k Ûk it follows that det(Âk ) = det(Ûk ) = [k] [k] a11 . . . akk 6= 0 (ii) Now suppose that det(Âk ) 6= 0. Then, using a induction argument, we have: [1] [1] (a) a11 = Â1 and therefore a11 6= 0. [1] [k] [k+1] (b) Assume that a11 , . . . , ak 6= 0. Then if A11 is the k + 1-th principle sub matrix of A[k+1] we can re-write it in the following block decomposed form # " [k+1] [k+1] A A 11 12 A[k+1] = [k+1] [k+1] . A21 A22 [k+1] It is easy to see that A11 upper triangular i.e. [1] [k+1] = (Lk )11 . . . (L1 )11 A11 . Therefore, det(A11 [k+1] A11 [1] a11 0 = .. . 0 [1] . . . a1,k+1 [2] . . . a2,k+1 , .. .. . . [k+1] . . . ak+1,k+1 [1] [k+1] ) = det(A11 ) 6= 0. But A11 is 4 CHAPTER 1. SYSTEMS OF LINEAR EQUATIONS [k+1] [1] [k+1] [k+1] and then det(A11 ) = a11 . . . ak+1,k+1 . This, together with the induction hypothesis yields that ak+1,k+1 6= 0 which completes the induction proof. Definition A matrix is strictly diagonally dominant if X |aii | > |aij | j6=i Corollary 1.2.4 If a matrix is strictly diagonally dominant then no pivoting is necessary. Definition Matrix A is symmetric positive definite (spd) if 1. A = AT 2. v T Av ≥ 0 for all v ∈ Rn 3. v T Av = 0 if and only if v ≡ 0 Corollary 1.2.5 If a matrix is spd then no pivoting is necessary. Definition The spectrum of a matrix A, S (A), is the set of all eigenvalues of the matrix. Theorem 1.2.6 If A ∈ Rn×n is symmetric the following statements are equivalent 1. A is positive definite, 2. S (A) contains only positive real numbers, and 3. Every principle sub-matrix is positive definite. Theorem 1.2.7 (Gershgorin) The spectrum S (A) is enclosed in ! ! n n \ [ [ ′′ ′ Di Di i=1 i=1 where X Di′ = z ∈ C : |z − aii | ≤ |aij | j6=i X Di′′ = z ∈ C : |z − aii | ≤ |aji | j6=i 1.3 1.3.1 Special Cases Cholesky Decomposition Theorem 1.3.1 If A ∈ Rn×n is spd then there exists a lower triangular matrix C such that CC T = A. Furthermore, the diagonal entries of C are positive. Proof Use induction on n: n=1 : C = [c11 ] = Assume the theorem holds for all matrices in Rn×n . A A = Tn a a an+1,n+1 √ a11 1.3. SPECIAL CASES 5 where an+1,n+1 > 0. By the induction hypothesis: An = Cn CnT Now try to find [X] such that Cn CC = XT T 0 c X A = Tn c a CnT 0 a an+1,n+1 or such that Cn X = a ⇒ X = Cn−1 a and X T X + c2 = an+1,n+1 But, is c > 0? c= q an+1,n1 − X T X T T X T X = Cn−1 a Cn−1 a = aT Cn−1 Cn−1 a = aT A−1 n a therefore Then define x := ⇒ h A−1 n a T , −1 iT T an+1,n+1 − aT A−1 n a = an+1,n+1 − X X 6= 0 so that xT Ax = an+1,n+1 − aT A−1 n a> 0 Cholesky Algorithm √ 1. c11 ← a11 2. For i = 2, 3, . . . , n 3. ci1 ← ai1 /c11 4. End i 5. For j = 2, 3, . . . , n − 1 6. cjj v u j−1 u X ← tajj − c2jk k=1 7. For i = j + 1, . . . , n 8. 1 cij ← cjj aij − j−1 X cik cjk k=1 9. End i 10. End j 11. cnn v u n−1 X u c2nk ← tann − k=1 12. End ! 6 CHAPTER 1. SYSTEMS OF LINEAR EQUATIONS 1.3.2 Thomas Algorithm Definition A band matrix is of the form a11 a1,2 ... a1,l 0 .. .. .. . . . a21 .. .. .. .. . . . . ak,1 . . . ak,k−1 akk ak,k+1 .. .. .. 0 . . . .. .. .. .. . . . . . . . .. .. .. . .. .. .. . . . .. .. . 0 .......................... ........................... .. . .. .. . . . . . ak,k+l−1 0 ... .. .. .. . . . .. .. . .. . .. .. .. . . .. . . .. . .. . .. . .. .. . . . . . an,n−1 . 0 an,n−k+1 0 .. . .. . 0 .. . 0 an−l+1,n .. . an−1,n an,n A band matrix may be stored compactly as an l + k − 1 × n matrix of the form: k−1 }| { z 0 ... 0 a11 a12 . . . a1l . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ak1 . . . ak,k−1 akk ak,k+1 . . . ak,k+l−1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . an,n−k+1 . . . an,n−1 ann .{z .. 0} |0 l−1 Thomas Algorithm For the linear system b1 a2 0 .. . .. . . .. 0 1. c1 b2 a3 .. . 0 c2 b3 .. . ................ 0 ........... c3 0 ... .. .. . . .. .. . .. .. . . ............ .. . an−1 0 . bn−1 an 2. cj−1 [f ] bj−1 x1 d1 x2 d2 .. = .. . . 0 xn dn cn−1 bn [f ] [f ] b j = b j − aj [f ] [f ] and dj = dj − aj [f ] [f ] xj−1 = dj−1 − bj xj , 1.4 0 0 0 .. . for dj−1 [f ] bj−1 , for j = 2, . . . , n j = n, . . . , 2 Matrix Norms Definition If A ∈ Rn×n then the mapping Rn×n → R is called a norm of A := ||A|| if and only if: 1. ||A|| ≥ 0 and ||A|| = 0 if and only if A = 0, 1.4. MATRIX NORMS 7 2. ||λA|| = |λ| ||A|| for all λ ∈ R, 3. ||A + B|| ≤ ||A|| + ||B||. If in adddition, ||AB|| ≤ ||A|| ||B|| the norm is called multiplicative. In the sequel, we make use of multiplicative norms only, and therefore, the notion of a norm presumes a multiplicative norm. Definition If A ∈ Rn×n , then for a vector norm ||·|| : Rn → R, ||Ax|| ||x||6=0 ||x|| ||A|| := sup is called a subordinate matrix norm. Theorem 1.4.1 Let A ∈ Rn×n . Then, 1. ||A||∞ := n X ||Ax||∞ |aij | = max 1≤i≤n ||x||∞ 6=0 ||x||∞ j=1 sup 2. n X ||Ax||1 |aij | = max 1≤j≤n ||x||1 6=0 ||x||1 i=1 ||A||1 := sup 3. ||Ax||2 = ||x||2 6=0 ||x||2 ||A||2 := sup Proof ||Ax||∞ therefore, q ρ (AT A) X n X n |aij | max |xj | = M ||x||∞ aij xj ≤ max = max 1≤j≤n 1≤i≤n 1≤i≤n j=1 j=1 ||A||∞ ≤ M Let i be the index for which max 1≤i≤n n X j=1 |aij | = M n is achieved. Choose x = {xj }j=1 xj := aij / |aij | : aij = 6 0 0 : aij = 0 so that ||x||∞ = 1 and ||Ax||∞ = M . Therefore, ||A||∞ = M 8 CHAPTER 1. SYSTEMS OF LINEAR EQUATIONS 1.5 Error Estimates Consider two vectors x1 and x2 where Ax1 = b and Ax2 = b + r then x1 = A−1 b and x2 = A−1 (b + r) and it follows immediately that Then, Therefore, ||x2 − x1 || = A−1 r −1 −1 A r A ||r|| ||x2 − x1 || ||x2 − x1 || ≤ = ≤ ||A|| ||x1 || ||Ax1 || ||b|| ||b|| ||r|| ||x2 − x1 || ≤ ||A|| A−1 ||x1 || | {z } ||b|| c(A) This provides means of estimating the error introduced in the solution of a linear system due to roundoff errors or uncertainty in the right-hand-side or the matrix of the system. 1.6 Iterative Method Preliminaries ∞ Definition If xk ∈ Rn then the sequence {xk }k=1 converges to x ∈ Rn in a norm ||·|| i.e. lim xk = x k→∞ if lim ||x − xk || = 0. k→∞ ∞ Definition If Ak ∈ Rn×n then the sequence {Ak }k=1 converges to A ∈ Rn×n in a norm ||·|| i.e. lim Ak = A k→∞ if lim ||A − Ak || = 0. k→∞ Consider the matrix power series f (A) ∼ ∞ X ak Ak (1.1) k=0 It is convergent if and only if lim K→∞ Consider also the numerical power series K X ak Ak = f (A) k=0 f (λ) ∼ ∞ X k=0 ak λk (1.2) 1.6. ITERATIVE METHOD PRELIMINARIES 9 Theorem 1.6.1 The matrix power series (1.1) is convergent if for all λi ∈ S (A) |λi | < ρ, where ρ is the radius of convergence of (1.2) and S (A) is the spectrum of A. It is divergent if |λi | > ρ for some i. Proof From Jordan’s theorem, there exists nonsingular C such that: λk 0 .. B = C −1 AC, B = diag (Bk ) , Bk = . . .. 0 1 λk .. . 0 1 .. . ... ... .. . .. .. . ....... . 0 0 0 1 λk where Bk ∈ Rnk ×nk and nk is the multiplicity of the eigenvalue λk . Then, f (A) is convergent if and only if f (B) = ∞ X k ak B = k=0 ∞ X ak C −1 k A C=C −1 k=0 ∞ X k=0 k ak A ! C is convergent as well. Now, let’s look at Bki : 2 λk 2λk 1 ... 0 0 λ2k 2λk . . . 0 Bk2 = . . . . . . . . . . . . . . . . . . . . . . . , 0 . . . . . . . . . . . . . . λ2k i−1 i i λk . . . . . . . . . . . . nki−1 λki−nk +1 1 λk i−1 i 0 λik . . . nki−2 λki−nk +2 1 λk Bki = . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 0 ...................... λik so that the i-th block of the m-th partial sum of f (B) is (n −1) 1 1 ′ fm (λi ) . . . (ni −1)! fm i (λi ) fm (λi ) 1! m X (n −2) 1 fm i (λi ) fm (λi ) . . . (ni −2)! fm (Bi ) = ak Bik = 0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . k=0 0 ... 0 fm (λi ) where f (B) = m X ak B k k=0 Therefore, fm (B) is convergent if and only if fm (λi ) = m X ak λki k=0 (l) (and therefore fm (λi )) is convergent for i = 1, . . . , n and l = 1, . . . , ni . fm (λi ) is convergent if |λi | is smaller than ρ, the radius of convergence of (1.2). Corollary 1.6.2 f (A) = I + A + A2 + · · · + Am + · · · is convergent if and only if |λk | < 1 for all k = 1, . . . , n and A ∈ Rn×n . Corollary 1.6.3 Since |λk | ≤ ||A|| for any subordinate norm ||·|| then f (A) is convergent if ||A|| < 1. 10 1.7 CHAPTER 1. SYSTEMS OF LINEAR EQUATIONS Iterative Methods Consider Ax = b If A = B + C then Ax = b x = x + C −1 (b − Ax) ⇔ and in this form we may solve the problem iteratively: xk+1 = xk + C −1 b − Axk (1.3) Theorem 1.7.1 The iteration (1.3) is convergent if and only if ρ I − C −1 A < 1. Proof xk+1 = I − C −1 A xk + C −1 b = D2 xk−1 + (I + D) C −1 b = . . . | {z } =D D k+1 0 x + I + D + D2 + . . . + Dk C −1 b If ρ I − C −1 A < 1 then I + D + . . . + Dk + . . . is convergent and therefore Dk →k→∞ 0 Definition R (D) = − log10 ρ (D) is called the convergence rate of the iteration. 1.7.1 Jacobi’s Method C = diag (A) 12 13 0 − aa11 − aa11 · · · − aa1n 11 .. .. D = I − C −1 A = ... . . an1 an2 an3 − ann − ann − ann · · · 0 Theorem 1.7.2 If A is strictly diagonally dominant then Jacobi is convergent. Proof ||D||∞ Then X aij < 1. = max aii i j6=i |λi | ≤ ||D||∞ < 1 1.7.2 . Gauss-Seidel (GS) Iteration a11 0 C = L = ... . . . a1n . . . ann B = U =A−L X k+1 = X k + C −1 b − AX k Theorem 1.7.3 If A is strictly diagonally dominant then Gauss-Seidel iteration is convergent. 1.7. ITERATIVE METHODS 11 Proof The formula for the i-th component of xk+1 is: xk+1 =− i X aij j<i aii xk+1 − j X aij j>i aii xkj + bi aii The exact solution clearly satisfies the iteration i.e. xi = − X aij j<i aii xj − X aij j>i aii xj + bi aii The error ǫk+1 = xk+1 − x for Ax = B is found as X aij X aij ǫk+1 =− ǫk+1 − ǫkj i j a a ii ii j<i j>i then Using induction on i we find that k+1 X aij k+1 X aij k ǫj + ǫ ≤ ǫ i aii aii j j>i j<i Therefore, X aij k+1 ≤ ǫ ǫk < ǫk i aii ∞ ∞ j6=i k+1 ǫ i ∞ < ǫk ∞ so that the error must strictly decrease with each iteration. Then the error must go to zero because the exact solution is the unique fixed point of the iteration (why is it unique?). 1.7.3 Successive Over-relaxation (SOR) Method 1 D ω A = L+D+U D = diag (A) C = L+ xk+1 = − ω −1 D + L −1 −1 b. 1 − ω −1 D + U xk + ω −1 D + L Note that in this case L has zeros on the main diagonal and is strictly lower triangular. Theorem 1.7.4 (Ostrowski-Reich) If A ∈ Rn×n is spd then SOR is convergent iff 0 < ω < 2. 1.7.4 Conjugate Gradient Method Theorem 1.7.5 If A ∈ Rn×n is spd then u is a solution to the linear system Au = b if and only if u minimizes the function: 1 F (v) = v T Av − bT v 2 Proof Let u∗ satisfy Au∗ = b then 1 1 T u Au − bT u − u∗ T Au∗ + bT u∗ 2 2 1 1 1 T = uT Au − u∗ T Au + u∗ T Au∗ = (u − u∗ ) A (u − u∗ ) ≥ 0 2 2 2 F (u) − F (u∗ ) = 12 CHAPTER 1. SYSTEMS OF LINEAR EQUATIONS then 1 (u − u∗ )T A (u − u∗ ) = 0 2 so that F (u) ≥ F (u∗ ) and F (u) = F (u∗ ) ⇔ ⇔ u = u∗ u = u∗ Definition If A is spd then for any two vectors u, v ∈ Rn , uT Av defines the energy inner product induced by A denoted by: hu, viA = uT Av. Note that in the rest of the notes at some occasions (u, v)A is also used to denote the energy inner product. 1.7.5 Steepest Descent Method Note that any iteration for finding a solution to a linear system Au = b can be written in the form: uk+1 = uk + αk pk where αk ∈ R is called iteration step, pk ∈ Rn is called the search direction. For example, in the case of matrix splitting methods the search direction is chosen to be pk = C −1 rk , with rk = b − Auk being called the residual of the iterate uk , and αk = 1. The step does not need to be constant and the gradient-type methods choose it at each iteration step so that this choice minimizes the function F (v). The basic gradient-type iterative algorithm therefore can be written as: 1. Choose an initial search direction p0 = r0 = b − Au0 2. For k = 1, 2, 3, . . . do: (a) Find αk ∈ R such that uk+1 = uk + αk pk minimizes F over the line uk + αpk (b) New iterate uk+1 = uk + αk pk (c) The new residual is rk+1 = b − Auk+1 (d) Find the new search direction pk+1 . (e) Let k = k + 1 and repeat item 2 until convergence. To find αk we minimize F along the search direction pk : d d F uk + αk pk = ∇F uk + αk pk · uk + αk pk dαk dαk T T k = A u + αk pk − b pk = Auk − b pk + αk pTk Apk = −rkT pk + αk pTk Apk = 0 so that αk = rkT pk pTk Apk 1.7. ITERATIVE METHODS 13 Note that for this choice of αk we have: T T k Auk+1 − b pk = A u + αk pk −b pk = −rkT pk + αk pTk Apk = 0 | {z } {z } | −rk+1 uk+1 or T −rk+1 pk = ∇F (uk+1 ) pk = 0 In the steepest descent method we choose pk = rk = b − Auk , because this is the direction of the negative gradient and in this direction the function F (v) is clearly nonincreasing. Unfortunately, this choice of search direction is not optimal. The basic reason for this is that the step αk is chosen so that the iterate uk+1 is a minimum of F (v) along the direction pk . However, the next minimum, along the direction pk+1 , is not guaranteed to be smaller than or equal to the current one. This issue is discussed in the next section. 1.7.6 Conjugate Gradient Method (CGM) Definition uk is optimal with respect to direction p 6= 0 iff F uk ≤ F uk + λp for all λ ∈ R. Lemma 1.7.6 uk is optimal with respect to p iff pT rk = 0. Proof uk is optimal with respect to p iff F uk + λp has a minimum at λ = 0, i.e. ∂F k = pT Auk − b + λpT Apλ=0 = 0 u + λp ∂λ | {z } λ=0 −rk This is obviously satisfied iff pT rk = 0. We know that the choice of αk guarantees that uk+1 is optimal with respect to pk and this applies in particular to the steepest descent method for which pk = −rk . However, uk+2 may not be optimal with respect to pk i.e. the next iteration can undo some of the minimization work of the previous iteration. Indeed: uk+2 = uk+1 + αk+1 rk+1 which gives rk+2 = rk+1 − αk+1 Ark+1 ⇒ rkT rk+2 = rkT rk+1 −αk+1 rkT Ark+1 | {z } ⇒ pTk rk+2 = −αk+1 rkT Ark+1 6= 0 =0 unless A = cI. Note that rkT rk+1 = 0 for the steepest descent method since rkT rk+1 = rkT (b − A(uk + αk rk )) = rkT (rk − αk Ark ) = rkT (rk − rkT rk Ark ) = 0. ||rk ||2A To find a different search direction pk+1 that maintains the optimality of uk+2 w.r.t. pk , we multiply by pk the equation for the residual: rk+2 = rk+1 − αk+1 Apk+1 14 CHAPTER 1. SYSTEMS OF LINEAR EQUATIONS and require the product to be zero: 0 = pTk rk+2 = pTk rk+1 +αk+1 pTk Apk+1 ⇔ pTk Apk+1 = 0. | {z } =0 For the conjugate gradient method we require that uk+2 is also optimal with respect to pk i.e. we choose the search direction pk+1 from the condition pTk Apk+1 = 0. Definition If u, v ∈ Rn are such that hu, viA = 0 for A (which is spd) then u, v are A-conjugate. This means that pk+1 should be A-conjugate to pk . We search for pk+1 in the form pk+1 = rk+1 + βk pk and hrk+1 , pk iA then from the condition pTk Apk+1 = 0 we easily obtain that βk = − 2 ||pk ||A Algorithm (Basic CGM procedure) If A ∈ Rn×n is spd and u0 is an initial guess: 1. r0 ← b − Au0 2. p0 = r0 3. For k = 1, 2, . . . , m: 2 T 4. αk−1 ← rk−1 pk−1 / ||pk−1 ||A 5. uk ← uk−1 + αk−1 pk−1 6. rk ← rk−1 − αk−1 Apk−1 7. pk ← rk − hrk , pk−1 iA 2 ||pk−1 ||A pk−1 8. Next k 9. End As it will be demonstrated below, if all arithmetic operations are exact, the algorithm computes the exact solution in a finite number of steps (finite termination property). Later, we will show that the error decreases as: "p #k k c (A) − 1 k u − u = ǫ ≤ 2 ||ǫ0 || A p A A c (A) + 1 If c (A) ≫ 1 then the convergence is slow. Therefore, the algorithm is often modified by formally multiplying the original system by a matrix that has a spectrum close to the spectrum of A and performing the CG method on the modified system which has the same solution as the original one. This process is called preconditioning. 1.7.7 Preconditioning Instead of Au = b we solve Ãũ = b̃ where à = B −1 T AB −1 , ũ = Bu, b̃ = B −1 T b Since the preconditioner B T B must be an approximation to the matrix A, a natural choice for B should be an approximation to the square root of A. Since A is s.p.d. its square root is also s.p.d. and therefore we assume that B T = B. The preconditioner B 2 must satisfy somewhat contradicting requirements since on one hand it 1.7. ITERATIVE METHODS 15 should be a good approximation of A and on the other, it should be much easier to solve a linear system with a matrix B 2 than with A. The second requirement follows from the fact that, as it will become clear below, on each iteration of the CG method with preconditioning, we will need to solve one system with B 2 . For the system Au = b we know that u : F (u) = minn F (v) v∈R Now, if ṽ = Bv then 1 T 1 −1 T −1 B ṽ A B ṽ − bT B −1 ṽ v Av − bT v = 2 2 1 T −1 −1 = ṽ |B {z AB } ṽ − B −1 b ṽ = F̃ (ṽ) 2 | {z } F (v) = à b̃ The preconditioned method should compute the new iterate from: ũk = ũk−1 + α̃k−1 p̃k−1 , with α̃k−1 = T r̃k−1 p̃k−1 ||p̃k−1 ||à . The search direction for the preconditioned method should be computed as p̃k = β̃k−1 p̃k−1 +r̃k where the residual hr̃ ,p̃k ià . These formulae are not convenient for practical r̃k is given by r̃k = r̃k−1 − α̃k−1 Ãp̃k−1 , and β̃k = − k+1 ||p̃ ||2 k à computations since they require to somehow compute the inverse of the preconditioner, (B 2 )−1 . Fortunately, it appears that the preconditioned method can be implemented almost exactly as the original CGM modifying only the computation of the search direction to: T zk z }| { 2 −1 rk Apk−1 B −1 rk − pk = B 2 pk−1 , (1.4) pTk−1 Apk−1 | {z } zk and defining the initial iterate as u0 = B −1 ũ0 and the initial search direction as p0 = (B 2 )−1 r0 . With this modification, the residuals rk , r̃k , search directions pk , p̃k , and iterates uk , ũk , of the modified CG algorithm and the preconditioned CG algorithm, verify the relations: rk = Br̃k , pk = B −1 p̃k , uk = B −1 ũk . Indeed, using the assumption that rk−1 = Br̃k−1 , pk−1 = B −1 p̃k−1 , and uk−1 = B −1 ũk−1 (these assumptions are trivially verifiable at k = 1), we obtain that: α̃k−1 = T r̃k−1 p̃k−1 2 ||p̃k−1 ||à = T rk−1 B −1 Bpk−1 T (Bpk−1 ) B −1 AB −1 Bpk−1 = T rk−1 pk−1 2 ||pk−1 ||A = αk−1 , and subsequently that: rk = Br̃k and uk = B −1 ũk . For example, ũk = ũk−1 + α̃k−1 p̃k−1 = B(uk−1 + αk−1 pk−1 ) = Buk . This immediately yields that rk = Br̃k . Then, if pk is computed from (1.4) we obtain, multiplying it by B, that: rT B −1 B −1 AB −1 Bpk−1 Bpk−1 , Bpk = B −1 rk − kT pk−1 BB −1 AB −1 Bpk−1 or, using the induction assumptions: Bpk = r̃k − hr̃k , p̃k−1 ià 2 ||p̃k−1 ||à p̃k−1 = p̃k Thus, using an induction argument we establish that rk = Br̃k , pk = B −1 p̃k , uk = B −1 ũk , and therefore, α̃k = αk , for all positive integer k, Then, it is clear that the modified CG algorithm does not require the knowledge of à to proceed. It needs only one extra step for the computation of the new search direction. Instead of using rk for its computation, we need to use zk that is a solution of the system B 2 zk = rk . The conjugate gradient method with preconditioning is given by: 16 CHAPTER 1. SYSTEMS OF LINEAR EQUATIONS Algorithm Preconditioned CGM: 1. k ← 0 2. r0 ← b − Au0 3. p0 = (B 2 )−1 r0 4. For k = 1, 2, . . . , m: 2 T 5. αk−1 ← rk−1 pk−1 / ||pk−1 ||A 6. uk ← uk−1 + αk−1 pk−1 7. rk ← rk−1 − αk−1 Apk−1 8. if ||rk ||2 ≥ τ then 9. solve B 2 zk = rk 10. pk ← z k − hzk , pk−1 iA 2 ||pk−1 ||A pk−1 11. k = k + 1 12. Go to (5) 13. End if 14. End Lemma 1.7.7 If A ∈ Rn×n is spd, then for m = 0, 1, 2, . . . we have span {p0 , p1 , . . . , pm } = span {r0 , r1 , . . . , rm } = span {r0 , Ar0 , . . . , Am r0 } Proof 1. For m = 0, this is trivial. 2. Suppose that for m = k we have span {p0 , p1 , . . . , pk } = span {r0 , r1 , . . . , rk } = span r0 , Ar0 , . . . , Ak r0 . 3. To complete the proof we should show that the same is true for m = k + 1 i.e. span {p0 , p1 , . . . , pk+1 } = span {r0 , r1 , . . . , rk+1 } = span r0 , Ar0 , . . . , Ak+1 r0 . We have shown that this implies so that rk+1 = rk − αk Apk pk ∈ span r0 , Ar0 , . . . , Ak r0 Apk ∈ span Ar0 , A2 r0 , . . . , Ak+1 r0 rk+1 ∈ span r0 , Ar0 , . . . , Ak+1 r0 span {r0 , r1 , . . . , rk+1 } ⊂ span r0 , Ar0 , . . . , Ak+1 r0 1.7. ITERATIVE METHODS 17 Now, Ak r0 ∈ span {p0 , p1 , . . . , pk } k+1 A So that Ak+1 r ∈ span Therefore r0 ∈ span {Ap0 , Ap1 , . . . , Apk } rk+1 = rk − αk Apk Ap0 , Ap1 , . . . , Apk−1 , rk+1 − rk ⊆ span {r0 , r1 , . . . rk+1 } | {z } |{z} |{z} (r1 −r0 ) (r2 −r1 ) c(rk −rk−1 ) Ak+1 r0 ∈ span {r0 , r1 , . . . , rk+1 } and by induction span r0 , Ar0 , . . . , Ak+1 r0 ⊂ span {r0 , . . . rk+1 } . span {p0 , p1 , . . . , pm } = span {r0 , . . . , rm } is proven using pm = rm + βm−1 pm−1 and the same arguments as above. Definition Km := span {r0 , r1 , . . . , rm−1 } is the m-th Krylov space of A. Theorem 1.7.8 If A is spd then: (A) hpk , pm iA = 0 for m 6= k and (B) rkT rm = 0 for m 6= k. Proof 1. For k, m ≤ 1: r1T r0 = 0 hp1 , p0 iA = 0 2. Assume: (a) hpk , pm iA = 0 for k 6= m and k, m ≤ l (b) rkT rm = 0 for k 6= m and k, m ≤ l 3. For m < l: rlT pm = rlT (c0 r0 + c2 r2 + · · · + cm rm ) = 0 from (b) and therefore T rl+1 pm = (rl − αl Apl )T pm = 0. For m = l we have T T rl+1 pl = (rl − αl Apl ) pl = rlT pl − αl hpl , pl iA = 0 by the definition of αl . But rm ∈ span {p0 , . . . , pm } for m = 1, . . . , l and therefore T rl+1 rm = 0, m = 0, 1, . . . , l, 18 CHAPTER 1. SYSTEMS OF LINEAR EQUATIONS which proves (B). To prove (A), we use: pl+1 = rl+1 + βl pl hpl+1 , pm iA = hrl+1 , pm iA + βl hpl , pm iA ; hpl , pm iA = 0, m = 0, . . . , l − 1. But Apm ∈ span {r0 , r1 , . . . , rm+1 } and from (B), hrl+1 , pm iA = 0 ⇒ hpl+1 , pm iA = 0 for m = 0, 1, . . . , l − 1. By the choice of pl+1 in the CGM, hpl+1 , pl iA = 0 and so (A) holds. From this theorem we can conclude that dim Kn ≤ n. Corollary 1.7.9 If A ∈ Rn×n is spd then for some m ≤ n, rm = 0. Theorem 1.7.10 If A is spd then uk+1 minimizes the error u − uk+1 = ǫk+1 over Kk+1 in ||·||A i.e. ||ǫk+1 ||A = min ||u − v||A (u : Au = b) v∈Kk+1 Proof Corollary 1.7.9 concluded that there exist some m + 1 ≤ n s.t. rm+1 = 0. Without loss of generality we can assume that u0 = 0 since any change of the initial iterate can be interpreted as a change in the right hand side vector b: A(u − u0 ) = b − Au0 , and therefore it does not affect the estimates that follow. Then the m P αi pi . Consider some exact solution must be a linear combination of the search directions p1 , . . . , pm i.e. u = i=0 v ∈ Kk+1 , k < m + 1. It must have the form v = ||u − v||2A = || k X i=0 k X i=0 k P γi pi . Then we have: i=0 (αi − γi )pi + |αi − γi |2 ||pi ||2A + m X i=k+1 m X i=k+1 αi pi ||2A = |αi |2 ||pi ||2A , since hpi , pj iA = 0 if i 6= j. Therefore, ||u − v||A has its minimum for αi = γi , i = 1, . . . , k. But uk+1 = which concludes the proof. Definition Πl denotes the set of all polynomials up to order l: Πl = pl : al xl + · · · + a0 n Corollary 1.7.11 If A ∈ Rn×n is symmetric positive definite with a spectrum {λi }i=1 then: ||ǫk+1 ||A ≤ ||ǫ0 ||A max |q (λj )| 1≤j≤n for all polynomials q ∈ Πk+1 such that q (0) = 1 k P i=0 αi pi 1.7. ITERATIVE METHODS 19 Proof From theorem 1.7.10 we have that: ||ǫk+1 ||A ≤ ||u − v||A , ∀v ∈ Kk+1 . But v ∈ Kk+1 = span r0 , Ar0 , . . . , Ak r0 Without loss of generality, we can assume that r0 = b i.e. u0 = 0 and then v = α0 b + α1 Ab + · · · + αk Ak b = p (A) b. Therefore ||ǫk+1 ||A ≤ ||u − p (A) b||A , ∀p (A) ∈ Πk but u0 = 0, b = Au, and ǫ0 = u so that ||ǫk+1 ||A ≤ ǫ0 − p (A) A u − u0 A = ||ǫ0 (I − p (A) A)||A ≤ ||ǫ0 || I − p (A) A = ||ǫ0 || ||q (A)||A , ∀q (A) ∈ Πk+1 s.t. q (0) = 1. | {z } ∈Πk+1 A Now we will prove that ||q (A)||A = max |q (λj )| . j Since A is symmetric, there exists an orthonormal set of eigenvectors {vj }nj=1 of A i.e. viT vj = Take x ∈ Rn then x= 0 1 if if i 6= j i=j n X ci vi i=1 2 and therefore ||x||A = n P i=1 2 ||q (A)||A = c2i λi . Then we have 2 sup ||q (A) x||A = ||x||A =1 n P i=1 n P i=1 sup c2i λi =1 n P i=1 sup c2i λi =1 q (A) n P j=1 cj vj !T n P ci vi = A q (A) i=1 q 2 (λi ) c2i λi ≤ max q 2 (λi ) . 1≤i≤n If max q 2 (λi ) is achieved for some index l then it is clear that the equality in the last relation would be achieved 1≤i≤n √ if we take x = vl / λl , and this concludes the proof of the corollary. From corollary 1.7.11 it is clear that the sharpest error estimate would be obtained if we pick q to be such that max |q (λj )| is minimal with respect to q. Let us denote the minimum and maximum eigenvalues of A 1≤j≤n by λmin and λmax correspondingly. Before we construct such a polynomial on the interval [λmin , λmax ] we first consider polynomials on [−1, 1] whose maximum is minimal among all possible polynomials of the same degree i.e. polynomials with a minimal infinity norm. To remind you: 20 CHAPTER 1. SYSTEMS OF LINEAR EQUATIONS Definition The infinity norm of a polynomial on [−1, 1] is defined as ||p||L∞ [−1,1] = sup |p (z)| z∈[−1,1] If p (0) = 1 then ||p||L∞ [−1,1] ≥ 1 so the minimum for such polynomials would be 1. The polynomials that have a unit infinity norm on [−1, 1] are the Chebyshev polynomials. They are defined recursively as follows: Definition T0 (z) = 1 T1 (z) = z Tn+1 (z) = 2zTn (z) − Tn−1 (z) Alternatively, we may represent the Chebyschev polynomials as k k p p 1 2 2 z+ z −1 + z− z −1 Tk (z) = , 2 or Tk (z) = cos(k arccos(z)). The polynomials with a minimum infinity norm on [λmin , λmax ], such that they equal 1 at 0, are given by the Shifted Chebyshev polynomials: Definition Lemma 1.7.12 z−λmin Tk 1 − 2 λmax −λmin T̃k (z) = λmin Tk 1 + 2 λmax −λmin T̃k (z) L∞ [λmin ,λmax ] = min p∈Πk ,p(0)=1 ||p(z)||L∞ [λmin ,λmax ] . Proof Since Tk (z) = cos(k arccos(z)) it is clear that Tk has k + 1 extrema Tk (zi ) = (−1)i , i = 0, 1, . . . , k in the nodes zi = cos(iπ/k) i.e it has k + 1 alternating minima and maxima on [−1, 1]. T̃k is just a scaled and shifted version of Tk so it must have exactly the same number of alternating maxima and minima on [λmin , λmax ]. Now assume towards contradiction, that there is some pk (z) such that pk (0) = 1 and ||pk (z)||L∞ [λmin ,λmax ] < . Then, pk − T̃k must also have at least the same number of alternating extrema as T̃k T̃k (z) L∞ [λmin ,λmax ] on [λmin , λmax ] since max |pk | < max |Tk | on this interval. This means that pk − T̃k has at least k zeros on [λmin , λmax ]. But pk (0) − T̃k (0) = 0 and 0 < λmin . So we reach the contradiction that the polynomial of k-th degree pk − T̃k has at least k + 1 zeros which concludes the proof. From the corollary we have that ||ǫk ||A ≤ ||ǫ0 ||A max |p (λi )| λi for any p (x) ∈ Πk such that p (0) = 1. Therefore, ||ǫk ||A ≤ ||ǫ0 ||A sup λmin <z<λmax |pk (z)| = ||ǫ0 ||A kpkL∞ [λmin ,λmax ] . Since pk can be any polynomial such that p (0) = 1, the sharpest estimate would be obtained for polynomials with a minimum infinity norm i.e. shifted Chebishev polynomials: ||ǫk ||A ≤ ||ǫ0 ||A max T̃k (z) λmin <z<λmax 1.7. ITERATIVE METHODS 21 Then, T 1 − 2 z−λmin k λ −λ max min max T̃k (z) = max λ λmin <z<λmax Tk 1 + 2 λ min −λ max min 1 ≤ 2 ≤ λmin Tk 1 + 2 λmax −λmin √ k √ λ − λ √ max √ min λmax + λmin where we have used Tk λmax + λmin = 1 λmax − λmin 2 √ √ √ k √ k ! λmax + λmin λmax − λmin √ √ √ + √ λmax − λmin λmax + λmin √ √ k 1 λmax + λmin √ √ ≥ . 2 λmax − λmin Then ||ǫk ||A ≤ 2 ||ǫ0 ||A If A is spd then √ k √ λ − λ √ max √ min λmax + λmin λmax c2 (A) = ||A||2 A−1 2 = λmin and ||ǫk ||A ≤ 2 ||ǫ0 ||A !k p c2 (A) − 1 p . c2 (A) + 1 Let p(γ) be the smallest integer k > 0 such that ||ǫk ||A ≤ γ ||ǫ0 ||A . Then, from the estimate above (which is sharp) it follows that k p 1−z 2 ≤ γ, ∀z ∈ [0, 1), where z = 1/ c2 (A), 1+z or k 1+z 2 . ≤ γ 1−z Note that z = 0 corresponds to the limit c2 (A) → ∞. Then we get ln or 2 1+z 1 1 ≤ k ln = 2k(z + z 3 + z 5 + . . . ), γ 1−z 3 5 z ∈ [0, 1) 1 2 1 1 arctanh(z) ln ≤ k(1 + z 2 + z 4 + . . . ) = k . 2z γ 3 5 z But arctanh(z)/z is a monotonically increasing function on [0, 1) and its minimum is 1. Therefore, we have; 2 1p c2 (A) ln ≤ k, 2 γ p p which means that p(γ) ≤ 1/2 c2 (A) ln(2/γ) + 1, i.e. for all practical purposes, we need of order of c2 (A) iterations for satisfying the convergence condition ||ǫk ||A ≤ γ ||ǫ0 ||A . 22 CHAPTER 1. SYSTEMS OF LINEAR EQUATIONS Chapter 2 Solutions to Partial Differential Equations 2.1 Classification of Partial Differential Equations Suppose we have a pde of the form ∂u ∂u ∂ 2 u ∂ 2 u ∂ 2 u φ x, y, u, , , 2, , 2 =0 ∂x ∂y ∂x ∂x∂y ∂y then we classify the equation in terms of the second order derivatives, if they are not present we then classify in terms of the first order derivatives. 2.1.1 First Order linear PDEs ∂u ∂u + β (t, x) = γ (t, x) ∂t ∂x Let us try to reduce it to an ODE over some path (t (s) , x (s)) in the domain of the equation. We have that: α (t, x) du ∂u dt ∂u dx = + ds ∂t ds ∂x ds If we can choose a path α = dt ds and β = ∂x ∂s then du =γ ds These paths are called characteristics. 2.1.2 Second Order PDE ∂ 2u ∂u ∂u ∂2u ∂ 2u α 2 +β + γ 2 = ψ x, y, u, , ∂x ∂x∂y ∂y ∂x ∂y We use a similar idea as in the first order case but now we try to reduce the equation to a system of ODEs for the first derivatives of the solution i.e. ∂ 2 u dx ∂ 2 u dy d ∂u = + 2 ds ∂x ∂x ds ∂x∂y ds 2 ∂ u dx ∂ 2 u dy d ∂u = + 2 ds ∂y ∂x∂y ds ∂y ds 23 24 CHAPTER 2. SOLUTIONS TO PARTIAL DIFFERENTIAL EQUATIONS Then we may write this system as α β dx ds 0 dy ds dx ds ∂2u ψ γ ∂x2 ∂ 2 u d ∂u 0 ∂x∂y = ds ∂x dy ds ∂2u ∂y 2 d ds ∂u ∂y If we require, similarly to the first-order case, that the original equation is a linear combination of the two equations for the first partial derivative, i. e. this system is linearly dependent, then 2 2 2 dy dx dy dy dy dx α −β = 0 and α −β +γ + γ = 0. ds ds ds ds dx dx We classify the equations based on the discriminant β 2 − 4αγ > 0 2 β − 4αγ = 0 β 2 − 4αγ < 0 2.2 ⇒ ⇒ ⇒ hyperbolic (2 real characteristics) parabolic (1 real characteristic) elliptic (0 real characteristics) Difference Operators Suppose that we have a grid of nodes in an interval [a, b], ∆ = {a = x0 < x1 < · · · < xN −1 < xN = b}. The following mapping: uh : ∆ → R is called a grid function: uh = (uh,0 , . . . , uh,N )T with uh,j = uh (xj ). A classical function u : [a, b] → R gives rise to an associated grid function and, abusing notation somewhat, we denote this grid function by u and its value at a given node xj by uj . In addition we will make use of the following operators on such functions, that are used to approximate the corresponding derivatives: 1 (uj+1 − uj ) , hj+1 = xj+1 − xj δ + uj = hj+1 1 (uj − uj−1 ) δ − uj = hj 1 δuj+ 21 = (uj+1 − uj ) , hj+1 1 uj − uj−1 uj+1 − uj 1 2 , hj+ 21 = (hj + hj+1 ) . − δ uj = hj+ 21 hj+1 hj 2 We will also make use of the averaging operator: 1 (uj + uj+1 ) . 2 Assuming that hj = h, ∀j, and enough regularity of the function u, and using Taylor expansions, we can derive the following estimates for the error of these approximations: ∂uj 1 ∂ 2 u j h2 1 + uj + h+ + . . . − uj δ uj = (uj+1 − uj ) = h h ∂x ∂x2 2 ∂uj ∂uj ∂ 2 uj = + h 2 + ... = + O (h) ∂x ∂x ∂x ∂uj + O (h) δ − uj = ∂x uj+1 − 2uj + uj−1 ∂ 2 uj δ 2 uj = = + O h2 h2 ∂x2 1 uj+ 21 = (uj + uj+1 ) = u xj+ 21 + O h2 . 2 Similar estimates can be derived for non-constant grid size hj , however, in the rest of the notes we will consider only equidistant grids. The non-equidistant case requires more careful examination. uj+ 12 = 2.2. DIFFERENCE OPERATORS 2.2.1 25 Poisson Equation −∇2 u (u, y) = f (x, y) , (x, y) ∈ (0, 1) × (0, 1) = Ω u (x, y) = g (x, y) , (x, y) ∈ ∂Ω Consider the grid: Ω̄h ≡ {(xj , yl ) ∈ [0, 1] × [0, 1] , xj = jh, yl = lh, i, l ∈ 0, . . . , N } Ωh ≡ {(xj , yl ) ∈ ∆, 1 ≤ j, l ≤ N − 1} ∂Ωh ≡ {(x0 , yl ) , (xN , yl ) , (xj , y0 ) , (xj , yN ) , 1 ≤ j, l ≤ N − 1} A second order scheme for an approximation to the solution uh is given by: − δx2 uh,j,l + δy2 uh,j,l = fh,j,l , (xj , yl ) ∈ Ωh uj,l = gh,j,l , (xj , yl ) ∈ ∂Ωh where fh,j,l = fj,l , (xj , yl ) ∈ Ωh ; gh,j,l = gj,l , (xj , yl ) ∈ ∂Ωh So that − (uh,j−1,l + uh,j,l−1 − 4uh,j,l + uh,j,l+1 + uh,j+1,l ) = h2 fh,j,l The left hand side is a discretization of ∇2 u on the following stencil yl+1 yl xj−1 where T I I T I .. . I T I I T yl−1 xj+1 xj uh,1 uh,2 .. . r1 r2 .. . = −h2 uh,N −2 rN −2 uh,N −1 rN −1 −4 1 fj,1 uh,j,1 1 −4 1 fj,2 uh,j,2 bj .. T ≡ , fj ≡ .. , rj ≡ fj + 2 , uh,j ≡ .. . h . . 1 −4 1 fj,N −1 uh,j,N −1 1 −4 gj,0 gN,1 + gN −1,0 g0,1 + g1,0 0 gN,2 g0,2 .. . . .. .. bj ≡ . for j = 2, . . . , N − 2, b1 ≡ , bN −1 ≡ 0 gN,N −2 g0,N −2 gN,N −1 + gN −1,N g0,N −1 + g1,N gj,N 26 2.2.2 CHAPTER 2. SOLUTIONS TO PARTIAL DIFFERENTIAL EQUATIONS Neumann Boundary Conditions y4 y3 y2 y1 x−1 x0 Suppose that we need to impose a Neumann condition ∂u = g|ij ∂n ij on the left boundary of the domain sketched above. Then, we write the scheme for all nodes on this boundary and discretize the Neumann condition using a central difference scheme, introducing an additional layer of points (corresponding to level -1 in x direction) uh,1,l − uh,−1,l = g0,l , 2h uh,−1,l + uh,0,l−1 − 4uh,0,l + uh,0,l+1 + uh,1,l = −h2 f0,l . In order to eliminate the additional layer of points, we combine these results to get uh,0,l−1 − 4uh,0,l + uh,0,l+1 + 2uh,1,l = −h2 f0,l + 2hg0,l 2.3 Consistency and Convergence Consider the continuous problem Lu = f u=g in Ω on ∂Ω Now, consider the corresponding discrete problem L h uh = fh uh = gh in Ωh on ∂Ωh . fh , gh are some sort of approximations of f, g and these notes we assume that fh = f, gh = g in the nodes of the discretization grid. Definition For a given φ ∈ C ∞ (Ω) τh (x) ≡ (L − Lh ) φ (x) is called a truncation error of Lh . The scheme Lh is consistent with L if lim τh (x) = 0 h→0 for all x ∈ Ω and φ ∈ C ∞ (Ω). If τh (x) = O (hp ), then Lh is consistent to order p. 2.3. CONSISTENCY AND CONVERGENCE 27 Proposition 2.3.1 −∇2h = −δx2 − δy2 is consistent to order 2 with −∇2 = − ∂2 ∂2 − 2 2 ∂x ∂y Definition uh converges to u in a given norm ||.|| if ||ǫh || = ||u − uh || satisfies lim ||ǫh || = 0. h→0 If ||ǫh || = O (hp ), then p is the order of convergence. In the section on finite difference methods we use exclusively the infinity (or maximum) norm: ||ǫh || = max |ǫh,j |. j Note that consistency does not guarantee that the difference between the solutions of the continuous and discrete problems also tends to zero. It only guarantees that the action of the difference between the differential and discrete operators on a smooth enough function tends to zero. Let us go back to the boundary value problem for the Poisson equation −∇2 u = f u=g in Ω in ∂Ω (2.1) in Ωh in ∂Ωh . (2.2) with the corresponding discrete problem −∇2h uh = fh uh = gh From the elliptic PDEs theory we know that the solution to (2.1) satisfies a maximum principle i.e. the solution achieves its maximum on the boundary of the domain. Similarly (2.2) satisfies a Discrete Maximum Principle: Theorem 2.3.2 (Discrete Maximum Principle) If ∇2h vj,l ≥ 0 for all (xj , yl ) ∈ Ωh , then max (xj ,yl )∈Ωh vj,l ≤ max (xj ,yl )∈∂Ωh vj,l Proof ∇2h vj,l = vj+1,l + vj−1,l + vj,l+1 + vj,l−1 − 4vj,l ≥0 h2 ⇒ vj,l ≤ 1 (vj+1,l + vj−1,l + vj,l+1 + vj,l−1 ) 4 ∀(xj , yl ) ∈ Ωh Suppose that the maximum of v is achieved in an internal point (xJ , yL ) ∈ Ωh , then: vJ ,L ≥ vJ −1,L , vJ ,L ≥ vJ +1,L , vJ ,L ≥ vJ ,L−1 , vJ ,L ≥ vJ ,L+1 so that 4vJ ,L ≥ vJ −1,L + vJ +1,L + vJ ,L−1 + vJ ,L+1 ⇒ 4vJ ,L = vJ −1,L + vJ +1,L + vJ ,L−1 + vJ ,L+1 Therefore, v must attain its maximum in all neighbouring points too. Applying this argument repeatedly to the neighbours we eventually will reach a boundary point. Corollary 2.3.3 The following two results follow from the discrete maximum principle: 28 CHAPTER 2. SOLUTIONS TO PARTIAL DIFFERENTIAL EQUATIONS 1. ∇2h uh = 0 uh = 0 in Ωh on ∂Ωh has a unique solution uh = 0 in Ωh ∪ ∂Ωh . 2. For given fh and gh , ∇2h uh = fh uh = gh in Ωh on ∂Ωh has a unique solution. Definition If v : Ωh ∪ ∂Ωh → R then ||v||Ω ≡ ||v||∂Ω ≡ max (xj ,yl )∈Ωh |vj,l | max (xj ,yl )∈∂Ωh |vj,l | Lemma 2.3.4 If vj,l = 0 for all (xj , yl ) ∈ ∂Ωh then ||v||Ω ≤ M∆ ∇2h v Ω for some M∆ > 0, that depends on Ω but not on h. Proof Let and consider the function 2 ∇h v = ν Ω w : wj,l = Denote its norm on the boundary by M∆ i.e. ⇒ −ν ≤ ∇2h v ≤ ν, 1 2 xj + yl2 ≥ 0. 4 ||w||∂Ω = M∆ > 0. Note that ∇2h w = 1. Then ∇2h (v + νw) ≥ 0 ∇2h (v − νw) ≤ 0 From the maximum principle and since vh |∂Ωh = 0 we have: in Ω vj,l ≤ vj,l + νwj,l ≤ ν ||wj,l ||∂Ω = M∆ ∇2h vj,l Ω vj,l ≥ vj,l − νwj,l ≥ −ν ||wj,l || = −M∆ ∇2h vj,l ∂Ω ∀(xj , yl ) ∈ Ωh Ω Theorem 2.3.5 (error estimate) Let u be the solution of the boundary value problem (2.1) and uh is a solution to the discrete analog (2.2). Assuming that u ∈ C 4 Ω̄ , then there exists a constant k > 0 such that: ||u − uh ||Ω ≤ kM h2 where ) ( 4 ∂ u ∂ 4 u M ≡ max 4 , ∂x L∞ (Ω) ∂y 4 L∞ (Ω) 2.4. ADVECTION EQUATION 29 Proof From the proposition: h2 ∂ 4 u ∂ 4u ∇2h − ∇2 uj,l = (ξ , y ) + (x , η ) j l j l 12 ∂x4 ∂y 4 for some ξj , ηl s.t. xj−1 ≤ ξj ≤ xj+1 , yl−1 ≤ ηl ≤ yl+1 . Taking into account that u is the exact solution we have: h2 ∂ 4 u ∂4u −∇2h uj,l = fj,l − (ξ , y ) + (x , η ) j l j l 12 ∂x4 ∂y 4 Subtract −∇2h uh,j,l = fh,j,l = fj,l and note that u − uh = 0 on ∂Ωh . Then ∇2h ∂4u h2 ∂ 4 u (ξj , yl ) + 4 (xj , ηl ) (uj,l − uh,j,l ) = 12 ∂x4 ∂y From the Lemma we get 2M∆ ||u − uh ||Ω ≤ M∆ ∇2h (u − uh )Ω ≤ M h2 = kM h2 12 | {z } =k 2.4 Advection Equation Consider the initial value problem ∂u ∂u +v = 0, ∂t ∂x −∞ < x < ∞, t>0 with v > 0, u (0, x) = f (x) for −∞ < x < ∞. The exact solution is given by u (t, x) = f (x − vt) Consider a grid ∆ ≡ ∆t × ∆x : ∆t ≡ {nk : n = 0, 1, 2, . . .} ∆x ≡ {jh : j = 0, ±1, ±2, . . .} One possible discretization is or equivalently n un+1 unh,j − unh,j−1 h,j − uh,j δt+ + vδx− unh,j = +v =0 k h n n n un+1 h,j = uh,j − C uh,j − uh,j−1 This is called the FB (Forwards Backwards) scheme where the Courant number is defined as C≡ vk h 30 CHAPTER 2. SOLUTIONS TO PARTIAL DIFFERENTIAL EQUATIONS x .. . 5h C >1 4h C=1 3h 2h h k 2k 3k t Define S (shift operator) such that Sunh,j = unh,j+1 n n n−1 2 n−2 = 1 − C + CS −1 uh,j = · · · = 1 − C + CS −1 u0h,j = 1 − C + CS −1 fj0 unh,j = 1 − C + CS −1 uh,j n n X X n n m −1 n−m (1 − C)m C n−m fj−n+m fj = (1 − C) CS = m m m=0 m=0 Now, suppose that the initial condition is perturbed with an error ǫj i.e. instead of fj the initial condition is given byfˆj = fj + ǫj . Then, the perturbed solution ûh is given by: ûh,j n X n m (1 − C) C n−m (fj−n+m + ǫj−n+m ) = m m=0 and we have n n X n m n n−m uh,j − ûh,j = (1 − C) C ǫj−n+m m m=0 n n X X n n m n m n−m |1 − C| C n−m = ||ǫ||∞ (|1 − C| + C) |1 − C| C |ǫj−n+m | ≤ ||ǫ||∞ ≤ m m m=0 m=0 So that Then, if C ≤ 1: However, if C > 1: n u − ûn ≤ ||ǫ|| (|1 − C| + C)n h,j h,j ∞ n u − ûn ≤ ||ǫ|| h,j h,j ∞ n u − ûn ≤ ||ǫ|| (|1 − C| + C)n = ||ǫ|| (2C − 1)n n→∞ → ∞ h,j h,j ∞ ∞ Theorem 2.4.1 If C ≤ 1 and u ∈ C 2 ((0, t] × R) then the FB scheme is convergent. That is E n = ||un − unh ||∞ ≤ tn O (k + h) where tn = kn. 2.5. VON NEUMANN STABILITY ANALYSIS 31 Proof k ∂2u ∂u (t, ξ) (t, x) + ∂t 2 ∂t2 h ∂2u ∂u (η, x) (t, x) − δx− u (t, x) = ∂x 2 ∂x2 + − δt u + vδx u = −τk,h (t, x) = O (k + h) δt+ u (t, x) = δt+ uh + vδx− uh = 0 and letting ǫh = u − uh we get the following estimate: n+1 ǫh,j ≤ (1 − C) ǫnh,j + C ǫnh,j−1 + k |τk,h (tn , xj )| Let E n = ||ǫnh ||∞ , T n = max |τk,h (tn , xj )| j Then we can rewrite the inequality in the form E n+1 ≤ (1 − C) E n + CE n + k |τk,h (tn , xj )| ≤ E n + kT n ≤ E n−1 + kT n + kT n−1 ≤ E 0 + k n X Tm m=1 0 If E = 0 then E n+1 ≤ k 2.5 n X m=1 T m ≤ nk max T m = tn max T m = tn O (k + h) m 1≤m≤n Von Neumann Stability Analysis Definition Discrete L2 norm: ∞ X 2 ||v||2 ≡ h j=−∞ |vj | 2 Definition A FD (Finite Difference) scheme for a time dependent PDE is stable if for any time T > 0, there exists K > 0 such that for any initial data u0h , the sequence {unh } satisfies: ||unh ||2 ≤ K u0h 2 for 0 ≤ nk ≤ T . The constant K is independent of the time and space mesh size! Consider the following grid on the real axes ∆x = {. . . , −2h, −h, 0, h, 2h, . . .} . Then the Fourier Transform of a given grid function v is defined as ∞ h X −ijhτ e vj (F v) (τ ) ≡ √ 2π j=−∞ Inverse transform: 1 vj = √ 2π Z π h eijhξ (F v) (ξ) dξ −π h Theorem 2.5.1 (Parseval Identity) ||F v||2L2 [− π , π ] := h h Z π h −π h (F v)2 dξ = ||v||22 32 CHAPTER 2. SOLUTIONS TO PARTIAL DIFFERENTIAL EQUATIONS Corollary 2.5.2 A finite difference scheme is stable if and only if ∃K > 0, independent of k, h, s.t.: ||F unh ||L2 [− π , π ] ≤ K F u0h L2 [− π , π ] h h h h Consider the forward backwards scheme. Since we have: n−1 n−1 unh,j = (1 − C) uh,j + Cuh,j−1 Z πh h n √ eijhξ (F unh ) (ξ) dξ uh,j = 2π − πh Z πh h n−1 uh,j−1 = √ ei(j−1)hξ F uhn−1 (ξ) dξ, 2π − πh we obtain: F unh = F uhn−1 (ξ) 1 − C + Ce−ihξ {z } | Amplification Factor A (hξ) = 1 − C + Ce−ihξ Then, since we have 2 ||F unh ||L2 [− π , π ] h h if |A (hξ)| ≤ 1, we obtain that: = A (hξ) 2 F uhn−1 2 dξ, ||F unh ||L2 [− π , π ] ≤ F uhn−1 L2 [− π , π ] h h h ||F unh ||L2 [− π , π ] ≤ F u0h L2 [− π , π ] . h Since π h −π h h and recursive application gives: Z h h h A (θ) = 1 − C + Ce−iθ 2 2 |A (θ)| = (1 − C + C cos θ) + (−C sin θ) = 1 − 2C + C 2 + 2C cos θ − 2C 2 cos θ + C 2 cos2 θ + C 2 sin2 θ = 1 − 2C (1 − cos θ) + 2C 2 (1 − cos θ) = 1 + 2C (1 − cos θ) (C − 1) ≤ 1 + 4C (C − 1) it follows that if C > 1 then |A (θ)| > 1, and if C ≤ 1 then |A (θ)| ≤ 1. Also, 1 − C + Ce−iθ ≤ |1 − C| + C If a numerical scheme requires a solution of a linear system with a non-diagonal matrix then we call it an implicit scheme, otherwise it is called an explicit scheme. Consider the following “Backwards Backwards” scheme n−1 unh,j − uh,j unh,j − unh,j−1 =0 k h n−1 n n (1 + C) uh,j − Cuh,j−1 = uh,j +v This is an implicit scheme. In order to study its stability (and the stability of any other FD scheme), it is enough, instead of considering the full Fourier transform of the solution, to consider the ”action” of the scheme on a single mode of the form: wn eijhξ 2.6. SUFFICIENT CONDITIONS FOR CONVERGENCE 33 then plugging into the BB scheme: wn 1 + C 1 − e−ihξ = wn−1 | {z } A−1 (hξ) wn = A (hξ) wn−1 −1 A (θ) = 1 + C 1 − e−iθ Then, 1 + C 1 − e−ihξ ≥ |1 + C| − Ce−ihξ = 1 so that |A (θ)| ≤ 1 ∀θ therefore the BB scheme is unconditionally stable. We may outline a procedure for von Neumann stability analysis: 1. Consider a single mode wn eijhξ . 2. Derive conditions for A to satisfy |A (hξ)| ≤ 1 2.6 Sufficient Conditions for Convergence Consider ∂u = Lu, (t, x) ∈ (0, T ) × R ∂t where L is a linear spatial operator, with an initial condition u (0, x) = f (x) . We assume that the boundary conditions, in case of a finite spatial domain, are incorporated in the operator L. Then after discretization using a scheme that involves two time levels, in general we obtain a discrete system that can be written as: un+1 − unh C h = Dun+1 + F unh h k where C, D, F are linear finite dimensional operators. The consistency error is obtained when we substitute the exact solution u in the discrete system i.e. C The last equation can be rewritten as: un+1 − un n+1 = Dun+1 + F un + τk,h . k n+1 Bun+1 = Aun + kτk,h where B = C − kD, A = kF − C If det (B) 6= 0 (otherwise the scheme would be inconsistent if the original problem is well posed) then n+1 n+1 un+1 = B −1 Aun + kB −1 τk,l = Eun + kB −1 τk,l n+1 where τk,h is the truncation error. Theorem 2.6.1 If the scheme is consistent to order (p, q) in time and space and is also stable then it is convergent of order (p, q), that is n+1 u = O (k p + hq ) − un+1 h 2 provided that the initial error is zero: 0 u − u0h = 0 2 34 CHAPTER 2. SOLUTIONS TO PARTIAL DIFFERENTIAL EQUATIONS Proof n+1 un+1 = Eun + kB −1 τk,h , un+1 = Eunk h so that n+1 X 0 0 n+1 n n −1 n+1 n+1 m u − u un+1 − u = E (u − u ) − kB τ = . . . = E −k E n−m+1 B −1 τk,h h h h k,h | {z } m=1 =0 Then, n+1 n+1 X X m n+1 m n+1 . E n−m+1 B −1 τk,h u E n−m+1 B −1 τk,h − u ≤ k ≤ k h 2 2 2 2 | {z }2 m=1 m=1 ≤M If the scheme is stable then integer l, ulh 2 ≤ K u0h 2 for some positive K. Since from l for any positive the scheme we have that uh 2 ≤ E l 2 u0h 2 and this is a sharp inequality we can conclude that ∃k2 > 0 n−m+1 E ≤ k2 and we have 2 n+1 X n+1 m m n+1 u ≤ RT max τk,h = RT O (k p + hq ) − u ≤ k k2 M τk,h h 2 2 2 m m=1 where T = k(n + 1) is the final time and R = k2 M . Theorem 2.6.2 (Lax Theorem) Given a consistent scheme for a well-posed initial boundary value problem, then stability is a necessary and sufficient condition for convergence. 2.7 Parabolic PDEs – Heat Equation ∂u ∂t Which has the solution 2.7.1 2 = D ∂∂xu2 , (t, x) ∈ (0, T ] × R u (0, x) = f (x) , x ∈ R 1 u (t, x) = √ 4πDt Z ∞ f (ξ) e− (x−ξ)2 4Dt dξ −∞ FC Scheme This is an explicit scheme. δt+ unh,j − Dδx2 unh,j = 0 n n n un+1 h,j = (1 − 2ΓD) uh,j + ΓD uh,j+1 + uh,j−1 , where Γ is called a grid ratio. tn+1 tn tn−l xj−1 xj xj+1 Γ= k h2 2.7. PARABOLIC PDES – HEAT EQUATION 35 The consistency error is found using ∂t u − δt+ u = O (k) ∂xx u − δx2 u = O h2 so that τk,h = O k + h2 . Stability is a central concern of parabolic and hyperbolic PDEs. Using the Neumann analysis idea we substitute the Fourier mode: vjn = wn eijθ into the scheme to get: 1 D wn+1 − wn eijθ = 2 wn eiθ − 2wn + wn e−iθ eijθ k h This gives: w n+1 2 θ = 1 − 4DΓ sin wn 2 | {z } A(θ)≤1 So that 0 < DΓ ≤ 1 2 ⇔ |A (θ)| ≤ 1 ⇒ stability and we require: 1 Dk ≤ 2 h 2 ⇒ k≤ 1 h2 2D and we note that k and h are not of the same order. So, this is a conditionally stable scheme, but it is explicit and this is to be expected. In this problem the domain Ω is the entire real axes and therefore it is convenient to use for measuring the error the following ∞-norm: ||vh ||Ω = max |vh,j |. Then the following theorem provides the convergence −∞<j<∞ estimate for the FC scheme. Theorem 2.7.1 If DΓ ≤ 1 2 and if uh satisfies exactly the initial condition then: ||u − uh ||Ω = ||ǫh ||Ω = O k + h2 Proof From the scheme we have: n n n 2 2 un+1 h,j = DΓuh,j−1 + (1 − 2DΓ) uh,j + DΓuh,j+1 + O k + kh so that: DΓ ǫnh,j+1 + O k 2 + kh2 ǫn+1 DΓ ǫnh,j−1 + (1 − 2DΓ) ǫnh,j + |{z} h,j = |{z} | {z } ≥0 ≥0 ≥0 by the triangle inequality, n+1 ǫ ≤ ||ǫn || + O k 2 + kh2 ≤ ǫ0 + (n + 1) O k 2 + kh2 = (n + 1) k O k + h2 . h Ω h Ω h Ω | {z } tn+1 Note that the the right hand side of this estimate will blow up as t → ∞ for fixed k, h. 36 2.7.2 CHAPTER 2. SOLUTIONS TO PARTIAL DIFFERENTIAL EQUATIONS BC scheme tn+1 tn tn−l xj−1 xj xj+1 2 n+1 δt− un+1 h,j = Dδx uh,j This is a O k + h2 consistent scheme. It can be rewritten as: n+1 n+1 n −DΓun+1 h,j−1 + (1 + 2ΓD) uh,j − DΓuh,j+1 = uh,j So we must invert a tridiagonal matrix. Now we will study its stability. Substituting vjn = wn eijθ into the scheme we have: wn+1 − wn = DΓ eiθ − 2 + e−iθ wn+1 So that: wn+1 = 1 wn . 1 + 4DΓ sin2 θ2 | {z } A(θ)≤1 So, it is clear that the BC scheme is unconditionally stable. 2.7.3 Crank-Nicolson Scheme It is given by δt− un+1 h,j = or n un+1 h,j − uh,j k Consistency: = 1 2 n+1 D δx uh,j + δx2 unh,j , 2 D n+1 n n n uh,j−1 − 2un+1 h,j + uh,j+1 + uh,j−1 − 2uh,j + uh,j+1 2 2h ! n+1 ∂ 2 u n ∂ 2 u 2 2 +O h + +O h ∂x2 j ∂x2 j n+ 1 ∂ 2 u 2 + O k 2 + h2 =D 2 ∂x j D D 2 n+1 δ u + δx2 unj = 2 x j 2 n+ 1 un+1 − unj ∂u 2 j = + O k2 k ∂t j 2.7. PARABOLIC PDES – HEAT EQUATION 37 tn+1 tn tn−l xj−1 xj xj+1 Neumann stability analysis: First, rewrite the scheme in the following two-stage form n+ 1 uh,j 2 − unh,j k 2 =D n+ 1 2 un+1 h,j − uh,j k 2 Then, =D unh,j−1 − 2unh,j + unh,j+1 Γ θ → A1 (θ) = 1 − 4D sin2 2 h 2 2 n+1 n+1 un+1 h,j−1 − 2uh,j + uh,j+1 h2 1 wn+ 2 = A1 (θ) wn → A2 (θ) = 1 . 1 + 4D Γ2 sin2 2θ 1 wn+1 = A2 (θ) wn+ 2 = A1 A2 wn | {z } A(θ) So that A (θ) ≤ 1, so the system is unconditionally stable, and O h + k . 2 2.7.4 2 Leapfrog Scheme tn+1 tn tn−l xj−1 xj xj+1 As we use central difference in both directions, the scheme is O k 2 + h2 . We may write the scheme as: n−1 un+1 h,j − uh,j 2k =D unh,j−1 − 2unh,j + unh,j+1 h2 Now we preform stability analysis: vjn = wn eijθ Then, assume that: wn+1 − wn−1 = 2DΓ e−iθ − 2 + eiθ wn wn = A (θ) wn−1 so that and wn+1 = A (θ) wn A2 − 1 wn−1 = 4DΓ (cos θ − 1) Awn−1 38 CHAPTER 2. SOLUTIONS TO PARTIAL DIFFERENTIAL EQUATIONS this gives r θ θ A1,2 = −4DΓ sin ± 1 + 16D2 Γ2 sin2 2 2 clearly |A2 | ≥ 1, and |A2 | > 1 for some θ, so the scheme is unconditionally unstable. 2 2.7.5 DuFort-Frankel Scheme δt unh = Since i D h n n+1 n−1 n u − u + u + u h,j+1 h,j−1 h,j h,j h2 1 n+1 n−1 uh,j + uh,j = unh,j + O k 2 2 this scheme can be considered as a stabilized version of the Leapfrog scheme. The truncation error is then: 2 2 k ∂ u k3 ∂ 2 u h2 ∂ 2 u τk,h (t, x) = 2 (t , x ) + (t , x ) − (tn , xj ) + O k 3 + h3 n j n j 2 2 2 h ∂t 3 ∂t 12 ∂x k h and the system is conditionally consistant if → 0. If k, h go to zero but k h → c2 , then DF is consistent with: ∂u ∂ 2u ∂2u + c2 2 = D 2 . ∂t ∂t ∂x The stability analysis reveals that this scheme is unconditionally stable. 2.8 Advection-Diffusion Equation ∂u ∂2u ∂u (t, x) ∈ (0, T ] × (0, L) +v =D 2 ∂t ∂x ∂x x vL and v, D > 0. Then define τ = vt L , ξ = L , and Peclet number P e = D . So the PDE becomes: vT 1 ∂2u ∂u ∂u × (0, 1) , (τ, ξ) ∈ 0, + = ∂τ ∂ξ P e ∂ξ 2 L 2.8.1 FC Scheme n un+1 h,j − uh,j unh,j+1 − unh,j−1 1 unh,j−1 − 2unh,j + unh,j+1 =0 − k 2h Pe h2 This scheme is consistant to O k + h2 . Now preform stability analysis: vjn = wn eijθ + 1 e−iθ − 2 + eiθ eiθ − e−iθ n =0 w − k 2h Pe h2 2Γ n = w −Ci sin θ + (cos θ − 1) + 1 Pe | {z } wn+1 − wn + k wn+1 A(θ) So we may write, 2 |A (θ)| = 2 θ Γ + C 2 sin2 θ sin2 1−4 Pe 2 If Γ/P e ≤ 1/2 the first term ≤ 1 and since C 2 = kΓ then C 2 sin2 θ ≤ P ek/2 = M k, M > 0, i.e. 2 |A| ≤ 1 + M k 2.8. ADVECTION-DIFFUSION EQUATION 39 2 Theorem 2.8.1 The FC scheme is stable if Γ ≤ P e/2 i.e. if |A (θ)| ≤ 1 + M k for some M > 0. Proof 2 (F unh ) (ξ) = A2 (hξ) F unn−1 Therefore, 2 (ξ) = A2n (hξ) F u0h 2 n (ξ) ≤ (1 + M k) F u0h 2 (ξ) Mtn 2 2 1 nkM F u0h 2 ≤ eMtn F u0h 2 , ||F unh ||22 ≤ (1 + M k)n F u0h ≤ (1 + M k) kM F u0h 2 = (1 + M k) kM 2 2 where tn = nk. Theorem 2.8.2 Assume: 1 Γ ≤ Pe 2 Then the solution of the FC scheme satisfies the maximum principle: n max un+1 h,j ≤ max uh,j j j for all n ≥ 1, if and only if h ≤ Proof Let h ≤ 2 Pe, 2 Pe. then: n Γ 1 1 Γ n+1 −1 n 1 + hP e uh,j−1 + 1 − 2ΓP e 1 − hP e unh,j+1 ≤ max unh,j uh,j + uh,j ≤ j Pe 2 Pe 2 because h ≤ P2e , PΓe ≤ 12 . Now, assume that n max un+1 h,j ≤ max uh,j . j j Aiming towards a contradiction, assume that h > u0h,j so that u1h,1 = Γ Pe 2 Pe. = Then, choose the initial data as follows 1, j = 0, 1 0, j > 1 1 1 2 Γ Γ 1 + hP e + 1 − 2 1+ > Pe − 2 + 1 = 1 2 Pe Pe 2 Pe and this gives: max u1h,j > 1 = max u0h,j j so, by contradiction, the theorem holds. j Since violation of the maximum principle leads to unphysical solution then we must choose h ≤ 2/P e and so, if the Peclet number P e is very large then we need to use very small h, of the order of P e−1 . This corresponds to problems with boundary layers that need to be resolved by h. However, it is known from the properties of the corresponding boundary value problem that the thickness of such boundary layers is of the order of P e−1/2 i.e. it can be resolved by h being of order of P e−1/2 . So, this scheme requires the use of a spatial step h much less than what is needed to resolve the actual solution. The next scheme cures this problem at the expense of loss of accuracy. 40 2.8.2 CHAPTER 2. SOLUTIONS TO PARTIAL DIFFERENTIAL EQUATIONS Upwinding Scheme (FB) n un+1 h,j − uh,j k Theorem 2.8.3 Assume that Γ Pe + unh,j − unh,j−1 1 unh,j−1 − 2unh,j + unh,j+1 = h Pe h2 ≤ 12 . Then the FB scheme satisfies a maximum principle provided that: k Γ + C ≤ 1, C= 2 Pe h The proof is left as an exercise. This theorem implies that the maximum principle is satisfied if k ≤ (P eh2 )/(2 + P eh). Note that in the limit P e → ∞, this restriction on k tends to h. On the other hand, the stability condition k ≤ P eh2 /2 implies that if P e → ∞ but P eh2 → const (i.e. h = O(P e−1/2 )), k is restricted by a constant. In conclusion, if the Peclet number P e is very large, the scheme guarantees that the solution satisfies a maximum principle, similar to the exact solution, and is stable if k ≤ h, but k can be of the order of h. And this is true even if h = O(P e−1/2 ) that is sufficient to resolve boundary layers. The scheme has a consistency error τk,h (t, x) = O (k + h) as δt+ unj = and δx− unj = and h ∂2u ∂u (tn , xj ) + O h2 (tn , xj ) − 2 ∂x 2 ∂x δx2 unj = Combining these results we have: δt+ unj + δx− unj ∂u (tn , xj ) + O (k) ∂t ∂2u (tn , xj ) + O h2 . 2 ∂x ∂u ∂u 1 2 n δ u = + − − Pe x j ∂t ∂x h 1 + Pe 2 ∂2u + O k + h2 ∂x2 i.e. up to terms that are O(k + h2 ), the scheme is consistent with the equation ∂u ∂u h ∂2u 1 = 0. + − + ∂t ∂x P e 2 ∂x2 This means that the scheme adds an extra diffusion term that is of order of h. If the leading order term in the spatial error is proportional to a derivative of an even order, we call the scheme dissipative since such schemes usually dissipate sharp changes (large gradients) of the solution. If the leading order term of the error is proportional to a derivative of an odd order, the scheme is called dispersive. This is because odd derivatives in PDEs lead to the so called dispersion in the solution. In terms of a Fourier decomposition of the solution, this effect is manifested in the fact that due to the presence of odd derivatives different Fourier modes travel with a different speed. Chapter 3 Introduction to Finite Elements 3.1 3.1.1 Weighted Residual Methods Sobolev spaces Sobolev space: Hm (Ω) ≡ v : Ω→R : then Z 2 2 <∞ v 2 + v (1) + . . . + v (m) Ω H0 (Ω) = L2 (Ω) and Hm (Ω) is a Hilbert space: (u, v)m = Z Ω and ||u||m = Proposition 3.1.1 uv + u(1) v (1) + . . . + u(m) v (m) dx sZ u2 + . . . + u(m) Ω 1. Hm+1 (Ω) ⊂ Hm (Ω) 2. If v ∈ Hm (Ω) then: 2 2 2 2 2 ||v||m = ||v||0 + v (1) + . . . + v (m) 0 3. ||v||m+1 ≥ ||v||m and ||v||m+1 dx 0 ≥ v (1) m . 4. If v, w ∈ Hm (Ω) then |(v, w)m | ≤ ||v||m ||v||m , this is the Cauchy-Schwartz inequality. 5. If v, w ∈ Hm (Ω) then ||v + w||m ≤ ||v||m + ||w||m Proof We will prove the Cauchy-Schwartz inequality |(u, v)m | ≤ ||u||m ||v||m for u, v ∈ Hm (Ω). For any s ∈ R we have 2 2 0 ≤ (u − sv, u − sv)m = ||u||m + s2 ||v||m − 2s (u, v)m . 41 42 CHAPTER 3. INTRODUCTION TO FINITE ELEMENTS Choose (u, v)m s= 2 ||v||m so that: 0 ≤ (u − sv, u − sv)m = ||u||2m + 2 2 (u, v)m and, upon rearranging, we have: −2 ||v||2m (u, v)m = ||u||2m − ||v||2m 2 (u, v)m ||v||2m (u, v)m ≤ ||u||m ||v||m 3.1.2 Weighted Residual Formulations Consider Lu = f, in Ω with u = 0, on ∂Ω where L=− so that: − d2 u =f dx2 d2 dx2 in Ω = (0, 1) with u = 0, on x = 0, 1 Define the “trial function” space, for example it can be chosen to be U = u : u ∈ H2 (Ω) , u = 0 on ∂Ω and we introduce the “weight functions” (also called test) space, for this choice of U it can be chosen to be W = w : w ∈ L2 (Ω) , w = 0 on ∂Ω Now, find u ∈ U such that Z d2 u + f w dΩ = 0, f ∈ L2 (Ω) 2 dx Ω for all w ∈ W . This is one “weighted residual formulation” of the original problem. This particular formulation is also usually called a strong formulation. It still defines a continuous problem and in order to discretize it we need to discretize the corresponding functional spaces i.e. to define appropriate approximations for each function in them. We select a subset Uh of U : and Wh , a subset of W : Uh ⊂ U, Uh = span {φ0 , . . . , φn−1 } Wh ⊂ W, Wh = span {ψ0 , . . . , ψn−1 } and then, search for the discrete approximation uh in the form: uh = n−1 X ci φi i=0 so that: Z Ω n−1 d2 X ci φi − f dx2 i=0 ! ψj dΩ = 0 ⇒ n−1 X ci i=0 This is a linear algebraic system Lc = F where: Z 2 d φi Lij = ψ dΩ, 2 j Ω dx Z Ω d2 φi ψj = dx2 Fi = Z Ω Z f ψj dΩ, Ω f ψi dΩ j = 0, 1, . . . , n − 1 3.2. WEAK METHODS 3.1.3 43 Collocation Methods Let ψj (x) = δ (x − xj ) 2 so that: Lij = ddxφ2i (xj ) and Fj = f (xj ) . Note that ψj are not in L2 however the formulation still makes sense d2 2 if φi ∈ C 0 because the integrals are well defined. Next, we may approximate dx 2 ∼ δx , and this gives us a finite difference method or we may select Uh as the space spanned by certain sets of orthogonal polynomials (Legendre, Tchebishev etc.), and this gives us spectral collocation. 3.2 Weak Methods By far, the most popular weighted residual methods are based on the so called weak formulation of the classical problem. To obtain it we start from Z 2 d u + f w dΩ = 0 ∀ w ∈ W dx2 Ω and choose the discrete test functions space to be W = w : w ∈ H10 (Ω) . Here we define . Then, H10 (Ω) = u ∈ H1 (Ω) , u = 0 on ∂Ω − Z Ω du dw + dx dx Z | ∂Ω du w ds = − dx {z } Z f w dΩ Ω =0 Then the problem is reformulated as: Find u ∈ U such that: Z Z du dw f w dΩ = Ω Ω dx dx A natural choice for U which makes all integrals well defined and incorporates the Dirichlet boundary conditions in the solution is U = W . Such formulation is called a Galerkin formulation. The space U can be discretized by means of piecewise polynomial functions, orthogonal polynomials, trigonometric polynomials, spline functions etc. These choices yield various weak methods. In the remainder of the notes we will focus on the piecewise polynomial approximations which yield the so called finite element methods. Before we proceed with the discretization we shall prove some important results for the continuous Galerkin formulation: Find u ∈ U = H10 (Ω) such that Z Z u′ v ′ dΩ = f v dΩ, ∀ v ∈ U. Ω Ω Theorem 3.2.1 If u is a solution of the classical formulation −u′′ = f u=0 in Ω on ∂Ω then u is a solution of the Galerkin formulation. If u is a solution to the Galerkin Formulation then it is a solution to the classical formulation if u ∈ C 2 ([0, 1]) and f ∈ C 0 ([0, 1]). 44 CHAPTER 3. INTRODUCTION TO FINITE ELEMENTS Proof Suppose that u solves the differential formulation: −u′′ = f then take v ∈ U so that: − Z ′′ u v dΩ = ZΩ u′ v ′ dΩ = Z f v dΩ ZΩ f v dΩ Ω Ω (u′ , v ′ )0 = (f, v)0 Suppose that u solves the Galerkin formulation so that: (u′ , v ′ ) = (f, v) then Z Ω (u′ v ′ − f v) dΩ = 0 ⇒ Z (u′′ + f ) v dΩ = 0 Ω Assume towards contradiction that u′′ + f 6= 0 in some point x ∈ Ω. However, u′′ ∈ C 0 (Ω) and f ∈ C 0 (Ω). Therefore, u′′ + f ∈ C 0 (Ω) =⇒ u′′ + f 6= 0 in an entire open interval contained in Ω. Therefore, there exists (x0 , x1 ) ⊂ Ω such that u′′ + f > 0 for all x ∈ (x0 , x1 ) or u′′ + f < 0 for all x ∈ (x0 , x1 ). We consider the first possibility and the second one can be considered in exactly the same way. Let us choose − (x − x0 ) (x − x1 ) in (x0 , x1 ) v= 0 otherwise. For this choice it is clear that v ∈ H10 (Ω) and v ≥ 0 in Ω. Now since u′′ + f > 0 in (x0 , x1 ) then (u′′ + f, v) > 0 but this contradicts the fact that u is a solution to (u′′ + f, v) = 0 for all v ∈ U . Therefore, u′′ + f = 0. Example Consider the deformation of a rod from it’s equilibrium position under a given load f : f (x) u Then E (u) = 1 ′ ′ (u , u )0 − (f, u)0 2 is the total energy of the system. Therefore, E is called the energy functional of any system described by a second order elliptic differential equation. Theorem 3.2.2 u ∈ H10 (Ω) is a solution to the Galerkin formulation if and only if u minimizes E (v) over H10 (Ω). That is, E (u) ≤ E (v) for all v ∈ H10 (Ω). 3.2. WEAK METHODS 45 Proof 1. Suppose that u is a solution to the Galerkin formulation so that: (u′ , v ′ )0 = (f, v) for all v ∈ H10 (Ω) so that: 1 1 1 ′ ′ (u + w) , (u + w) 0 −(f, u + w)0 = (u′ , u′ )0 −(f, u)0 + (w′ , w′ )0 −(f, w)0 +(u′ , w′ )0 E (v) = E (u + w) = 2 2 2 Since (u′ , w′ )0 = (f, w)0 we have E (v) = E (u) + 1 2 ||w′ ||0 ≥ E (u) 2 2. Suppose that E (u) ≤ E (v) for all v ∈ H10 (Ω). Then if we take s ∈ R, E (u + sw) has a min at s = 0, and therefore: d E (u + sw) =0 ds s=0 which we expand as: 1 (u + sw)′ , (u + sw)′ 0 − (f, u + sw)0 =0 2 s=0 d 1 ′ ′ 2 ′ ′ ′ ′ ′ ′ =0 E (u) + s (u , w )0 − s (f, w)0 + s (w , w ) = (u , w )0 − (f, w)0 + s (w , w ) {z } | ds 2 s=0 d ds =0 s=0 So that: (u′ , w′ )0 = (f, w)0 for all w ∈ H10 (Ω). Therefore, u solves the Galerkin Formulation. Consider a Hilbert space V equipped with the product (·, ·)V i..e. V is complete with respect to the norm ||·||V induced by this product. Now, consider the mapping a : V × V → R (further called a bilinear form) such that: 1. a (αu + βv, w) = αa (u, w) + βa (v, w), 2. a (w, αu + βv) = αa (w, u) + βa (w, v), 3. there exists β > 0 such that |a (u, v)| ≤ β ||u||V ||v||V (such bilinear form is called bounded w.r.t. ||.||V ), and 4. there exists ρ > 0 such that 2 a (u, u) ≥ ρ ||u||V (such bilinear form is called coercive w.r.t. ||.||V ) Now, suppose that G (v) is a functional such that: 1. G (αu + βv) = αG (u) + βG (v) (such a functional is called linear) and 2. there exists δ > 0 such that |G (u)| ≤ δ ||u||V (such functional is called bounded). The following theorem is one version of a very well known result from functional analysis. Theorem 3.2.3 (Riesz Theorem) If V is a Hilbert space with a given inner product (., .)∗ and if G (v) is a bounded linear functional on V (w.r.t. the norm induced by its inner product), then there is a unique element û ∈ V such that (û, v)∗ = G (v) for all v ∈ V . This theorem almost directly yields the proof of a somewhat restricted version of the following fundamental result in PDE theory and their numerical analysis. 46 CHAPTER 3. INTRODUCTION TO FINITE ELEMENTS Lemma 3.2.4 (Lax-Milgram) If a (u, v) and G (v) satisfy the six conditions stated above then there exists a unique solution û ∈ V for the following problem: Find u ∈ V s.t. a (u, v) = G (v) ∀v ∈ V If a (u, v) = a (v, u) (symmetric form) then û is the minimizer of E (v) = 1 a (v, v) − G (v) 2 Proof We shall consider only the case where a (u, v) = a (v, u). • The proof of the first claim is a direct consequence of the Riesz theorem if we prove that a (u, v) defines an inner product on V . We know that: 2 a (u, u) ≥ ρ ||u||V which guarantees that p a (u, u) = ||u||a is a norm on V . It is not difficult then to verify that a(u, v) defines an inner product on V . Also, it is straightforward to show that the norm ||.||a is equivalent to ||.||V i.e.: p √ ρ ||u||V ≤ ||u||a ≤ β ||u||V . Therefore, since V is complete with respect to ||·||V then it is complete with respect to the norm ||.||a i.e. V is a Hilbert space with respect to the product a(., .). As G (v) is bounded with respect to ||·||V , G (v) is bounded with respect to ||·||a . Since a(., .) is an inner product on V and G (v) is bounded w.r.t. the norm induced by this inner product, the Riesz theorem implies the first claim of the lemma. • Next we show that û minimizes E (v) = 12 a (v, v) − G (v). E (v) = E (û + w) = E (û) + 1/2 a (w, w) ≥ E (û) | {z } ≥0 Example −∇2 u = f u=0 in Ω . on ∂Ω Its Galerkin formulation reads: Find u ∈ H10 (Ω) s.t.: (∇u, ∇v)0 = (f, v)0 ∀v ∈ H10 (Ω) . (3.1) Below we prove the so-called Poincaré inequality which will guarantee that a(u, v) = (∇u, ∇v)0 defines a bilinear form satisfying the conditions of the Lax-Milgram lemma. Lemma 3.2.5 (Poincaré Inequality) If Ω is a bounded domain and v ∈ H10 (Ω) then ∃C > 0, depending only on the domain Ω, such that 2 2 ||v||0 ≤ C ||∇v||0 . Proof Consider first v ∈ C10 (Ω) i.e. v is a continuously differentiable function defined on Ω that vanishes on its boundary. We can always find a ∈ R large enough so that the cube Q = {x ∈ Rn : |xj | < a, 1 ≤ j ≤ n} contains Ω. Integrating by parts in the x1 direction and taking into account that the surface integral vanishes since v = 0 on ∂Ω we obtain: R R R 2 ∂v 2 ||v||0 = v 2 dx = 1 · v 2 dx = − x1 ∂x dx 1 Ω Ω Ω R R ∂v ∂v = −2 x1 v ∂x dx ≤ 2a |v| ∂x dx. 1 1 Ω Ω 3.3. FINITE ELEMENT METHOD (FEM) 47 Using the Cauchy-Schwarz inequality for the L2 inner product on Ω, we obtain: R ∂v 2 ∂v ||v||0 ≤ 2a |v| ∂x k0 ≤ 2akvk0 k∇vk0 . dx ≤ 2akvk0 k ∂x 1 1 Ω Dividing by kvk0 gives the desired result with C = (2a)2 and v ∈ C10 (Ω). Due to some classical results in functional analysis (see for example Ern, A. and Guermond, J.-L., Theory and Practice of Finite Elements, Applied Mathematical Sciences, v. 159, Springer, 2004, p. 485), we can claim that for each v ∈ H10 (Ω) we can 1 1 choose a sequence of functions in {vk }∞ k=1 ⊂ C0 (Ω) such that the sequence converges to v in the H Ω) norm i.e. kv − vk k0 ≤ kv − vk k1 → 0 as k → ∞ and similarly k∇v − ∇vk k0 ≤ kv − vk k1 → 0. Using the triangular inequality in the form ku − wk0 ≥ kuk0 − kwk0 yields that kvk k0 → kvk0 and k∇vk k0 → k∇vk0 . This allows us to take the limit in the Poincaré inequality for vk ∈ C10 (Ω) to derive the inequality for any function in H10 (Ω). Using this inequality we easily prove the following proposition. Proposition 3.2.6 a (u, v) = (∇u, ∇v)0 is a bilinear, bounded, and coercive form in H01 (Ω); G (v) = (f, v)0 is bounded in H01 (Ω) if ||f ||0 < ∞. Proof Z |a (u, v)| = ∇u∇vdx ≤ ||∇u||0 ||∇v||0 ≤ ||u||1 ||v||1 Ω as ||u||1 = ||∇u||0 + ||u||0 ≥ ||∇u||0 . Then, 1 1 1 2 2 2 2 2 ||∇u||0 + ||∇u||0 ≥ ||u||0 + ||∇u||0 = ||u||1 2 2 2 R The last proposition guarantee us that the bilinear form a (u, v) = ∇u∇vdx and the functional G (v) = 2 a (u, u) = (∇u, ∇u) = ||∇u||0 = Ω (f, v)0 satisfy the conditions of the Lax-Milgram lemma if ||f ||0 < ∞. This automatically guarantees that the corresponding weak formulation (3.1) has a unique solution in H01 (Ω). 3.3 Finite Element Method (FEM) Let us define a grid: ∆ = {x0 , . . . , xM } and the following linear space of continuous piecewise linear functions on ∆: M01 (∆) = span {l0 , l1 , . . . , lM } ∈ H1 (Ω) where Define the discrete space: which is equivalent to x−x i−1 xi −xi−1 , xi−1 ≤ x ≤ xi i+1 li = xx−x , xi ≤ x ≤ xi+1 . −x i 0i+1 otherwise. Vh = v ∈ M01 (∆) , v (0) = v (1) = 0 ⊂ H10 (Ω) Vh = span {l1 , . . . , lM−1 } Let us now consider the following discrete problem: Find uh ∈ Vh such that: (u′h , vh′ )0 = (f, vh )0 48 CHAPTER 3. INTRODUCTION TO FINITE ELEMENTS for all vh ∈ Vh . It is clear that: uh = M−1 X u j lj j=1 and (u′h , li′ )0 = (f, li )0 , so that: M−1 X j=1 uj lj′ , li′ = (f, li )0 M−1 X ⇒ 1≤i≤M −1 uj lj′ , li′ j=1 0 0 = (f, li )0 , 1 ≤ i ≤ M − 1. The last set of equations clearly constitute a linear algebraic system: Auh = F where Aij = li′ , lj′ Now, as 0 , Fi = (f, li )0 , and uh,i = ui . 0 ≤ (vh′ , vh′ )0 = M−1 X vi li′ , i=1 M−1 X vj lj′ = M−1 X M−1 X vi li′ , lj′ i=1 j=1 j=1 v 0 j = v T Av the Lax-Milgram lemma guarantees that this system has a unique solution if f (x) ∈ L2 (Ω). 3.4 Gaussian Quadrature We may approximate the integral of function φ (x) over the range x ∈ (−1, 1) using Gaussian integration: Z 1 −1 φ (x) dx ≈ n X i=1 (n) (n) Ai φ xi (n) (n) where Ai are the weights and xi are the nodes of the quadrature. n is a positive integer that controls the accuracy. The weights and nodes are determined from the condition that the quadrature is exact for all polynomials of the highest possible degree. We have 2n undetermined coefficients (n weights and n nodes) and so we can make the quadrature exact for polynomials of 2n − 1 degree (that have 2n coefficients). The following table contains the weights and Gauss points for the first fourGauss quadratures: (n) n 1 2 Ai 2 1, 1 3 4 5 8 5 9, 9, 9 (n) xi 0 − √13 , √13 q q − 35 , 0, 35 −0.861136, −0.339981, 0.339981, 0.861136 0.347854, 0.652145, 0.347854, 0.652145 . 2n − 1 1 3 5 7 Remark We may approximate the integral over arbitrary bounds a and b using Gaussian Quadrature by applying the substitution: y (x) a y x b −1 1 3.5. ERROR ESTIMATES 49 y (x) = So that y (a) = −1, y (b) = 1, and dx = Z b a 3.5 b−a 2 a+b 2 x+ b−a a−b dy. Therefore: b−a φ (x) dx = 2 Z 1 φ −1 (b − a) y + a + b 2 dy Error Estimates Theorem 3.5.1 If uh is the finite element approximation to the exact solution u of the boundary value problem −u′′ = f, u (0) = u (1) = 0 then there exists k > 0 such that: ||u − uh ||1 ≤ k ||u − vh ||1 for all vh ∈ Vh . Proof u is the exact solution, so that a (u, wh ) = (f, wh ) for all wh ∈ Vh , and a (uh , wh ) = (f, wh ) So that a u − uh , wh = 0 | {z } Then, ǫh ⇒ a (ǫh , wh ) = 0 1 ||ǫh ||21 ≤ a (ǫh , ǫh ) + a (ǫh , wh ) ≤ a (ǫh , ǫh + wh ) = a (ǫh , u − uh + wh ) = a (ǫh , u − vh ) ≤ ||ǫh ||1 ||u − vh ||1 2 So that ||ǫh ||1 ≤ 2 ||u − vh ||1 Corollary 3.5.2 If u and uh are as in the theorem, then we have ||u − uh ||1 ≤ ch where c > 0 is independent of h. Proof For the time being we will assume that u ∈ C 2 (Ω) i.e. u is a solution to the classical problem for PM−1 f ∈ C 0 (Ω). Let û be the piecewise linear interpolant of u in Vh i.e. û = i=1 ui li , where ui are the values of u in the nodes of the grid xi . Using Taylor expansion it is possible to prove the following estimate for the interpolation error: ||u′ − û′ ||L∞ ≤ ||u′′ ||L∞ h Recall that: ||u − uh ||21 ≤ 4 ||u − v||21 for all v ∈ Vh and as û ∈ Vh we have: 2 2 2 2 2 ||u − uh ||1 ≤ 4 ||u − û||1 = 4 ||u − û||0 + ||u′ + û′ ||0 ≤ 8 ||u′ − û′ ||0 50 CHAPTER 3. INTRODUCTION TO FINITE ELEMENTS where, in the last step, we applied the Poincaré inequality. Continuing, we have: Z 1 2 2 2 2 (u′ − û′ ) dx ≤ 8 ||u′ − û′ ||L∞ ≤ 8 ||u′′ ||L∞ h2 ||u − uh ||1 ≤ 8 0 2 2 where we have used the interpolation estimate ||u′ − û′ ||L∞ ≤ ||u′′ ||L∞ h2 . This yields √ ||u − uh ||1 ≤ 8 ||u′′ ||L∞ h = O (h) And, applying Poincaré’s inequality, we have: 2 ||u − uh ||21 = ||u − uh ||20 + ||u′ − u′h ||0 ≥ 2 ||u − uh ||20 so that: ||u − uh ||0 ≤ ch ||u′′ ||L∞ . This is a suboptimal estimate because from interpolation theory we know that u can be approximated with a piecewise linear function with a second order accuracy in the L2 norm. 3.6 Optimal Error Estimates Lemma 3.6.1 If û is the finite element interpolant of u ∈ H2 (Ω), i.e. û = 1. ||u − û||0 ≤ 2. ||u′ − û′ ||0 ≤ h 2 π h π ||u′′ ||0 PM−1 j=1 u(xj )lj , then: ||u′′ ||0 Proof The domain is subdivided into elements at points: x0 = 0, x1 , x2 , . . . xi−1 , xi , . . . xM−1 , xM = 1 Consider the difference u − û 1 in [0, h] = [x0 , x1 ] and define η (x) = u − û ∈ H ([0, h]). Clearly, η (0) = η (h) = 0. Since η ∈ C 0 [0, h] it can be expanded in a uniformly convergent Fourier sine series: η (x) = ∞ X ηn sin n=1 nπx h Then, using the Parceval identity, Z 0 h 2 η 2 dx = ||η||L2 [0,h] = Differentiating term wise we get: η ′ (x) = ∞ X ηn nπx nπ cos h h ηn nπx n2 π 2 sin h2 h n=1 ⇒ ∞ hX 2 η . 2 n=1 n 2 ||η ′ ||L2 [0,h] = ∞ h X nπ 2 ηn 2 n=1 h and differentiating again we have: η ′′ (x) = − ∞ X n=1 ⇒ 2 ||η ′′ ||L2 [0,h] = ∞ h X 2 nπ 4 η 2 n=1 n h 3.6. OPTIMAL ERROR ESTIMATES 51 So that 2 ||η ′ ||L2 [0,h] = ∞ ∞ h X 2 nπ 2 nπ 2 h2 h2 ′′ 2 h2 h X 2 nπ 4 h2 ′′ 2 = ≤ ||η || = ||u ||L2 [0,h] ηn η 2 L [0,h] 2 n=1 h h n2 π 2 π 2 2 n=1 n h π2 π2 and 2 ||η||L2 [0,h] 4 ∞ ∞ h2 ′ 2 h2 h X 2 nπ 2 h hX 2 2 = 2 ||η ||L2 [0,h] = ||u′′ ||L2 [0,h] η ≤ 2 η = 2 n=1 n π 2 n=1 n h π π Applying the same for any subsequent subinterval and summing the resulting inequalities we complete the proof. Corollary 3.6.2 2. ||u − uh ||1 ≤ Proof √ 1. ||u − û||1 ≤ 8h π √ h 2 π ||u′′ ||0 ||u′′ ||0 1. 2 2 ||u − û||21 = ||u − û||20 + ||u′ − û′ ||0 ≤ 2 ||u′ − û′ ||0 ≤ 2 2 h 2 ||u′′ ||0 π 2. for all vh ∈ Vh in the finite element space we have: ||u − uh ||1 ≤ 2 ||u − vh ||1 Take vh = û, then: √ h 8 ′′ ||u ||0 ||u − uh ||1 ≤ 2 ||u − û||1 ≤ π So that ||u − uh ||1 ≤ Ch ||u′′ ||0 where C is independent of h. Theorem 3.6.3 (L2 lifting theorem) If uh ∈ Vh is the Galerkin finite element approxitamion of: −u′′ = f (x) , x ∈ (0, 1) u : u (0) = u (1) = 0 then there exists Γ > 0 such that ||u − uh ||0 ≤ Γh2 ||u′′ ||0 Proof Define: ǫh = u − u h then, (u′ , vh′ )0 = (f, vh )0 for all vh ∈ Vh or, so that Now, consider the auxiliary problem a (u, vh ) = (u′ , vh′ )0 a (u, vh ) = (f, vh )0 a (uh , vh ) = (f, vh )0 a (ǫh , vh ) = 0 −φ′′ = ǫh , 0 < x < 1 φ (0) = φ (1) = 0 (3.2) 52 CHAPTER 3. INTRODUCTION TO FINITE ELEMENTS Then, 2 ||ǫh ||0 = (ǫh , ǫh )0 = − (ǫh , φ′′ )0 = (ǫ′h , φ′ )0 = a (ǫh , φ) Now, taking vh = φ̂ in (3.2), we get: a ǫh , φ̂ = 0, and subtracting it from the above equation we obtain: 2 ||ǫh ||0 Z 1 ′ ′ ′ = ǫ′ φ′ − φ̂′ dx = a (ǫh , φ) = a (ǫh , φ) − a ǫh , φ̂ = a ǫh , φ − φ̂ = ǫh , φ − φ̂ 0 0 √ √ h 2h 2 ≤ ||ǫ′h ||0 φ′ − φ̂′ ≤ ||ǫ′h ||0 φ − φ̂ ≤ ||ǫ′h ||0 ||φ′′ ||0 = ||ǫ′h ||0 ||ǫh ||0 π π 0 1 √ h 2 ||ǫh ||0 ≤ Γh2 ||u′′ ||0 ||ǫh ||0 ≤ ||ǫh ||1 π So that ||ǫh ||0 ≤ Γh2 ||u′′ ||0 for some Γ > 0, independent of h. 3.6.1 Other Boundary Conditions Let us consider an elliptic problem with more general boundary conditions: −u′′ (x) = f, 0 < x < 1 u (0) = β1 , u′ (1) = β2 . Since we need to satisfy non-homogeneous boundary conditions of Dirichlet and Neumann type, this time we discretize the solution with the expansion: uh = β1 l0 (x) + M X uj lj (x) j=1 so that uh (0) = β1 l0 (0) = β1 . Now, consider the original equation: −u′′ = f multiply it by v ∈ V = H1 (Ω), and integrate over Ω to obtain: (−u′′ , v)0 = (f, v)0 (u′ , v ′ )0 + u′ (0) v (0) − u′ (1) v (1) = (f, v)0 . ⇒ If we approximate u with (3.3) and take the test functions to be vh ∈ Vh = span {l1 , . . . , lM }, we have: (u′h , lj )0 − u′h (1) v (1) = (f, lj )0 , or β1 l0′ + M X un ln′ , lj′ n=1 This yields: β1 l0′ + M X n=1 ! j = 1, . . . , M, − β2 lj (1) = (f, lj )0 . un ln′ , l1′ ! = (f, l1 )0 , (3.3) 3.7. TRANSIENT PROBLEMS 53 M X un ln′ , lj′ n=1 for j = 2, . . . , M − 1 and finally, M X ′ un ln′ , lM n=1 3.7 ! 0 ! = (f, lj ) − β2 = (f, lM )0 . Transient Problems Consider the Initial Boundary Value Problem given by: ∂u ∂ a (x) ∂u (t, x) ∈ (0, T ] × (0, 1) ∂t − ∂x ∂x = f (x) for u (t, 0) = u (t, 1) = 0 for t > 0 u (0, x) = g (x) for x ∈ (0, 1) For a (x) ∈ C 1 Ω̄ such that 0 < α ≤ a (x) ≤ A < ∞ and ||f (x) ||0 ≤ L < ∞. The Galerkin formulation of the problem is given by: Find u ∈ H10 (Ω) such that ∂u , v + b (u, v) = (f (x) , v)0 ∂t 0 ∂v for all v ∈ H10 (Ω), where b (u, v) = a (x) ∂u ∂x , ∂x 0 . Note that b(u, v) is a coercive and bounded bilinear form on 1 H0 (Ω) (why?). In order to discretize it we define Vh = span {l1 , l2 , . . . lM−1 } and search for uh ∈ Vh such that: ∂uh , vh ∂t 0 + b (uh , vh ) = (f, vh )0 for all vh ∈ Vh . Then, taking into account that uh (t, x) = M−1 X uj (t) lj (x) j=1 we obtain the linear system where for each i, l ∈ 1, . . . , M − 1, and ∂ (Mu) + Su = F ∂t Z Z ∂li ∂lj a (x) (li , lj ) dx + Aij = dx ∂x ∂x } {z } | Ω {z |Ω mass matrix M stiffness S Fj = Z f lj dx Ω Then, as M is time-independent, ∂ (u) + Su = F ∂t We may discretize this ODE system using for example a backward difference scheme (a good choice for such problems, as we know from the previous section) M M un − un−1 + Sun = Fn + τk k 54 CHAPTER 3. INTRODUCTION TO FINITE ELEMENTS ∂u n un − un−1 is the truncation error of the approximation of the time derivative of u by a where τk = − + ∂t k backward difference. Define M A= +S , k where : Mij = Z li lj Ω and Z ∂li ∂lj dx ≥ α Sij = a (x) ∂x ∂x Ω Z Ω ∂li ∂lj dx. ∂x ∂x M, A are are spd (why?). If u is the exact solution, then ∂u ∂u ∂ a (x) =f − ∂t ∂x ∂x Now we multiply by vh ∈ Vh and integrate: ∂u , vh + b (u, vh ) = (f, vh )0 ∂t 0 Applying the discretization, we get: n n n u − un−1 ∂u u − un−1 n n , vh + b (u , vh )0 = (f , vh )0 − , vh , vh + k ∂t k 0 0 0 which we may rewrite as: un − un−1 , vh k 0 + b (un , vh ) = (f n , vh )0 + (τk , vh )0 . On the other hand we have for the finite element approximation to the solution that: n uh − uhn−1 , vh + b (unh , vh ) = (f n , vh )0 . k 0 Letting un − unh = ǫnh and subtracting the two preceding equations equations, we get: n ǫh − ǫhn−1 , vh + b (ǫnh , vh ) = (τk , vh )0 k Now, define wh such that b (wh , vh ) = b (u, vh ) (3.4) for all vh ∈ Vh . wh is called the elliptic projection of u onto Vh . Then we split the error ǫnh into ǫnh = η n + ξ n where η n = un − whn is the error of the elliptic projection of u. Therefore, as we have already established for the solution of (3.4) 2 n n 2 ∂ u ||u − wh ||0 ≤ Γh 2 ∂x 0 But then, as ∂wh ∂t ∂u ∂t is the elliptic projection of onto vh ∈ Vh , n 3 n ∂u ∂wh 2 ∂ u ≤ Γh − ∂t 2 ∂t ∂t∂ x 0 0 3.7. TRANSIENT PROBLEMS 55 Now, ξ n − ξ n−1 , vh But, as ξ n ∈ Vh , ξ n − ξ n−1 , ξ n 0 + kb (ξ n , vh ) = k (τkn , vh )0 − η n − η n−1 , vh 0 − kb (η n , vh ) 0 + kb (ξ n , ξ n ) = k (τkn , ξ n )0 − η n − η n−1 , ξ n 0 − kb (η n , ξ n ) and, since since b (ξ n , ξ n ) ≥ 0 and b (η n , ξ n ) = 0, this leads to the inequality: ξ n − ξ n−1 , ξ n 0 ≤ k (τkn , ξ n )0 − η n − η n−1 , ξ n 0 which we may rewrite as: 2 + k (τkn , ξ n )0 − η n − η n−1 , ξ n ≤ ||ξ n ||0 ≤ ξ n−1 , ξ n Lemma 3.7.1 ξ n−1 , ξ n Proof 0 0 1 n−1 2 1 n a ξ + ||ξ ||0 0 2 2 0 2 2 n 0 ≤ ξ n − ξ n−1 0 = ||ξ n ||0 − 2 ξ n , ξ n−1 0 + ξ n−1 0 Upon rearranging, this yields the desired result. Theorem 3.7.2 (Young Inequality) 1 ||v||20 4α2 (u, v)0 ≤ α2 ||u||20 + Lemma 3.7.3 k2 k (τ , ξ ) ≤ 2 n Proof n 2 2 ∂ u k n 2 ∂t2 dt + 2 ||ξ ||0 (n−1)k 0 Z nk k n 2 k n 2 ||τ ||0 + ||ξ ||0 . 2 2 Z n un − un−1 1 nk ∂u ∂2u Now using the integral form of the truncation error: τ n = − = [t − (n − 1) k] 2 dt, ∂t k k (n−1)k ∂t and the Cauchy-Schwartz inequality for the integral in time, we obtain: Z 2 n 1 nk n n−1 2 2 ∂u u − u ∂ u n 2 = ||τ ||0 = dt − [t − (n − 1) k] 2 ∂t k k ∂t (n−1)k 0 0 2 s s Z nk Z nk 2 2 1 ∂ u 2 dt ≤ [t − (n − 1)] dt 2 ∂t (n−1)k (n−1)k k 0 √ s 2 Z Z Z 2 2 nk nk k 3 ∂2u ∂2u = k ≤ dt dt dx ∂t2 ∂t2 Ω (n−1)k (n−1)k k k (τ n , ξ n ) ≤ 0 2 2 ∂ u =k ∂t2 dt (n−1)k 0 Z So then, nk k k k2 2 2 k τ k , ξ n ≤ ||τ n ||0 + ||ξ n ||0 ≤ 2 2 2 2 2 ∂ u k n 2 ∂t2 dt + 2 ||ξ ||0 (n−1)k Z nk 56 CHAPTER 3. INTRODUCTION TO FINITE ELEMENTS Lemma 3.7.4 2 4 n η − η n−1 , ξ n ≤ Γ h 0 2 3 2 ∂ u k n 2 ∂t∂x2 dt + 2 ||ξ ||0 (n−1)k 0 Z nk Proof Using the Cauchy-Schwartz inequality for the integral in time and subsequently, we obtain Z ! sZ 2 nk nk √ n ∂η ∂η η − η n−1 , ξ n = dt, kξ n dt, ξ n ≤ 0 (n−1)k ∂t ∂t (n−1)k 0 nk 2 k ∂η 1 2 dt + ||ξ n ||0 = ∂t 2 2 Ω (n−1)k Z nk 2 ∂η 1 dt + k ||ξ n ||2 = 0 2 (n−1)k ∂t 0 2 ≤ Then, applying the result 1 2 Z Z Z nk (n−1)k Z Ω ∂η ∂t 2 the Young’s inequality 0 dt + k n 2 ||ξ ||0 2 3 2 2 2 2 ∂wh ∂u ∂u 2 ∂ u 2 ∂ = Γh − ≤ Γh , ∂t 2 2 ∂t 0 ∂x ∂t ∂x ∂t 0 0 we obtain the final claim of the lemma. Lemma 3.7.5 Let k ≤ 14 , then: 3 2 2 n−1 X ∂ u 2 ∂ u + 2h4 Γ2 2 + ||ξ m ||20 ||ξ n ||20 ≤ 2 ξ 0 0 + 2k 2 2 ∂t L2 (0,T,L2 (Ω)) ∂x ∂t L2 (0,T,L2 (Ω)) m=0 Where we define ||f (t, x)||L2 (0,T,L2 (Ω)) = Z 0 T Z f 2, Ω f ∈ L2 ((0, T ) × Ω) . Proof Using the results of the last three lemmas we easily obtain: 2 2 Z Z 1 m 2 Γ2 h4 mk ∂ 3 u 1 m−1 2 k 2 mk ∂ 2 u m m m 2 dt + ||ξ ||0 + kb (ξ , ξ ) ≤ ξ + ∂x2 ∂t dt + k ||ξ ||0 0 2 2 2 (m−1)k ∂t2 0 2 (m−1)k 0 and, as b (ξ m , ξ m ) ≥ 0, 2 ||ξ m ||0 2 ≤ ξ m−1 0 + k 2 2 2 Z mk 3 2 ∂ u ∂ u m 2 dt + Γ2 h4 2 dt + 2k ||ξ ||0 ∂t2 (m−1)k ∂x ∂t 0 (m−1)k 0 Z mk Now we sum for m = 1, . . . , n to get: Z nk 2 2 Z nk 3 2 n−1 X 2 ∂ u ∂ u 2 2 2 dt + Γ2 h4 dt + 2k ||ξ n ||0 ≤ ξ 0 0 + k 2 ||ξ m ||0 + 2k ||ξ n ||0 ∂t2 ∂t∂x2 0 0 0 0 m=1 Z nk 3 2 Z nk 2 2 n−1 X ∂ u 2 ∂ u 2 2 2 4 ||ξ m ||0 . (1 − 2k) ||ξ n ||0 ≤ ξ 0 0 + k 2 ∂x2 ∂t dt + 2k ∂t2 + Γ h 0 0 0 0 m=1 2 Since k ≤ 1/4 then 1 − 2k ≥ 1/2. Also ξ 0 0 ≥ 0 so we can multiply it by 4k and add to the right hand side of the last inequality to obtain: Z nk 2 2 Z nk 3 2 n−1 X 0 2 ∂ u ∂ u n 2 2 2 4 dt + 4k ||ξ ||0 ≤ 2 ξ 0 + 2k + 2Γ h ||ξ m ||20 . ∂t2 ∂x2 ∂t 0 0 0 0 m=0 But k ≤ 1/4 and this competes the proof. 3.7. TRANSIENT PROBLEMS 57 N Lemma 3.7.6 (Gronwall Lemma) If the sequence {ξ n }n=1 satisfies |ξ n | ≤ ν + µ n−1 X m=1 |ξ m | , ν, µ ≥ 0 then, Proof Consider the sequence zn defined by zn = ν + µ |ξ n | ≤ eµN ν + µ ξ 0 n X m=1 |ξ m | ⇒ z n − z n−1 = µ |ξ n | so that, z n ≤ (1 + µ) z n−1 z n−1 ≤ (1 + µ) z n−2 ≤ . . . ≤ (1 + µ) (n−1)µ µ z 0 ≤ eµ(n−1) z 0 ≤ eµN z 0 because: 1 (1 + µ) µ ≤ eµ . Then, |ξ n | ≤ z n−1 ≤ eµN z 0 = eµN ν + µ ξ 0 for n = 1, . . . , N . Now let µ = 4k and 2 2 2 2 0 2 2 ∂ u 2 4 ∂ u ν = 2 ξ 0 + 2k 2 + 2Γ h . 2 ∂t L2 (0,T ;L2 (Ω)) ∂t∂x L2 (0,T ;L2 (Ω)) Denoting the final time instant by T = nk and using the results of lemma 3.7.5 we find that 2 ||ξ n ||0 ≤e 2 2 0 2 2 ∂ u 3 ξ 0 + 2k 2 ∂t 2 4T L (0,T ;L2 (Ω)) Let’s choose u0h = wh0 i..e. ≤e 4T L (0,T ;L2 (Ω)) ! 2 2 3 2 ∂ u 2 4 ∂ u 2k 2 = O k 2 + h4 + 2Γ h 2 ∂t L2 (0,T ;L2 (Ω)) ∂t∂x L2 (0,T ;L2 (Ω)) 2 Theorem 3.7.7 Consider the scheme: M If k < 1 4 ! b u0h , vh = a (g, vh ) for all vh ∈ Vh . Then ξ 0 = 0. Therefore, 2 ||ξ n ||0 3 2 ∂ u + 2h Γ ∂t∂x2 2 4 2 un − un−1 + Aun = f k and u0h = wh0 , then ||ǫnh ||0 ≤ O k + h2 58 CHAPTER 3. INTRODUCTION TO FINITE ELEMENTS Proof h 2 ξ ≤ a2 k 2 + b2 h4 ≤ ak + lh2 2 0 where we define: 2 2 ∂ u a2 = 2e4T 2 ∂t L2 (0,T ;L2 (Ω)) 3 2 2 4T 2 ∂ u b = 2e Γ ∂t∂x2 L2 (0,T ;L2 (Ω)) Then, n n ||ǫn h ||0 ≤ ||η ||0 + ||ξ ||0 ≤ 3.8 √ 2 ∂ u 2e2T 2 ∂t k + h2 Γ L2 (0,T ;L2 (Ω)) " 2 n √ ∂ u + 2e2T max 2 0≤n≤N ∂x 0 Finite Elements in Two Dimensions Consider the model problem: − (pux )x − (puy )y + qu = g u=γ 1 3 ∂ u ∂t∂x2 L2 (0,T ;L2 (Ω)) in Ω on ∂Ω 2 e1 e2 7 4 3 e3 e5 e4 5 9 e7 e6 6 e8 8 10 This is the global numbering, the local numbering (for i = 1, 2, 3) is: 3 → ik3 ik1 ← 1 2 → ik2 where k is the element number and ikj is the global number of the j-th local point in the k-th element. # 3.8. FINITE ELEMENTS IN TWO DIMENSIONS 3.8.1 59 Finite Element Basis Functions With each global mode i we associate one basis funcion φi (x, y). If Nj is the j-th point in the grid, we require: φi (Nj ) = δij where φj |ek ∈ P 1 , the space of all first degree polynomials. Then, Vh = span {φ1 , . . . , φM } where M is the total number of points in the grid. uh (x, y) = M X ui φi (x, y) i=1 1 1 φi (x, y) (k) φ r (x, y) Ni ek r ek where ikr = Ni and r can be 1, 2, or 3 (local numbers). ξ 3 1 ek Tk y 2 1 x 1 Now, we define: Tk ≡ Then, So that: η x = x (ξ, η) = a1 ξ + a2 η + a3 y = y (ξ, η) = b1 ξ + b2 η + b3 φk 1 (ξ, η) = 1 − ξ − η, Z φk 2 (ξ, η) = ξ, φi (x, y) φj (x, y) dΩ = Ω Ke Z X k=1 φk 3 (ξ, η) = η φi (ξ, η) φj (ξ, η) dek ek Where Ke is the total number of finite elements, dek = |JTk | dξdη, and JTk is the Jacobian of Tk . 60 3.8.2 CHAPTER 3. INTRODUCTION TO FINITE ELEMENTS Discrete Galerkin Formulation Find uh ∈ Vh such that: Z puh,x vh,x + Ω Z puh,y vh,y + Ω Z quh vh = Ω Z gvh Ω and, for simplicity, we choose γ ≡ 0, p = constant and q = constant. Then, substitutiting uh = M X ui φi i=1 and vh = φj for j = 1, . . . , M yields the linear system: (pS + qM) uh = G where: Sij = p Mij = q Z ZΩ Ω ∇φi ∇φj dΩ φi φj dΩ (the stiffness matrix) (the mass matrix)