Iterative Methods for Sparse Linear Systems

Transcription

1 Iterative Methods for Sparse Linear Systems Luca Bergamaschi berga Department of Mathematical Methods and Models for Scientific Applications University of Padova Iterative Methods p.1/31

2 Discretized diffusion equation After discretization in the space, the PDE is converted in a system of ODEs like P u t = Hu + b where P = I in FD discretization. Stiffness matrix H is symmetric definite negative, capacity or mass matrix P is symmetric positive definite. Finally discretization in time transforms the system of ODEs to a linear system of size equal to the number of gridpoints to be solved at each time step. Time discretization divides the time interval (0,T] into subintervals of length t i so that solution in time is computed at t k = P k i=1 i and uses Finite Differences to approximate the time derivative. For example applying the implicit Euler method (u u(t k+1) u(t k ) t k ). (I + t k H)u(t k+1 ) = t k P u(t k ) + t k b Iterative Methods p.2/31

3 Eigenvalues and condition number of the Laplacian The 2D FD discretization of the Laplace equation takes a general form (in the case x = y = h): H = 1 h 2 0 S I I S I I S I I S 1 C A where S = C A Eigenvalues of S and hence those of H are explicitly known: λ j,k (H) = 4 «jπh h 2 sin2 + 4 «kπh 2 h 2 sin2, j, k =1,..., n x = 1 2 h 1 Iterative Methods p.3/31

4 Eigenvalues and condition number of the Laplacian Theorem. The smallest and largest eigenvalues of H behave like: λ n =2π 2 + O(h 2 ),λ 1 =8h 2 + O(1). Proof. λ n = 8 «πh h 2 sin2 = 8 2 h 2 (hπ/2+o(h3 )) 2 =2π 2 + O(h 2 ). λ 1 = 8 «nx πh h 2 sin2 = 8 π 2 h 2 sin2 2 πh «= 2 = 8 «πh h 2 cos2 = 8 2 h (1 2 O(h2 )) 2 =8h 2 + O(1). Corollary. The condition number of H behaves like κ(h) = 4 π 2 h 2 + O(1) (κ(h) = 3d π d h 2 when Ω R d ) Corollary. The number of iteration of the CG for solving Hx = b is proportional to 8 < n 1D discretizations h 1 = n 2D discretizations. : 3 n 3D discretizations Iterative Methods p.4/31

5 Iterative methods on the diffusion equations Again stationary iterative methods 1. Jacobi iteration. x k+1 = D 1 (L + U)x k + D 1 b = E J x k + q J. 2. Gauss-Seidel iteration. x k+1 = (D + L) 1 Ux k +(D + L) 1 b = E GS x k + q GS 3. Acceleration of Gauss-Seidel = SOR method depending on ω R. x k+1 = (D+ωL) 1 ((1 ω)d ωu) x k +ω(d+ωl) 1 b = E SOR x k +q SOR Theorem. (Young & Varga). If all the eigenvalues of the Jacobi iteration matrix are real then 1. λ(e GS )=λ(e J ) 2 2. If in addition ρ(e J ) < 1 then there is an optimal value of ω: 2 ω opt = 1+ p 1 ρ(e J ) ρ(e SOR (ω opt )) = ω opt 1 Iterative Methods p.5/31

6 Iterative method for the diffusion equation Recall H = 1 h 2 0 S I I S I I S I I S 1 C A where S = Since D = diag(h) = 4h 2 I, then eigenvalues of E J = I D 1 A are ««jπh kπh λ j,k (E) = 1 sin 2 sin C A λ max (E J ) 1 π2 h 2 2 «λ max (E GS ) 1 π2 h π 2 h 2 2 λ max (E SOR ) 2 1+πh 1 1 2πh Iterative Methods p.6/31

7 Iterative methods on the diffusion equations What about CG? Asymptotically the reduction factor is µ = κ 1 κ +1 where κ(h) = 4 π 2 h 2. Hence µ 2 1 hπ =1 2hπ +1 2 hπ 2+hπ 1 πh Asymptotic behaviour: method error reduction factor type Jacobi 1 5h 2 sharp Gauss-Seidel 1 10h 2 sharp SOR with ω opt 1 6h sharp SD (no precond) 1 5h 2 tight CG (no precond) 1 3h tight Iterative Methods p.7/31

8 Some results 2D discretization of the diffusion equation by FD, with h =0.005, tol = 10 8 Results: Method iterations CPU time comments Jacobi r k > 10 6 Gauss Seidel SOR ω =1.969 ω opt SOR ω =1.95 SOR ω =1.98 CG no prec CG IC(0) prec 2D discretization of the diffusion equation by FD, with h =0.0025, tol = 10 8 Results Method iterations CPU time comments SOR ω =1.984 ω opt CG no prec CG IC(0) prec Iterative Methods p.8/31

9 Results and Comments Results with h = , tol = 10 8 Method iterations CPU time comments SOR ω =1.992 ω opt CG no prec CG IC(0) prec Comments Theoretically SOR(ω opt ) is the fastest method with CG very close. ω opt is always very difficult to assess. Bounds on SOR is sharp while that on CG is tight. CG can be preconditioned while SOR can not. CG better than the estimates also when eigenvalues are spread into the spectral interval. Dependence of condition number on h 2 (and the number of iterations on h 1 ) only alleviated by preconditioning. SOR is dramatically dependent on ω. Solving the FD-discretized diffusion equation, it takes advantage on the a priori knowledge of the spectral interval. Iterative Methods p.9/31

10 Iterative methods for general systems PCG is expected to converge only on spd matrices. Why? One reason is that for non spd matrices it is impossible to have at the same time orthogonality and a minimization property by means of a short term recurrence. PCG at iteration k minimizes the error on a subspace of size k and constructs an orthogonal basis using a short-term recurrence. Extensions of PCG for general nonsymmetric matrices: 1. PCG method applied to the (spd) system A T Ax = A T b (Normal equations). This system is often ill-conditioned. It is the case of symmetric matrices where κ(a T A)=κ(A 2 ) = (κ(a)) 2. Also difficult to find a preconditioner if A T A could not be explicitly formed. 2. Methods that provide orthogonality + minimization by using a long-term recurrence (GMRES) 3. Methods that provide (bi) orthogonality. Examples: BiCG, BiCGstab 4. Methods that provide some minimization properties. Examples: QMR, TFQMR. Iterative Methods p.10/31

11 The GMRES method Given an arbitrary nonzero vector v, a Krylov subspace of dimension m is defined as K m (v) =span `v,av,a 2 v,..., A m 1 v The GMRES (Generalized Minimal RESidual) method finds the solution of the linear system Ax = b by minimizing the norm of the residual r m = b Ax m over all the vectors x m written as x m = x 0 + y, y K m (r 0 ) where x 0 is an arbitrary initial vector and K m is the Krylov subspace generated by the initial residual. First note that the basis of K m is little linearly independent. {r 0,Ar 0,A 2 r 0,..., A m 1 r 0 } Iterative Methods p.11/31

12 Orthogonalization of the Krylov basis Theorem. If the eigenvalues of A are such that λ 1 > λ 2 then A k λ2 k, r 0 = v 1 + O λ 1 with v1 the eigenvector corresponding to λ 1. To compute a really independent basis for K m we have to orthonormalize such vectors using the Gram-Schmidt procedure. β = r 0, v 1 = r 0 β DO k =1,m 1. w k+1 = Av k 2. DO j =1,k 1 3. h jk = w T k+1 v j 4. w k+1 := w k+1 h jk v j 5. END DO 6. h k+1,k = w k+1 ; v k+1 = w k+1 /h k+1,k END DO Once the Krylov subspace basis is computed, the GMRES method minimizes the norm of the residual onto mx x 0 + K m = x 0 + z i v i (1) j=1 Iterative Methods p.12/31

13 Minimizing the residual Theorem. The new vectors v k satisfy: Av k = k+1 X j=1 h jk v j, k =1,..., m Proof. In fact from step 4. of the Gram Schmidt algorithm we have w k+1 = Av k kx h jk v j j=1 Now substituting w k+1 = h k+1,k v k+1 we obtain h k+1,k v k+1 = Av k kx h jk v j j=1 and the thesis holds. Iterative Methods p.13/31

14 Minimizing the residual Now define V m =[v 1, v 2,... v m ] and H m =(h jk ). The statement of the theorem reads AV m = V m+1 H m H m is a rectangular m +1 m matrix. It is a Hessenberg matrix since it has the following nonzero pattern: H m = 0 h 11 h h 21 h 22 h h 32 h 33 h h m,m 1 h mm h m+1,m 1 C A Iterative Methods p.14/31

15 Minimizing the residual = A V m = V m+1 H m Iterative Methods p.15/31

16 Minimizing the residual Now we would like to minimize r m among all the x m : or Now x m = x 0 + x m = x 0 + V m z, mx z i v i j=1 z =(z 1,..., z m ) T r m = b Ax m = b A (x 0 + V m z)= = r 0 AV m z = = βv 1 V m+1 H m z = = V m+1 (βe 1 H m z) (2) where e 1 = [1, 0,..., 0] T. Iterative Methods p.16/31

17 Minimizing the residual Recalling that V T m+1v m+1 = I m+1 : r m = p r T mr m = q = (βe 1 H m z) T Vm+1 T V m+1 (βe 1 H m z)= q = (βe 1 H m z) T (βe 1 H m z)= and hence z = argmin βe 1 H m z. Comments: = (βe 1 H m z) (3) H m has more rows than columns then H m z = βe 1 has (in general) no solutions. Minimization problem z = argmin βe 1 H m z is very small m = 20, 50, 100 as compared to the original size n. Computational solution of this problem will be very cheap. Iterative Methods p.17/31

18 Least square minimization Any (even rectangular) matrix H can be factorized as a product of an orthogonal matrix Q and an upper triangular matrix R (known as QR factorization. When H is not square the resulting R has as many final zero rows as the gap between rows and columns. Let us now factorize our H m as: H m = QR (computational cost O(m 3 )). Then, in view of the orthogonality of Q: min βe 1 H m z = min Q βq T e 1 Rz = min g Rz where g = βq T e 1. Iterative Methods p.18/31

19 End of minimization Matrix R takes on the form: R = 0 r 11 r r 22 r r mm 0 1 C A (4) The solution to min g Rz is simply accomplished by solving Rz = g where R is obtained from R by dropping the last row and g the first m components of g. This last system being square, small, and upper triangular, is easily and cheaply solved. Finally note that min r m = g m+1. Iterative Methods p.19/31

20 Algorithm ALGORITHM: GMRES Input: x 0, A, b, k max, toll r 0 = b Ax 0,k=0,ρ 0 = r 0,β = ρ 0, v 1 = r 0 β WHILE ρ k > toll b AND k < k max DO 1. k = k v k+1 = Av k 3. DO j =1,k h jk = v T k+1v j v k+1 = v k+1 h jk v j END DO 4. h k+1,k = v k+1 5. v k+1 = v k+1 /h k+1,k 6. z k = argmin βe 1 H k z 7. ρ k = βe 1 H k z k END WHILE x k = x 0 + V k z k Iterative Methods p.20/31

21 Practical GMRES implementations. GMRES is optimal in the sense that it minimizes the residual over a subspace of increasing size (finite termination property). GMRES is a VERY computationally costly method. 1. Storage: It needs to keep in memory all the vectors of Krylov basis. When the number of iterations becomes large (order of hundreds) the storage may be prohibitive. 2. Computational cost. The cost of a Gram-Schmidt orthogonalization increases with the iteration number (O(mn)) so again problems arise when the number of iteration is high. Practical implementations fix a maximum number of vectors to be kept in memory, say p. After p iterations, x p is computed and a new Krylov subspace is begin constructed starting from r p. In this way however, we loose optimality with consequent slowing down of the method. Iterative Methods p.21/31

22 Convergence Assume for simplicity that A is diagonalizable that is A = V ΛV 1 where Λ is the diagonal matrix of eigenvalues and columns of V are normalized eigenvectors of A. Then r k = min P k Vp k (Λ)V 1 r 0 κ(v ) min P k p k (Λ) r 0 or r k r 0 κ(v ) min P k max p k (λ i ) Matrix V may be ill-conditioned We cannot simply relate convergence of GMRES with the eigenvalue distribution of A. Iterative Methods p.22/31

23 Preconditioning Also for general systems preconditioning is premultiplying the original system by a nonsingular matrix. Here preconditioning, more than reducing condition number of A is aimed at clustering eigenvalues away from the origin. Common general preconditioners. 1. Diagonal: M = diag( A i ), with A i the ith row of A. 2. ILU decomposition (based on pattern and/or on dropping tolerance). 3. Approximate inverse preconditioners Computational cost. Instead of applying A one should apply M 1 A to a vector (step 2. of algorithm) as t = Av k, Mv k+1 = t Iterative Methods p.23/31

24 Numerical example Convergence profile (semi-logarithmic plot) of GMRES as applied to a very small (n = 27) matrix with two preconditioners and some choices of the restart parameter r_k diagonale, nrest=1 diagonale, nrest=10 diagonale, nrest=50 ILU, nrest=1 ILU, nrest= numero di iterazioni Iterative Methods p.24/31

25 Other iterative methods for nonsymmetric systems They are based on a fixed amount of work per iteration (short term recurrence) No a priori theoretical estimates on the error. Possibility of failure (breakdown) that can be alleviated through the use of look-ahead strategies. Iterative Methods p.25/31

26 The Lanczos method. It is useful to underline connection between symmetric and nonsymmetric problems. 1. GMRES for spd matrices produces a tridiagonal matrix instead of an Hessenberg one (a symmetric Hessenberg is tridiagonal). This means that there is a three-term recurrence which implements Gram-Schmidt orthogonalization on a Krylov subspace generated by an spd matrix. 2. For A spd V T m AV m = T m and it can be proved that the extremal eigenvalues of T m converge quickly to the extremal eigenvalues of A. Moreover there is a strict connection between the Lanczos process and the CG method (entries of T m can be computed from CG iterates). Knowing approximately the largest and the smallest eigenvalues is useful for an appropriate exit test for the CG method (recall 1st lecture on reliability of exit tests). 3. If A is nonsymmetric the two-sided Lanczos method can be employed to construct a biorthogonal basis for the Krylov subspaces corresponding to A and A T by using a short term recurrence. Iterative Methods p.26/31

27 The Biconjugate Gradient method (BiCG). Based on two-sided Lanczos process: 1. Given an arbitrary cr 0 non orthogonal to r 0, it constructs two sets of vectors v 1,...,v n K(A, r 0 ) and w 1,..., w n K(A T, cr 0 ) s. t. v T i w j =0,i j. In matrix form AV k = V k+1 T k+1,k, A T W k = W k+1 b Tk+1,k, V T k W k = I k 2. Write x m = x 0 + v k y k. From the several choices for vector y k BiCG forces r k = r 0 AV k y k to be orthogonal to W k. This leads to the equation W T k r 0 W T k AV k y k =0 By noticing that W T k AV k = T k and that W T k r 0 = βe 1,β = r 0, equation for y k becomes T k y k = βe 1 3. Practical implementation of BiCG algorithm implicitly performs an LDU factorization of matrix T k (Also CG for spd matrices is equivalent to Lanczos process with LDL T factorization of T k ), Iterative Methods p.27/31

28 BiCG Algorithm ALGORITHM: BICONJUGATE GRADIENT Input: x 0, cr 0, A, b, k max, tol r 0 = p 0 = b Ax 0,k=0, cp 0 = cr 0 WHILE r k > tol b AND k < k max DO 1. z = Ap k, u = A T cp k 2. α k = rt k cr k z T p k 3. x k+1 = x k + α k p k 4. r k+1 = r k α k z r k+1 = r k α k u 5. β k = rt k+1r k+1 r T k r k 6. p k+1 = r k+1 + β k p k, p k+1 = r k+1 + β k cp k 7. k = k +1 END WHILE Iterative Methods p.28/31

29 BiCGstab Drawbacks of the BiCG method: 1. Need to implement a matrix vector product also for the transpose of the matrix (this can be a problem: sometimes the matrix is not available explicitly but only the result of its application to a vector) 2. No local minimization property fulfilled in BiCG. This leads to sometimes large oscillations in the convergence profile. First CGS (Conjugate Squared method) avoids using A T then a subsequent improvement: the BiConjugate Gradient stabilized (BiCGstab) method is aimed at avoiding such an erratic behavior trying to produce a residual of the form r k =Ψ k (A)φ k (A)r 0 where Ψ k (z) = ky (1 ω j z) and the ω s are chosen to minimize j=1 r j = (1 ω j A)Ψ j 1 (A)φ j (A)r 0 Iterative Methods p.29/31

30 BiCGstab ALGORITHM: BICGSTAB Input: x 0, A, b, k max, toll r 0 = b Ax 0,k=0, choose ˆr 0 s. t. ˆr T 0 ˆr 0 0 ρ 0 = α 0 = η 1 = 1; WHILE r k > toll b AND k < k max DO 1. k = k ρ k =ˆr T 0 r k 3. β k =(ρ k /ρ k 1 )(α k 1 /η k ) 4. p k = r k + β k (p k 1 η k v k 1 ) 5. v k = Ap k 6. α k = ρ k /ˆr T 0 v k 7. s = r k α k v k 8. t = As 9. η k+1 = s T t/t T t 10. c k+1 = c k + α k p k + η k+1 s 11. r k+1 = s η k+1 t END WHILE Iterative Methods p.30/31

31 Bibliography Books on iterative methods Anne Greenbaum. Iterative Methods for Solving Large Linear systems. C. T. Kelley. Iterative Methods for Linear and Nonlinear Equations Preconditioning M. Benzi. Preconditioning Techniques for Large Linear Systems: A Survey, Journal of Computational Physics, 182 (2002), pp Conjugate gradient J. R. Shewchuk. An Introduction to the CG method without the agonizing pain. Nonsymmetric linear systems Y. Saad and M. H. Schultz. Gmres: A generalized minimal residual algorithm for solving nonsymmetric linear systems. SIAM J. Sci. Stat. Comput., pages , (7) H. A. van der Vorst. Bi-CGstab: A fast and smoothly converging variant of BI-CG for the solution of nonsymmetric linear systems. SIAM J. Sci. Stat. Comput., pages , (13) Iterative Methods p.31/31