Introduction to Multigrid Methods Sampling Theory and Elements of Multigrid

Size: px
Start display at page:

Download "Introduction to Multigrid Methods Sampling Theory and Elements of Multigrid"

Transcription

1 Introduction to Multigrid Methods Sampling Theory and Elements of Multigrid Gustaf Söderlind Numerical Analysis, Lund University

2 Contents V3.15 What is a multigrid method? Sampling theory Digital filters Downsampling and aliasing Anti-aliasing A twogrid algorithm Spectral analysis Error estimation A convection diffusion problem

3 1. What is a multigrid method? Recall that standard iterative methods u m+1 = u m M 1 (T h u m f ) are slow unless the preconditioner M 1 T 1 h In a multigrid method, the preconditioner is constructed from an inverse of T H on a coarse grid At least two grids are used, a fine grid h and a coarse grid H The objective is to achieve a convergence rate independent of h

4 Solution, error and residual Given a linear system T h u = f construct sequence u m u Now recall the following definitions The Good The Bad The Ugly The solution is defined by T h u = f The error is defined by e m = u m u The residual is defined by r m = T h u m f Error equation relates error to residual T h e m = r m

5 Error (defect) correction Note that u = u m e m. If we can estimate the error, ẽ e m, then ũ u m ẽ is an improved solution, so we can put u m+1 u m ẽ Question Is it simpler to solve T h e m = r m than T h u = f? Answer YES! Because we only need moderate accuracy, solve the error equation on coarse grid This will still recover the persistent low frequency modes!

6 Elements of multigrid MG is a highly efficient iterative method for various operator equations, such as the Poisson equation Lu = f Key assumptions L is elliptic u, Lu M 2 [L] u 2 2 < 0 L 1 is regularizing u is smoother than the data f Basic idea Solve Lu = f on (at least) two grids L h u h = f h L H u H = f H Coarse grid mesh width H, fine grid h, mesh ratio ρ = H/h

7 Multigrid pattern algorithm Using at least two grids, iterate u m+1 h = u m h Ph H L 1 H RH h (L hu m h f h) For preconditioner M 1 = P h H L 1 H RH h we require I M 1 L h < 1 The operator P h H L 1 H RH h L h is key As L h and L H are given, success depends crucially on the operators R H h, which should be smoothing, and Ph H Downsampling R H h and oversampling P h H

8 Eigenvalues and eigenvectors u = f T h u = f To introduce multigrid, we return to damped Jacobi r m = T h u m f u m+1 = u m γd 1 r m Theorem Let T h with homogeneous Dirichlet conditions have eigenvectors given by T h v k = λ k v k. Then the eigenvalues and eigenvectors of the Jacobi matrix P J (γ) = (1 γ)i + γp J are v k = sin kπx h λ k [P J (γ)] = 1 γ + γ cos kπh k = 1 : N where grid and mesh width are given by x h = {jh} N 1 (0, 1)

9 High frequency damping in Jacobi iteration N = 31 Damping parameter γ = 1, 2/3, 1/2, 1/3 1 Eigenvalues of damped Jacobi matrix Frequency

10 Effect of damped Jacobi in the frequency domain Error recursion e m+1 = P J (1/2)e m, with F π = P J (1/2) = Apply to eigenvector v k = sin kπx h of normalized frequency θ k = kπh (0, π) to get ( 1 P J (1/2)v k = λ k [P J (1/2)] v k = ) cos kπh v k = cos 2 kπh 2 2 v k Theorem Discrete Fourier transform of P J (1/2) is ˆF π (θ) = cos 2 θ 2

11 Separate treatment of HF and LF in Multigrid A few damped Jacobi iterations (γ = 1/2) suppress HF error components, while LF errors remain unaffected Evaluate residual r h = T h u m f and restrict to coarse grid by r H R H h r h (downsampling) Solve error equation T H e H = r H for error estimate e H and note that e H = T 1 H r H is smoother than r H Prolong to fine grid, ẽ h P h H e H (oversampling), e.g. by using linear interpolation Remove LF error u m+1 u m ẽ h on fine grid and repeat

12 A simple two-grid algorithm Solve T h u h = f h Solve T h u h = f h iteratively using the following algorithm 1. Run a few damped γ = 1/2 Jacobi iteration on fine grid u m+1 h u m h γd 1 h r m h m = 0 : M 1 2. Compute residual and restrict r H I H h F π(t h u M h f h) 3. Solve error equation T H e H = r H 4. Prolong e M h I h H e H 5. Correct u M+1 h u M h em h 6. Run one damped Jacobi iteration on fine grid 7. Repeat from 1 until convergence

13 Matlab code segment The two-grid V step rf = Tdx*v - f; v = v - gamma*ddx\rf; rf = Tdx*v - f; rf = lowpass(rf); rc = FMGrestrict(rf); ec = Tdxc\rc; ef = FMGprolong(ec); v = v - ef; rf = Tdx*v - f; v = v - gamma*ddx\rf; % Compute residual % Pre-smoothing Jacobi % Compute residual % Anti-alias filter % Restrict to coarse grid % Solve error equation % Prolong to fine grid % Correct: remove error % Compute residual % Post-smoothing Jacobi

14 Key points to analyze in MG algorithms There is a large number of technical issues to consider in MG 1. Sampling theory 2. Aliasing 3. Digital filters and anti-aliasing 4. Downsampling (restriction) 5. Oversampling (prolongation) 6. Fourier analysis of error 7. Regularizing effect of T 1 H 8. Grid resonance and AA-filter design

15 2. Sampling theory Fine grid x h = {x h j }N f +1 j=0 = {jh} N f +1 j=0 with h = 1/(N f + 1) Coarse grid x H = {x H j } Nc+1 j=0 with H = 2h Gridpoint relation x H j = x 2h j = x h 2j for j = 0(1)N c + 1 Discrete Fourier modes v k (x h j ) = sin kπx h j = sin ω k jh = sin θ h k j Wave number k for k = 1(1)N f Frequency ω k = kπ Normalized frequency θk h = hω k [0, π]

16 Nyquist frequency hω = π Define Nyquist wave number k f = N f + 1 = 1/h Note As hω kf = π sin ω kf jh 0, the Nyquist mode cannot be represented on the fine grid ω h := π/h = ω kf is called the (fine grid) Nyquist frequency Normalized Nyquist frequency θ h = hω h = π Since Hω H = π, if coarse mesh width H = 2h, then ω h = 2ω H

17 Downsampling LF N f = LF modes k = 1 (top) and k = 3 (bottom)

18 Downsampling LF N c = LF modes k = 1 (top) and k = 3 (bottom)

19 HF to LF aliasing HF N f = HF mode k = 31

20 HF to LF aliasing HF N c = LF mode k = 1

21 Nyquist frequency ω N f = Nyquist HF mode k = 32

22 Aliasing Downsampling h H = 2h Due to the Nyquist Shannon sampling theorem, sin 2ω H x h 2j = sin ω h x h 2j 0 cos 2ω H x h 2j = cos ω h x h 2j 1 Hence, using sin(α β) = sin α cos β cos α sin β ( ) ( ) sin (ω H + ω m )x2j h = sin (2ω H (ω H ω m ))x2j h ( ) = sin (ω H ω m )xj H HF modes ω H + ω m > ω H are aliased to LF ω H ω m < ω H

23 Aliasing Downsampling h H = 2h LF + HF spectral content on fine grid

24 Aliasing Downsampling h H = 2h Downsampling: HF fold-over to LF

25 Aliasing Downsampling h H = 2h HF fold-over corrupts spectral content on coarse grid

26 Linear filter theory Analogue and digital filters An analogue linear filter is an integral operator F : v u u = Fv u(x) = 1 0 f (x, y)v(y) dy Its discrete counterpart is a digital filter u = Fv u i = N f i,j v j j=1 The linear operator F, as defined by its kernel f (x, y) or the matrix F = {f i,j }, determines filter characteristics

27 Toeplitz operators and convolution filters A convolution filter is an integral operator F : v u u = Fv u(x) = 1 0 f (x y) v(y) dy A discrete convolution is of the form u = Fv u i = N f i j v j j=1 F is a Toeplitz matrix, as i j = const implies f i,j = f i j = const. Toeplitz filters are discrete convolutions. Bdry conditions may apply

28 Anti-aliasing HF content must be suppressed before downsampling One (inefficient) way to do this in MG is to apply a smoother often selected as a (few) damped Jacobi iterations Better is to use digital filter theory and apply an anti-aliasing filter u F θ v F θ is a linear operator designed to suppress HF presence in v, and can be specifically designed to quench a given frequency θ [0, π] Example The filter F π is designed to block the frequency θ = π, corresponding to the undesired HF oscillation e iπn = ( 1) n

29 The lowpass filter F π Note F π = P J (1/2) for T h To filter out HF, repeated averaging can be used u = v = F π v Then u is smoother than v, and F π is a 2 nd order lowpass filter with DFT ˆF π (θ) = cos 2 θ 2 ˆF π (π) = 0 F π removes HF oscillation ( 1) n ˆF π (0) = 1 F π allows LF through virtually unchanged

30 The lowpass filter F π... A continuous DFT Let v = a k v k, where v k = sin kπx h Reflecting that a k is associated with the normalized frequency θ k, we write a k = a(θ k ) and consider the function a(θ) for θ (0, π) to represent the spectral content of v, with DFT ˆv(θ) = a(θ) Then u = F π v = k a(θ k )F π sin kπx h = k a(θ k ) cos 2 θ k 2 v k Thus, if spectral content of v is ˆv(θ) = a(θ), then u = F π v has spectral content û(θ) = cos 2 θ 2 ˆv(θ) Convolutions in space are multiplications in the frequency domain

31 Matlab matrix-free implementation of convolutions Given vector v and convolution filter kernel f π = 1 4 ] [ the operation u F π v is implemented by the Matlab commands ker = [1 2 1]/4; u = conv(v,ker, same ); Likewise ker = [1-2 1]/dx2; r = conv(u,ker, same ) - f; computes residual r = T h u f without ever using matrices

32 Downsampling with and without anti-aliasing filter Due to aliasing HF folds back into LF and must be suppressed by AA-filter before downsampling Essential in MG is to obtain error estimate e H from error equation L H e H = r H where r H is downsampled from fine grid residual r h Without HF suppression the error estimate e H will be poor, causing slow convergence or even divergence Unlike smoothers anti-aliasing is only grid dependent

33 Downsampling without anti-aliasing (retaining even nodes only) Corrupted spectral content on coarse grid

34 Doubly averaging AA-filter F π (3-node kernel, 2nd order filter) Good HF suppression, little effect on LF

35 AA-convolution filter with 5-node kernel support (4th order) Strong HF suppression, some effect also on LF

36 Fine grid, coarse grid Fine grid with odd number N of internal grid points x h = linspace(0,1,n+2) h = 1/(N + 1) Coarse grid with (N 1)/2 internal grid points x H = linspace(0,1,(n+3)/2) H = 2/(N + 1) Mesh widths h and H = 2h, respectively

37 Downsampling from fine grid to coarse Example N h = 7 internal points on fine grid x h map to N H = 3 internal points on coarse grid via restriction map v H = v h = Ih H v h The restriction drops every second element An alternative is a smoothing map, aka weighted restriction 1/4 1/2 1/ v H = 0 0 1/4 1/2 1/4 0 0 v h /4 1/2 1/4

38 Weighted restriction = plain restriction AA-filter = I H h F π 1/4 1/2 1/ /4 1/2 1/ /4 1/2 1/4 = This is proper downsampling with necessary anti-aliasing

39 From fine grid to coarse, including boundaries Example N h = 3 internal points on fine grid x h map to N H = 1 internal point on coarse grid x H via restriction map v H = v h = Īh H v h A weighted counterpart, with unchanged boundaries, is v H = 0 1/4 1/2 1/4 0 v h

40 Oversampling from coarse grid to fine Note The maps I H h do not have inverses (information loss) I h H is an interpolation operator generating new vector elements Left and right inverses IH h I h H I I H h I h H = I The second identity holds if Ih H is a plain restriction, but not if it is a weighted restriction (smoothing operator)

41 Prolongation through linear interpolation Example The interpolation operator IH h piecewise linear interpolation can be defined by v j+1/2 (v j + v j+1 )/2 This defines a prolongation of the grid function by inserting an interpolated value between previous coarse grid points Advanced interpolation (splines) is possible but often too expensive Choose linear interpolation with linear FEM basis functions

42 Matrix representation of the prolongation Example Prolongation from 3 internal points to 7 v h = 1/ /2 1/ /2 1/ /2 v H = IH h v H Compare using weighted restriction I h H = 2F π(i H h )T

43 Tools for restriction R H h and prolongation Ph H Two tools needed, the Toeplitz AA-filter F π = 1 4 tridiag(1 2 1) and the plain restriction ( injection ) Ih H = General restriction map R H h = I H h F π Linear interpolation prolongation map P h H = 2F π(i H h )T

44 Complete multigrid pattern algorithm L h u h = f h 1. Pre-smoothing u h u h γd 1 h (L hu h f h ) 2. Restriction r H R H h (L hu h f h ) 3. Solve for error L H e H = r H 4. Prolongation e h P h H e H 5. Correction u h u h e h 6. Post-smoothing u h u h γd 1 h (L hu h f h ) How does L h act on e h Im(P h H )? Galerkin condition L H = R H h L hp h H with Ph H = ρd (R H h ) Congruence transformation preserves ellipticity on all grids

45 4. Spectral analysis of MG u = f on [0, 1] 2nd order FDM discretization of u = f yields T h u = f, with 2 1 T h = h The N f N f matrix is negative definite, with eigenvalues λ k [T h ] = 4 h 2 sin2 θh k 2, k = 1 : N f and Fourier mode eigenvectors v k j [T h] = sin ω k x h j = sin θ h k j

46 Eigenvalue locations λ k [T h ] = 2 h h 2 cos θh k = 4 h 2 sin2 θh k 2

47 MG algorithm Consider T H u = f on coarse grid and T h u = f on fine grid 1. Jacobi smoother u u γd 1 h (T hu f ) 2. Restriction r H R H h (T hu f ) 3. Solve for error T H e H = r H 4. Prolongation e h 2(R H h )T e H 5. Correction u u e h 6. Jacobi smoother u u γd 1 h (T hu f ) Damped Jacobi and Galerkin condition downsampling share eigenvectors with T h spectral analysis is possible

48 Error analysis in the frequency domain Assume fine grid error is a linear combination of eigenvectors e h = 2N c+1 k=1 ê(θ h k ) v k [T h ] Derive expression for the computable coarse grid error estimate N c e H = ẽ(θk H ) A(θH k ) v k [T H ] k=1 with ẽ(θk H) ê(θh k ), and where A(θ) 1 at least for θ π Accuracy A(θ) 1 e H e h

49 Step 1. Damped Jacobi smoother Let D h = diag(t h ) and iterate u m+1 = u m γd 1 h (T hu m f ) with damping 0 < γ 1 to get u m+1 = PJ h(γ)um + γd 1 h f with 1 γ γ/2 γ/2 1 γ γ/2 PJ h (γ) =... γ/2 γ/2 1 γ Symmetric Toeplitz with eigenvalues λ k [P h J (γ)] = 1 γ + γ cos θh k, k = 1 : N f

50 Frequency response Damped Jacobi Frequency response 1 γ + γ cos θ for γ = 0.5(0.1)1

51 Error after one damped Jacobi iteration Denote the damped Jacobi frequency response δ γ (θ) = 1 γ + γ cos θ After one damped Jacobi iteration, the error is ē h := P h J (γ) e h = 2N c+1 k=1 ê(θ h k ) δ γ(θ h k ) v k [T h ]

52 Step 2a. Evaluate residual Because r h = T h ē h, the residual is r h = 2N c+1 k=1 ê(θ h k ) λ k[t h ] δ γ (θ h k ) v k [T h ] Note that λ k [T h ] is huge λ k [T h ] = 4 h 2 sin2 θh k 2

53 Step 2b. Apply Galerkin condition anti-aliasing filter Apply doubly averaging AA-filter r h := F π r h before downsampling 2 1 F π = Symmetric Toeplitz convolution filter Identical to damped Jacobi (γ = 1/2) with frequency response σ π (θ) = δ 1/2 (θ) = cos 2 θ 2

54 Doubly averaged residual and downsampling After application of F π the residual is r h = 2N c+1 k=1 ê(θ h k ) σ π(θ h k ) λ k[t h ] δ γ (θ h k ) v k [T h ] Downsampling h H retains only even vector components ( r h ) 2j = 2N c+1 k=1 ê(θ h k ) σ π(θ h k ) λ k[t h ] δ γ (θ h k ) v k 2j[T h ]

55 Interlude Aliasing in terms of vectors Recall HF aliasing to LF according to ( ) ( ) sin (ω H + ω m )x2j h = sin (2ω H (ω H ω m ))x2j h ( ) = sin (ω H ω m )xj H Aliasing in terms of Fourier mode eigenvector representation v kc+m 2j [T h ] = v kc m j [T 2h ] with Nyquist wave numbers k c = N c + 1 and k f = 2k c corresponding to Nyquist frequencies ω H and ω h, respectively

56 Step 2c. Aliasing due to downsampling Eigenvector aliasing v kc+m 2j [T h ] = v kc m j [T 2h ] = v kc m j [T H ] implies coarse grid residual has LF as well as offending fold-over HF N c r H = ê(θk h ) σ π(θk h ) λ k[t h ] δ γ (θk h ) v k [T H ] k=1 N c ê(θk h f k ) σ π(θk h f k ) λ k f k[t h ] δ γ (θk h f k ) v k [T H ] k=1

57 Step 3. Solving for coarse grid error e H = T 1 H r H = k=1 N c k=1 ê(θ h k ) σ π(θ h k ) δ γ(θ h k ) λk[t h ] λ k [T H ] v k [T H ] N c ê(θk h f k ) σ π(θk h f k ) δ γ(θk h f k ) λk f k[t h ] λ k [T H ] Attenuation factors depend on eigenvalue ratios v k [T H ] A LF k = σ π (θ h k ) δ γ(θ h k ) λ k[t h ] λ k [T H ] A HF k f k = σ π(π θ h k ) δ γ(π θ h k ) λ k f k[t h ] λ k [T H ]

58 Interlude Eigenvalue ratios In the LF band we have, for H = 2h, ( ) ( λ k [T h ] 4 H 2 λ k [T H ] = h 4 4 ) sin 2 θk h/2 sin 2 θk H/2 = 4 sin2 θk h/2 sin 2 θk h = 1 cos 2 θ h k /2 Likewise, in the HF band we have λ kf k[t h ] λ k [T H ] = 1 cos 2 (π θ h k )/2 Note resonance at θ = π/2

59 Eigenvalue locations T h

60 Eigenvalue locations T h, T H

61 Attenuation factors on (0, π/2) and full band attenuation on (0, π) A LF (θ) = σ π(θ) δ γ (θ) cos 2 θ/2 A HF (π θ) = σ π(π θ) δ γ (π θ) cos 2 (π θ)/2 Note A HF (π θ) equals foldover of A LF (θ) for θ (π/2, π) Unfold A HF to study full-band attenuation A(θ) = σ π(θ) δ γ (θ) cos 2 θ/2 θ (0, π) of coarse grid error estimate e H = ê(θk h ) A(θh k ) v k [T h ] N f k=1

62 Full-band attenuation on (0, π) The full band attenuation factor is A(θ) = σ π(θ) δ γ (θ) cos 2 θ/2 Double pole at θ = π must be annihilated by σ π (π) or δ γ (π) Damped Jacobi smoother helps iff γ = 1/2 Better Design dedicated AA-filter, putting a sufficiently heavy zero on top of θ = π to make A(π) = 0

63 However... it s your lucky day! Double averaging used in Galerkin condition restriction has so σ π (θ) = cos 2 θ 2 A(θ) = σ π(θ) δ γ (θ) cos 2 θ/2 = δ γ (θ) = 1 γ(1 cos θ) Hence A(0) = 1 but A(π) = 1 2γ := 0 requires γ = 1/2 Coarse grid error estimate e H has accurate LF content if γ = 1/2 and AA-filter F π is used

64 But... wait a minute! It is not the smoother, but the AA-filter that does the trick! Multigrid algorithm design Looking for a (problem dependent) smoother is likely looking the wrong way Looking for a (grid dependent) AA-filter is necessary to guarantee success The AA-filter is necessary to suppress eigenvalue grid resonance, but that the wrong smoother (γ 1/2) will still corrupt LF error, as A(π) 0 causes aliasing effects

65 Key ingredients that make MG work 1. L is elliptic This guarantees that grid resonances occur at regular intervals 2. L 1 is regularizing This implies that the error is smoother than the residual 3. A standard, smoothing iterative method reduces HF error This suppresses HF error on fine grid 4. Anti-aliasing is used before downsampling This guarantees that coarse grid error estimate is accurate

66 Eigenvalue locations T h, T H Non-elliptic Helmholtz equation Non-elliptic case u + βu = 0 with β > π 2 (principle sketch)

67 Conclusions Analysis of damaging effects of fold-over in residual Specific design of anti-aliasing filters are key to MG Smoothers are of secondary importance Arbitrary downsampling ratios possible Multitude of Toeplitz AA-filters available Works on non-selfadjoint operators as well Works even when eigenmodes are not of Fourier type Works on regular meshes (tensor product in 2D) Can be adapted to non-uniform meshes if smooth

68 Numerical test: Two-grid method N = Exact (g) & initial approx (b) 10 0 Error estimate progression x Remaining errors Exact (g) & numerical (b) V-cycles on u = f on on fine grid with 2048 points, error estimation on 1024, single P J (1/2) pre- and post-smoothing

69 Numerical test: Two-grid method N = Exact (g) & initial approx (b) 10 0 Error estimate progression x Remaining errors Exact (g) & numerical (b) V-cycles on u = f on on fine grid with 256 points, error estimation on 128, single P J (1/2) pre- and post-smoothing

70 Numerical test: Two-grid method N = Exact (g) & initial approx (b) 10 0 Error estimate progression x Remaining errors Exact (g) & numerical (b) V-cycles on u = f on on fine grid with 32 points, error estimation on 16, single P J (1/2) pre- and post-smoothing

71 Observations in the numerical tests One can almost use the same, small number of iterations By and large, convergence is unaffected by mesh width h 0 One can solve the problem to high accuracy so that only the global error remains The basic iterative method must be smoothing but is less crucial Large initial errors are not a problem The method is robust even with respect to noise

72 Why is the Multigrid Method faster? All iterative methods work with the residual y m+1 = y m + γmr m The difference is that multigrid estimates the error The residual is large and nonsmooth The error is moderate and smooth Basic residual-reducing iterative methods fail to eliminate low frequency error components Therefore only error estimating methods can be fast

73 6. Multigrid on nested grids. The V-cycle We want to solve T K u K = f K on a fine grid Ω K of mesh width h K Construct an embedding of grids Ω 0 Ω 1 Ω K with mesh widths h 0 > h 1 > > h K On regular grids one typically takes h k = 2h k+1 to make restrictions and prolongations simple We want to formulate multigrid methods for solving on Ω K using all available grids T K u K = f K

74 Assumptions There are restriction operators R k 1 k : Ω k Ω k 1 There are prolongation operators P k+1 k : Ω k Ω k+1 The restrictions and prolongations must be very fast and run in O(N) operations There is an iterative method with the smoothing property on every grid Ω k, i.e., it must attenuate high frequencies by a constant factor independent of h k T 0 u 0 = f 0 can be solved fast to provide a crude estimate u 0

75 The operators {T k } It may be difficult to construct the T k needed on different grids Standard approach: if possible use the same discretization Use restriction and prolongation operators! Definition The operator triple (T k, R k 1 k, Pk 1 k ) satisfies the Galerkin condition if T k 1 = R k 1 k T k Pk 1 k where R k 1 k is the restriction and P k k 1 is the prolongation The Galerkin condition preserves ellipticity on all grids

76 The full, recursive multigrid V-cycle 1. Procedure v := FMGV (v k, T k, f k ) 2. If k = 0 return v T 0 \f 0 and STOP else: 3. Pre-smoothing v k v k γd 1 k (T kv k f k ) 4. Restrict residual to coarse grid r k 1 R k 1 k (T k v k f k ) 5. Recursive error estimation e k 1 FMGV (0, T k 1, r k 1 ) 6. Prolong error to fine grid e k P k k 1 e k 1 7. Correct by removing error v k v k e k 8. Post-smoothing v k v k γd 1 k (T kv k f k ) 9. Output v v k

77 Starting procedure Obtain initial approximation v 0 = T 0 \f 0 using direct solver on coarse grid Sweep down on successively finer grids: 1. Compute v 0 = T 0 \f 0 2. Prolong to next finer fine grid v k P k k 1 v k 1 3. Compute residual r k = T k v k f k 4. Smoothing iteration(s) v k v k γd 1 k r k 5. Prolong to next grid &c. until first v K has been obtained

78 Matlab code segment The recursive V-cycle rf = Tdx*v - f; v = v - gamma*ddx\rf; rf = Tdx*v - f; rf = lowpass(rf); rc = FMGrestrict(rf); ec = FMGV(0,T2dx,rc); ef = FMGprolong(ec); v = v - ef; rf = Tdx*v - f; v = v - gamma*ddx\rf; % Compute residual % Pre-smoothing Jacobi % Compute residual % Anti-alias filter % Restrict to coarse grid % Solve error equation % Prolong to fine grid % Remove error estimate % Compute residual % Post-smoothing Jacobi

79 Numerical test: Full MG V-cycles N = Error progression Solution: exact (g), numerical (b), err/dx 2 /10 (r) V-cycles on u = f on on fine grid with 1023 points Coarsest grid with 255 points and γ = 1/2, 2/3, 1 Jacobi

80 N = Numerical test: Full MG V-cycles Error progression Solution: exact (g), numerical (b), err/dx2 /10 (r) V-cycles on u 00 = f on on fine grid with 16k points Coarsest grid with 255 points and γ = 1/2, 2/3, 1 Jacobi c G So derlind FMNN15 Multigrid V3.15 Sampling Theory and Elements of Multigrid

81 Numerical test: Full MG V-cycles N = Error progression Solution: exact (g), numerical (b), err/dx 2 /10 (r) V-cycles on u = f on on fine grid with 256k points Coarsest grid with 255 points and γ = 1/2, 2/3, 1 Jacobi

82 7. Efficiency. Memory and execution time Let the K th grid have N internal points Then grid K 1 has (N + 1)/2 1 < N/2 internal points Total storage requirement is therefore S = N ( K ) < 2N So, independent of recursion depth memory requirement does not exceed that of using a single finer mesh

83 Memory requirement in d dimensions Let the K th grid have N d points (curse of dimension) Then grid K 1 has (N 1) d /2 d < (N/2) d points Total storage requirement in d dimensions is S d < N d (1 + 2 d dk ) N d S 1 < 2N S 2 < 4N2 3 S 3 < 8N3 7 So, for once, the curse of dimension works our way! Multigrid is always cheaper than using a single finer grid

84 Execution time in theory Work unit U K is the cost of one sweep on the fine K th grid W 0 2 dk N 2d is cost for direct solution on coarsest grid Neglect restriction and prolongation costs Cost W of V-cycle with one pre- and one post-smoothing W d < 2U K (1 + 2 d dk ) + 2W 0 2U K + 2W 0 W 1 < 4U K + 2W 0 ; W 2 < 8U K 3 + 2W 0 ; W 3 < 16U K 7 + 2W 0 Execution time less than that of a single sweep on next finer grid

85 So it s cheap, what accuracy do we get? Solving T h u h = f we obtain v h with Algebraic error e h = v h u h Global error g h = u h u(x h ) Total error δ h = v h u(x h ) Note that δ h = e h g h, therefore v h u(x h ) x v h u h x + u h u(x h ) x ε MG + C h 2 Try to make ε MG C x 2 to make global error dominate

86 Convergence analysis a brief sketch Assume linear convergence rate ρ for each V-cycle Run M V-cycles, reducing error to global error level if ρ M h 2 N 2 for a 2 nd order method, implying number of iterations M = O(log N) Because the cost of each V-cycle is U K = O(N d ) total cost to reduce the error to the global error level is O(N d log N) = C cost(fft)

87 6. Convection diffusion u t = u xx + αu x f Stationary problem, homogeneous Dirichlet conditions on [0, 1] L α u = u + αu = f Convection dominated for Péclet numbers α 1 Boundary layer at x = 0 for α > 0, otherwise at x = 1 L α is not self-adjoint, but w(x) = e αx is a symmetrizer: wl α u = wu + αwu = (wu ) so (wl α ) = wl α. Pre- and postmultiply by w 1/2 produces A α = w 1/2 L α w 1/2 = w 1/2 L αw 1/2 = (w 1/2 L α w 1/2 ) = A α

88 Eigenfunctions L α u = λu Put u = w 1/2 v. Then L α w 1/2 v = λw 1/2 v, and L α w 1/2 v = (w 1/2 v) + α(w 1/2 v) = λw 1/2 v simplifies to so ( ) v k = λ k + α2 v k 4 λ k [L α ] = (kπ) 2 α2 4 u k (x) = e αx/2 sin kπx Note L α is equi-elliptic wrt Péclet number as M 2 [L α ] = π 2

89 FDM discretization Convection diffusion N N Toeplitz matrix T h with h = 1/(N + 1) hα/2 T h = 1 1 hα/ hα/2 h hα/2 1 hα/2 2 Unsymmetric system T h u h = f Discrete diagonal exponential symmetrizer exists

90 Eigenvalues of unsymmetric Toeplitz matrices An N N Toeplitz matrix T =. d b a d b.. b a d has eigenvalues λ k [T ] = d + 2 ab cos kπ N + 1, k = 1 : N Note Real eigenvalues if ab > 0, otherwise complex conjugate

91 Discrete eigenvalues The mesh Péclet number For the FDM approximation of the convection diffusion operator L α the discrete eigenvalues are λ k [T h ] = 2 h h 2 1 h2 α 2 cos θk h 4, k = 1 : N f Note λ k [L α ] is always real, but λ k [T h ] is real only if h α < 2 The mesh Péclet number limits H which must satisfy H α < 2

92 Computational experiment High Péclet numbers x 10 4 Original C grid error (exact) x 10 5 C grid error estimate x 10 5 Exact F grid error (r) and scaled, exact C grid error (b) Classical MG α = 200, H = 1/500, ρ = 2

93 Computational experiment High Péclet numbers 10 4 Convergence history: discretization (r) and algebraic errors (b) Iteration number Classical MG α = 200, H = 1/500, ρ = 2

94 Eigenvalue locations T h, T H Convection diffusion Mesh Péclet condition satisfied (principle sketch)

Kasetsart University Workshop. Multigrid methods: An introduction

Kasetsart University Workshop. Multigrid methods: An introduction Kasetsart University Workshop Multigrid methods: An introduction Dr. Anand Pardhanani Mathematics Department Earlham College Richmond, Indiana USA pardhan@earlham.edu A copy of these slides is available

More information

Introduction to Multigrid Methods

Introduction to Multigrid Methods Introduction to Multigrid Methods Chapter 9: Multigrid Methodology and Applications Gustaf Söderlind Numerical Analysis, Lund University Textbooks: A Multigrid Tutorial, by William L Briggs. SIAM 1988

More information

1. Fast Iterative Solvers of SLE

1. Fast Iterative Solvers of SLE 1. Fast Iterative Solvers of crucial drawback of solvers discussed so far: they become slower if we discretize more accurate! now: look for possible remedies relaxation: explicit application of the multigrid

More information

Computational Linear Algebra

Computational Linear Algebra Computational Linear Algebra PD Dr. rer. nat. habil. Ralf-Peter Mundani Computation in Engineering / BGU Scientific Computing in Computer Science / INF Winter Term 2018/19 Part 4: Iterative Methods PD

More information

Aspects of Multigrid

Aspects of Multigrid Aspects of Multigrid Kees Oosterlee 1,2 1 Delft University of Technology, Delft. 2 CWI, Center for Mathematics and Computer Science, Amsterdam, SIAM Chapter Workshop Day, May 30th 2018 C.W.Oosterlee (CWI)

More information

Introduction to Multigrid Method

Introduction to Multigrid Method Introduction to Multigrid Metod Presented by: Bogojeska Jasmina /08/005 JASS, 005, St. Petersburg 1 Te ultimate upsot of MLAT Te amount of computational work sould be proportional to te amount of real

More information

Iterative Methods and Multigrid

Iterative Methods and Multigrid Iterative Methods and Multigrid Part 1: Introduction to Multigrid 1 12/02/09 MG02.prz Error Smoothing 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 Initial Solution=-Error 0 10 20 30 40 50 60 70 80 90 100 DCT:

More information

Multigrid Methods and their application in CFD

Multigrid Methods and their application in CFD Multigrid Methods and their application in CFD Michael Wurst TU München 16.06.2009 1 Multigrid Methods Definition Multigrid (MG) methods in numerical analysis are a group of algorithms for solving differential

More information

6. Multigrid & Krylov Methods. June 1, 2010

6. Multigrid & Krylov Methods. June 1, 2010 June 1, 2010 Scientific Computing II, Tobias Weinzierl page 1 of 27 Outline of This Session A recapitulation of iterative schemes Lots of advertisement Multigrid Ingredients Multigrid Analysis Scientific

More information

INTRODUCTION TO MULTIGRID METHODS

INTRODUCTION TO MULTIGRID METHODS INTRODUCTION TO MULTIGRID METHODS LONG CHEN 1. ALGEBRAIC EQUATION OF TWO POINT BOUNDARY VALUE PROBLEM We consider the discretization of Poisson equation in one dimension: (1) u = f, x (0, 1) u(0) = u(1)

More information

Stabilization and Acceleration of Algebraic Multigrid Method

Stabilization and Acceleration of Algebraic Multigrid Method Stabilization and Acceleration of Algebraic Multigrid Method Recursive Projection Algorithm A. Jemcov J.P. Maruszewski Fluent Inc. October 24, 2006 Outline 1 Need for Algorithm Stabilization and Acceleration

More information

AMS526: Numerical Analysis I (Numerical Linear Algebra)

AMS526: Numerical Analysis I (Numerical Linear Algebra) AMS526: Numerical Analysis I (Numerical Linear Algebra) Lecture 24: Preconditioning and Multigrid Solver Xiangmin Jiao SUNY Stony Brook Xiangmin Jiao Numerical Analysis I 1 / 5 Preconditioning Motivation:

More information

Adaptive algebraic multigrid methods in lattice computations

Adaptive algebraic multigrid methods in lattice computations Adaptive algebraic multigrid methods in lattice computations Karsten Kahl Bergische Universität Wuppertal January 8, 2009 Acknowledgements Matthias Bolten, University of Wuppertal Achi Brandt, Weizmann

More information

Numerical Programming I (for CSE)

Numerical Programming I (for CSE) Technische Universität München WT 1/13 Fakultät für Mathematik Prof. Dr. M. Mehl B. Gatzhammer January 1, 13 Numerical Programming I (for CSE) Tutorial 1: Iterative Methods 1) Relaxation Methods a) Let

More information

Multigrid absolute value preconditioning

Multigrid absolute value preconditioning Multigrid absolute value preconditioning Eugene Vecharynski 1 Andrew Knyazev 2 (speaker) 1 Department of Computer Science and Engineering University of Minnesota 2 Department of Mathematical and Statistical

More information

Introduction to Scientific Computing II Multigrid

Introduction to Scientific Computing II Multigrid Introduction to Scientific Computing II Multigrid Miriam Mehl Slide 5: Relaxation Methods Properties convergence depends on method clear, see exercises and 3), frequency of the error remember eigenvectors

More information

Constrained Minimization and Multigrid

Constrained Minimization and Multigrid Constrained Minimization and Multigrid C. Gräser (FU Berlin), R. Kornhuber (FU Berlin), and O. Sander (FU Berlin) Workshop on PDE Constrained Optimization Hamburg, March 27-29, 2008 Matheon Outline Successive

More information

Solving PDEs with Multigrid Methods p.1

Solving PDEs with Multigrid Methods p.1 Solving PDEs with Multigrid Methods Scott MacLachlan maclachl@colorado.edu Department of Applied Mathematics, University of Colorado at Boulder Solving PDEs with Multigrid Methods p.1 Support and Collaboration

More information

Robust solution of Poisson-like problems with aggregation-based AMG

Robust solution of Poisson-like problems with aggregation-based AMG Robust solution of Poisson-like problems with aggregation-based AMG Yvan Notay Université Libre de Bruxelles Service de Métrologie Nucléaire Paris, January 26, 215 Supported by the Belgian FNRS http://homepages.ulb.ac.be/

More information

Solving Symmetric Indefinite Systems with Symmetric Positive Definite Preconditioners

Solving Symmetric Indefinite Systems with Symmetric Positive Definite Preconditioners Solving Symmetric Indefinite Systems with Symmetric Positive Definite Preconditioners Eugene Vecharynski 1 Andrew Knyazev 2 1 Department of Computer Science and Engineering University of Minnesota 2 Department

More information

Elliptic Problems / Multigrid. PHY 604: Computational Methods for Physics and Astrophysics II

Elliptic Problems / Multigrid. PHY 604: Computational Methods for Physics and Astrophysics II Elliptic Problems / Multigrid Summary of Hyperbolic PDEs We looked at a simple linear and a nonlinear scalar hyperbolic PDE There is a speed associated with the change of the solution Explicit methods

More information

Algebraic Multigrid as Solvers and as Preconditioner

Algebraic Multigrid as Solvers and as Preconditioner Ò Algebraic Multigrid as Solvers and as Preconditioner Domenico Lahaye domenico.lahaye@cs.kuleuven.ac.be http://www.cs.kuleuven.ac.be/ domenico/ Department of Computer Science Katholieke Universiteit Leuven

More information

Geometric Multigrid Methods

Geometric Multigrid Methods Geometric Multigrid Methods Susanne C. Brenner Department of Mathematics and Center for Computation & Technology Louisiana State University IMA Tutorial: Fast Solution Techniques November 28, 2010 Ideas

More information

Multigrid finite element methods on semi-structured triangular grids

Multigrid finite element methods on semi-structured triangular grids XXI Congreso de Ecuaciones Diferenciales y Aplicaciones XI Congreso de Matemática Aplicada Ciudad Real, -5 septiembre 009 (pp. 8) Multigrid finite element methods on semi-structured triangular grids F.J.

More information

Geometric Multigrid Methods for the Helmholtz equations

Geometric Multigrid Methods for the Helmholtz equations Geometric Multigrid Methods for the Helmholtz equations Ira Livshits Ball State University RICAM, Linz, 4, 6 November 20 Ira Livshits (BSU) - November 20, Linz / 83 Multigrid Methods Aim: To understand

More information

6. Iterative Methods for Linear Systems. The stepwise approach to the solution...

6. Iterative Methods for Linear Systems. The stepwise approach to the solution... 6 Iterative Methods for Linear Systems The stepwise approach to the solution Miriam Mehl: 6 Iterative Methods for Linear Systems The stepwise approach to the solution, January 18, 2013 1 61 Large Sparse

More information

First-order overdetermined systems. for elliptic problems. John Strain Mathematics Department UC Berkeley July 2012

First-order overdetermined systems. for elliptic problems. John Strain Mathematics Department UC Berkeley July 2012 First-order overdetermined systems for elliptic problems John Strain Mathematics Department UC Berkeley July 2012 1 OVERVIEW Convert elliptic problems to first-order overdetermined form Control error via

More information

Scientific Computing: An Introductory Survey

Scientific Computing: An Introductory Survey Scientific Computing: An Introductory Survey Chapter 11 Partial Differential Equations Prof. Michael T. Heath Department of Computer Science University of Illinois at Urbana-Champaign Copyright c 2002.

More information

The Conjugate Gradient Method

The Conjugate Gradient Method The Conjugate Gradient Method Classical Iterations We have a problem, We assume that the matrix comes from a discretization of a PDE. The best and most popular model problem is, The matrix will be as large

More information

University of Illinois at Urbana-Champaign. Multigrid (MG) methods are used to approximate solutions to elliptic partial differential

University of Illinois at Urbana-Champaign. Multigrid (MG) methods are used to approximate solutions to elliptic partial differential Title: Multigrid Methods Name: Luke Olson 1 Affil./Addr.: Department of Computer Science University of Illinois at Urbana-Champaign Urbana, IL 61801 email: lukeo@illinois.edu url: http://www.cs.uiuc.edu/homes/lukeo/

More information

INTERGRID OPERATORS FOR THE CELL CENTERED FINITE DIFFERENCE MULTIGRID ALGORITHM ON RECTANGULAR GRIDS. 1. Introduction

INTERGRID OPERATORS FOR THE CELL CENTERED FINITE DIFFERENCE MULTIGRID ALGORITHM ON RECTANGULAR GRIDS. 1. Introduction Trends in Mathematics Information Center for Mathematical Sciences Volume 9 Number 2 December 2006 Pages 0 INTERGRID OPERATORS FOR THE CELL CENTERED FINITE DIFFERENCE MULTIGRID ALGORITHM ON RECTANGULAR

More information

Chapter Two: Numerical Methods for Elliptic PDEs. 1 Finite Difference Methods for Elliptic PDEs

Chapter Two: Numerical Methods for Elliptic PDEs. 1 Finite Difference Methods for Elliptic PDEs Chapter Two: Numerical Methods for Elliptic PDEs Finite Difference Methods for Elliptic PDEs.. Finite difference scheme. We consider a simple example u := subject to Dirichlet boundary conditions ( ) u

More information

Simple Examples on Rectangular Domains

Simple Examples on Rectangular Domains 84 Chapter 5 Simple Examples on Rectangular Domains In this chapter we consider simple elliptic boundary value problems in rectangular domains in R 2 or R 3 ; our prototype example is the Poisson equation

More information

An Introduction of Multigrid Methods for Large-Scale Computation

An Introduction of Multigrid Methods for Large-Scale Computation An Introduction of Multigrid Methods for Large-Scale Computation Chin-Tien Wu National Center for Theoretical Sciences National Tsing-Hua University 01/4/005 How Large the Real Simulations Are? Large-scale

More information

Numerical algorithms for one and two target optimal controls

Numerical algorithms for one and two target optimal controls Numerical algorithms for one and two target optimal controls Sung-Sik Kwon Department of Mathematics and Computer Science, North Carolina Central University 80 Fayetteville St. Durham, NC 27707 Email:

More information

New Multigrid Solver Advances in TOPS

New Multigrid Solver Advances in TOPS New Multigrid Solver Advances in TOPS R D Falgout 1, J Brannick 2, M Brezina 2, T Manteuffel 2 and S McCormick 2 1 Center for Applied Scientific Computing, Lawrence Livermore National Laboratory, P.O.

More information

An Introduction to Algebraic Multigrid (AMG) Algorithms Derrick Cerwinsky and Craig C. Douglas 1/84

An Introduction to Algebraic Multigrid (AMG) Algorithms Derrick Cerwinsky and Craig C. Douglas 1/84 An Introduction to Algebraic Multigrid (AMG) Algorithms Derrick Cerwinsky and Craig C. Douglas 1/84 Introduction Almost all numerical methods for solving PDEs will at some point be reduced to solving A

More information

The Fourier spectral method (Amath Bretherton)

The Fourier spectral method (Amath Bretherton) The Fourier spectral method (Amath 585 - Bretherton) 1 Introduction The Fourier spectral method (or more precisely, pseudospectral method) is a very accurate way to solve BVPs with smooth solutions on

More information

EFFICIENT MULTIGRID BASED SOLVERS FOR ISOGEOMETRIC ANALYSIS

EFFICIENT MULTIGRID BASED SOLVERS FOR ISOGEOMETRIC ANALYSIS 6th European Conference on Computational Mechanics (ECCM 6) 7th European Conference on Computational Fluid Dynamics (ECFD 7) 1115 June 2018, Glasgow, UK EFFICIENT MULTIGRID BASED SOLVERS FOR ISOGEOMETRIC

More information

2.4 Eigenvalue problems

2.4 Eigenvalue problems 2.4 Eigenvalue problems Associated with the boundary problem (2.1) (Poisson eq.), we call λ an eigenvalue if Lu = λu (2.36) for a nonzero function u C 2 0 ((0, 1)). Recall Lu = u. Then u is called an eigenfunction.

More information

Iterative Methods and Multigrid

Iterative Methods and Multigrid Iterative Methods and Multigrid Part 1: Introduction to Multigrid 2000 Eric de Sturler 1 12/02/09 MG01.prz Basic Iterative Methods (1) Nonlinear equation: f(x) = 0 Rewrite as x = F(x), and iterate x i+1

More information

Notes for CS542G (Iterative Solvers for Linear Systems)

Notes for CS542G (Iterative Solvers for Linear Systems) Notes for CS542G (Iterative Solvers for Linear Systems) Robert Bridson November 20, 2007 1 The Basics We re now looking at efficient ways to solve the linear system of equations Ax = b where in this course,

More information

Name: INSERT YOUR NAME HERE. Due to dropbox by 6pm PDT, Wednesday, December 14, 2011

Name: INSERT YOUR NAME HERE. Due to dropbox by 6pm PDT, Wednesday, December 14, 2011 AMath 584 Name: INSERT YOUR NAME HERE Take-home Final UWNetID: INSERT YOUR NETID Due to dropbox by 6pm PDT, Wednesday, December 14, 2011 The main part of the assignment (Problems 1 3) is worth 80 points.

More information

DELFT UNIVERSITY OF TECHNOLOGY

DELFT UNIVERSITY OF TECHNOLOGY DELFT UNIVERSITY OF TECHNOLOGY REPORT 06-09 Multigrid for High Dimensional Elliptic Partial Differential Equations on Non-equidistant Grids H. bin Zubair, C.W. Oosterlee, R. Wienands ISSN 389-6520 Reports

More information

Optimal multilevel preconditioning of strongly anisotropic problems.part II: non-conforming FEM. p. 1/36

Optimal multilevel preconditioning of strongly anisotropic problems.part II: non-conforming FEM. p. 1/36 Optimal multilevel preconditioning of strongly anisotropic problems. Part II: non-conforming FEM. Svetozar Margenov margenov@parallel.bas.bg Institute for Parallel Processing, Bulgarian Academy of Sciences,

More information

Von Neumann Analysis of Jacobi and Gauss-Seidel Iterations

Von Neumann Analysis of Jacobi and Gauss-Seidel Iterations Von Neumann Analysis of Jacobi and Gauss-Seidel Iterations We consider the FDA to the 1D Poisson equation on a grid x covering [0,1] with uniform spacing h, h (u +1 u + u 1 ) = f whose exact solution (to

More information

Multigrid Method for 2D Helmholtz Equation using Higher Order Finite Difference Scheme Accelerated by Krylov Subspace

Multigrid Method for 2D Helmholtz Equation using Higher Order Finite Difference Scheme Accelerated by Krylov Subspace 201, TextRoad Publication ISSN: 2090-27 Journal of Applied Environmental and Biological Sciences www.textroad.com Multigrid Method for 2D Helmholtz Equation using Higher Order Finite Difference Scheme

More information

Notes on Multigrid Methods

Notes on Multigrid Methods Notes on Multigrid Metods Qingai Zang April, 17 Motivation of multigrids. Te convergence rates of classical iterative metod depend on te grid spacing, or problem size. In contrast, convergence rates of

More information

Comparison of V-cycle Multigrid Method for Cell-centered Finite Difference on Triangular Meshes

Comparison of V-cycle Multigrid Method for Cell-centered Finite Difference on Triangular Meshes Comparison of V-cycle Multigrid Method for Cell-centered Finite Difference on Triangular Meshes Do Y. Kwak, 1 JunS.Lee 1 Department of Mathematics, KAIST, Taejon 305-701, Korea Department of Mathematics,

More information

Lab 1: Iterative Methods for Solving Linear Systems

Lab 1: Iterative Methods for Solving Linear Systems Lab 1: Iterative Methods for Solving Linear Systems January 22, 2017 Introduction Many real world applications require the solution to very large and sparse linear systems where direct methods such as

More information

Chapter 7 Iterative Techniques in Matrix Algebra

Chapter 7 Iterative Techniques in Matrix Algebra Chapter 7 Iterative Techniques in Matrix Algebra Per-Olof Persson persson@berkeley.edu Department of Mathematics University of California, Berkeley Math 128B Numerical Analysis Vector Norms Definition

More information

Theory of Iterative Methods

Theory of Iterative Methods Based on Strang s Introduction to Applied Mathematics Theory of Iterative Methods The Iterative Idea To solve Ax = b, write Mx (k+1) = (M A)x (k) + b, k = 0, 1,,... Then the error e (k) x (k) x satisfies

More information

An efficient multigrid solver based on aggregation

An efficient multigrid solver based on aggregation An efficient multigrid solver based on aggregation Yvan Notay Université Libre de Bruxelles Service de Métrologie Nucléaire Graz, July 4, 2012 Co-worker: Artem Napov Supported by the Belgian FNRS http://homepages.ulb.ac.be/

More information

Time-dependent variational forms

Time-dependent variational forms Time-dependent variational forms Hans Petter Langtangen 1,2 1 Center for Biomedical Computing, Simula Research Laboratory 2 Department of Informatics, University of Oslo Oct 30, 2015 PRELIMINARY VERSION

More information

Lecture 9 Approximations of Laplace s Equation, Finite Element Method. Mathématiques appliquées (MATH0504-1) B. Dewals, C.

Lecture 9 Approximations of Laplace s Equation, Finite Element Method. Mathématiques appliquées (MATH0504-1) B. Dewals, C. Lecture 9 Approximations of Laplace s Equation, Finite Element Method Mathématiques appliquées (MATH54-1) B. Dewals, C. Geuzaine V1.2 23/11/218 1 Learning objectives of this lecture Apply the finite difference

More information

Numerical Analysis of Differential Equations Numerical Solution of Elliptic Boundary Value

Numerical Analysis of Differential Equations Numerical Solution of Elliptic Boundary Value Numerical Analysis of Differential Equations 188 5 Numerical Solution of Elliptic Boundary Value Problems 5 Numerical Solution of Elliptic Boundary Value Problems TU Bergakademie Freiberg, SS 2012 Numerical

More information

Numerical Solutions to Partial Differential Equations

Numerical Solutions to Partial Differential Equations Numerical Solutions to Partial Differential Equations Zhiping Li LMAM and School of Mathematical Sciences Peking University Variational Problems of the Dirichlet BVP of the Poisson Equation 1 For the homogeneous

More information

Bootstrap AMG. Kailai Xu. July 12, Stanford University

Bootstrap AMG. Kailai Xu. July 12, Stanford University Bootstrap AMG Kailai Xu Stanford University July 12, 2017 AMG Components A general AMG algorithm consists of the following components. A hierarchy of levels. A smoother. A prolongation. A restriction.

More information

Solving the stochastic steady-state diffusion problem using multigrid

Solving the stochastic steady-state diffusion problem using multigrid IMA Journal of Numerical Analysis (2007) 27, 675 688 doi:10.1093/imanum/drm006 Advance Access publication on April 9, 2007 Solving the stochastic steady-state diffusion problem using multigrid HOWARD ELMAN

More information

Multilevel Preconditioning of Graph-Laplacians: Polynomial Approximation of the Pivot Blocks Inverses

Multilevel Preconditioning of Graph-Laplacians: Polynomial Approximation of the Pivot Blocks Inverses Multilevel Preconditioning of Graph-Laplacians: Polynomial Approximation of the Pivot Blocks Inverses P. Boyanova 1, I. Georgiev 34, S. Margenov, L. Zikatanov 5 1 Uppsala University, Box 337, 751 05 Uppsala,

More information

MULTIGRID METHODS FOR NONLINEAR PROBLEMS: AN OVERVIEW

MULTIGRID METHODS FOR NONLINEAR PROBLEMS: AN OVERVIEW MULTIGRID METHODS FOR NONLINEAR PROBLEMS: AN OVERVIEW VAN EMDEN HENSON CENTER FOR APPLIED SCIENTIFIC COMPUTING LAWRENCE LIVERMORE NATIONAL LABORATORY Abstract Since their early application to elliptic

More information

An Introduction to the Discontinuous Galerkin Method

An Introduction to the Discontinuous Galerkin Method An Introduction to the Discontinuous Galerkin Method Krzysztof J. Fidkowski Aerospace Computational Design Lab Massachusetts Institute of Technology March 16, 2005 Computational Prototyping Group Seminar

More information

An Efficient Low Memory Implicit DG Algorithm for Time Dependent Problems

An Efficient Low Memory Implicit DG Algorithm for Time Dependent Problems An Efficient Low Memory Implicit DG Algorithm for Time Dependent Problems P.-O. Persson and J. Peraire Massachusetts Institute of Technology 2006 AIAA Aerospace Sciences Meeting, Reno, Nevada January 9,

More information

Electromagnetic wave propagation. ELEC 041-Modeling and design of electromagnetic systems

Electromagnetic wave propagation. ELEC 041-Modeling and design of electromagnetic systems Electromagnetic wave propagation ELEC 041-Modeling and design of electromagnetic systems EM wave propagation In general, open problems with a computation domain extending (in theory) to infinity not bounded

More information

Background. Background. C. T. Kelley NC State University tim C. T. Kelley Background NCSU, Spring / 58

Background. Background. C. T. Kelley NC State University tim C. T. Kelley Background NCSU, Spring / 58 Background C. T. Kelley NC State University tim kelley@ncsu.edu C. T. Kelley Background NCSU, Spring 2012 1 / 58 Notation vectors, matrices, norms l 1 : max col sum... spectral radius scaled integral norms

More information

Solving Boundary Value Problems (with Gaussians)

Solving Boundary Value Problems (with Gaussians) What is a boundary value problem? Solving Boundary Value Problems (with Gaussians) Definition A differential equation with constraints on the boundary Michael McCourt Division Argonne National Laboratory

More information

6. Iterative Methods: Roots and Optima. Citius, Altius, Fortius!

6. Iterative Methods: Roots and Optima. Citius, Altius, Fortius! Citius, Altius, Fortius! Numerisches Programmieren, Hans-Joachim Bungartz page 1 of 1 6.1. Large Sparse Systems of Linear Equations I Relaxation Methods Introduction Systems of linear equations, which

More information

Outline. 1 Partial Differential Equations. 2 Numerical Methods for PDEs. 3 Sparse Linear Systems

Outline. 1 Partial Differential Equations. 2 Numerical Methods for PDEs. 3 Sparse Linear Systems Outline Partial Differential Equations Numerical Methods for PDEs Sparse Linear Systems 1 Partial Differential Equations 2 Numerical Methods for PDEs 3 Sparse Linear Systems Michael T Heath Scientific

More information

K.S. Kang. The multigrid method for an elliptic problem on a rectangular domain with an internal conductiong structure and an inner empty space

K.S. Kang. The multigrid method for an elliptic problem on a rectangular domain with an internal conductiong structure and an inner empty space K.S. Kang The multigrid method for an elliptic problem on a rectangular domain with an internal conductiong structure and an inner empty space IPP 5/128 September, 2011 The multigrid method for an elliptic

More information

Jim Lambers ENERGY 281 Spring Quarter Lecture 5 Notes

Jim Lambers ENERGY 281 Spring Quarter Lecture 5 Notes Jim ambers ENERGY 28 Spring Quarter 27-8 ecture 5 Notes These notes are based on Rosalind Archer s PE28 lecture notes, with some revisions by Jim ambers. Fourier Series Recall that in ecture 2, when we

More information

Preface to the Second Edition. Preface to the First Edition

Preface to the Second Edition. Preface to the First Edition n page v Preface to the Second Edition Preface to the First Edition xiii xvii 1 Background in Linear Algebra 1 1.1 Matrices................................. 1 1.2 Square Matrices and Eigenvalues....................

More information

Diffusion on the half-line. The Dirichlet problem

Diffusion on the half-line. The Dirichlet problem Diffusion on the half-line The Dirichlet problem Consider the initial boundary value problem (IBVP) on the half line (, ): v t kv xx = v(x, ) = φ(x) v(, t) =. The solution will be obtained by the reflection

More information

Iterative Methods for Solving A x = b

Iterative Methods for Solving A x = b Iterative Methods for Solving A x = b A good (free) online source for iterative methods for solving A x = b is given in the description of a set of iterative solvers called templates found at netlib: http

More information

Math 46, Applied Math (Spring 2008): Final

Math 46, Applied Math (Spring 2008): Final Math 46, Applied Math (Spring 2008): Final 3 hours, 80 points total, 9 questions, roughly in syllabus order (apart from short answers) 1. [16 points. Note part c, worth 7 points, is independent of the

More information

arxiv: v1 [math.na] 6 Nov 2017

arxiv: v1 [math.na] 6 Nov 2017 Efficient boundary corrected Strang splitting Lukas Einkemmer Martina Moccaldi Alexander Ostermann arxiv:1711.02193v1 [math.na] 6 Nov 2017 Version of November 6, 2017 Abstract Strang splitting is a well

More information

Chapter 5. Methods for Solving Elliptic Equations

Chapter 5. Methods for Solving Elliptic Equations Chapter 5. Methods for Solving Elliptic Equations References: Tannehill et al Section 4.3. Fulton et al (1986 MWR). Recommended reading: Chapter 7, Numerical Methods for Engineering Application. J. H.

More information

Aggregation-based algebraic multigrid

Aggregation-based algebraic multigrid Aggregation-based algebraic multigrid from theory to fast solvers Yvan Notay Université Libre de Bruxelles Service de Métrologie Nucléaire CEMRACS, Marseille, July 18, 2012 Supported by the Belgian FNRS

More information

Linear Solvers. Andrew Hazel

Linear Solvers. Andrew Hazel Linear Solvers Andrew Hazel Introduction Thus far we have talked about the formulation and discretisation of physical problems...... and stopped when we got to a discrete linear system of equations. Introduction

More information

Sparse Matrix Techniques for MCAO

Sparse Matrix Techniques for MCAO Sparse Matrix Techniques for MCAO Luc Gilles lgilles@mtu.edu Michigan Technological University, ECE Department Brent Ellerbroek bellerbroek@gemini.edu Gemini Observatory Curt Vogel vogel@math.montana.edu

More information

INTRODUCTION TO DELTA-SIGMA ADCS

INTRODUCTION TO DELTA-SIGMA ADCS ECE37 Advanced Analog Circuits INTRODUCTION TO DELTA-SIGMA ADCS Richard Schreier richard.schreier@analog.com NLCOTD: Level Translator VDD > VDD2, e.g. 3-V logic? -V logic VDD < VDD2, e.g. -V logic? 3-V

More information

DISCRETE MAXIMUM PRINCIPLES in THE FINITE ELEMENT SIMULATIONS

DISCRETE MAXIMUM PRINCIPLES in THE FINITE ELEMENT SIMULATIONS DISCRETE MAXIMUM PRINCIPLES in THE FINITE ELEMENT SIMULATIONS Sergey Korotov BCAM Basque Center for Applied Mathematics http://www.bcamath.org 1 The presentation is based on my collaboration with several

More information

Scientific Computing I

Scientific Computing I Scientific Computing I Module 8: An Introduction to Finite Element Methods Tobias Neckel Winter 2013/2014 Module 8: An Introduction to Finite Element Methods, Winter 2013/2014 1 Part I: Introduction to

More information

Numerical Methods I Orthogonal Polynomials

Numerical Methods I Orthogonal Polynomials Numerical Methods I Orthogonal Polynomials Aleksandar Donev Courant Institute, NYU 1 donev@courant.nyu.edu 1 Course G63.2010.001 / G22.2420-001, Fall 2010 Nov. 4th and 11th, 2010 A. Donev (Courant Institute)

More information

Preconditioned Locally Minimal Residual Method for Computing Interior Eigenpairs of Symmetric Operators

Preconditioned Locally Minimal Residual Method for Computing Interior Eigenpairs of Symmetric Operators Preconditioned Locally Minimal Residual Method for Computing Interior Eigenpairs of Symmetric Operators Eugene Vecharynski 1 Andrew Knyazev 2 1 Department of Computer Science and Engineering University

More information

A multilevel algorithm for inverse problems with elliptic PDE constraints

A multilevel algorithm for inverse problems with elliptic PDE constraints A multilevel algorithm for inverse problems with elliptic PDE constraints George Biros and Günay Doǧan School of Engineering and Applied Science, University of Pennsylvania, Philadelphia, PA 19104 E-mail:

More information

Multigrid Algorithms for Inverse Problems with Linear Parabolic PDE Constraints

Multigrid Algorithms for Inverse Problems with Linear Parabolic PDE Constraints University of Pennsylvania ScholarlyCommons Departmental Papers (MEAM) Department of Mechanical Engineering & Applied Mechanics 10-16-008 Multigrid Algorithms for Inverse Problems with Linear Parabolic

More information

Lecture 18 Classical Iterative Methods

Lecture 18 Classical Iterative Methods Lecture 18 Classical Iterative Methods MIT 18.335J / 6.337J Introduction to Numerical Methods Per-Olof Persson November 14, 2006 1 Iterative Methods for Linear Systems Direct methods for solving Ax = b,

More information

Domain decomposition on different levels of the Jacobi-Davidson method

Domain decomposition on different levels of the Jacobi-Davidson method hapter 5 Domain decomposition on different levels of the Jacobi-Davidson method Abstract Most computational work of Jacobi-Davidson [46], an iterative method suitable for computing solutions of large dimensional

More information

Lecture4 INF-MAT : 5. Fast Direct Solution of Large Linear Systems

Lecture4 INF-MAT : 5. Fast Direct Solution of Large Linear Systems Lecture4 INF-MAT 4350 2010: 5. Fast Direct Solution of Large Linear Systems Tom Lyche Centre of Mathematics for Applications, Department of Informatics, University of Oslo September 16, 2010 Test Matrix

More information

Key words. preconditioned conjugate gradient method, saddle point problems, optimal control of PDEs, control and state constraints, multigrid method

Key words. preconditioned conjugate gradient method, saddle point problems, optimal control of PDEs, control and state constraints, multigrid method PRECONDITIONED CONJUGATE GRADIENT METHOD FOR OPTIMAL CONTROL PROBLEMS WITH CONTROL AND STATE CONSTRAINTS ROLAND HERZOG AND EKKEHARD SACHS Abstract. Optimality systems and their linearizations arising in

More information

Notes on PCG for Sparse Linear Systems

Notes on PCG for Sparse Linear Systems Notes on PCG for Sparse Linear Systems Luca Bergamaschi Department of Civil Environmental and Architectural Engineering University of Padova e-mail luca.bergamaschi@unipd.it webpage www.dmsa.unipd.it/

More information

Poisson Solvers. William McLean. April 21, Return to Math3301/Math5315 Common Material.

Poisson Solvers. William McLean. April 21, Return to Math3301/Math5315 Common Material. Poisson Solvers William McLean April 21, 2004 Return to Math3301/Math5315 Common Material 1 Introduction Many problems in applied mathematics lead to a partial differential equation of the form a 2 u +

More information

Multigrid Methods for Maxwell s Equations

Multigrid Methods for Maxwell s Equations Multigrid Methods for Maxwell s Equations Jintao Cui Institute for Mathematics and Its Applications University of Minnesota Outline Nonconforming Finite Element Methods for a Two Dimensional Curl-Curl

More information

Multilevel low-rank approximation preconditioners Yousef Saad Department of Computer Science and Engineering University of Minnesota

Multilevel low-rank approximation preconditioners Yousef Saad Department of Computer Science and Engineering University of Minnesota Multilevel low-rank approximation preconditioners Yousef Saad Department of Computer Science and Engineering University of Minnesota SIAM CSE Boston - March 1, 2013 First: Joint work with Ruipeng Li Work

More information

Numerical Approximation Methods for Non-Uniform Fourier Data

Numerical Approximation Methods for Non-Uniform Fourier Data Numerical Approximation Methods for Non-Uniform Fourier Data Aditya Viswanathan aditya@math.msu.edu 2014 Joint Mathematics Meetings January 18 2014 0 / 16 Joint work with Anne Gelb (Arizona State) Guohui

More information

ECE Digital Image Processing and Introduction to Computer Vision

ECE Digital Image Processing and Introduction to Computer Vision ECE592-064 Digital Image Processing and Introduction to Computer Vision Depart. of ECE, NC State University Instructor: Tianfu (Matt) Wu Spring 2017 Outline Recap, image degradation / restoration Template

More information

Recovery-Based A Posteriori Error Estimation

Recovery-Based A Posteriori Error Estimation Recovery-Based A Posteriori Error Estimation Zhiqiang Cai Purdue University Department of Mathematics, Purdue University Slide 1, March 2, 2011 Outline Introduction Diffusion Problems Higher Order Elements

More information

Introduction. Math 1080: Numerical Linear Algebra Chapter 4, Iterative Methods. Example: First Order Richardson. Strategy

Introduction. Math 1080: Numerical Linear Algebra Chapter 4, Iterative Methods. Example: First Order Richardson. Strategy Introduction Math 1080: Numerical Linear Algebra Chapter 4, Iterative Methods M. M. Sussman sussmanm@math.pitt.edu Office Hours: MW 1:45PM-2:45PM, Thack 622 Solve system Ax = b by repeatedly computing

More information

Parallel Discontinuous Galerkin Method

Parallel Discontinuous Galerkin Method Parallel Discontinuous Galerkin Method Yin Ki, NG The Chinese University of Hong Kong Aug 5, 2015 Mentors: Dr. Ohannes Karakashian, Dr. Kwai Wong Overview Project Goal Implement parallelization on Discontinuous

More information

A FAST SOLVER FOR ELLIPTIC EQUATIONS WITH HARMONIC COEFFICIENT APPROXIMATIONS

A FAST SOLVER FOR ELLIPTIC EQUATIONS WITH HARMONIC COEFFICIENT APPROXIMATIONS Proceedings of ALGORITMY 2005 pp. 222 229 A FAST SOLVER FOR ELLIPTIC EQUATIONS WITH HARMONIC COEFFICIENT APPROXIMATIONS ELENA BRAVERMAN, MOSHE ISRAELI, AND ALEXANDER SHERMAN Abstract. Based on a fast subtractional

More information