Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay...

76
Jay Gopalakrishnan Multigrid methods Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days Centre de Recherches Math ´ ematiques February, 2005 Minicourse: Part I Department of Mathematics [Slide 1 of 36]

Transcript of Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay...

Page 1: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Multigrid methodsJay Gopalakrishnan

University of Florida

Montreal Scientific Computing Days

Centre de Recherches Mathematiques

February 2005

Minicourse Part I

Department of Mathematics [Slide 1 of 36]

Jay Gopalakrishnan

Why multigrid

Multigrid techniques give algorithms that solve sparse linearsystems

Ax = b

of N unknowns with O(N) work and storage for large classesof problems It applies typically to systems arising fromdiscretization of partial differential equations

Iterative method Convergence rate estimates

Gauszlig-Seidel δ lt 1minus Ch2

SOR δ lt 1minus Ch

ADI δ lt (1minus Ch)2

k-parameter ADI δ lt (1minus Ch1k)2

Multigrid δ lt 1 independent of h

Department of Mathematics [Slide 2 of 36]

Jay Gopalakrishnan

Why multigrid

Iterative method Convergence rate estimates

Gauszlig-Seidel δ lt 1minus Ch2

SOR with best parameter δ lt 1minus Ch

ADI δ lt (1minus Ch)2

k-parameter ADI δ lt (1minus Ch1k)2

Multigrid δ lt 1 independent of h

Application minus∆u = f in Ω equiv (minus1 1)2 u = 0 on partΩ

Discretization method Linear finite elements on uniform grid of mesh-size h

Meaning of δ For many iterative methods one can prove that the iterates x(i)

satisfy x(i+1) minus x le δx(i) minus x for some 0 lt δ lt 1 in some norm middot

Department of Mathematics [Slide 2 of 36]

Jay Gopalakrishnan

Why multigrid

Iterative method Convergence rate estimates

Gauszlig-Seidel [classical] δ lt 1minus Ch2

SOR [Young 1950] δ lt 1minus Ch

ADI [Peaceman amp Rachford 1955] δ lt (1minus Ch)2

k-parameter ADI [1960rsquos] δ lt (1minus Ch1k)2

Multigrid [see below] δ lt 1 independent of h

Source

Wesselingrsquos book 1992

60rsquos [Fedorenko 1964] [Bachvalov 1966]70rsquos [Brandt 1973] [Nicolaides 1975] [Brandt 1977]80rsquos [Bank amp Dupont1981] [Braess amp Hackbusch1983] [Bramble amp Pasciak1987]

Department of Mathematics [Slide 2 of 36]

Jay Gopalakrishnan

Structure of multigrid algorithmsMultigrid algorithms are based on a sequence of meshesobtained by successive refinement

Whenever it is possible to solve on the coarsest mesh fastmultigrid algorithms allow fast solution on the finest mesh

A 2D example

k = 1 k = 2

k = J

Highlyrefined

Mesh 1 Mesh 2 Mesh J

(Coarsest mesh) (Finest mesh)

Multigrid algorithms have a recursive structure Eachmultigrid iteration typically consists of the following steps

1 Smooth errors at current grid

2 Transfer residual to next coarser grid

3 Correct iterate using the coarser residual (recursively)

Department of Mathematics [Slide 3 of 36]

Jay Gopalakrishnan

Structure of multigrid algorithmsMultigrid algorithms are based on a sequence of meshesobtained by successive refinement

Whenever it is possible to solve on the coarsest mesh fastmultigrid algorithms allow fast solution on the finest mesh

A 2D example

k = 1 k = 2

k = J

Highlyrefined

Mesh 1 Mesh 2 Mesh J

(Coarsest mesh) (Finest mesh)

Multigrid algorithms have a recursive structure Eachmultigrid iteration typically consists of the following steps

1 Smooth errors at current grid

2 Transfer residual to next coarser grid

3 Correct iterate using the coarser residual (recursively)

Department of Mathematics [Slide 3 of 36]

Jay Gopalakrishnan

Structure of multigrid algorithmsMultigrid algorithms are based on a sequence of meshesobtained by successive refinement

Whenever it is possible to solve on the coarsest mesh fastmultigrid algorithms allow fast solution on the finest mesh

Multigrid algorithms have a recursive structure Eachmultigrid iteration typically consists of the following steps

1 Smooth errors at current grid

2 Transfer residual to next coarser grid

3 Correct iterate using the coarser residual (recursively)

Department of Mathematics [Slide 3 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator Ak

To compute u = Aminus1J b iteratively we use multigrid iterations

u(i+1) = MgJ(u(i) b) i = 0 1 2

starting with some initial guess u(0) where the routineMgJ(middot middot) recursively invokes MgJminus1(middot middot) MgJminus2(middot middot)

We set Mg1(v b) equiv Aminus11 b

Idea

1 Reduce fine grid components of error

2 Reduce coarse grid components of error

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of error

2 Reduce coarse grid components of error

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorWe donrsquot have error e but we have residual r = bminusAJu(i) = AJe

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorNeed an approximation for the coarse components of Aminus1

Jr

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorApply the routine Mg

Jminus1 to r projected to the next coarser grid

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smooth errors v = SmoothJ(u(i) b)

2 Transfer residual to coarser gridr = RestrictJminus1(bminus AJv)

3 Correct by recursion w = MgJminus1(0 r)

u(i+1) = v + ProlongJ(w)Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + ProlongJ(MgJminus1(0 RestrictJminus1(bminus AJv)))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Prolong2

Restrict1

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + ProlongJ(MgJminus1(0 RestrictJminus1(bminus AJv)))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

L2

Lt2

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + LJMgJminus1(0 LtJ(bminus AJv))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Operator

Rewrite discrete problem as the operator eq

Ahuh = fh

where Ah Vh 7rarr Vh is defined by

(Ahwh vh) = (nablawhnablavh) forallwh vh isin Vh

Need multigrid to solve for uh equiv Aminus1h fh efficiently

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

Multilevel spacesVk = vh isin H

10(Ω) vh|K isin P1(K) for all elements K in

the kth level mesh

Multilevel operators At each level we also have operatorsgenerated by (nablamiddotnablamiddot) namely Ak Vk 7rarr Vk defined by

(Akv w) = (nablavnablaw) forallv w isin VkDepartment of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

Computationally this means we simply implement a change ofbasis matrix

Ω

v1 isin V1 L2v1 isin V2

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Elliptic eigenfunctionsThe smoothing component of multigrid relies on the fact thatthe eigenfunctions of elliptic operators corresponding to highereigenvalues are increasingly oscillatory

minus∆φ` = λ`φ` φ`L2(Ω) = 1

Eg here are the 1st 50th and 700th eigenfunctions of adiscrete Laplacian on an L-shaped domain

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 9 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

First observe the propagation of errors e(i)

x(i+1) = x(i) + (1λ(k)max)(Akxminus Akx

(i))

x = x + (1λ(k)max)(Akxminus Akx)

=rArr e(i+1) = e(i) minus (1λ(k)max)Ake

(i)

Hence an equivalent question is

why is I minus (1λ(k)max)Ak a smoothing operator

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλn

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated

+ ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλ

(k)max

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Eg 1 The algorithmThus all components of the algorithm are now well defined

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 2: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Why multigrid

Multigrid techniques give algorithms that solve sparse linearsystems

Ax = b

of N unknowns with O(N) work and storage for large classesof problems It applies typically to systems arising fromdiscretization of partial differential equations

Iterative method Convergence rate estimates

Gauszlig-Seidel δ lt 1minus Ch2

SOR δ lt 1minus Ch

ADI δ lt (1minus Ch)2

k-parameter ADI δ lt (1minus Ch1k)2

Multigrid δ lt 1 independent of h

Department of Mathematics [Slide 2 of 36]

Jay Gopalakrishnan

Why multigrid

Iterative method Convergence rate estimates

Gauszlig-Seidel δ lt 1minus Ch2

SOR with best parameter δ lt 1minus Ch

ADI δ lt (1minus Ch)2

k-parameter ADI δ lt (1minus Ch1k)2

Multigrid δ lt 1 independent of h

Application minus∆u = f in Ω equiv (minus1 1)2 u = 0 on partΩ

Discretization method Linear finite elements on uniform grid of mesh-size h

Meaning of δ For many iterative methods one can prove that the iterates x(i)

satisfy x(i+1) minus x le δx(i) minus x for some 0 lt δ lt 1 in some norm middot

Department of Mathematics [Slide 2 of 36]

Jay Gopalakrishnan

Why multigrid

Iterative method Convergence rate estimates

Gauszlig-Seidel [classical] δ lt 1minus Ch2

SOR [Young 1950] δ lt 1minus Ch

ADI [Peaceman amp Rachford 1955] δ lt (1minus Ch)2

k-parameter ADI [1960rsquos] δ lt (1minus Ch1k)2

Multigrid [see below] δ lt 1 independent of h

Source

Wesselingrsquos book 1992

60rsquos [Fedorenko 1964] [Bachvalov 1966]70rsquos [Brandt 1973] [Nicolaides 1975] [Brandt 1977]80rsquos [Bank amp Dupont1981] [Braess amp Hackbusch1983] [Bramble amp Pasciak1987]

Department of Mathematics [Slide 2 of 36]

Jay Gopalakrishnan

Structure of multigrid algorithmsMultigrid algorithms are based on a sequence of meshesobtained by successive refinement

Whenever it is possible to solve on the coarsest mesh fastmultigrid algorithms allow fast solution on the finest mesh

A 2D example

k = 1 k = 2

k = J

Highlyrefined

Mesh 1 Mesh 2 Mesh J

(Coarsest mesh) (Finest mesh)

Multigrid algorithms have a recursive structure Eachmultigrid iteration typically consists of the following steps

1 Smooth errors at current grid

2 Transfer residual to next coarser grid

3 Correct iterate using the coarser residual (recursively)

Department of Mathematics [Slide 3 of 36]

Jay Gopalakrishnan

Structure of multigrid algorithmsMultigrid algorithms are based on a sequence of meshesobtained by successive refinement

Whenever it is possible to solve on the coarsest mesh fastmultigrid algorithms allow fast solution on the finest mesh

A 2D example

k = 1 k = 2

k = J

Highlyrefined

Mesh 1 Mesh 2 Mesh J

(Coarsest mesh) (Finest mesh)

Multigrid algorithms have a recursive structure Eachmultigrid iteration typically consists of the following steps

1 Smooth errors at current grid

2 Transfer residual to next coarser grid

3 Correct iterate using the coarser residual (recursively)

Department of Mathematics [Slide 3 of 36]

Jay Gopalakrishnan

Structure of multigrid algorithmsMultigrid algorithms are based on a sequence of meshesobtained by successive refinement

Whenever it is possible to solve on the coarsest mesh fastmultigrid algorithms allow fast solution on the finest mesh

Multigrid algorithms have a recursive structure Eachmultigrid iteration typically consists of the following steps

1 Smooth errors at current grid

2 Transfer residual to next coarser grid

3 Correct iterate using the coarser residual (recursively)

Department of Mathematics [Slide 3 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator Ak

To compute u = Aminus1J b iteratively we use multigrid iterations

u(i+1) = MgJ(u(i) b) i = 0 1 2

starting with some initial guess u(0) where the routineMgJ(middot middot) recursively invokes MgJminus1(middot middot) MgJminus2(middot middot)

We set Mg1(v b) equiv Aminus11 b

Idea

1 Reduce fine grid components of error

2 Reduce coarse grid components of error

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of error

2 Reduce coarse grid components of error

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorWe donrsquot have error e but we have residual r = bminusAJu(i) = AJe

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorNeed an approximation for the coarse components of Aminus1

Jr

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorApply the routine Mg

Jminus1 to r projected to the next coarser grid

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smooth errors v = SmoothJ(u(i) b)

2 Transfer residual to coarser gridr = RestrictJminus1(bminus AJv)

3 Correct by recursion w = MgJminus1(0 r)

u(i+1) = v + ProlongJ(w)Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + ProlongJ(MgJminus1(0 RestrictJminus1(bminus AJv)))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Prolong2

Restrict1

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + ProlongJ(MgJminus1(0 RestrictJminus1(bminus AJv)))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

L2

Lt2

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + LJMgJminus1(0 LtJ(bminus AJv))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Operator

Rewrite discrete problem as the operator eq

Ahuh = fh

where Ah Vh 7rarr Vh is defined by

(Ahwh vh) = (nablawhnablavh) forallwh vh isin Vh

Need multigrid to solve for uh equiv Aminus1h fh efficiently

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

Multilevel spacesVk = vh isin H

10(Ω) vh|K isin P1(K) for all elements K in

the kth level mesh

Multilevel operators At each level we also have operatorsgenerated by (nablamiddotnablamiddot) namely Ak Vk 7rarr Vk defined by

(Akv w) = (nablavnablaw) forallv w isin VkDepartment of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

Computationally this means we simply implement a change ofbasis matrix

Ω

v1 isin V1 L2v1 isin V2

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Elliptic eigenfunctionsThe smoothing component of multigrid relies on the fact thatthe eigenfunctions of elliptic operators corresponding to highereigenvalues are increasingly oscillatory

minus∆φ` = λ`φ` φ`L2(Ω) = 1

Eg here are the 1st 50th and 700th eigenfunctions of adiscrete Laplacian on an L-shaped domain

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 9 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

First observe the propagation of errors e(i)

x(i+1) = x(i) + (1λ(k)max)(Akxminus Akx

(i))

x = x + (1λ(k)max)(Akxminus Akx)

=rArr e(i+1) = e(i) minus (1λ(k)max)Ake

(i)

Hence an equivalent question is

why is I minus (1λ(k)max)Ak a smoothing operator

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλn

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated

+ ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλ

(k)max

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Eg 1 The algorithmThus all components of the algorithm are now well defined

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 3: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Why multigrid

Iterative method Convergence rate estimates

Gauszlig-Seidel δ lt 1minus Ch2

SOR with best parameter δ lt 1minus Ch

ADI δ lt (1minus Ch)2

k-parameter ADI δ lt (1minus Ch1k)2

Multigrid δ lt 1 independent of h

Application minus∆u = f in Ω equiv (minus1 1)2 u = 0 on partΩ

Discretization method Linear finite elements on uniform grid of mesh-size h

Meaning of δ For many iterative methods one can prove that the iterates x(i)

satisfy x(i+1) minus x le δx(i) minus x for some 0 lt δ lt 1 in some norm middot

Department of Mathematics [Slide 2 of 36]

Jay Gopalakrishnan

Why multigrid

Iterative method Convergence rate estimates

Gauszlig-Seidel [classical] δ lt 1minus Ch2

SOR [Young 1950] δ lt 1minus Ch

ADI [Peaceman amp Rachford 1955] δ lt (1minus Ch)2

k-parameter ADI [1960rsquos] δ lt (1minus Ch1k)2

Multigrid [see below] δ lt 1 independent of h

Source

Wesselingrsquos book 1992

60rsquos [Fedorenko 1964] [Bachvalov 1966]70rsquos [Brandt 1973] [Nicolaides 1975] [Brandt 1977]80rsquos [Bank amp Dupont1981] [Braess amp Hackbusch1983] [Bramble amp Pasciak1987]

Department of Mathematics [Slide 2 of 36]

Jay Gopalakrishnan

Structure of multigrid algorithmsMultigrid algorithms are based on a sequence of meshesobtained by successive refinement

Whenever it is possible to solve on the coarsest mesh fastmultigrid algorithms allow fast solution on the finest mesh

A 2D example

k = 1 k = 2

k = J

Highlyrefined

Mesh 1 Mesh 2 Mesh J

(Coarsest mesh) (Finest mesh)

Multigrid algorithms have a recursive structure Eachmultigrid iteration typically consists of the following steps

1 Smooth errors at current grid

2 Transfer residual to next coarser grid

3 Correct iterate using the coarser residual (recursively)

Department of Mathematics [Slide 3 of 36]

Jay Gopalakrishnan

Structure of multigrid algorithmsMultigrid algorithms are based on a sequence of meshesobtained by successive refinement

Whenever it is possible to solve on the coarsest mesh fastmultigrid algorithms allow fast solution on the finest mesh

A 2D example

k = 1 k = 2

k = J

Highlyrefined

Mesh 1 Mesh 2 Mesh J

(Coarsest mesh) (Finest mesh)

Multigrid algorithms have a recursive structure Eachmultigrid iteration typically consists of the following steps

1 Smooth errors at current grid

2 Transfer residual to next coarser grid

3 Correct iterate using the coarser residual (recursively)

Department of Mathematics [Slide 3 of 36]

Jay Gopalakrishnan

Structure of multigrid algorithmsMultigrid algorithms are based on a sequence of meshesobtained by successive refinement

Whenever it is possible to solve on the coarsest mesh fastmultigrid algorithms allow fast solution on the finest mesh

Multigrid algorithms have a recursive structure Eachmultigrid iteration typically consists of the following steps

1 Smooth errors at current grid

2 Transfer residual to next coarser grid

3 Correct iterate using the coarser residual (recursively)

Department of Mathematics [Slide 3 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator Ak

To compute u = Aminus1J b iteratively we use multigrid iterations

u(i+1) = MgJ(u(i) b) i = 0 1 2

starting with some initial guess u(0) where the routineMgJ(middot middot) recursively invokes MgJminus1(middot middot) MgJminus2(middot middot)

We set Mg1(v b) equiv Aminus11 b

Idea

1 Reduce fine grid components of error

2 Reduce coarse grid components of error

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of error

2 Reduce coarse grid components of error

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorWe donrsquot have error e but we have residual r = bminusAJu(i) = AJe

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorNeed an approximation for the coarse components of Aminus1

Jr

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorApply the routine Mg

Jminus1 to r projected to the next coarser grid

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smooth errors v = SmoothJ(u(i) b)

2 Transfer residual to coarser gridr = RestrictJminus1(bminus AJv)

3 Correct by recursion w = MgJminus1(0 r)

u(i+1) = v + ProlongJ(w)Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + ProlongJ(MgJminus1(0 RestrictJminus1(bminus AJv)))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Prolong2

Restrict1

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + ProlongJ(MgJminus1(0 RestrictJminus1(bminus AJv)))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

L2

Lt2

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + LJMgJminus1(0 LtJ(bminus AJv))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Operator

Rewrite discrete problem as the operator eq

Ahuh = fh

where Ah Vh 7rarr Vh is defined by

(Ahwh vh) = (nablawhnablavh) forallwh vh isin Vh

Need multigrid to solve for uh equiv Aminus1h fh efficiently

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

Multilevel spacesVk = vh isin H

10(Ω) vh|K isin P1(K) for all elements K in

the kth level mesh

Multilevel operators At each level we also have operatorsgenerated by (nablamiddotnablamiddot) namely Ak Vk 7rarr Vk defined by

(Akv w) = (nablavnablaw) forallv w isin VkDepartment of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

Computationally this means we simply implement a change ofbasis matrix

Ω

v1 isin V1 L2v1 isin V2

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Elliptic eigenfunctionsThe smoothing component of multigrid relies on the fact thatthe eigenfunctions of elliptic operators corresponding to highereigenvalues are increasingly oscillatory

minus∆φ` = λ`φ` φ`L2(Ω) = 1

Eg here are the 1st 50th and 700th eigenfunctions of adiscrete Laplacian on an L-shaped domain

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 9 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

First observe the propagation of errors e(i)

x(i+1) = x(i) + (1λ(k)max)(Akxminus Akx

(i))

x = x + (1λ(k)max)(Akxminus Akx)

=rArr e(i+1) = e(i) minus (1λ(k)max)Ake

(i)

Hence an equivalent question is

why is I minus (1λ(k)max)Ak a smoothing operator

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλn

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated

+ ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλ

(k)max

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Eg 1 The algorithmThus all components of the algorithm are now well defined

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 4: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Why multigrid

Iterative method Convergence rate estimates

Gauszlig-Seidel [classical] δ lt 1minus Ch2

SOR [Young 1950] δ lt 1minus Ch

ADI [Peaceman amp Rachford 1955] δ lt (1minus Ch)2

k-parameter ADI [1960rsquos] δ lt (1minus Ch1k)2

Multigrid [see below] δ lt 1 independent of h

Source

Wesselingrsquos book 1992

60rsquos [Fedorenko 1964] [Bachvalov 1966]70rsquos [Brandt 1973] [Nicolaides 1975] [Brandt 1977]80rsquos [Bank amp Dupont1981] [Braess amp Hackbusch1983] [Bramble amp Pasciak1987]

Department of Mathematics [Slide 2 of 36]

Jay Gopalakrishnan

Structure of multigrid algorithmsMultigrid algorithms are based on a sequence of meshesobtained by successive refinement

Whenever it is possible to solve on the coarsest mesh fastmultigrid algorithms allow fast solution on the finest mesh

A 2D example

k = 1 k = 2

k = J

Highlyrefined

Mesh 1 Mesh 2 Mesh J

(Coarsest mesh) (Finest mesh)

Multigrid algorithms have a recursive structure Eachmultigrid iteration typically consists of the following steps

1 Smooth errors at current grid

2 Transfer residual to next coarser grid

3 Correct iterate using the coarser residual (recursively)

Department of Mathematics [Slide 3 of 36]

Jay Gopalakrishnan

Structure of multigrid algorithmsMultigrid algorithms are based on a sequence of meshesobtained by successive refinement

Whenever it is possible to solve on the coarsest mesh fastmultigrid algorithms allow fast solution on the finest mesh

A 2D example

k = 1 k = 2

k = J

Highlyrefined

Mesh 1 Mesh 2 Mesh J

(Coarsest mesh) (Finest mesh)

Multigrid algorithms have a recursive structure Eachmultigrid iteration typically consists of the following steps

1 Smooth errors at current grid

2 Transfer residual to next coarser grid

3 Correct iterate using the coarser residual (recursively)

Department of Mathematics [Slide 3 of 36]

Jay Gopalakrishnan

Structure of multigrid algorithmsMultigrid algorithms are based on a sequence of meshesobtained by successive refinement

Whenever it is possible to solve on the coarsest mesh fastmultigrid algorithms allow fast solution on the finest mesh

Multigrid algorithms have a recursive structure Eachmultigrid iteration typically consists of the following steps

1 Smooth errors at current grid

2 Transfer residual to next coarser grid

3 Correct iterate using the coarser residual (recursively)

Department of Mathematics [Slide 3 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator Ak

To compute u = Aminus1J b iteratively we use multigrid iterations

u(i+1) = MgJ(u(i) b) i = 0 1 2

starting with some initial guess u(0) where the routineMgJ(middot middot) recursively invokes MgJminus1(middot middot) MgJminus2(middot middot)

We set Mg1(v b) equiv Aminus11 b

Idea

1 Reduce fine grid components of error

2 Reduce coarse grid components of error

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of error

2 Reduce coarse grid components of error

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorWe donrsquot have error e but we have residual r = bminusAJu(i) = AJe

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorNeed an approximation for the coarse components of Aminus1

Jr

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorApply the routine Mg

Jminus1 to r projected to the next coarser grid

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smooth errors v = SmoothJ(u(i) b)

2 Transfer residual to coarser gridr = RestrictJminus1(bminus AJv)

3 Correct by recursion w = MgJminus1(0 r)

u(i+1) = v + ProlongJ(w)Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + ProlongJ(MgJminus1(0 RestrictJminus1(bminus AJv)))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Prolong2

Restrict1

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + ProlongJ(MgJminus1(0 RestrictJminus1(bminus AJv)))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

L2

Lt2

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + LJMgJminus1(0 LtJ(bminus AJv))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Operator

Rewrite discrete problem as the operator eq

Ahuh = fh

where Ah Vh 7rarr Vh is defined by

(Ahwh vh) = (nablawhnablavh) forallwh vh isin Vh

Need multigrid to solve for uh equiv Aminus1h fh efficiently

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

Multilevel spacesVk = vh isin H

10(Ω) vh|K isin P1(K) for all elements K in

the kth level mesh

Multilevel operators At each level we also have operatorsgenerated by (nablamiddotnablamiddot) namely Ak Vk 7rarr Vk defined by

(Akv w) = (nablavnablaw) forallv w isin VkDepartment of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

Computationally this means we simply implement a change ofbasis matrix

Ω

v1 isin V1 L2v1 isin V2

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Elliptic eigenfunctionsThe smoothing component of multigrid relies on the fact thatthe eigenfunctions of elliptic operators corresponding to highereigenvalues are increasingly oscillatory

minus∆φ` = λ`φ` φ`L2(Ω) = 1

Eg here are the 1st 50th and 700th eigenfunctions of adiscrete Laplacian on an L-shaped domain

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 9 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

First observe the propagation of errors e(i)

x(i+1) = x(i) + (1λ(k)max)(Akxminus Akx

(i))

x = x + (1λ(k)max)(Akxminus Akx)

=rArr e(i+1) = e(i) minus (1λ(k)max)Ake

(i)

Hence an equivalent question is

why is I minus (1λ(k)max)Ak a smoothing operator

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλn

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated

+ ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλ

(k)max

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Eg 1 The algorithmThus all components of the algorithm are now well defined

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 5: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Structure of multigrid algorithmsMultigrid algorithms are based on a sequence of meshesobtained by successive refinement

Whenever it is possible to solve on the coarsest mesh fastmultigrid algorithms allow fast solution on the finest mesh

A 2D example

k = 1 k = 2

k = J

Highlyrefined

Mesh 1 Mesh 2 Mesh J

(Coarsest mesh) (Finest mesh)

Multigrid algorithms have a recursive structure Eachmultigrid iteration typically consists of the following steps

1 Smooth errors at current grid

2 Transfer residual to next coarser grid

3 Correct iterate using the coarser residual (recursively)

Department of Mathematics [Slide 3 of 36]

Jay Gopalakrishnan

Structure of multigrid algorithmsMultigrid algorithms are based on a sequence of meshesobtained by successive refinement

Whenever it is possible to solve on the coarsest mesh fastmultigrid algorithms allow fast solution on the finest mesh

A 2D example

k = 1 k = 2

k = J

Highlyrefined

Mesh 1 Mesh 2 Mesh J

(Coarsest mesh) (Finest mesh)

Multigrid algorithms have a recursive structure Eachmultigrid iteration typically consists of the following steps

1 Smooth errors at current grid

2 Transfer residual to next coarser grid

3 Correct iterate using the coarser residual (recursively)

Department of Mathematics [Slide 3 of 36]

Jay Gopalakrishnan

Structure of multigrid algorithmsMultigrid algorithms are based on a sequence of meshesobtained by successive refinement

Whenever it is possible to solve on the coarsest mesh fastmultigrid algorithms allow fast solution on the finest mesh

Multigrid algorithms have a recursive structure Eachmultigrid iteration typically consists of the following steps

1 Smooth errors at current grid

2 Transfer residual to next coarser grid

3 Correct iterate using the coarser residual (recursively)

Department of Mathematics [Slide 3 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator Ak

To compute u = Aminus1J b iteratively we use multigrid iterations

u(i+1) = MgJ(u(i) b) i = 0 1 2

starting with some initial guess u(0) where the routineMgJ(middot middot) recursively invokes MgJminus1(middot middot) MgJminus2(middot middot)

We set Mg1(v b) equiv Aminus11 b

Idea

1 Reduce fine grid components of error

2 Reduce coarse grid components of error

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of error

2 Reduce coarse grid components of error

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorWe donrsquot have error e but we have residual r = bminusAJu(i) = AJe

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorNeed an approximation for the coarse components of Aminus1

Jr

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorApply the routine Mg

Jminus1 to r projected to the next coarser grid

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smooth errors v = SmoothJ(u(i) b)

2 Transfer residual to coarser gridr = RestrictJminus1(bminus AJv)

3 Correct by recursion w = MgJminus1(0 r)

u(i+1) = v + ProlongJ(w)Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + ProlongJ(MgJminus1(0 RestrictJminus1(bminus AJv)))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Prolong2

Restrict1

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + ProlongJ(MgJminus1(0 RestrictJminus1(bminus AJv)))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

L2

Lt2

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + LJMgJminus1(0 LtJ(bminus AJv))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Operator

Rewrite discrete problem as the operator eq

Ahuh = fh

where Ah Vh 7rarr Vh is defined by

(Ahwh vh) = (nablawhnablavh) forallwh vh isin Vh

Need multigrid to solve for uh equiv Aminus1h fh efficiently

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

Multilevel spacesVk = vh isin H

10(Ω) vh|K isin P1(K) for all elements K in

the kth level mesh

Multilevel operators At each level we also have operatorsgenerated by (nablamiddotnablamiddot) namely Ak Vk 7rarr Vk defined by

(Akv w) = (nablavnablaw) forallv w isin VkDepartment of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

Computationally this means we simply implement a change ofbasis matrix

Ω

v1 isin V1 L2v1 isin V2

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Elliptic eigenfunctionsThe smoothing component of multigrid relies on the fact thatthe eigenfunctions of elliptic operators corresponding to highereigenvalues are increasingly oscillatory

minus∆φ` = λ`φ` φ`L2(Ω) = 1

Eg here are the 1st 50th and 700th eigenfunctions of adiscrete Laplacian on an L-shaped domain

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 9 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

First observe the propagation of errors e(i)

x(i+1) = x(i) + (1λ(k)max)(Akxminus Akx

(i))

x = x + (1λ(k)max)(Akxminus Akx)

=rArr e(i+1) = e(i) minus (1λ(k)max)Ake

(i)

Hence an equivalent question is

why is I minus (1λ(k)max)Ak a smoothing operator

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλn

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated

+ ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλ

(k)max

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Eg 1 The algorithmThus all components of the algorithm are now well defined

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 6: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Structure of multigrid algorithmsMultigrid algorithms are based on a sequence of meshesobtained by successive refinement

Whenever it is possible to solve on the coarsest mesh fastmultigrid algorithms allow fast solution on the finest mesh

A 2D example

k = 1 k = 2

k = J

Highlyrefined

Mesh 1 Mesh 2 Mesh J

(Coarsest mesh) (Finest mesh)

Multigrid algorithms have a recursive structure Eachmultigrid iteration typically consists of the following steps

1 Smooth errors at current grid

2 Transfer residual to next coarser grid

3 Correct iterate using the coarser residual (recursively)

Department of Mathematics [Slide 3 of 36]

Jay Gopalakrishnan

Structure of multigrid algorithmsMultigrid algorithms are based on a sequence of meshesobtained by successive refinement

Whenever it is possible to solve on the coarsest mesh fastmultigrid algorithms allow fast solution on the finest mesh

Multigrid algorithms have a recursive structure Eachmultigrid iteration typically consists of the following steps

1 Smooth errors at current grid

2 Transfer residual to next coarser grid

3 Correct iterate using the coarser residual (recursively)

Department of Mathematics [Slide 3 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator Ak

To compute u = Aminus1J b iteratively we use multigrid iterations

u(i+1) = MgJ(u(i) b) i = 0 1 2

starting with some initial guess u(0) where the routineMgJ(middot middot) recursively invokes MgJminus1(middot middot) MgJminus2(middot middot)

We set Mg1(v b) equiv Aminus11 b

Idea

1 Reduce fine grid components of error

2 Reduce coarse grid components of error

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of error

2 Reduce coarse grid components of error

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorWe donrsquot have error e but we have residual r = bminusAJu(i) = AJe

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorNeed an approximation for the coarse components of Aminus1

Jr

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorApply the routine Mg

Jminus1 to r projected to the next coarser grid

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smooth errors v = SmoothJ(u(i) b)

2 Transfer residual to coarser gridr = RestrictJminus1(bminus AJv)

3 Correct by recursion w = MgJminus1(0 r)

u(i+1) = v + ProlongJ(w)Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + ProlongJ(MgJminus1(0 RestrictJminus1(bminus AJv)))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Prolong2

Restrict1

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + ProlongJ(MgJminus1(0 RestrictJminus1(bminus AJv)))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

L2

Lt2

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + LJMgJminus1(0 LtJ(bminus AJv))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Operator

Rewrite discrete problem as the operator eq

Ahuh = fh

where Ah Vh 7rarr Vh is defined by

(Ahwh vh) = (nablawhnablavh) forallwh vh isin Vh

Need multigrid to solve for uh equiv Aminus1h fh efficiently

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

Multilevel spacesVk = vh isin H

10(Ω) vh|K isin P1(K) for all elements K in

the kth level mesh

Multilevel operators At each level we also have operatorsgenerated by (nablamiddotnablamiddot) namely Ak Vk 7rarr Vk defined by

(Akv w) = (nablavnablaw) forallv w isin VkDepartment of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

Computationally this means we simply implement a change ofbasis matrix

Ω

v1 isin V1 L2v1 isin V2

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Elliptic eigenfunctionsThe smoothing component of multigrid relies on the fact thatthe eigenfunctions of elliptic operators corresponding to highereigenvalues are increasingly oscillatory

minus∆φ` = λ`φ` φ`L2(Ω) = 1

Eg here are the 1st 50th and 700th eigenfunctions of adiscrete Laplacian on an L-shaped domain

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 9 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

First observe the propagation of errors e(i)

x(i+1) = x(i) + (1λ(k)max)(Akxminus Akx

(i))

x = x + (1λ(k)max)(Akxminus Akx)

=rArr e(i+1) = e(i) minus (1λ(k)max)Ake

(i)

Hence an equivalent question is

why is I minus (1λ(k)max)Ak a smoothing operator

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλn

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated

+ ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλ

(k)max

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Eg 1 The algorithmThus all components of the algorithm are now well defined

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 7: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Structure of multigrid algorithmsMultigrid algorithms are based on a sequence of meshesobtained by successive refinement

Whenever it is possible to solve on the coarsest mesh fastmultigrid algorithms allow fast solution on the finest mesh

Multigrid algorithms have a recursive structure Eachmultigrid iteration typically consists of the following steps

1 Smooth errors at current grid

2 Transfer residual to next coarser grid

3 Correct iterate using the coarser residual (recursively)

Department of Mathematics [Slide 3 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator Ak

To compute u = Aminus1J b iteratively we use multigrid iterations

u(i+1) = MgJ(u(i) b) i = 0 1 2

starting with some initial guess u(0) where the routineMgJ(middot middot) recursively invokes MgJminus1(middot middot) MgJminus2(middot middot)

We set Mg1(v b) equiv Aminus11 b

Idea

1 Reduce fine grid components of error

2 Reduce coarse grid components of error

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of error

2 Reduce coarse grid components of error

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorWe donrsquot have error e but we have residual r = bminusAJu(i) = AJe

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorNeed an approximation for the coarse components of Aminus1

Jr

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorApply the routine Mg

Jminus1 to r projected to the next coarser grid

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smooth errors v = SmoothJ(u(i) b)

2 Transfer residual to coarser gridr = RestrictJminus1(bminus AJv)

3 Correct by recursion w = MgJminus1(0 r)

u(i+1) = v + ProlongJ(w)Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + ProlongJ(MgJminus1(0 RestrictJminus1(bminus AJv)))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Prolong2

Restrict1

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + ProlongJ(MgJminus1(0 RestrictJminus1(bminus AJv)))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

L2

Lt2

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + LJMgJminus1(0 LtJ(bminus AJv))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Operator

Rewrite discrete problem as the operator eq

Ahuh = fh

where Ah Vh 7rarr Vh is defined by

(Ahwh vh) = (nablawhnablavh) forallwh vh isin Vh

Need multigrid to solve for uh equiv Aminus1h fh efficiently

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

Multilevel spacesVk = vh isin H

10(Ω) vh|K isin P1(K) for all elements K in

the kth level mesh

Multilevel operators At each level we also have operatorsgenerated by (nablamiddotnablamiddot) namely Ak Vk 7rarr Vk defined by

(Akv w) = (nablavnablaw) forallv w isin VkDepartment of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

Computationally this means we simply implement a change ofbasis matrix

Ω

v1 isin V1 L2v1 isin V2

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Elliptic eigenfunctionsThe smoothing component of multigrid relies on the fact thatthe eigenfunctions of elliptic operators corresponding to highereigenvalues are increasingly oscillatory

minus∆φ` = λ`φ` φ`L2(Ω) = 1

Eg here are the 1st 50th and 700th eigenfunctions of adiscrete Laplacian on an L-shaped domain

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 9 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

First observe the propagation of errors e(i)

x(i+1) = x(i) + (1λ(k)max)(Akxminus Akx

(i))

x = x + (1λ(k)max)(Akxminus Akx)

=rArr e(i+1) = e(i) minus (1λ(k)max)Ake

(i)

Hence an equivalent question is

why is I minus (1λ(k)max)Ak a smoothing operator

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλn

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated

+ ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλ

(k)max

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Eg 1 The algorithmThus all components of the algorithm are now well defined

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 8: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator Ak

To compute u = Aminus1J b iteratively we use multigrid iterations

u(i+1) = MgJ(u(i) b) i = 0 1 2

starting with some initial guess u(0) where the routineMgJ(middot middot) recursively invokes MgJminus1(middot middot) MgJminus2(middot middot)

We set Mg1(v b) equiv Aminus11 b

Idea

1 Reduce fine grid components of error

2 Reduce coarse grid components of error

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of error

2 Reduce coarse grid components of error

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorWe donrsquot have error e but we have residual r = bminusAJu(i) = AJe

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorNeed an approximation for the coarse components of Aminus1

Jr

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorApply the routine Mg

Jminus1 to r projected to the next coarser grid

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smooth errors v = SmoothJ(u(i) b)

2 Transfer residual to coarser gridr = RestrictJminus1(bminus AJv)

3 Correct by recursion w = MgJminus1(0 r)

u(i+1) = v + ProlongJ(w)Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + ProlongJ(MgJminus1(0 RestrictJminus1(bminus AJv)))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Prolong2

Restrict1

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + ProlongJ(MgJminus1(0 RestrictJminus1(bminus AJv)))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

L2

Lt2

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + LJMgJminus1(0 LtJ(bminus AJv))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Operator

Rewrite discrete problem as the operator eq

Ahuh = fh

where Ah Vh 7rarr Vh is defined by

(Ahwh vh) = (nablawhnablavh) forallwh vh isin Vh

Need multigrid to solve for uh equiv Aminus1h fh efficiently

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

Multilevel spacesVk = vh isin H

10(Ω) vh|K isin P1(K) for all elements K in

the kth level mesh

Multilevel operators At each level we also have operatorsgenerated by (nablamiddotnablamiddot) namely Ak Vk 7rarr Vk defined by

(Akv w) = (nablavnablaw) forallv w isin VkDepartment of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

Computationally this means we simply implement a change ofbasis matrix

Ω

v1 isin V1 L2v1 isin V2

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Elliptic eigenfunctionsThe smoothing component of multigrid relies on the fact thatthe eigenfunctions of elliptic operators corresponding to highereigenvalues are increasingly oscillatory

minus∆φ` = λ`φ` φ`L2(Ω) = 1

Eg here are the 1st 50th and 700th eigenfunctions of adiscrete Laplacian on an L-shaped domain

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 9 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

First observe the propagation of errors e(i)

x(i+1) = x(i) + (1λ(k)max)(Akxminus Akx

(i))

x = x + (1λ(k)max)(Akxminus Akx)

=rArr e(i+1) = e(i) minus (1λ(k)max)Ake

(i)

Hence an equivalent question is

why is I minus (1λ(k)max)Ak a smoothing operator

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλn

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated

+ ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλ

(k)max

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Eg 1 The algorithmThus all components of the algorithm are now well defined

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 9: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of error

2 Reduce coarse grid components of error

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorWe donrsquot have error e but we have residual r = bminusAJu(i) = AJe

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorNeed an approximation for the coarse components of Aminus1

Jr

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorApply the routine Mg

Jminus1 to r projected to the next coarser grid

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smooth errors v = SmoothJ(u(i) b)

2 Transfer residual to coarser gridr = RestrictJminus1(bminus AJv)

3 Correct by recursion w = MgJminus1(0 r)

u(i+1) = v + ProlongJ(w)Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + ProlongJ(MgJminus1(0 RestrictJminus1(bminus AJv)))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Prolong2

Restrict1

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + ProlongJ(MgJminus1(0 RestrictJminus1(bminus AJv)))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

L2

Lt2

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + LJMgJminus1(0 LtJ(bminus AJv))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Operator

Rewrite discrete problem as the operator eq

Ahuh = fh

where Ah Vh 7rarr Vh is defined by

(Ahwh vh) = (nablawhnablavh) forallwh vh isin Vh

Need multigrid to solve for uh equiv Aminus1h fh efficiently

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

Multilevel spacesVk = vh isin H

10(Ω) vh|K isin P1(K) for all elements K in

the kth level mesh

Multilevel operators At each level we also have operatorsgenerated by (nablamiddotnablamiddot) namely Ak Vk 7rarr Vk defined by

(Akv w) = (nablavnablaw) forallv w isin VkDepartment of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

Computationally this means we simply implement a change ofbasis matrix

Ω

v1 isin V1 L2v1 isin V2

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Elliptic eigenfunctionsThe smoothing component of multigrid relies on the fact thatthe eigenfunctions of elliptic operators corresponding to highereigenvalues are increasingly oscillatory

minus∆φ` = λ`φ` φ`L2(Ω) = 1

Eg here are the 1st 50th and 700th eigenfunctions of adiscrete Laplacian on an L-shaped domain

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 9 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

First observe the propagation of errors e(i)

x(i+1) = x(i) + (1λ(k)max)(Akxminus Akx

(i))

x = x + (1λ(k)max)(Akxminus Akx)

=rArr e(i+1) = e(i) minus (1λ(k)max)Ake

(i)

Hence an equivalent question is

why is I minus (1λ(k)max)Ak a smoothing operator

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλn

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated

+ ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλ

(k)max

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Eg 1 The algorithmThus all components of the algorithm are now well defined

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 10: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorWe donrsquot have error e but we have residual r = bminusAJu(i) = AJe

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorNeed an approximation for the coarse components of Aminus1

Jr

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorApply the routine Mg

Jminus1 to r projected to the next coarser grid

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smooth errors v = SmoothJ(u(i) b)

2 Transfer residual to coarser gridr = RestrictJminus1(bminus AJv)

3 Correct by recursion w = MgJminus1(0 r)

u(i+1) = v + ProlongJ(w)Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + ProlongJ(MgJminus1(0 RestrictJminus1(bminus AJv)))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Prolong2

Restrict1

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + ProlongJ(MgJminus1(0 RestrictJminus1(bminus AJv)))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

L2

Lt2

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + LJMgJminus1(0 LtJ(bminus AJv))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Operator

Rewrite discrete problem as the operator eq

Ahuh = fh

where Ah Vh 7rarr Vh is defined by

(Ahwh vh) = (nablawhnablavh) forallwh vh isin Vh

Need multigrid to solve for uh equiv Aminus1h fh efficiently

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

Multilevel spacesVk = vh isin H

10(Ω) vh|K isin P1(K) for all elements K in

the kth level mesh

Multilevel operators At each level we also have operatorsgenerated by (nablamiddotnablamiddot) namely Ak Vk 7rarr Vk defined by

(Akv w) = (nablavnablaw) forallv w isin VkDepartment of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

Computationally this means we simply implement a change ofbasis matrix

Ω

v1 isin V1 L2v1 isin V2

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Elliptic eigenfunctionsThe smoothing component of multigrid relies on the fact thatthe eigenfunctions of elliptic operators corresponding to highereigenvalues are increasingly oscillatory

minus∆φ` = λ`φ` φ`L2(Ω) = 1

Eg here are the 1st 50th and 700th eigenfunctions of adiscrete Laplacian on an L-shaped domain

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 9 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

First observe the propagation of errors e(i)

x(i+1) = x(i) + (1λ(k)max)(Akxminus Akx

(i))

x = x + (1λ(k)max)(Akxminus Akx)

=rArr e(i+1) = e(i) minus (1λ(k)max)Ake

(i)

Hence an equivalent question is

why is I minus (1λ(k)max)Ak a smoothing operator

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλn

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated

+ ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλ

(k)max

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Eg 1 The algorithmThus all components of the algorithm are now well defined

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 11: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorNeed an approximation for the coarse components of Aminus1

Jr

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorApply the routine Mg

Jminus1 to r projected to the next coarser grid

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smooth errors v = SmoothJ(u(i) b)

2 Transfer residual to coarser gridr = RestrictJminus1(bminus AJv)

3 Correct by recursion w = MgJminus1(0 r)

u(i+1) = v + ProlongJ(w)Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + ProlongJ(MgJminus1(0 RestrictJminus1(bminus AJv)))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Prolong2

Restrict1

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + ProlongJ(MgJminus1(0 RestrictJminus1(bminus AJv)))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

L2

Lt2

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + LJMgJminus1(0 LtJ(bminus AJv))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Operator

Rewrite discrete problem as the operator eq

Ahuh = fh

where Ah Vh 7rarr Vh is defined by

(Ahwh vh) = (nablawhnablavh) forallwh vh isin Vh

Need multigrid to solve for uh equiv Aminus1h fh efficiently

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

Multilevel spacesVk = vh isin H

10(Ω) vh|K isin P1(K) for all elements K in

the kth level mesh

Multilevel operators At each level we also have operatorsgenerated by (nablamiddotnablamiddot) namely Ak Vk 7rarr Vk defined by

(Akv w) = (nablavnablaw) forallv w isin VkDepartment of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

Computationally this means we simply implement a change ofbasis matrix

Ω

v1 isin V1 L2v1 isin V2

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Elliptic eigenfunctionsThe smoothing component of multigrid relies on the fact thatthe eigenfunctions of elliptic operators corresponding to highereigenvalues are increasingly oscillatory

minus∆φ` = λ`φ` φ`L2(Ω) = 1

Eg here are the 1st 50th and 700th eigenfunctions of adiscrete Laplacian on an L-shaped domain

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 9 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

First observe the propagation of errors e(i)

x(i+1) = x(i) + (1λ(k)max)(Akxminus Akx

(i))

x = x + (1λ(k)max)(Akxminus Akx)

=rArr e(i+1) = e(i) minus (1λ(k)max)Ake

(i)

Hence an equivalent question is

why is I minus (1λ(k)max)Ak a smoothing operator

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλn

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated

+ ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλ

(k)max

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Eg 1 The algorithmThus all components of the algorithm are now well defined

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 12: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

The multigrid idea

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkIdea

1 Reduce fine grid components of errorBy smoothing error (without knowing the error)

2 Reduce coarse grid components of errorApply the routine Mg

Jminus1 to r projected to the next coarser grid

Difficulty Donrsquot know exact solution so donrsquot know the error

Department of Mathematics [Slide 4 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smooth errors v = SmoothJ(u(i) b)

2 Transfer residual to coarser gridr = RestrictJminus1(bminus AJv)

3 Correct by recursion w = MgJminus1(0 r)

u(i+1) = v + ProlongJ(w)Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + ProlongJ(MgJminus1(0 RestrictJminus1(bminus AJv)))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Prolong2

Restrict1

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + ProlongJ(MgJminus1(0 RestrictJminus1(bminus AJv)))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

L2

Lt2

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + LJMgJminus1(0 LtJ(bminus AJv))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Operator

Rewrite discrete problem as the operator eq

Ahuh = fh

where Ah Vh 7rarr Vh is defined by

(Ahwh vh) = (nablawhnablavh) forallwh vh isin Vh

Need multigrid to solve for uh equiv Aminus1h fh efficiently

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

Multilevel spacesVk = vh isin H

10(Ω) vh|K isin P1(K) for all elements K in

the kth level mesh

Multilevel operators At each level we also have operatorsgenerated by (nablamiddotnablamiddot) namely Ak Vk 7rarr Vk defined by

(Akv w) = (nablavnablaw) forallv w isin VkDepartment of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

Computationally this means we simply implement a change ofbasis matrix

Ω

v1 isin V1 L2v1 isin V2

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Elliptic eigenfunctionsThe smoothing component of multigrid relies on the fact thatthe eigenfunctions of elliptic operators corresponding to highereigenvalues are increasingly oscillatory

minus∆φ` = λ`φ` φ`L2(Ω) = 1

Eg here are the 1st 50th and 700th eigenfunctions of adiscrete Laplacian on an L-shaped domain

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 9 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

First observe the propagation of errors e(i)

x(i+1) = x(i) + (1λ(k)max)(Akxminus Akx

(i))

x = x + (1λ(k)max)(Akxminus Akx)

=rArr e(i+1) = e(i) minus (1λ(k)max)Ake

(i)

Hence an equivalent question is

why is I minus (1λ(k)max)Ak a smoothing operator

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλn

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated

+ ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλ

(k)max

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Eg 1 The algorithmThus all components of the algorithm are now well defined

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 13: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smooth errors v = SmoothJ(u(i) b)

2 Transfer residual to coarser gridr = RestrictJminus1(bminus AJv)

3 Correct by recursion w = MgJminus1(0 r)

u(i+1) = v + ProlongJ(w)Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + ProlongJ(MgJminus1(0 RestrictJminus1(bminus AJv)))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Prolong2

Restrict1

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + ProlongJ(MgJminus1(0 RestrictJminus1(bminus AJv)))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

L2

Lt2

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + LJMgJminus1(0 LtJ(bminus AJv))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Operator

Rewrite discrete problem as the operator eq

Ahuh = fh

where Ah Vh 7rarr Vh is defined by

(Ahwh vh) = (nablawhnablavh) forallwh vh isin Vh

Need multigrid to solve for uh equiv Aminus1h fh efficiently

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

Multilevel spacesVk = vh isin H

10(Ω) vh|K isin P1(K) for all elements K in

the kth level mesh

Multilevel operators At each level we also have operatorsgenerated by (nablamiddotnablamiddot) namely Ak Vk 7rarr Vk defined by

(Akv w) = (nablavnablaw) forallv w isin VkDepartment of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

Computationally this means we simply implement a change ofbasis matrix

Ω

v1 isin V1 L2v1 isin V2

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Elliptic eigenfunctionsThe smoothing component of multigrid relies on the fact thatthe eigenfunctions of elliptic operators corresponding to highereigenvalues are increasingly oscillatory

minus∆φ` = λ`φ` φ`L2(Ω) = 1

Eg here are the 1st 50th and 700th eigenfunctions of adiscrete Laplacian on an L-shaped domain

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 9 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

First observe the propagation of errors e(i)

x(i+1) = x(i) + (1λ(k)max)(Akxminus Akx

(i))

x = x + (1λ(k)max)(Akxminus Akx)

=rArr e(i+1) = e(i) minus (1λ(k)max)Ake

(i)

Hence an equivalent question is

why is I minus (1λ(k)max)Ak a smoothing operator

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλn

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated

+ ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλ

(k)max

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Eg 1 The algorithmThus all components of the algorithm are now well defined

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 14: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + ProlongJ(MgJminus1(0 RestrictJminus1(bminus AJv)))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Prolong2

Restrict1

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + ProlongJ(MgJminus1(0 RestrictJminus1(bminus AJv)))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

L2

Lt2

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + LJMgJminus1(0 LtJ(bminus AJv))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Operator

Rewrite discrete problem as the operator eq

Ahuh = fh

where Ah Vh 7rarr Vh is defined by

(Ahwh vh) = (nablawhnablavh) forallwh vh isin Vh

Need multigrid to solve for uh equiv Aminus1h fh efficiently

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

Multilevel spacesVk = vh isin H

10(Ω) vh|K isin P1(K) for all elements K in

the kth level mesh

Multilevel operators At each level we also have operatorsgenerated by (nablamiddotnablamiddot) namely Ak Vk 7rarr Vk defined by

(Akv w) = (nablavnablaw) forallv w isin VkDepartment of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

Computationally this means we simply implement a change ofbasis matrix

Ω

v1 isin V1 L2v1 isin V2

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Elliptic eigenfunctionsThe smoothing component of multigrid relies on the fact thatthe eigenfunctions of elliptic operators corresponding to highereigenvalues are increasingly oscillatory

minus∆φ` = λ`φ` φ`L2(Ω) = 1

Eg here are the 1st 50th and 700th eigenfunctions of adiscrete Laplacian on an L-shaped domain

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 9 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

First observe the propagation of errors e(i)

x(i+1) = x(i) + (1λ(k)max)(Akxminus Akx

(i))

x = x + (1λ(k)max)(Akxminus Akx)

=rArr e(i+1) = e(i) minus (1λ(k)max)Ake

(i)

Hence an equivalent question is

why is I minus (1λ(k)max)Ak a smoothing operator

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλn

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated

+ ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλ

(k)max

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Eg 1 The algorithmThus all components of the algorithm are now well defined

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 15: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

Prolong2

Restrict1

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + ProlongJ(MgJminus1(0 RestrictJminus1(bminus AJv)))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

L2

Lt2

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + LJMgJminus1(0 LtJ(bminus AJv))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Operator

Rewrite discrete problem as the operator eq

Ahuh = fh

where Ah Vh 7rarr Vh is defined by

(Ahwh vh) = (nablawhnablavh) forallwh vh isin Vh

Need multigrid to solve for uh equiv Aminus1h fh efficiently

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

Multilevel spacesVk = vh isin H

10(Ω) vh|K isin P1(K) for all elements K in

the kth level mesh

Multilevel operators At each level we also have operatorsgenerated by (nablamiddotnablamiddot) namely Ak Vk 7rarr Vk defined by

(Akv w) = (nablavnablaw) forallv w isin VkDepartment of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

Computationally this means we simply implement a change ofbasis matrix

Ω

v1 isin V1 L2v1 isin V2

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Elliptic eigenfunctionsThe smoothing component of multigrid relies on the fact thatthe eigenfunctions of elliptic operators corresponding to highereigenvalues are increasingly oscillatory

minus∆φ` = λ`φ` φ`L2(Ω) = 1

Eg here are the 1st 50th and 700th eigenfunctions of adiscrete Laplacian on an L-shaped domain

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 9 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

First observe the propagation of errors e(i)

x(i+1) = x(i) + (1λ(k)max)(Akxminus Akx

(i))

x = x + (1λ(k)max)(Akxminus Akx)

=rArr e(i+1) = e(i) minus (1λ(k)max)Ake

(i)

Hence an equivalent question is

why is I minus (1λ(k)max)Ak a smoothing operator

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλn

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated

+ ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλ

(k)max

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Eg 1 The algorithmThus all components of the algorithm are now well defined

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 16: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

A typical pseudocode

k = 1 k = 2

k = J

Highlyrefined

L2

Lt2

Need to solveAJu = b

Suppose discretization of some continuum problem at meshlevel k gives a linear operator AkALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ(u(i) b)

2 Correction

u(i+1) = v + LJMgJminus1(0 LtJ(bminus AJv))

Department of Mathematics [Slide 5 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Operator

Rewrite discrete problem as the operator eq

Ahuh = fh

where Ah Vh 7rarr Vh is defined by

(Ahwh vh) = (nablawhnablavh) forallwh vh isin Vh

Need multigrid to solve for uh equiv Aminus1h fh efficiently

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

Multilevel spacesVk = vh isin H

10(Ω) vh|K isin P1(K) for all elements K in

the kth level mesh

Multilevel operators At each level we also have operatorsgenerated by (nablamiddotnablamiddot) namely Ak Vk 7rarr Vk defined by

(Akv w) = (nablavnablaw) forallv w isin VkDepartment of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

Computationally this means we simply implement a change ofbasis matrix

Ω

v1 isin V1 L2v1 isin V2

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Elliptic eigenfunctionsThe smoothing component of multigrid relies on the fact thatthe eigenfunctions of elliptic operators corresponding to highereigenvalues are increasingly oscillatory

minus∆φ` = λ`φ` φ`L2(Ω) = 1

Eg here are the 1st 50th and 700th eigenfunctions of adiscrete Laplacian on an L-shaped domain

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 9 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

First observe the propagation of errors e(i)

x(i+1) = x(i) + (1λ(k)max)(Akxminus Akx

(i))

x = x + (1λ(k)max)(Akxminus Akx)

=rArr e(i+1) = e(i) minus (1λ(k)max)Ake

(i)

Hence an equivalent question is

why is I minus (1λ(k)max)Ak a smoothing operator

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλn

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated

+ ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλ

(k)max

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Eg 1 The algorithmThus all components of the algorithm are now well defined

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 17: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Operator

Rewrite discrete problem as the operator eq

Ahuh = fh

where Ah Vh 7rarr Vh is defined by

(Ahwh vh) = (nablawhnablavh) forallwh vh isin Vh

Need multigrid to solve for uh equiv Aminus1h fh efficiently

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

Multilevel spacesVk = vh isin H

10(Ω) vh|K isin P1(K) for all elements K in

the kth level mesh

Multilevel operators At each level we also have operatorsgenerated by (nablamiddotnablamiddot) namely Ak Vk 7rarr Vk defined by

(Akv w) = (nablavnablaw) forallv w isin VkDepartment of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

Computationally this means we simply implement a change ofbasis matrix

Ω

v1 isin V1 L2v1 isin V2

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Elliptic eigenfunctionsThe smoothing component of multigrid relies on the fact thatthe eigenfunctions of elliptic operators corresponding to highereigenvalues are increasingly oscillatory

minus∆φ` = λ`φ` φ`L2(Ω) = 1

Eg here are the 1st 50th and 700th eigenfunctions of adiscrete Laplacian on an L-shaped domain

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 9 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

First observe the propagation of errors e(i)

x(i+1) = x(i) + (1λ(k)max)(Akxminus Akx

(i))

x = x + (1λ(k)max)(Akxminus Akx)

=rArr e(i+1) = e(i) minus (1λ(k)max)Ake

(i)

Hence an equivalent question is

why is I minus (1λ(k)max)Ak a smoothing operator

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλn

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated

+ ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλ

(k)max

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Eg 1 The algorithmThus all components of the algorithm are now well defined

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 18: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Weakform

Find u isin H10 (Ω) satisfying

(nablaunablav) = (f v) forallv isin H10(Ω)

BVPminus∆u = f on Ω

u = 0 on partΩ

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Operator

Rewrite discrete problem as the operator eq

Ahuh = fh

where Ah Vh 7rarr Vh is defined by

(Ahwh vh) = (nablawhnablavh) forallwh vh isin Vh

Need multigrid to solve for uh equiv Aminus1h fh efficiently

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

Multilevel spacesVk = vh isin H

10(Ω) vh|K isin P1(K) for all elements K in

the kth level mesh

Multilevel operators At each level we also have operatorsgenerated by (nablamiddotnablamiddot) namely Ak Vk 7rarr Vk defined by

(Akv w) = (nablavnablaw) forallv w isin VkDepartment of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

Computationally this means we simply implement a change ofbasis matrix

Ω

v1 isin V1 L2v1 isin V2

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Elliptic eigenfunctionsThe smoothing component of multigrid relies on the fact thatthe eigenfunctions of elliptic operators corresponding to highereigenvalues are increasingly oscillatory

minus∆φ` = λ`φ` φ`L2(Ω) = 1

Eg here are the 1st 50th and 700th eigenfunctions of adiscrete Laplacian on an L-shaped domain

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 9 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

First observe the propagation of errors e(i)

x(i+1) = x(i) + (1λ(k)max)(Akxminus Akx

(i))

x = x + (1λ(k)max)(Akxminus Akx)

=rArr e(i+1) = e(i) minus (1λ(k)max)Ake

(i)

Hence an equivalent question is

why is I minus (1λ(k)max)Ak a smoothing operator

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλn

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated

+ ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλ

(k)max

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Eg 1 The algorithmThus all components of the algorithm are now well defined

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 19: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Eg 1 Laplace equation

FEM

Find uh isin Vh satisfying

(nablauhnablavh) = (f vh) forallvh isin Vh

Vh sub H10(Ω) is any standard fe space

Operator

Rewrite discrete problem as the operator eq

Ahuh = fh

where Ah Vh 7rarr Vh is defined by

(Ahwh vh) = (nablawhnablavh) forallwh vh isin Vh

Need multigrid to solve for uh equiv Aminus1h fh efficiently

Department of Mathematics [Slide 6 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

Multilevel spacesVk = vh isin H

10(Ω) vh|K isin P1(K) for all elements K in

the kth level mesh

Multilevel operators At each level we also have operatorsgenerated by (nablamiddotnablamiddot) namely Ak Vk 7rarr Vk defined by

(Akv w) = (nablavnablaw) forallv w isin VkDepartment of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

Computationally this means we simply implement a change ofbasis matrix

Ω

v1 isin V1 L2v1 isin V2

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Elliptic eigenfunctionsThe smoothing component of multigrid relies on the fact thatthe eigenfunctions of elliptic operators corresponding to highereigenvalues are increasingly oscillatory

minus∆φ` = λ`φ` φ`L2(Ω) = 1

Eg here are the 1st 50th and 700th eigenfunctions of adiscrete Laplacian on an L-shaped domain

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 9 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

First observe the propagation of errors e(i)

x(i+1) = x(i) + (1λ(k)max)(Akxminus Akx

(i))

x = x + (1λ(k)max)(Akxminus Akx)

=rArr e(i+1) = e(i) minus (1λ(k)max)Ake

(i)

Hence an equivalent question is

why is I minus (1λ(k)max)Ak a smoothing operator

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλn

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated

+ ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλ

(k)max

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Eg 1 The algorithmThus all components of the algorithm are now well defined

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 20: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

Multilevel spacesVk = vh isin H

10(Ω) vh|K isin P1(K) for all elements K in

the kth level mesh

Multilevel operators At each level we also have operatorsgenerated by (nablamiddotnablamiddot) namely Ak Vk 7rarr Vk defined by

(Akv w) = (nablavnablaw) forallv w isin VkDepartment of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

Computationally this means we simply implement a change ofbasis matrix

Ω

v1 isin V1 L2v1 isin V2

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Elliptic eigenfunctionsThe smoothing component of multigrid relies on the fact thatthe eigenfunctions of elliptic operators corresponding to highereigenvalues are increasingly oscillatory

minus∆φ` = λ`φ` φ`L2(Ω) = 1

Eg here are the 1st 50th and 700th eigenfunctions of adiscrete Laplacian on an L-shaped domain

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 9 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

First observe the propagation of errors e(i)

x(i+1) = x(i) + (1λ(k)max)(Akxminus Akx

(i))

x = x + (1λ(k)max)(Akxminus Akx)

=rArr e(i+1) = e(i) minus (1λ(k)max)Ake

(i)

Hence an equivalent question is

why is I minus (1λ(k)max)Ak a smoothing operator

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλn

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated

+ ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλ

(k)max

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Eg 1 The algorithmThus all components of the algorithm are now well defined

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 21: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Eg 1 Multigrid setting

Assume that Vh is a fe space on a highly refined mesh

Ω

middot middot middot

V1 V2 VJ equiv Vh

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 7 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

Computationally this means we simply implement a change ofbasis matrix

Ω

v1 isin V1 L2v1 isin V2

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Elliptic eigenfunctionsThe smoothing component of multigrid relies on the fact thatthe eigenfunctions of elliptic operators corresponding to highereigenvalues are increasingly oscillatory

minus∆φ` = λ`φ` φ`L2(Ω) = 1

Eg here are the 1st 50th and 700th eigenfunctions of adiscrete Laplacian on an L-shaped domain

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 9 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

First observe the propagation of errors e(i)

x(i+1) = x(i) + (1λ(k)max)(Akxminus Akx

(i))

x = x + (1λ(k)max)(Akxminus Akx)

=rArr e(i+1) = e(i) minus (1λ(k)max)Ake

(i)

Hence an equivalent question is

why is I minus (1λ(k)max)Ak a smoothing operator

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλn

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated

+ ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλ

(k)max

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Eg 1 The algorithmThus all components of the algorithm are now well defined

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 22: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

Computationally this means we simply implement a change ofbasis matrix

Ω

v1 isin V1 L2v1 isin V2

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Elliptic eigenfunctionsThe smoothing component of multigrid relies on the fact thatthe eigenfunctions of elliptic operators corresponding to highereigenvalues are increasingly oscillatory

minus∆φ` = λ`φ` φ`L2(Ω) = 1

Eg here are the 1st 50th and 700th eigenfunctions of adiscrete Laplacian on an L-shaped domain

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 9 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

First observe the propagation of errors e(i)

x(i+1) = x(i) + (1λ(k)max)(Akxminus Akx

(i))

x = x + (1λ(k)max)(Akxminus Akx)

=rArr e(i+1) = e(i) minus (1λ(k)max)Ake

(i)

Hence an equivalent question is

why is I minus (1λ(k)max)Ak a smoothing operator

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλn

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated

+ ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλ

(k)max

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Eg 1 The algorithmThus all components of the algorithm are now well defined

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 23: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Eg 1 Prolongation

The multilevel spaces in this example are nested

V1 sub V2 sub middot middot middot sub VJ

Hence we choose Lk to be the imbedding operator

Vkminus1 rarr Vk

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 8 of 36]

Jay Gopalakrishnan

Elliptic eigenfunctionsThe smoothing component of multigrid relies on the fact thatthe eigenfunctions of elliptic operators corresponding to highereigenvalues are increasingly oscillatory

minus∆φ` = λ`φ` φ`L2(Ω) = 1

Eg here are the 1st 50th and 700th eigenfunctions of adiscrete Laplacian on an L-shaped domain

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 9 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

First observe the propagation of errors e(i)

x(i+1) = x(i) + (1λ(k)max)(Akxminus Akx

(i))

x = x + (1λ(k)max)(Akxminus Akx)

=rArr e(i+1) = e(i) minus (1λ(k)max)Ake

(i)

Hence an equivalent question is

why is I minus (1λ(k)max)Ak a smoothing operator

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλn

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated

+ ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλ

(k)max

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Eg 1 The algorithmThus all components of the algorithm are now well defined

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 24: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Elliptic eigenfunctionsThe smoothing component of multigrid relies on the fact thatthe eigenfunctions of elliptic operators corresponding to highereigenvalues are increasingly oscillatory

minus∆φ` = λ`φ` φ`L2(Ω) = 1

Eg here are the 1st 50th and 700th eigenfunctions of adiscrete Laplacian on an L-shaped domain

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 9 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

First observe the propagation of errors e(i)

x(i+1) = x(i) + (1λ(k)max)(Akxminus Akx

(i))

x = x + (1λ(k)max)(Akxminus Akx)

=rArr e(i+1) = e(i) minus (1λ(k)max)Ake

(i)

Hence an equivalent question is

why is I minus (1λ(k)max)Ak a smoothing operator

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλn

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated

+ ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλ

(k)max

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Eg 1 The algorithmThus all components of the algorithm are now well defined

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 25: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

First observe the propagation of errors e(i)

x(i+1) = x(i) + (1λ(k)max)(Akxminus Akx

(i))

x = x + (1λ(k)max)(Akxminus Akx)

=rArr e(i+1) = e(i) minus (1λ(k)max)Ake

(i)

Hence an equivalent question is

why is I minus (1λ(k)max)Ak a smoothing operator

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλn

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated

+ ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλ

(k)max

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Eg 1 The algorithmThus all components of the algorithm are now well defined

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 26: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

First observe the propagation of errors e(i)

x(i+1) = x(i) + (1λ(k)max)(Akxminus Akx

(i))

x = x + (1λ(k)max)(Akxminus Akx)

=rArr e(i+1) = e(i) minus (1λ(k)max)Ake

(i)

Hence an equivalent question is

why is I minus (1λ(k)max)Ak a smoothing operator

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (I minus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1 + middot middot middot+ cn()ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλn

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated

+ ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλ

(k)max

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Eg 1 The algorithmThus all components of the algorithm are now well defined

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 27: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλn

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated

+ ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλ

(k)max

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Eg 1 The algorithmThus all components of the algorithm are now well defined

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 28: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Richardson smoother

Let λ(k)max be the largest eigenvalue of (spd) Ak Then

ALGORITHM x(i+1) larrminus Smoothk(x(i) b)

x(i+1) = x(i) + (1λ(k)max)(bminus Akx

(i))

is the Richardson smootherWhy does this iteration smooth errors

If the eigenvalues of Ak are 0 lt λ1 le λ2 le middot middot middotλn equiv λ(k)max

and the eigenfunction corresponding to λ` is ψ` then

e = c1ψ1 + c2ψ2 + middot middot middot+ cnψn

=rArr (Iminus1

λ(k)max

Ak)e = c1(1minusλ1

λ(k)max

)ψ1+middot middot middot+cn(1minusλ

(k)max

λ(k)max

)ψn

=rArr The components in ψn-direction are annihilated + ellipticity

Department of Mathematics [Slide 10 of 36]

Jay Gopalakrishnan

Eg 1 The algorithmThus all components of the algorithm are now well defined

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 29: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Eg 1 The algorithmThus all components of the algorithm are now well defined

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing v = SmoothJ (u(i) b)

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJ v))

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 30: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Eg 1 The algorithm

ALGORITHM u(i+1) larrminus MgJ(u(i) b)

1 Smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

u(i+1) = v + LJ MgJminus1(0 LtJ (bminus AJv))

This algorithm is a ldquobackslash multigrid cyclerdquo (or ldquondashcyclerdquo)

Department of Mathematics [Slide 11 of 36]

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 31: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Eg 1 A V-cycle algorithm

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 Pre-smoothing

v = u(i) +1

λ(k)max

(bminus AJu(i))

2 Correction

w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 Post-smoothing

u(i+1) = w +1

λ(k)max

(bminus AJw)

Department of Mathematics [Slide 12 of 36]

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 32: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Multigrid cyclesSchedule of multilevel grids in standard multigrid algorithms

-cycle

FMG schedule

F-cycle

W-cycle

V-cycle

hJ

hJminus1

h1

hJ

hJminus1

h1

All these cycles satisfy the optimal O(N) work estimateDepartment of Mathematics [Slide 13 of 36]

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 33: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Braess-Hackbusch theoremConsider the error reduction operator Ek given by

uminus u(i+1) = uminus Vcyclek(u(i) b) equiv Ek (uminus u(i))

Let a(u v) equiv (nablaunablav) and |||v|||a equiv a(v v)12

[Braess amp Hackbusch1983]

THEOREM If Ω is a convex polygon then the V-cycle algorithmfor Eg 1 converges at a rate δ lt 1 independent of mesh sizes

|||Ek|||a le δ

Department of Mathematics [Slide 14 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 34: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

where Pkminus1 is the orthogonal projection into Vkminus1 with respectto the a(middot middot)ndashinnerproduct

Vk = Pkminus1Vk︸ ︷︷ ︸

oplus (I minus Pkminus1)Vk︸ ︷︷ ︸

Coarse grid components Fine grid components

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 35: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus if a v isin Vk is left undamped by the smoother ie if

|||v|||a asymp |||Kkv|||a

then v must be a coarse grid function (roughly)

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 36: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

ProofThe proof is difficult to motivate since it demands considerabletechnical ingenuity Here is an outline of its ldquomodernrdquo version

Step 1 With Kk = I minus (1λ(k)max)Ak we first prove

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 The error reduction operator Ek of the algorithmsatisfies the recursion

a(Eku u)= |||(I minus Pkminus1)Kku|||2a+a(Ekminus1Pkminus1Kku Pkminus1Kku)

Using Step 1 and estimating we eventually prove the theorem

Department of Mathematics [Slide 15 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 37: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic(AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 38: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

= w minus Pkminus1wL2(Ω)

radic

λ(k)max

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 39: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(1

λ(k)max

AkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 40: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

(

w minusKkwAkw

)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 41: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

radic

a(ww)minus a(Kkww)

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 42: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||2a le

C hminus1k w minus Pkminus1wL2(Ω)

︸ ︷︷ ︸

radic

a(ww)minus a(Kkww)

le C|||w minus Pkminus1w|||aby the Aubin-Nitscheduality argument forfinite elements

Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 43: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Proof

Step 1 elaborated |||(I minus Pkminus1)w|||2a =

= (w minus Pkminus1wAkw)

le w minus Pkminus1wL2(Ω)

radic

Chminus2k

[

a(ww)minus a(Kkww)

]

Thus |||(I minus Pkminus1)w|||a le Cradic

a(ww)minus a(Kkww)

Using also the convergence properties of the smoothingiteration we finally have

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

Step 2 elaborated Later Department of Mathematics [Slide 16 of 36]

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 44: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Regularity amp ApproximationA critical inequality in the previous proof is

w minus Pkminus1wL2(Ω) le Chk|||w minus Pkminus1w|||a

This is proved using the Aubin-Nitsche duality argument whichrequires that the exact solution φ of

minus∆φ = f on Ω φ = 0 on partΩ

has an approximation φk isin Vk satisfying

|||φminus φk|||a le ChkfL2(Ω)

This is known to hold when Ω is a convex polygon

|φ|H2(Ω) le CfL2(Ω) |||φminus φk|||a le Chk|φ|H2(Ω)

( REGULARITY + APPROXIMATION)Department of Mathematics [Slide 17 of 36]

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 45: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Practical smoothers

The Richardson smoother requires λ(k)max at every level k

These numbers are not easy to obtain in practice even forsimple examples

Fortunately many other classical iterative methods possessthe smoothing property

x(i+1) larrminus Jacobi(x(i) b)

x(i+1) larrminus Gauszlig-Seidel(x(i) b)

Department of Mathematics [Slide 18 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 46: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

x(i+1) = x(i) + R(bminus Ax(i))

x = x + R(bminus Ax)

e(i+1) = e(i) minus RAe(i)

(Hence smoothing iterations smooth errors)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 47: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

The smoothing effectA classical linear iteration for a matrix A of the form

x(i+1) = x(i) + R(bminus Ax(i))

with some matrix R is a smoothing iteration if

(Iminus RA)e is smoother than e for any e

If D is the diagonal and L is the lower triangular part of A then

Jacobi iteration R = Dminus1

Gauszlig-Seidel iteration R = (L + D)minus1

The smoothing effect of these iterations is easy to demonstratecomputationally (Click to see code)

Department of Mathematics [Slide 19 of 36]

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 48: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

The smoothing effect

The smoothing effect on errors of Gauszlig-Seidel iteration

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

xy

A random vector After 7 Gauszlig-Seidel iterations

Department of Mathematics [Slide 20 of 36]

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 49: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

The smoothing conditionThe mathematical statement of the smoothing effect of anyiteration of the form

x(i+1) = x(i) +Rk(bminus Akx(i))

that is useful for multigrid analysis is as before

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

but now with Kk = I minusRkAk

As in the case of the Richardson iteration proof of thistypically requires not only an analysis of smoother Rk but alsoelliptic regularity and approximation estimates

Department of Mathematics [Slide 21 of 36]

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 50: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Abstract V-cycleAssume general nested spaces and inherited bilinear forms

V1 sub V2 sub middot middot middot sub VJ sub middot middot middot ( V a(middot middot) )

A1 A2 middot middot middot AJ middot middot middot A

R1 R2 middot middot middot RJ

All spaces have a base innerproduct 〈middot middot〉 and norm middot Operator Ak satisfies 〈Akv w〉 = a(v w) for all v w isin Vk

ALGORITHM u(i+1) larrminus VcycleJ(u(i) b)

1 v = u(i) +RJ(bminus AJu(i))

2 w = v + LJ VcycleJminus1(0 LtJ (bminus AJv))

3 u(i+1) = w +RtJ(bminus AJw)

Department of Mathematics [Slide 22 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 51: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le |||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 52: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le

[

(1minus δ) + δ

]

|||(I minus Pkminus1)Kkv|||2a + δ|||Pkminus1Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 53: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)|||(I minus Pkminus1)Kkv|||2a + δ|||Kkv|||

2a

(1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Department of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 54: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Abstract theory

THEOREM If Kk equiv I minusRkAk satisfies

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

then|||Ek|||a le δ equiv

C

C + 1

PROOF a(Ekv v) =

= |||(I minus Pkminus1)Kku|||2a + a(Ekminus1Pkminus1Kku Pkminus1Kku)

le (1minus δ)C

[

|||v|||2a minus |||Kkv|||2a

]

+ δ|||Kkv|||2a

Result follows since (1minus δ)C = δDepartment of Mathematics [Slide 23 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 55: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

This is the REGULARITY amp APPROXIMATION property[Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 56: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

It holds for our Eg 1 on domains admitting full regularity

|||φminus φk|||a le Chk|φ|H2(Ω) |φ|H2(Ω) le CfL2(Ω)Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 57: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Applying the theorem

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

We verified the condition of the abstract theorem in the case ofthe Richardson smoother Similar arguments prove it for theGauszlig-Seidel iteration and a scaled Jacobi iteration providedwe have the following For every f isin Vk the exact solution of

a(φ v) = 〈f v〉 forallv isin V

has an approximation φk isin Vk satisfying

|||φminus φk|||a le Chkf

The theory can be extended to situations with less than fullregularity [Bramble amp Pasciak 1987]

Department of Mathematics [Slide 24 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 58: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Therefore the intuition with which we built the multigridalgorithm previously seems to fail

While higher eigenfunctions become more oscillatory forelliptic operators for degenerate elliptic operators even thelower eigenfunctions can be oscillatory in which case thelower eigenfunction components cannot be resolved by thecoarse grid

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 59: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Eg 2 A degenerate elliptic eq

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ

Here 0 lt b0 le b(x y) but 0 lt ε(x y) can be degenerateHence the problem is not uniformly elliptic

Here are the 1st 5th and 10th eigenfunctions of adiscretization of this operator (when ε = 0001 b = 1)

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

Department of Mathematics [Slide 25 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 60: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

We pick a random vector apply the Gauszlig-Seidelreducer a few times and observe the resultscomputationally (Click to see code)

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 61: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Eg 2 Failure of smootherAs expected the performance of the standard Gauszlig-Seideliteration as a smoother is not satisfactory for this example

minus1

minus05

0

05

1

minus1

minus05

0

05

10

02

04

06

08

1

xy minus1

minus05

0

05

1

minus1

minus05

0

05

10

01

02

03

04

05

06

07

08

xy

A random vector After 6 Gauszlig-Seidel iterations

While the resulting vector is smooth in the y-direction it is os-

cillatory in the x-direction

Department of Mathematics [Slide 26 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 62: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Eg 2 Modifying the smootherRecall

|||(I minus Pkminus1)Kkv|||2a︸ ︷︷ ︸le C

[

|||v|||2a minus |||Kkv|||2a

]

︸ ︷︷ ︸

forallv isin Vk

Norm of fine grid componentsafter smoothing

le CA measure of damp-ing by the smoother

Thus for multigrid to work the undamped components aftersmoothing must be representable on coarser grids

Therefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 63: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 64: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Eg 2 Modifying the smootherTherefore the smoother must damp not only the highereigenfunctions but also some of the lower ones like

minus1 minus08 minus06 minus04 minus02 0 02 04 06 08 1minus1

minus08

minus06

minus04

minus02

0

02

04

06

08

1

as these are not well represented on coarse grid

Such functions are damped if we use a block Jacobi iterationwith each block consisting of unknowns along y-lines rarr

Department of Mathematics [Slide 27 of 36]

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 65: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Eg 2 Line smoother

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

The 10th eigenfunction Result after 6 point Gauszlig-Seidel iterations

minus1

minus05

0

05

1

minus1

minus05

0

05

1minus01

minus005

0

005

01

xy

larrminus Result after one iteration of the

line Jacobi smoother

x(i+1) = x(i) + R(bminus Ax(i))

(Now R is a block diagonal matrix in an

appropriate ordering) Department of Mathematics [Slide 28 of 36]

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 66: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Eg 2 TheoryThe V-cycle with line smoothers were analyzed by[Bramble amp Zhang 2000] Here are their results If u solves

minuspartx(ε(x y)partxu)minus party(b(x y)partyu) = f on Ω

u = 0 on partΩ thenLEMMA (REGULARITY)

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω) le Cεminus12f2L2(Ω)

LEMMA (APPROXIMATION) There is a uk isin Vk such that

|||uminus uk|||ale Chk

[

ε12partxxu2L2(Ω)+partyyu

2L2(Ω)+partxyu

2L2(Ω)

]

Hence |||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

Department of Mathematics [Slide 29 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 67: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Γ0

Γ1

DΩ 0

z

r

Ω minusrarr D

Significant computational savings (3D to 2D domain)

But numerical analyses face difficulties due to Γ0

Department of Mathematics [Slide 30 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 68: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 69: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

on Γ0

For smooth functions φ since partrφ is an even function of r

partrφ|Γ0= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 70: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 71: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Eg 3 Axisymmetric Laplace eq

Dirichlet problem on Ω Reduced problem on D

minus∆U = f on Ω

U = 0 on partΩ

(f is axisymmetric)

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f on D

u = 0 on Γ1

partru = 0 on Γ0

Weak formulation Find u isin V such thatint

D

r(partru)(partrv) + r(partzu)(partzv) drdz =

int

D

fv r drdz

forallv isin V equiv w isin L2r(D) partrw partzw isin L

2r(D)

︸ ︷︷ ︸

H1r (D)

w|Γ1= 0

Department of Mathematics [Slide 31 of 36]

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 72: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Eg 3 BackgroundIn view of the fact that line smoothings were necessary for thedegenerate problem of Eg 2 and because of the similarpresence of the singular coefficients in the current example ofthe axisymmetric equation

minus1

r

part

partr

(

rpartu

partr

)

minuspart2u

partz2= f

multigrid algorithms with line smoothing iterations andmultilevel grids with semicoarsening have usually beensuggested for this problem until recently

We will prove that neither is necessary

Department of Mathematics [Slide 32 of 36]

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 73: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Eg 3 Bilinear elements

Mesh D by square elements Vh = bilinear FE subspace of VExact solution u isin V

ar(u v) = (f v)r forallv isin V

Finite element approximation uh isin Vh

ar(uh vh) = (f vh)r forallvh isin Vh

Error analysis

|uminus uh|H1r(D) le inf

vhisinVh

|uminus vh|H1r(D) (standard)

le |uminus Πhu|H1r(D) (non-standard Πh)

le Ch|u|H2r (D) le ChfL2

r(D) (wtd regularity)

Department of Mathematics [Slide 33 of 36]

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 74: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Eg 3 Convergence of V-cycle

We apply the abstract theory with form a(middot middot) = ar(middot middot) baseinnerproduct 〈middot middot〉 = (middot middot)r and use the REGULARITY (whichholds on convex domains) and APPROXIMATION property to get

|||(I minus Pkminus1)Kkv|||2a le C

[

|||v|||2a minus |||Kkv|||2a

]

forallv isin Vk

THEOREM If Ω is convex the multigrid V-cycle with pointGauszlig-Seidel or Jacobi smoother applied to the bilinear finiteelement discretization of the axisymmetric Laplace equationconverges at a rate independent of mesh sizes [G amp Pasciak 2005]

Thus not all problems with degenerate coefficients requiresemicoarsening and line relaxations

Department of Mathematics [Slide 34 of 36]

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 75: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

Eg 3 Numerical results

We numerically estimated the operator norm |||EJ |||a by using afew iterations of the Lanczos method

J 2 3 4 5 6 7 8 9 10

|||EJ |||a 012 016 017 017 017 017 017 017 017

These values show that the practical convergence rates in|||middot|||andashnorm of the V-cycle are excellent

Ω = (minus1 1)2

Bilinear finite elements used

Meshes are obtained by dividing Ω into ntimes n square with n = 2k

Department of Mathematics [Slide 35 of 36]

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion
Page 76: Jay Gopalakrishnan University of Florida Montreal ...web.pdx.edu/~gjay/talks/montrealI.pdf · Jay Gopalakrishnan University of Florida Montreal Scientific Computing Days ... Iterative

Jay Gopalakrishnan

ConclusionWe saw how one builds multigrid algorithms using thebasic ingredients of smoothing prolongation andmultilevel recursion

We discussed the theory of multigrid convergence andsaw how elliptic regularity and approximation propertiesplayed a crucial role

We applied the multigrid paradigm to the followingexamples

Eg 1 Laplacersquos equation

Eg 2 A degenerate anisotropic elliptic problem

Eg 3 Axisymmetric Poisson equation

In Part II of this course we will see newer multigrid theoriesand advanced applications in electromagnetics

Department of Mathematics [Slide 36 of 36]

  • Why multigrid
  • Structure of multigrid algorithms
  • The multigrid idea
  • A typical pseudocode
  • Eg 1 Laplace equation
  • Eg 1 Multigrid setting
  • Eg 1 Prolongation
  • Elliptic eigenfunctions
  • Richardson smoother
  • Eg 1 The algorithm
  • Eg 1 A V-cycle algorithm
  • Multigrid cycles
  • Braess-Hackbusch theorem
  • Proof
  • Proof
  • Regularity amp Approximation
  • Practical smoothers
  • The smoothing effect
  • The smoothing effect
  • The smoothing condition
  • Abstract V-cycle
  • Abstract theory
  • Applying the theorem
  • Eg 2 A degenerate elliptic eq
  • Eg 2 Failure of smoother
  • Eg 2 Modifying the smoother
  • Eg 2 Line smoother
  • Eg 2 Theory
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Axisymmetric Laplace eq
  • Eg 3 Background
  • Eg 3 Bilinear elements
  • Eg 3 Convergence of V-cycle
  • Eg 3 Numerical results
  • Conclusion