(Slide) Advances in NA2003

vn ~ VP~Rq$s n r r r n } ~RqRpèRfd } snb$eRpEfR } r d n o r d r r t d n `n } 3ihf~rR8 z } R~R `n~~ { n dr r gz E`n~ kn$ r u r d l o r Rn~ } R~r`fzn RVn`ne$`d r r n `n } zs` Rqn h`s no snrj
d l o r Rn~ } R~r`fzn RVn`ne$`d r r n `n } zs` Rqn h`s no snrj $sr } p Re$s { l n r d n d n r r Rwn $sr`n Rl d d t r } d } i } ~ 1| { 2 %Dx 4 PzHyU9 w u t n r o n l i g vRs`qnpRwmkjhfed$Hp 3 y x s &w0t s t vu ' BD 4 ' 9 7 B Cerq(C(# "h2p# ihS389 gD 3&# f&e&db# `X VU&(9 T 4 Y " 2 @ % % @ % ' 2 @D 9 c 2 a Y W 9 4 % # " 7 ' % 7D D @ % 7 A " # @ % 9 7 4 % 2 " # ) ' % # SRQ PIHGF4E3CB $1&3886 53$110(&$" !
u t r r I~rfRf H nR ~R~ { u n ' r n d r ~yr zs`&fF wn $`n~gz { n
xC q T H
Cu~srndrn$r } n RRH w d r n WWY f 7 W X C v8tusq r f d ac q Pip h ge7 b` VWWW C 7IHC EC A (yxGFDB9 $ w 7 UTS RPFGFDB@865~h`qRe 31 Q I H C E C A 9 7 4 4 4 r n 2
r r )r ' ng ss0n`(q
&
$ r n n r n %~Rv$mq~`
ef } r } Rnsr } j o n n r r } pzs } ~ 1| n dr n qp } Pe } } VRn p~R q { o o r

" #!
C D
B 05 9 7 5 3 1 'A@86420 !
gvRP~ srf `n~g } PIsrfR no s`n n n d r n r ~ }& Rs no jV r r o n r r n r } Rhn } zs`j `nz 1| jGg ~Rv f`q } rR qnpo i r g r gr r
( & $ " )'%#
I EC A9 xB@7

o vn
r`nr } oi t ~
i~ r r ReRkoR~r r eg xB@7 I EC A9 r ot `nr } i ovn } } n d rd r wFsR&`qnq

#
G "
g g
vn ~ VP~Rq$s n r r r n } ~RqRpèRfd } snb$eRpEfR } r d n o r d r r t d n `n } 3ihf~rR8 z } R~R `n~~ { n dr r gz E`n~ kn$ r ur d l o r Rn~ } R~r`fzn RVn `ne$`d r r n `n } zs` Rqn h`s no snrj $sr } p Re$s { l n r d n d n r r Rwn $sr`n Rl d d t r } d } i } ~ 1| { 2 %Dx 4 PzHyU9 w
' BD 4 ' 9 7 B Cerq(C(# "h2p# ihS389 gD 3&# f&e&db# `X VU&(9 T 4 Y " 2 @ % % @ % ' 2 @D 9 c 2 a Y W 9 4 % # " 7 ' % 7D D @ % 7 A " # @ % 9 7 4 % 2 " # ) ' % # SRQ PIHGF4E3CB $1&3886 53$110(&$" !
d f zh`' e sr } sfhrf PPi~Pi n nd t t

C
&
r r jy1~Re$~
U V
ko$v
T
( f 8C A f 8C A
F d "h f F F f d F
S R Q I PH G Q I PH G F I d 9 I d 9 w h
f F d F f h d
r t n~ } hi$ {
( C E
f u~ A
y u$ f
C A 7 D% B@987
d } qn$Rq
5 3 6B 0 42
rv nfdn~g } 64 I 1B9 08 } r n 44 ( C ) 7 WWY ( ' Ax%&$ f PH CDAB9 $ !$ # Ax f yH DB9 $ W X C I C I C A w h f d 7 ` VWWW 7

FC T Q A H
! 7I C A "(yH DB9 $
UTS Q yH DB@7 `nhr } i I C A9 r ot r d l `n~ } GR~r`fzn R 2 v {
g "
" ) Gg " g

&
g $ h%"
z~r`n } `nn Rn~eR 2 kn$$v`n r tr o n r r r RR z~r $ } srS md { hsr$ef } $ ew $R``~R sG g 1 cb i mk d g tn n tn d u svsn H r`~n ~r V { rn
$ %" C C " # C
P I
` a%
P X YV R
h H G XV T R YWUSH I H 9 h f yH ( 9 7 I C I ` h IPQ( @7 d d % C 9 f u~ n`s8$r } n ~Re$~ A r d r n d r r
E(
P I
H
I I
P h H G 9 I PC QB9 7 d
` h
f 7d ` h
f PH B9 7 I C
f u~ A
G Q
r drn d n`s8$r } n ~Re$~ { r r

` h G h 9 f d (I ` b9 h 7 7 h f
R I PH G
7 Id
` h
S I PH
G
w 9 Id
E FD
A B@
g 97 8
G0 4 6

1 1 # ' % # )4320)$(&$"!
d e
$r } n ~s`nh`' e~R } r d r sR VvF z`n``r } si~qnzd on r dr n V s hd~riqR`d Rn s 2 { r o

C
zsr }
bcq I
z
9
V WU
a Y `
9 Cq I
a Y ` X
q v A v H
T
gz sReRp`hf ur d r r n n r i tn `g` PF1 R k Ri$$p dRq

C
P Q
( '
P G
P I
8 9
8 9 " 4 2 0 ( & $ " 76531)'%#! v I d C Q T d A d 7 PH T G 9 A I 9 7 h 7 9 T
P I

P G
I P h
HC xB9 7

P I
( $ )' "
b c V
dR`sq `n~heR`d } nhe$ r t

R H G

P h H G 9 I PC QB9 7 d
f PFB@7 IHC 9
o r o hu n~R$t } H V

1 1 # ' % # )4320)$(&$"!
g !D"
1 1 __ __ 16 9
H C F E C A I3GDB@
1 __ 4
1 C
t __ x 2
I "9 ) #0
C I " !7 I "9 8#9 9 #) f #9 ! 7 I "

' (h
u
8 U
&
a
o r n nn } `gIsrd knpo

&
#
a
C f
!7 w U!7 I
T
V 2 nz~ d or v T w @" H v ` h Y h U ! 7 9 h " q A

U
v A
T
U
TUS Q &C " U 7 9 % $ f #9 U ! 7 f " I " I # a f 9 U B9 k$ nRid C 7 n d
7 f nzs or v
8 U
8 U
7 h q
f

Y U 8 U
7 C
7 w
q h
f

H 7 h v
a 7
7
U
` 7
v 2
C A
f
9 @7
q 7 0 f 0f gz `n``r } iPH 7 r dr n gI r d l o o `n } rRfsr`fzn $epGn5 Re
rnn r `g
C
y s i sVRS`n~qR`gvzzsr r o n r d o n d u d $Pe$`rno V eR`d VRS`n~$tRd { o n r qn$ nRiIeRd { o

I #
a 7
D"
1 1 # ' % # )4320)$(&$"!
"
` P I
d ! 9 d
f yFB9 7 IHC
o r Rn `ne$`d
"
` ns vd s kBA 9 h d ~R } r n r ' n s`n(e ~$ } r ~r
& &
U U
'
q Cq
9
7 0 # q
7

v q f q

7 0
r dr n f gz in`ssr } y zqR`Sf d r d l o f `n~ } VRfsrzn R0nP R 2 (

( & $ )'%"
H C F E C A I3GDB@
Iq
9 f
I &`nq } eU~r$ { 9 no r n T T 7 h 9 7 $ f 7 V eRd { #9 ) T f I 9 I " n r or t hu~R } r H n~ nzs qRir {

8
1 1 # ' % # )4320)$(&$"!
e V zheR`d r V h r on r v&n~eR`d
Y
t$R`` $ sG n n tn dRj
C C

4 % % f H
X
C 9
f A
T Q A
(
C b q I h B9 4 A

h A ` %
f GB9 ) 7 I A
b $ %"
0.5 x 0.6
r i l ` 2 jR g ~rRiH ~`g ~ g } t r
0.1 0.2 0 0.3 0.4 |psi(x,t)| 0.5 0.6 0.7 0.8 0.9 1
___ Re psi(x,t) , |psi(x,t)|
E FD
A Q@
g g8 G0 4 6
@
1 1 # ' % # )4320)$(&$"!
0.8
0.6
0.4
0.2
0.2
0.4
0.6
0.8
|psi(x,t)|
0
0
0.1
0.1
0.2
0.2
0.3
0.3
0.4
0.4
t=0.006
t=0.000
0.5 x
0.5 x
0.6
0.6
0.7
0.7
0.8
0.8
0.9
0.9
1
|psi(x,t)|
0.05
0.15
0.25
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0.1
0.2
0.1
0.1
0.2
0.2
0.3
0.3
0.4
0.4
t=0.002
t=0.008
0.5 x
0.6
3% reflection
17% reflection
0.7
0.7
0.8
0.8
0.9
0.9
Initial Function at t=0, X=1, Y=1, dx=0.05, dy=0.05, V=0
1 0.9
1
0.8 0.7 0.6 psi(x,y,0)

|psi(x,y)|, t=T
0.5 0.4 0.3 0.2 0.1 0
15
15
10
10 5 5 15 0 y x 10
5 0 y 18 16 14 12 x 10 8 6
Version I:solution: X=1, Y=1, T=150, dx=0.05, dy=0.05, dt=0.0002, X=20, Y=20, V=0
0.8 |psi(x,y)|, t=T

|psi(x,y)|, t=T
0.6
0.4
0.2
0 0 15 10 5 0 y 15 x 10 5
1 1 # ' % # )4320)$(&$"!
Version I:solution: X=1, Y=1, T=100, dx=0.05, dy=0.05, dt=0.0002, X=20, Y=20, V=0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0
Version I:solution: X=1, Y=1, T=180, dx=0.05, dy=0.05, dt=0.0002, X=20, Y=20, V=0
1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 15 0 10 5 0 y 15 x 10 5
E FD
0
A B@
(
l z } R sr RiH g ~ g } t o r
g 8

G0 4 6
f A
u V it~ kn$u V n( } $ dr o
P
0 1000 1500 500 elapsed cputime 0 10000 20000 time steps 30000 40000
kh I 9 & I #$HR n } ``Pz~ qn$ n 9 r r ntr n v { n r n `n } `g 2 z~ir d Rn } ` e$`H `n~ } $hepV n u r n t r n d r t s$zzP Vvn I q 9 ~R } r o n ur r o r t d ~n~ } ~3 s`n n fsg~$v d {

1 1 # ' % # )4320)$(&$"!
DTBC vs. exp.ansatz (L=10,20,30), 40000 time steps
gz nr } r dr n ' n qpo rn(e R } `g V R(sg~$vs { on r t d

P
T

T T
C 4 4 C 4
f
!
T w
T

4 9 T f
T
T
C

T

9

nk } e { r
P P
I (9 9
& 5I 9 n v`n } q on r d & I q 9 % zv~R } Cv`n } q n r o n r d r r r t s`n~ } ~P H$s`nno ~~RRd { f A
"
C A

f GB9 I A

o r d Rn n~ qn$ m qno

g
'
"
h "
w 9
f

( C
C T Q

T C
f

( & $ )'%"
o n r R RVhvqn v( } vn eRd I q T 9 f R`R Rn } Rn n~ qn$ ~ id { rr n o o r
" g ! g g 5 " # "
10 10 ||aref||L2 / ||I||L2 10
7
f (
r n fhn } g
E FD C A ' 3 1 1
g 8
10
10
relative L2Norm of reflected part for different number of coefficients
0.1
t 0.2 L = 20 L = 30 L = 40 L = 50 0.3
n sRz( 8` p } jy ~ kn$ g qnH$ q ( t ~Re

&
f (
r n fhn } g
G0 4 6
@
Solution calculated with approximation of convolution cofficients, 5 terms in the sum of exponentials, X=Y=1, T=40, dx=0.02, dy=0.02, dt=0.0002, J=50, K=50, V=0
Solution calculated with approximation of convolution coefficients, 20 terms in the sum of exponentials, X=Y=1, T=40, dx=0.02, dy=0.02, dt=0.0002, J=50, K=50, V=0
1 0.9 0.8 0.7 |psi(x,y)|, t=T 0.6

|psi(x,y)|, t=T 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 40 0 5 10 15 20 25 30 35 x
Solution calculated by approximation of convolution coefficients, 20 terms in the sum of exponentials, X=Y=1, T=90, dx=0.02, dy=0.02, dt=0.0002, J=50, K=50, V=0
0.5 0.4 0.3 0.2 0.1 20 0 0 30 10 20 30 40 40 x

0 10
50
50
0.25
0.25
0.2
0.2
|psi(x,y)|, t=T
0.15
|psi(x,y)|, t=T
0.15
0.1 10 0.05 30 0 0 5 10 15 20 25 30 35 x
0.1 10 0.05 30 0 0 40 5 10 15 20 25 30 35 x
20
40 40 45 50 50 y
0.014 0.012 0.01 |psi(x,y)|, t=T 0.008 0.006 0.004 20 0.002 30 0 0 5 10 15 20 25 30 35 x 40 50 50 40 45 y 10 |psi(x,y)|, t=T
0.014 0.012 0.01 0.008 0.006 0.004 20 0.002 30 0 0 5 10 15 20 25 30 35 x 40 50 50 40 45 y 10
E FD C B@ A ' 3 1 1
0 10 20 30 40 45 50 50 y
0 20 40 45 50 50 y
0
z } Rlp sr$Fpg ~ g } f A V t r r n `n } `g f ( fhn } g r n
f (
g 8
G0 4 6
p g Ag q shc h 0he2t dc g c t g g c e e w dsrg nscdb|z{tdwrghe2t tSjwAyHn2xyi2 mCvt et Sjy%jgwx w 5b c b t l t w t e t p eAskgscd|{tRrsc f c n gjw uhy2y w b z t ew t rYbd hc ie eAsrghc f w g tg t y w w ` q
n 2 p
t d ojtA c n
f gwyw q PY %eAsrhys c cg w g c pw ~ }2edjwkghcRb|z{t tfuhywAyxY26 xwuveAswksssrt sAy e l g c tw g w eAswrg b YA g e p C2jt g e c e s2t dfc q ml e w w y e djkjgid h f c y e b g edS
` ` V
2fIp fAf2 t y t u e w y t yw b u t r vfxAvRsc q fihfda p b g e c b $W 0Y(X5
"$ " #HUT# QSF 321 @ RQ3 P$ $ E B IHGFDC @ A9 7 865 4 $ $ 0)('&%#" !
t g l w e t y z s p gg d q Yhc 2 f jgi2VV~ ~ iy)Asrg c ot VVs t y ew e w l z 5c c p tg t p g c e vthdwAy e c w iy fo ejy2 kYA eVV { s o y ew w t w g b c z h t wy A fueAf 2 VV zs k mz w s w d i w v d l ~ ~ } h {|o z yxv u tr i qo d w d s p kn lmj i hf ru k g fe d I I
` ` `
} u Asrhd|vRsc f c n e w g c b z t t r } p 0S 22jt Rsc g u e t r f t w g c x f tRrwih2~ t g Rr f c w 5fe u dsrhdd{)t g A p 2t e t t p b e w g c b z t t y s 2 } p 02jt q % d{yYu t t r ew eAsrghb A2 ghc de 2 2Asrhsjtc T u AAsrhc f w t l e w g c w t e w g w t AP hy2 wdCp c notCywkgcyieA c y2A rYA A 6 b g ew e p b t w g b c c } g e t pw y t k h2CyA{wu
$ " %#
s tqrS TpRQP gPfQS 7 QRP edb c b`PYU&i4ehX W cVUTSRQP B ( ( c a ( ( ( 00( (0 0 ( H 34 I@ 4(CC( F

p
( 0 (0
p gg A2 q h c t 8 11 1 1 ))))) 02) ) ) ) @ ) )h ( ( ( ( c t 3F 4( ( 4C C )) )) F 6 4C C @ DC( EAC( B G 4( ( ( 3
yxv rwu
( yv u ( ( ( (
F 4C C pYXP4T3 Q ( ( h Pp@ h (( " $" '&%#!

$
67 3 (4( ((
((
89
l Q
w v u d
edjwkghcRb|{d{c kYh2t jgw yfd z t t r w g c u ew b 8 ml f g t At 2t SjA0jx w y e t w y t gw e `
k j h e d igfYu Qr
y t e t 72{t
} v x
q u q ru
~ c u e c t y s ftdfsw ~ u e c c p { c jtAo
b i
8 8
y t x } v t 79g 2c 0vwu
VV ` p v u ` d VV k ` d p v u d
e fU g hU
6
d d d b cT
h g jt w w
! "
` XWH F H aF YQIw GE v s dz yU b g 2 s T rh2 p t z cw g e t g R PH F SQIw GE zD B d s k C A @ l k w d i v i z i { s o H o d v 9r i p o h d 8s { z 7 s { tYA i ~ t 6 z ge g ! % ('& #$# $ $% # 5 3 42 1 0 )
2 3
h {
} 1 c t w y & t ml hfc sA(u 07y jyA f c x t t Asrh yw x w jrg c ic ( 0c o} )#u h 5G{srt 2 sy e w g c c w y e ( ( ( y t tw g w & 'y p h % xy c ( #n x $ g " e w g Asrhb AP t d(n l c jgw vw y o} !u h 5 t 2 jAy e c t y ew tg w t
` `
8p
f i
hy hc PvdjiPsQc g rYA c g t g e w w p e w g b
! "
t i h { e i ie i b { ' i { f i '% i e w g c b z t t r o w v k d i p o Asrhd|vRsc f c n "'&$%#!# # h% ! % ('$ " ' $ s
100 0 90 Transmission Loss 10*log_10 |p|^2 80 70 60 50 40 30
10 12 Range r [km] Example 1
14
16
18
20
Acoustic Wave Field |psi(r,z)| in stratified Ocean 0 0.2 0.18 0.05 0.16 0.14 Depth z [km] 0.1 0.12 0.1 0.15 0.08 0.06 0.2 0.04 0.02 0 0.5 1 1.5 2 Range r [km] 2.5 3 3.5 4 0
Workshop Advances in Numerical Algorithms Advanced Simulation Technologies
Graz, Austria, September 10-13, 2003
Turbulence modelling from the perspective of the commercial CFD
Dr. B. Basara AVL List GmbH
AST-UMW_CFD_Name-1
TM
Product Logo
Content
- Introduction (motivation, some examples) - Computational grids - Numerical method (control volume method) - Turbulence models - Conclusions
AST-UMW_CFD_Name -2
TM
Product Logo
Motivation
- Computational Fluid Dynamics (CFD) is used to solve fluid flow and associated transport processes. - Complex real-life flows are predominantly turbulent - Turbulence models are considered as an Achilles heel of modern CFD - Plugged on numerical algorithms and used in various calculations of fluid flow, the turbulence models are the largest source of error - However, numerical algorithms often define the range of usability of various complex turbulence model
AST-UMW_CFD_Name -3
TM
Product Logo
Some real-life examples
AST-UMW_CFD_Name -4
TM Product Logo Turbulence models
DU i P = + xi x j Dt
uiu j
U U j 2 U i k ij ui u j + x j xi 3 xk
-known as Reynolds stresses are modelled
- robust ? - accurate ? - where is the optimum?
AST-UMW_CFD_Name -5
TM
Product Logo
Automatic Tet Mesh
Meshing Options
Automatic Hex Mesh Block Structured Mesh Multi-Domain Meshing
hUser Defined Boundary Layer hEntirely Conform
h User Defined Boundary Layer h Entirely Hex h Entirely Conform
h Semi-Automatic h Entirely Conform
h Arbitrary Interfaces
AST-UMW_CFD_Name -6
TM
Product Logo
Calculation grids
AST-UMW_CFD_Name -7
TM
Product Logo
Finite Volume Method
- polyhedral control volume - co-located variable arrangement - connectivity is defined for cell-facecell Solution algorithm directly includes: hArbitrary Interfaces hlocal grid refinement hrearrangement of the cells due to moving/sliding
AST-UMW_CFD_Name -8
TM Product Logo
Control volumes
AST-UMW_CFD_Name -9
TM Product Logo
Discretization method
( dA) = ( dA i )
j k k
= ( ndA ) j
V (...)dV = A n (...)dA or V x k dV = A dAk
n (...)dA = n (...)dA
A j =1 Aj
n
nf
or
dA =
A k j =1
nf
Aj
dAk
f f f d kk V dV + (U k U bk ) dAk = dAk + s dV + sAk dAk Aj Aj V Aj dt V xk j =1 j =1 j =1
Rate of change: R
Convection: C
Diffusion: D
Sources: S
AST-UMW_CFD_Name -10
TM Product Logo
Surface and volume integrals
Aj
f idA = f k dAk f k Ak and

Aj
VP
dV PVP
nf nf nf d ( PVPP ) + m j j ( kk ) j Ak = ( sV ) P VP + ( sAk Ak ) j dt j =1 j =1 j =1 xk j
f f f d V or ( PVP P ) + C j D j = S + ( SA ) j dt j =1 j =1 j =1
TM Product Logo
Discretization method
All dependent variables are stored at the geometric center of a control volume (colocated or non-staggered variable arrangement); at boundaries they are defined at the centers of the CV boundary faces. The surface integrals are approximated by using the values of integrands that prevail at the geometric center (known as the midpoint rule approximation). For evaluation of dependent variables, their derivatives and fluid properties at locations other then cell centers, linear variation in space is assumed.
TM Product Logo Gauss formula, deferred correction approach ...
j = f j P + (1 f j ) P
j = j ' + j ' r j r j '

1 P VP
nf
)
( A )
j =1 j k nf j
1 j Aj or x V j =1 P k P
j = j +
P P j i d j Aj id j j
Aj
j = f j P + (1 f j ) P linear interpolation j j = arithmetic average j = 0.5 P + Pj
TM Product Logo
Central and upwind differencing schemes
j = f j P + (1 f j ) P
P if m j 0 = Pj if m j < 0
UDS j
UDS j
j =
+ f
* j
mj mj
Pj
1 f j if m j 0 f = if m j < 0 fj
* j
TM Product Logo
Higher order schemes
TM Discretization procedure Product Logo
-convection
C j = m j
D j = j Aj Aj id j
UDS j
+ m j j Pj P
P
-diffusion
Pj
A2 j dj + j j A j Aj id j
-Algebraic equations
a P P =
a
j =1
nf
j Pj
+ S
solved by the biconjugate gradient method in conjunction with the incomplete Cholesky preconditioning technique or by the Algebraic multigrid
TM Product Logo
Face Velocity (pressure-velocity coupling)
TM Product Logo Momentum equation and the standard k- model
DU i P = + xi x j Dt
2 ui u j = 2 t Sij ij k 3
t = C
k2
1 U i U j Sij = + xi 2 x j
Dk = ( Pk ) + x j Dt
t + k
k x j
+ t k x j
U k D = C 1 Pk + C 3 k C 2 + xk Dt k x j
K-
- robust ?
TM Product Logo
Reynolds-stress model - RSM (AVL Swift uses SSG)
ui u j ui u j U j ui u j U i ui u j k +Uk = ui uk + u j uk + Cs uk ul + t xk xk xk xk xk xl
p ' ui u j 2 ij + + x j xi 3
1 k = uiui 2
U k D = C 1 Pk + C 3 k C 2 + xk Dt k x j k C ui u j xi
RSM - accurate ?
U i , ui u j , p coupling ?
A vortex shedding around a square cylinder

Product Logo
(Re=21400). Predictions with RSM (SSG) model.
-velocity
-t.k.e
Table 1. Predictions and measurements of integral parameters. Cd Present RSM Measurements 2.28 2.16-2.28 Cl 1.39 1.1-1.4 Str 0.141 0.130-0.139
Product Logo
Example: VW model
Data
-In the case of RSM model, a description of the flow pattern over the slant is very close to the measured one as shown in Figures. -Predicted pressure distribution by RSM is in very good agreement with the measurements.
k-
1,5 1
RSM
DATA SWIFT k-eps SWIFT RSM
0,5
0 0 -0,5 0,2 0,4 0,6 0,8 1
-1
-1,5
Product Logo
Renault model
Product Logo
External Car Aerodynamics
Ford KA
K- turbulent model produces too less separation on the rear window Transient RSM approach has been used.Global
coefficients (lift and drag) fit better to the measurement RSM is closer to measured Cp curve specially on the rear end of the car The CPU time by RSM was 4 times longer, the same mesh has been used
Experiment
SWIFT k-
SWIFT RSM
AST-UMW_CFD_Name -23R
Product Logo
External Car Aerodynamics
AST-UMW_CFD_Name -24R
TM Product Logo AVL HTM model
DU i P = + Dt xi x j
ui u j ui u j k + Cs uk ul xk xl
ui u j ui u j U j Ui +Uk = ui uk + u j uk + t xk xk xk xk
p ' ui u j 2 ij + + x j xi 3
U k D = C 1 Pk + C 3 k C 2 + Dt xk k x j
k C ui u j xi
1 k = uiui 2
2 ui u j = 2 t Sij ij k 3
t = C
k2
C
TM Product Logo
AVL HTM model
-Therefore, AVL HTM model is more accurate than the k- model

-AVL HTM model is more robust than the RSM -AVL HTM model introduces following formula instead using the constant value in Boussinesqs relation, thus
U i k 2 2 C = ui u j / S x j S = 2 Sij Sij
- Equation above is checked by using DNS data.
TM Product Logo
Example: a vortex shedding flow
-measurements -measurements
-calculations
-calculations
-transient calculations show a large variations of
, Basara et al. 2001
TM Product Logo
Example: Backward-facing step
K- HTM
-the worst convergence rate achieved with RSM, see Figure above (the grid is orthogonal !). - Basara & Jakirlic (2003)
TM Product Logo
Example: 180 degree turn-around duct
K-
RSM 180 degree
AVL HTM
C = 0.02 0.2
-separation predicted with HTM and RSM but not with the k- model (Basara 2001).
Product Logo
Real-life application: EADE Benchmark
Prototype VRAK
Results Comparison Drag
Exp. Case 1 k- HTM 0.359 0.368 0.351
[%]
2.45 2.22
Lift
0.336 0.466 0.365
[]
0.13 0.029
TM Product Logo
Simplified Bus
K-
-A separation point is fixed by the shape of the body and therefore the flow pattern predicted by two models is very similar. -Note the difference of the turbulence kinetic energy at the front part of the body as well as behind the body. -Therefore, the acoustic module will get different sources !
HTM
Product Logo
A linearized Euler Method Based Prediction of Turbulence Induced Noise using Time-averaged Flow Properties.
The turbulence is regarded as the autonomous source of the noise and therefore it is very importan to get a proper intensity and the distribution of turbulence kinetic energy.The time-accurate sources can be extracted from the results of time-averaged RANS. The radiation of the acoustic sources is determined using a Linearized Euler Solver. Etc.
TM
Product Logo
Conclusions
-Courtesy of Dr. B.Niceno (AVL sponsored PhD work at TU Delft)
- Alternatives: DNS, LES, DES, PANS ? - LES and DNS are too costly for most engineering calculations and require additional modelling for other flow classes e.g. multiphase flow, combustion, etc., which still have to be resolved. - DES and PANS are under investigations - However, the RANS framework, in conjunction with turbulence models, can be expected to remain the main tool to solve practical industrial applications for a long time
} u o t o y o ~ } | lu { p t z yq u x wv i v l nq u o t s r m q p o n m l k j i h g f e d y x w u t t sr a` U` q V p T i W dh fTTaU U` W` UgTUY V` fv Y`RWWa setdsWr` UWQSTTcY ` a b a`Y Q T X W V Q R UU TS R Q P I @ H G @ B @ B A @ F E D C @ B @ B A @ 9 87 65 4 3 2 1 1 0
! ! ) ( ' $ ! ! ' & %$ # " ! ! !

10 10 Semilog Plot 10 10 10
6 3 0 3 6
500
Eigenvalues on scale 9
linear scale 1 logarithmic scale 0 1000 1500 2000 x 10 7 2 5 6 Linear Plot

5
3 4
' ) & ) 9& 6& 4 ) 9 C ( 7 B
4 ( 5 18
6 ) ( F (2 B F 9 C
73E CD

' ) & ) 9& 6& 4 ) 9 C ( 7 B
2 1 ) 9 ' 1& 4
4 ( 5 18
6 (2 ( 0 1 A

10
5
) 9& 6 3 8 & 7 6 5& 4 ( 3 2 1 0 ) ( '& %
$ #
"!

10
10
500
Logarithmic Scale
eigenvalues of X T eigenvalues of N X N
1000
1500
2000
u ed c b s w d s Vi Y e b X y T x
f ai g s b i g u w e g s g U T v
u g ed e t s r T q
e X pi f W V e h g f e b ed c b a ` Y X W V U T S
R PI Q PI H G
p q k q o p j p m o p j
pk
ui d
p p m o p j po q j q m l q k
h f f i v d g v g
ts r
i vw
q o p j no n j po q j q o q j l
~ ~ } y | { z y x v w f v f h i v u u i d
tsr
qm pm
qk pk
qo q j qo p j
po q j no n j
h fi g
m l k j
i hg fg fe d
7 6 A@
8 9
7
6
77 6
5 4 %$ 3 ) 2 ) 1 0 ! % (' ) & %$

# "
!

tv u
i i w
y xw t z sr q po k k k n j ~h g }| d { {
z y xw v u t sr q po n k k j ih g m l l k k fe d
v iw w xq f i g u v i y qi t i x f i uw
f x i v i x i
qi
tv u
f i g u v i y qi
f x i v i iw x f uw
y q i w xq r tr s r f
fq p
tv u
i h g ff e
d S EV U E c b
D F C F U C RC a F U Y E V ` Q E R H S C R QB
YC V R E S S X W
V UT
E P I H G F E DC B
G
0 1 2 3 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.05
A
S
R Q " ! %
A P
# " !
HGF E 9 3 8 76 76 @ 76
BD @
B 76 @ A
54 31 2
' & % ! # $ # " !

0.1 0.15 0.2 0.25 0.3 0.35 0.4
A
s p w x a W x y W Y b y W b q p i x x i r h f ie h p h b a w Y Y x h e ig e p s u y a Y` Y X W l e p s a b Y y W` z ed ed h g f p q i h h u h hp h
fi g f p p m r
l p pq f i h pg h g f f i
p f e e p h h p h d he s i r i g h e ig
p f e e p h h q d fq g i m i s p ed u p q i h i r g g d pe e h
} ~ ~ ~ } z y | { z y u x uw t v u t s r r p pq fq q p p h f p h m p m v k ed o h nq r g r h m l tj t s p fe p i i d r g i h pg v ed f ed u b Y w Y` W X Y y x w v u t s p r hq fe h i pe i e h g fe d c b a Y` Y X W V
9 8
0.04 0.02
7 6 5 4 3 2 0 0.2 0.4 0.6 0.8 1

0.04 0.06 0 0 0.02
7 6
7 6
scale j
scale j
5 4 3 2 1 0 0.5 locality k 1
5 4 3 2 1 0 0.5 locality k 1
1.5 1 0.5 0 0.5 1 1.5 0.5
1.5
scaling function
0.5
wavelet 0 0.5 1 1.5
0 wavelet 0.5 1 0 1 2

0.2 0.4 0.6 0.8 1
scaling function
0.35 0.3 0.25 0.2 0.15 0.1 0.05
9 85 6 & % 7 $0 % 65 $ ( % 4 $0 "& " 3 2 1 $& "& & 0 ( " ) ( % ' $ % $ & % $ # " !

g `c b r g p g Y c Y g c t g ` a h Yu u ` r Y X b hc t g f d ` Y p i f g Y c Y b h Yi i f d h Y X b hc j Yi i c p `c d u p b X bc Y f X Y vi g ` D E V E G D S B u `y
` s D V C G R R Q S D I E B E D G BR A B E D S D TF D g f h f X u ` q Q A E D C G D T A VR f h f X f c Y u fi hc d ~ } | { z x v g f u `y j x w v v u t s r q DA B D I p e q d k o n m k w kl x n m k l n q m k l n q w q m k w kl x
j i g f h g f e d u ` b fu Y a ` t a h Y X b X bc g `c b fi p d u ` f Y Y X b g c vi b g Yi f c p Y u `
w bf q q X bc xw w x y x w q q q vs t Y g c d u Y b Y t Y s g f r q g `c bp i ` h g `c b f ec d c ba ` Y X W A B E V A D C U T S F ER Q G D P D I E H B G D F E A D C B A @

500 450 400 350 300 250 200 150 100 50 0
50
100
150
200
250
300
350
400
450
500
H F % & !& ( & ) 1 f 5 04 3 @ 7 5 0 8 y x w `Y v v 1
u u u s t s r h q 2 7p
fe i h feg fe d
`Y 8 X4 ) ) 8 6 6 W ) c F b ' $ D @ 8 9 a 82 V ' G T U $ %$ & T
S R Q I B P I B 0 8 ) A 8 @ 9 4 8 5 0 H $ % F ( G F E& ! D % " % @ 9 1 C B 0 8 ) A 8 @ 9 4 8 )2 1 7 6 1 5 04 3 02 1 0 )
( ' & ! % $ !#
"!

t ys l p }s q po v r n q ~ n w v n m l ~ t p | ls o s p uq n w y y n q v w p n { n m l z o n ~ y { n m l y w w nt s l w z w n xs q l p { o n ~ y { n m l t y l t n ~ t n n ~ f d j
q y l p q n y o ps l t n q n
s ~ nm l y l nv ~
f d h
w n v o p u t n rs n ~ n l po y ws y t
z yq n } y l n w yo x |q n u n q p w n v o p u t n rs n l w no o p { w n m l
z no p x w y l no p x w n w p n q x t s w n v o p u t n rs n n rq po n m l
g g f g d i h g f e
g g j d
10 10 Semilog Plot 10 10 10
6 3 0 3 6
500
Eigenvalues on scale 9
linear scale 1 logarithmic scale 0 1000 1500 2000 x 10 7 2 5 6 Linear Plot

5
3 4
!0 v 5

8 ' # # # # A C # y

x w v ut s r q # " ' I dV p i h XV Pg` Y fT P eV P dX P`` Y W X c W P` b SQa ` YT SV XT P WV U PT S R PQ PIH ! 6 5 4444444 ) 6 F ' G F A C E D AC # " B A @ 9 # 8 7 6 # 5 %

4444444

3 3 3 1 2 1 0 ) ' ( % ( % & ( % ! ' % %& % " ! $ #

r g l tj qj l t gr e j s t g q d j tr e e x x w } y w u
10
5
zx w u } } y w y z| u v | } y { y } u z y w y u } y w| y u x x | } v zx ~ }| { y u z y x x w v u l t j q n s j r q p j o g n f m l g kj i h g f e d
10
10
0.5
0 5
eigenvalues of N X N T eigenvalues of B S B
500
500
Logarithmic Scale
Linear Scale
1000
1000
eigenvalues of X eigenvalues of S
1500
1500
2000
2000
w w
4v
7 u 8 $ d q
d u 8 d t$ p i0 i
9 8 $ 7 0 b
$
srqr $ F h g f B eD d
c
`
2 3 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.1 0.2
T F # E SD G F # E RD G F # E QD G F # E PD G F P E ID G H H H G F # E ! D C
9 8 7 1 0 ) ( ' & 5 6 5 3 2 1 0 ) ( ' & % 4
$ # "!
10
4 5 6
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
$ YX B W V U B A@
e e e e
kp
~ y ~
lf if f q } | zy { zy y k e p ri k ex wqf ed o h v
u s t s k i r l ed o h l l q ip k o n
m l l e h ik j d i h d g e f ed y x

Wavelet Methods for Linear-Quadratic Elliptic Control Problems

Carsten Burstedde
Institut fr Angewandte Mathematik u Universitt Bonn, Germany a
Advances in Numerical Algorithms Graz, September 13, 2003
Supported by Deutsche Forschungsgemeinschaft SFB 611
Motivation Evaluation of (fractional) norms Uniform condition numbers Fast fully iterative scheme Potential for adaptivity
Outline Modelling Minimum conditions Discretization Solution strategy Numerical results
Neumann Problem with Distributed Control Minimize 1 2 J(y, u) = y y0 H1s() + u 2 1t) () (H 2 2 a(y, v) := ( y v + yv)dx = f + u, v for all v for all y, v
subject to
y, Av := a(y, v) Target state y0, right hand side f , weight Non-trivial solution for y0 = A1f Smoothness parameters
0 s 1 for observation space Z = L2() . . . H1() 0 t 1 for control space U = L2() . . . (H1) ()
Conventions Weak formulation: nd a(y, v) = f + u, v where a(y, y) v 2 Y a(y, v) =: v, Ay for all y Y for all y, v Y y Y := H1() such that
for all v Y
Observation space Z := H1s() as regular as permitted by data, embedding Y Z, T : Y Z Control space U := (H1t) (), U Y , E : U Y Any combination of norms realized by s and t
Wavelet Bases H {Y, Z, U} Sobolev space with H, dual space H with , HH Wavelets H := {H, : I H I} H Riesz basis for H H v = vT H :=
I I
innite index set I I with v H v
vH,
I) 2 (I L2
Diagonal scaling for Sobolev space H
H = DHL2 , L2
T Dene Riesz operator RH = RH : H H by
u 2 =: u, RHu HH u, RHu H with
I) 2 (I
RH = RH 1 = L2 , L2
L2
=: R
independent of H
Lagrangian Multiplier Formulation Abstract formulation 1 L(y, u, p) = T y y0 2 + u 2 + p, Ay f Eu YY Z U 2 2 with embedding operators T and E
Ay = Dy,u,pL = 0 RY u =
f + Eu () ET p
AT p = T T RZ (T y y0)
All-in-one Formulation Eliminate y, p from (), solve for u () with Q= g = (T A1E)T RZ (T A1f y0) (T A1E)T RZ (T A1E) + RY Qu = g
Residual r = Qw g = RY u E T p last equation of () Both Q and g contain two applications of A1
Wavelet Discretization Cookbook Operators on function spaces innite-dimensional matrices Embedding operators diagonal scaling on wavelet coecients Riesz operators Gramian matrices L2 , L2 T :Y Z E:U Y RZ : Z Z RY : Y Y
DZ DU R R1
Norm changes by s, t realized by diagonal matrices DZ , DU Wavelet norm equivalences 2(A), 2(Q) = O(1)
Solution Strategy Inexact Conjugate Gradient Algorithm Nested context: inner and outer systems Construct conjugate gradient algorithm on Routine to compute residual
Qu = g
r = RES(u)
y (by inner CG) p (by inner CG) yields outer Residual
Ay = f + DU u AT p = DZ R(DZ y y0) r = R1u DU p
Computation of Qx in outer conjugate gradient algorithm by
Qx = g RES(x)
(g is computed to high accuracy on startup using two inner CG runs)
Nested Iteration Algorithm Compute exact solution u on coarsest discretization level Repeat Prolongate u to next higher level Compute g to high accuracy on this level Initialize outer CG (d = r = RES(u)) Repeat Iterate outer CG (using g RES()) Until r discretization error, yields solution u Until target accuracy is reached Complexity O(N ) instead of O(N log N )
Numerical examples - 1D Solution y and control u for y0 = 0, f = 1, = 1 for dierent norms

0.55 0.5 0.45 0.4 0.35 0.3 0 0.25 0.2 0.15 0.1 -1.5 0.05 0 0 0.2 0.4 0.6 0.8 1 -2 0 0.2 0.4 0.6 0.8 1 -0.5 "Y-D1j8u0z0B0.gnp" "Y-D1j8u20z20B0.gnp" "Y-D1j8u10z10B0.gnp" "Y-D1j8u5z5B0.gnp" "Y-D1j8u2z2B0.gnp" "Y-D1j8u1z1B0.gnp" 2 "U-D1j8u0z0B0.gnp" "U-D1j8u20z20B0.gnp" "U-D1j8u10z10B0.gnp" "U-D1j8u5z5B0.gnp" "U-D1j8u2z2B0.gnp"
1.5
0.5
-1
Numerical examples - 2D f , y and u for dierent right hand sides, y0 = 0, = 1, Z = U = L2

line 1 line 1 line 1
12 10 8 6 4 2 0 1 0.8 0 0.6 0.2 0.4 0.4 0.6 0.2 0.8 1 0
0.91 0.9 0.89 0.88 0.87 0.86 0.85 0.84 0.83 0.82 1 0.8 0 0.6 0.2 0.4 0.4 0.6 0.2 0.8 1 0
-0.8524 -0.8526 -0.8528 -0.853 -0.8532 -0.8534 -0.8536 -0.8538 -0.854 -0.8542 -0.8544 1 0.8 0 0.6 0.2 0.4 0.4 0.6 0.2 0.8 1 0
line 1
line 1
line 1
1.2 1 0.8 0.6 0.4 0.2 0 -0.2 1 0.8 0 0.6 0.2 0.4 0.4 0.6 0.2 0.8 1 0
0.18 0.17 0.16 0.15 0.14 0.13 0.12 0.11 0.1 0.09 1 0.8 0 0.6 0.2 0.4 0.4 0.6 0.2 0.8 1 0
-0.119 -0.12 -0.121 -0.122 -0.123 -0.124 -0.125 -0.126 -0.127 1 0.8 0 0.6 0.2 0.4 0.4 0.6 0.2 0.8 1 0
Numerical examples - 2D Solution y and control u for Z = L2 and dierent norms for U
0.91 0.9 0.89 0.88 0.87 0.86 0.85 0.84 0.83 0.82 1 0.8 0 0.6 0.2 0.4 0.4 0.6 0.2 0.8 1 0
0.69 0.68 0.67 0.66 0.65 0.64 0.63 0.62 0.61 0.6 1 0.8 0 0.6 0.2 0.4 0.4 0.6 0.2 0.8 1 0
0.07 0.06 0.05 0.04 0.03 0.02 0.01 0 -0.01 -0.02 1 0.8 0 0.6 0.2 0.4 0.4 0.6 0.2 0.8 1 0
line 1
line 1
line 1
-0.8524 -0.8526 -0.8528 -0.853 -0.8532 -0.8534 -0.8536 -0.8538 -0.854 -0.8542 -0.8544 1 0.8 0 0.6 0.2 0.4 0.4 0.6 0.2 0.8 1 0
0 -0.2 -0.4 -0.6 -0.8 -1 -1.2 -1.4 1 0.8 0 0.6 0.2 0.4 0.4 0.6 0.2 0.8 1 0
12 10 8 6 4 2 0 -2 -4 -6 1 0.8 0 0.6 0.2 0.4 0.4 0.6 0.2 0.8 1 0
Numerical examples - 2D Solution y and control u for Z = U = L2 and dierent choices for
0.91 0.9 0.89 0.88 0.87 0.86 0.85 0.84 0.83 0.82 1 0.8 0 0.6 0.2 0.4 0.4 0.6 0.2 0.8 1 0
0.21 0.2 0.19 0.18 0.17 0.16 0.15 0.14 0.13 0.12 1 0.8 0 0.6 0.2 0.4 0.4 0.6 0.2 0.8 1 0
0.0004 0.0002 0 -0.0002 -0.0004 -0.0006 -0.0008 -0.001 1 0.8 0 0.6 0.2 0.4 0.4 0.6 0.2 0.8 1 0
line 1
line 1
line 1
-0.8524 -0.8526 -0.8528 -0.853 -0.8532 -0.8534 -0.8536 -0.8538 -0.854 -0.8542 -0.8544 1 0.8 0 0.6 0.2 0.4 0.4 0.6 0.2 0.8 1 0
-1.542 -1.544 -1.546 -1.548 -1.55 -1.552 -1.554 -1.556 -1.558 -1.56 -1.562 1 0.8 0 0.6 0.2 0.4 0.4 0.6 0.2 0.8 1 0
0 -1 -2 -3 -4 -5 -6 -7 -8 -9 -10 -11 1 0.8 0 0.6 0.2 0.4 0.4 0.6 0.2 0.8 1 0
Numerical examples - 2D Residual error in 2 versus number of outer iterations

0.01 "r-D2j10u0z0B1.txt"
0.001
0.0001
1e-05
1e-06
1e-07
1e-08 0 20 40 60 80 100 120 140 160 180
Summary Evaluation of fractional norms in the error functional Uniform condition number for dierential operators Fast convergence (106 unknowns in 30 minutes on PC) Outlook Implementation of the adaptive wavelet method
Virtual Telemetry for DDDAS

Craig C. Douglas University of Kentucky and Yale University
douglas-craig@cs.yale.edu http://www.dddas.org
Martin Cole, Yalchin Efendiev, Richard Ewing, Victor Ginting, Chris Johnson, Greg Jones, Raytcho Lazarov, Chad Shannon, and Jenny Simpson Joint project of the University of Kentucky, Texas A&M, and the University of Utah Supported in part by the National Science Foundation (ITR-DDDAS)
Leaky Underground Storage Tanks
UNSATURATED ZONE SATURATED ZONE AQUIFER
NEED TO DEVELOP MONITORING AND CLEAN UP METHODS
Bioremediation Strategies
INJECTION RECOVERY
MACROSCALE
GROWTH MECHANISMS Attachment Detachment Reproduction Adsorption Desorption Filtration Interaction
MICROSCALE FLOW
MESOSCALE
INPUT Substrate Suspended Cells Oxygen
Forward versus Backwards in Time

! Forward
Standard pollutant tracking problem in some sense Start by knowing who is the polluter and where it began
! Backward
Might not know who the polluter is or where it began Determine it correctly !Report it so remediation can be started !Get sued !Spend 5-10 years dealing with lawyers Virtual telemetry solves legal problems until software used in the field by others who are used to lawsuits
Data to Drive Application

! Where is the contaminant?
Use remote sensing data to locate media, update positions, and find new spread directions. !Open water bodies !Buoys and/or submerged stations !Chemical sensors !Visible, near IR, and IR scanning !Radiation detection !Underground !Well sensors
Data to Drive Application (cont.)

! What is the terrain like in that area? What small-scale features are there?
New topography sets give world topography at 30 arcsec (~ 1 km), US at 3 arcsec (~100 m). Better local sources might be available.
! What are the changing weather conditions?

Large-scale data (current analyses or forecasts) used for initial conditions and for updating boundary conditions.
How a DDDAS Might Work (Research Mode)

! Use simulations: first use all available data for past (and eventually current) experimental contaminant tracking to direct collection at crucial times and places. ! Attempt to prove that the prediction of major contaminant behavior can be far more effective than the traditional method of tracking and intuition.
How a DDDAS Might Work (Operational Mode)

! Human or a sensor (possibly on a buoy or in a well) determines a contaminant has been located near position X. ! Need to determine severity and possible expansion. ! Produce a prediction and post it on a public, known web site.
While running model at large-scale over a region Use latest data to locate contaminant boundary.
! Determine communication methods for remediation workers.
How a DDDAS Might Work (Operational Mode; cont.) !Have application

Seek out contaminant classification data Collect recent large-scale data (analyses and forecast) for contaminant model initial and boundary conditions. Initialize and spawn smaller-scale domains, telescoping down to the contaminant area. Ignite contamination in the model at observed location. Simulate the next Y hours/months/years of spread behavior. Dispatch forecast to Web site.
! Easier and cheaper to write a new version of the application using modern programming environments and liberal code re-use.
Savannah River Site

! Difficult topography ! Highly Heterogeneous Soils ! Saturated and Unsaturated Flows ! Reactions with disparate time scale ! Transient/Mixed Boundary Conditions
Need for Simulation

!Develop better understanding of nonlinear behavior
Computational laboratory Experiments Understand sensitivities of parameters Isolate phenomena then combine
!Scale-Up information and understanding

Microscale Laboratory Field
!Obtain bounding calculations !Develop predictive capabilities

Optimization and control
Modeling Process
PHYSICAL PROCESS
PHYSICAL MODEL
MATHEMATICAL MODEL
OUTPUT VISUALIZATION
DISCRETE MODEL
NUMERICAL MODEL
Identification (Inverse) Problem

PHYSICAL PROCESS MEASUREMENTS INPUTS MATHEMATICAL MODEL OUTPUTS OUTPUTS
INPUTS
!Determine suitable mathematical model !Estimate parameters within mathematical model
Large Scale Interactive Applications on Remote Supercomputers

! Model Development and Formulation ! Coupled Codes with Complex Boundary Conditions ! Numerical Discretization and Parallel Algorithm Development ! MPP Code Development ! Field Testing and Production Runs ! User Environments and Visualization Tools
Need for interactive tracking and steering and possibly elimination of human in the loop
Issues of Perturbations from On-Line Data Inputs

! Solve: F(x+ x(t)) = 0 Choice of new approximation for x
Do not need a precise solve of equation at each step !Incomplete solves of a sequence of related models !Effects of perturbations (either data or model) !Convergence questions? Premium on quick approximate direction choices !Lower-rank updates !Continuation methods Interchanges between algorithms and simulations Local boundaries and conditions not known a priori
! Fault-tolerant algorithms
Many Different Components that We Want to Vary Quickly during Project

! We need graphics, solvers, and specialized codes
Matlab inappropriate though it has all of the features we need !Expensive (becoming more so as a result of antitrust suit) and keeps going dense instead of sparse !Matlab compiler does not solve speed issues completely SCIRun has great potential and will be used !Extensible since owners of code are project members !Easy to construct complicated codes using cut and paste plus connect the boxes techniques
SCIRun Methodology
SCIRun and Telemetry

! Streaming paradigm very useful
Add/delete remote clients trivially Well understood for audio and somewhat less for video
! Streaming implementations problematic for DDDAS

Standard formats lossy (filter out frequencies that cannot be heard or seen) Implementations of lossless formats (e.g., Ogg Vorbis) are lossy, too "
SCIRun and Telemetry

! Adding streaming modules
Line points to a stream, either incoming or outgoing, instead of another SCIRun module Multicasting is an alternative, but it is banned by almost all Internet providers and may be outlawed shortly
! Fill in data (interpolation as needed) as data comes in ! Timestamp data so that running backwards in time possible and reverting to an earlier time step (warm restart) is possible (locally or globally)
Why Interpolation Is Required

! Data available only at subset of mesh points ! Data not available on mesh at all
Why Virtual Telemetry?

! We are doing contaminant tracking. Suppose you have real telemetry data to offer us
Would you really admit you are a significant polluter? Do you want the extent of your pollution known?
! Oil simulation experts in academia rarely have access to long term actual data from reservoirs. The data is usually sanitized and only for starting simulations from simple datasets. ! It is hard to beat the cost of virtual telemetry. ! We can vary quantity and quality of data using multiple streams.
Out of Country Experience

! Offered real telemetry from Rio de Janeiro. ! Cannot afford the phone bill to get the data on current grants. ! Internet delivery not an option. ! Personal pickup is an option.
Virtual Telemetry: The Other Code

! We have another code running in real time with small time steps for possibly duration of project
Known to be very accurate from past history matching. Probably not close to state of the art in terms of runtime. Computes everything each time step. Is computing on a very different mesh than our DDDAS code is. Code streams out data to whoever is interested (and authorized). Data lands on the floor when no one is listening to it or can be saved.
Virtual Telemetry Meets Telemetry

! Design SCIRun modules for telemetry so that real telemetry data can be used in the future.
This is a data representation problem Requires a hook to do interpolation correctly into computational mesh !Both virtual and real telemetry need hook !Requires looking at several applications to do right !Volunteers needed to tell us what telemetry data looks like for your DDDAS applications !Greg McRae, ?
Conclusions
! Designing general purpose virtual telemetry system that will be useful by DDDAS community. ! Inexpensive to set up and use. ! Goals:
To make it transparent when switching between real and virtual telemetry. Allow us to investigate easily DDDAS aspects that are not in static dataset formulations (e.g., how to analyze situation like not knowing boundary conditions or where the local boundary is a priori).
! Dissemination through http://www.dddas.org

Links to other projects there, too.
Fast interpolation techniques and meshless methods

Csaba Gspr
Szchenyi Istvn University, Department of Mathematics Gyr, Hungary
gasparcs@mail.sze.hu
C Gaspar Advances in Numerical Algorithms, Graz, 2003 1
Motivations
surface approximation completion of incomplete data sets construction of meshfree methods
fC :=
f N + fW + f S + f E 4
f = 0
C Gaspar
Advances in Numerical Algorithms, Graz, 2003
The scattered data interpolation problem in R

Problem : Given the points x1,..., x N (irregularly spaced) in R 2 and the values u1,..., u N . Find a (sufficiently smooth) function u : R 2 R such that u ( xk ) = u k (k = 1,..., N )
Vectorial problems: some additional conditions are required div u = 0 rot u = 0
C Gaspar
Method of radial basis functions

The form of the interpolation function:
u ( x) := j j ( x x j ) + a j p j ( x)
j =1 j =1 N m
where the basis functions depend on ||x|| only. The coefficients are a priori unknown and can be determined by solving the interpolation equations:
u ( xk ) = u k (k = 1,2,..., N )
and the orthogonality conditions:

j =1
ak pk ( x j ) = 0 (k = 1,2,..., m)
C Gaspar
Some popular RBFs

Globally supported RBFs
Method of Multiquadrics (Hardy): Thin Plate Splines (Duchon): (biharmonic fundamental solution)
j (r ) := r 2 + c 2 j
j ( r ) := r 2 log r
Gaussian functions:
Traditional DRM (Partridge, Brebbia):
j (r ) := e
c 2r 2 j
j ( r ) := 1 + r
Compactly supported RBFs (Wendland):
j ( r ) := (1 c j r ) 2 + j ( r ) := (1 c j r ) 4 (4c j r + 1) + .........
C Gaspar
Why do we like the RBFs so much?

Main advantages: excellent approximation properties easy to implement
But what is the price to pay?

We have to face that: RBFs generally lead to a system with large, dense and ill-conditioned matrix memory requirement: O(N 2) computational cost: O(N 3)
C Gaspar
Remedies
domain decomposition preconditioning fast multipole evaluation methods compactly supported RBFs multi-level techniques direct multi-elliptic interpolation
C Gaspar
Methods based on iterated elliptic operators

Observation: Thin plate splines are fundamental solutions of the biharmonic operator Idea: Let us define the interpolation function in such a form: Lu = 0 in \ {x1, x2 ,..., x N } u ( xk ) := uk (interpolation conditions) u | := 0 (boundary conditions) Second-order operators: unusable! Fourth-order operators: biharmonic operator; bi-Helmholtz-operator; Laplace-Helmholtzoperator Higher order operators: triharmonic operator; tetraharmonic operator; iterated Helmholtzoperator, etc.
Harmonic interpolation: a counterexample

:= (0,1) (0,1), f 0 ( x, y ) := sin x sin y
u = 0 in \ {( x1, y1 ), ( x2 . y2 ),..., ( x N , y N )} u | := 0 u ( xk , yk ) := f 0 ( xk , yk ) (k = 1,2,..., N ) Numerical features: ill-posed problem poor interpolation surface logarithmic singularities at the interpolation points
C Gaspar
Biharmonic interpolation
2 Classical formulation: For a given function u0 H 0 () , find a function 2 u H 0 () such that u = 0 in \ {x1, x2 ,..., x N }
u ( xk ) = u0 ( xk ) (k = 1,2,..., N ) Equivalent formulations: Let us introduce the closed subspace: 2 W := {w H 0 () :w( x1 ) = = w( x N ) = 0}

2 Direct problem: For a given function u0 H 0 () , find a function v W such that (u0 + v ) = 0 in \ { x1 , x2 ,..., x N } 2 Variational problem: For a given function u0 H 0 (), find a function v W such that (u0 + v), w L () = 0 w W
2
Theorem: The direct and the variational problems are equivalent and have a unique solution in W.
Representation of the biharmonic interpolation

Theorem: u is uniquely represented in the form: u ( x) = w( x) + j ( x x j )
j =1 N
where w is an everywhere biharmonic function, and is the biharmonic fundamental solution:

(r ) = r 2 log r
This is an RBF-like interpolation... ...but the coefficients can be calculated from: without solving any system of equations!
C Gaspar Advances in Numerical Algorithms, Graz, 2003
u ( x) = j ( x x j )
j =1
11
Approximation
Separation distance: h:= sup
x y{ x1 ,..., x N }
inf
|| x y ||
Theorem: The following estimations hold:
|| u u0 || L2 () C1h 2 || u0 || || u u0 ||
1 H 0 ()
2 H 0 ()
C2 h || u0 ||
2 H 0 ()
C Gaspar
12
Bi-Helmholtz interpolation
2 Classical formulation: For a given function u0 H 0 () , find a function 2 u H 0 () such that ( c 2 I ) 2 u = 0 in \ {x1, x2 ,..., x N }
u ( xk ) = u0 ( xk ) (k = 1,2,..., N )
2 Direct problem: For a given function u0 H 0 () , find a function v W such that ( c 2 I ) 2 (u0 + v ) = 0N in \ { x1 , x2 ,..., x N } (r ) = rK1 (cr ) ) Representation formula: u ( x) = j ( x x j 2, j =1 Variational problem: For a given function u0 H 0 () , find a function N v W such that be (u0 + v), w = ( = Coefficients can computed directly: 0 uwx) W j ( x x j ) L ()
2
j =1
Now the fundamental solution is rapidly decreasing ("almost compactly supported").

The direct and the variational problems are equivalent, and have a unique solution in W. The representation and approximation theorems are still valid.
Further generalizations
The use of even higher order differential operators The use of different operators simultaneously
In each case, the representation automatically produces an RBF-like interpolation based on the fundamental solution of the applied operator.
Triharmonic operator: u = 0 Mixed Laplace-Helmholtz operators: ( c 2 I ) f = 0
( c 2 I ) f = 0 ( c 2 I ) 2 f = 0
C Gaspar
14
The main computational problem: How to solve the multi-elliptic interpolation equation?!
C Gaspar
15
Numerical techniques
Main features: No algebraic interpolation equation has to be solved. Instead, the problem is converted to an auxiliary differential equation. The domain is not predefined The boundary conditions are not strictly prescribed Numerical solution techniques: Traditional FDM or FEM ? FFT ? MGR ? quadtree-based multigrid techniques
C Gaspar
16
Quadtree cell systems

Non-uniform cell subdivision generated by a finite set of points Schemes based on finite volumes Multigrid solution technique Computational cost can be reduced to O(N log N) Demo application
C Gaspar
17
Biharmonic interpolation on a QT cell system: an example

: = (0,1) (0,1), f 0 ( x , y ): = sin 2 x sin 2 y
The QT-cell system

Relative L2 -errors (%): N \ lmax 7 8 32 17.314 7.759 64 4.159 4.026 128 0.770 1.332 256 0.213 0.154 512 0.095 0.062
The interpolation function

9 8.305 4.529 0.691 0.143 0.049 10 9.202 2.692 0.581 0.148 0.041
C Gaspar
18
Bi-Helmholtz interpolation: an example

:= (0,1) (0,1), 1 1 u0 ( x, y ) := sin 2 ( 2( x )) sin 2 (2( y )) 4 4
Relative L2 -errors (N := 200): c 10 15 20 25 30 40 50 Error 1.55 1.63 2.31 4.24 7.20 15.22 25.01
c = 10 c = 50
C Gaspar
19
Construction of meshless methods. General approaches

Original problem: Lu = f (in ) + boundary conditions
M
Method 1. The solution u is directly interpolated: u ( x) := j ( x x j )

M j =1 j =1
j L ( xk x j ) = f k
M
Method 2. First, the function f is interpolated: f ( x) := j ( x x j )

j =1
u ( x) := j ( xk x j ) with
j =1
L =
C Gaspar
20
Construction of meshless methods by direct multi-elliptic interpolation

Model problem: u = f (in ) + boundary conditions
Method 1. The solution u is directly interpolated by tetraharmonic interpolation: 4u = 0 in \ {x1, x2 ,..., x N }
u ( xk ) = f k
(k = 1,2,..., M )
u ( xk ) = uk (k = M + 1,..., N ) ...A bit uncomfortable...
4 W := {w H 0 () : u ( x1 ) = ... = u ( xM ) = 0,
u ( xM +1 ) = ... = u ( x N ) = 0}
Method 2. First, the function f is interpolated by biharmonic interpolation. Then, using the same QT cell system, the model problem is solved, without boundary conditions.
This procedure mimics the computation of Newtonian potential Reduces the domain problem to a boundary problem
How to solve the boundary problem?

Traditional tool: the Boundary Element Method
Advantages: reduction of dimension reduction of mesh Drawbacks: large, dense and ill-conditioned matrices
Direct Multi-Elliptic Boundary Interpolation
C Gaspar
22
Boundary interpolation 1. The basic idea

The use of the mixed Laplace-Helmholtz operator: ( c 2 I )u = 0 1 Fundamental solution: ( x) = ( K 0 (c || x ||) + log(c || x ||)) 2 2 c Model problem: Laplace equation with Dirichlet boundary condition
The interpolation points are located at the boundary only, and...

'Harmonic' interpolation: u = 0 Biharmonic interpolation: u = 0 'Quasi-harmonic' interpolation:
( c 2 I )u = 0
Boundary interpolation 2. Application to the Laplace equation

The original problem: 1 c u* = 0, u* | = u0 u | = u0 , u | = v0 n The associated boundary interpolation problem: u u = 0, 2
Key issue: the proper choice of the scaling constant c Theorem 1: || u u* ||L2 () C (|| u0 || 1/ 2 + || v0 || 1/ 2 ) H () H ( ) c C (|| u0 || 1 + || v0 || ) H ( ) L2 ( ) c
With more regular data: Theorem 2: || u u* ||

H 1 ()
C Gaspar
24
Dirichlet boundary condition Examples
c = 500 c = 250 c=0
(1.718%) (0.149%) (26.102%)
C Gaspar
25
Neumann boundary condition The naive approach

u * | = v0 n The associated boundary interpolation problem: The original problem: u* = 0, ( c 2 I ) 2 u = 0, u ( xk ) = v0 ( xk ) n in \ {x1,..., x N } (k = 1,2,..., N )
3 u H 0 ()
The method fails...

u approximates v0 but fails to approximate u* !
C Gaspar
26
Neumann boundary condition An improved approach

Idea: Locate additional Dirichlet points along Neumann boundary in outward normal direction; Define Dirichlet data to enforce Neumann conditions ~ u := u + d v0 or: u ~ u := u + d (v0 ) n This procedure mimics the method of fundamental solutions
Neumann boundary condition Examples
d = 0.025 d = 0.050 d = 0.100
C Gaspar
28
Boundary interpolation 3. Application to the biharmonic equation

The original problem: u* = 0, u* | = u0 , The associated boundary interpolation problem:
u 2 u = 0, u | = u0 , c 1 u = v0 , n 2u n
2
u * = v0 n
= w0
Theorem 3: || u u* ||
H 1 ()
C (|| u0 || 3/ 2 + || v0 || 1/ 2 + || w0 || 1/ 2 ) H H () H () () c
Further applications: 2D Stokes flow
C Gaspar
29
Vectorial interpolation problems

Problem : Given the points x1,..., x N in R 2 and the values u1,..., u N R 2 . Find a function u : R 2 R 2 such that u( xk ) = u k
Componentwise interpolation works well...
(k = 1,..., N )
... but often additional conditions are required: rot u = 0, div u = 0

Use the potential/stream function approach D1 ( xk ) = vk , D2 ( xk ) = uk
RBF-method fails
C Gaspar Advances in Numerical Algorithms, Graz, 2003
or
(k = 1,2,..., N )
30
Tri-Helmholtz-interpolation for vectorial problems ( x, y ) = sin 2x sin 2y

2 2 Classical formulation: For a given function u = (u , v) H () H () 3 find a function H 0 () such that
D1 ( xk ) = v( xk ), D2 ( xk ) = u ( xk ) (k = 1,2,..., N )
3 Variational formulation: Define W := {w H 0 () : grad w( xk ) = 0, k = 1,..., N } 3 3 For a given function 0 H 0 () find a function H 0 () such that
grad( c 2 I )( 0 + ), grad( c 2 I ) w c 2 ( c 2 I )( 0 + ), ( c 2 I ) w
L2 ( )
L2 ( )
= 0 w W
Theorem: The variational problem has a unique solution in W. The vectorial interpolation function: (grad ) is divergence-free.
Representation and RBF-like solution

Representation formula: If c > 0 then is represented in the form: ( x) = j D1 ( x x j ) + j D2 ( x x j )
j =1 j =1 N N
where is the fundamental solution of the operator ( c I ) : 1 (r ) = (r 2c 2 K 0 (rc) + 2rcK1 (rc)) 16c 4 Vectorial tri-Helmholtz Exact Componentwise TPS
2 3
The theory assures existence and uniqueness also for the RBF-like interpolation: interpolation
j =1 N
j D11 ( xk x j ) + j D12 ( xk x j ) = vk
j =1 N
j =1
j D12 ( xk x j ) + j D22 ( xk x j ) = uk
j =1
( k = 1,2,..., N )
32
C Gaspar
Summary and conclusions

Interpolation based on iterated elliptic equations Boundary interpolation by singularly perturbed equations Special RBF- and RBF-like methods are automatically generated
No explicit use of RBFs is needed No large and dense algebraic equations are to be solved Quadtree-based multigrid methods are used
C Gaspar
33
Thank you for your attention!
C Gaspar
34
A two term recursive calculation of coarse matrices

Gundolf Haase and Stefan Reitzinger
ghaase@numa.uni-linz.ac.at Johannes Kepler University Linz, Austria Institute of Computational Mathematics SFB F013 Numerical and Symbolic Scientic Computing
Workshop on Advances in Numerical Algorithms Sept. 10-13, 2003 at the University of Graz
Johannes Kepler University Linz

Institute of Computational Mathematics
Considered Problem Classes

(linear) 2. order problem. Find u such that Lu(x) = f (x) lu(x) = g(x)
formulation
x x
Poisson equation (temperature) Lame equations (deection) Maxwells equations (magnetic eld)
Variational
Find u V :
a(u, v) = F, v
FVM, FIT
v V
FEM, FDM
Matrix Kh is sparse, positive denite (symmetric, large)
Solve
K h uh = f h
uh RNh
non-linear and time dependent problems, inverse problems.

Fast solver: (Classical) Multigrid V(F ,B )-cycle: MG(ul , fl ,l)

if l = COARSELEV EL then ul = (Kl)1fl is solved with a direct solver else Smooth F -times with Klul = fl Calculate defect dl = fl Klul Restrict defect dl+1 = PTdl l Set wl+1 0 MG(wl+1 ,dl+1 ,l+1) Interpolate the correction wl = Plwl+1 Update solution ul = ul + wl Smooth B -times with Klul = fl end if


Main Steps in AMG

(i) Coarsening procedure :
h h h = C F .

(ii) Interpolation weights : P = {ij }ih, jh : RH Rh ,

C
( Ps ) .
(iii) Coarse grid matrix:

P
= P K P
(or K
=
s=1
PsT Ks Ps)
(iv) (i)-(iii) recursively. (v) Apply mg-procedure.

Magnetostatic Valve
curl u curl v + uv dx
= 104, = 108
100 10 1 0.1 0.01 rel. residuum 0.001 0.0001 1e-05 1e-06 1e-07 1e-08 1e-09 0 50 100 150 200 250 300 350 number of iterations standard AMG IC AMG with point GS smoother edge element AMG
Nh 8714 65219 504246
it 14 28 63
setup solver 0.53 2.49 4.17 44.01 34.49 792.49

Parallel Results - Inverse Source Reconstruction

grad u grad v dx
12 matrix setup solver 1iter 10
= humain head = 108 , SGI ORIGIN 200

speedup 12 matrix setup solver 1iter 10
speedup
0 0 2 4 6 processors 8 10 12
Tetrahedra, Nh = 118299 Hexahedra, Nh = 325384
0 0 2 4 6 processors 8 10 12

Simultaneous Right-Hand Sides

Speedup relative to new code
grad u grad v dx
3.5 "LINUX_rs.data" "SGI_rs.data" "ORIGIN_rs.data" "LINUX_Hand_rs.data" "SGI_Hand_rs.data" "ORIGIN_Hand_rs.data"
= (0, 1)2 = 10
5.5 5 4.5 4 Speedup 3.5 3 2.5 2 1.5 1 0 5 10 15 20 25 DOFNUMBER 30 35 40 "LINUX.data" "SGI.data" "ORIGIN.data" "LINUX_Hand.data" "SGI_Hand.data" "ORIGIN_Hand.data"
2.5
8
2
Total speedup in simultaneous AMG
1.5
1 0 5 10 15 20 25 DOFNUMBER 30 35 40
Simultaneous right-hand sides Total performance gain

Classical AMG and PEBBLES

AMG is stable for M-matrices, list of parameters , = Coarsening: determine locally the eigenvector of the largest eigenvalue Local interpolation on ne patches best interpolation coarse grid matrices get denser and denser in AMG = Ensure local support of coarse base functions Problems in preserving of the constant = New idea ! AMGe [LLNL] Two term recursion [Linz]

Coarsening in AMG
i,T (C , F ) Coarse({Sh }, C , F )
while C F h do i,T k Pick({Sh }, h \ (C F ), k,T k,T if |Sh | + |Sh F | = 0 then F F (h \ C ) else C C {k} k,T F F (Sh \ C ) end if end while

Patch based AMG

Finte elements ne grid patch Fine grid patches coarse grid patch Determine coarse nodes from EV investigations of ne patches P overlapping patches s:
P P
43 49
36
42
K =
s=1
Ks =
s=1 (r) s
1 FEM (r) Kr W
29
32
35
35
22
25
26
27
28
Recursive coarsening ? Local support of coarse grid functions ? Preserving of the constant (null space of K) on coarse grid ?
15
18
19
20
21
patch 26 patch 12 patch 18 patch 20
10
11
12
14
10

Standard support of coarse base functions
11

Long couplings in standard approach
H i

j j
H
12

Decouplings of patches
H i

j j
H
13

Main Idea for New Approach

0. Fine f.e.-basis h = h =
r: (r)
(r)
.
h
1. Determine classical interpolation P = { i} h,ih .

C
2. Classical coarse basis H = 3. Redistribute i h New basis H :=

(r)
H = i
i h =
h h r: (r)
i h
(r)
.
h iC
s.t. local support can be achieved. H = i i h

(r) (r)
.
h iC
(r) H (r) i
Constant (null space) is preserved, i.e.,

h iC
H (x) = i
(r) (r)
h iC
H (x) i
4. Coarse matrix
[i,j] KH
=
s=1 (r) H ,k (r) s
i K k kj
(r)
14

Redistribution of shape functions I
Distribution of shape function h (r) to 2 ne patches and 6 coarse patches ( i (and it will contribute to base functions in H ).
15

Redistribution of shape functions II
Both shape functions h (r) are distributed to the same 4 coarse patches ( i (and it will contribute to base functions in H ).
i i

16

(r) Calculation of interpolation weights i

h i for C (r) h := i + s li for F s h : (r) H
C s
i li
(r)
(r)
h i C ,
(r)
= i/
j .
j: (r) H j
weights:
(r) li (r) lt
The weights li are chosen such (r) li = 1 , e.g., that

i: (r) H i (r) li (r) lt
s l
= li/ (li + lt ) = lt / (li + lt )
17

Modfying interpolation weights per element - algorithm

h Calculate i , set i i h, i C , (r) h h for all s C do {coarse nodes} h H for all F s do {ne nodes in coarse patch s} for all r {q : (q) h \ H } do {elements of ne patch - outside coarse patch s} s [k] h SF := {k F : xh (r) } {ne patches sharing element (r)} h h SC := {i C : i k , k SF } {coarse patches using those ne patches} d jSC j for all i SC do (r) i i/d (r) (r) (r) i i +si end for end for end for end for (r)
18

New shape functions - new base functions

H = i h H i
(r)
h kF : h H (r) h (r) i k k
i h
(r)
(r)
: shape functions on 0 (elements) (r) i : new shape functions on 1 = h (ne patches)

h kF
H : new base function on 2 = H (coarse patches) i overlapping patches weighting of shape function on 0 overlapping patches new shape functions are discontinuous on coarse patch
19

New restriction / interpolation

Need to divide ne grid values among elements: uh =
(r) : (r)
(r)
uh
(r) (r)
Require
(r) : (r)
(r)
= 1 , e.g.,
(r)
rK
uH = i
(r) :i (r) (r) P P
(r) i
(r)
uh
1 (r) W
T
=
s=1 (r) :i (r) (r) (r) H s
(r)
(r)
uh W1(r)
=
s=1
uH s
=
s=1
PsT
uh s
= P uh
Accumulation required!
20

Repetition: Non-overlapping Domain Decomposition

accumulated distributed
us = A s u Ms = AsMAT s
r =
s=1 P
AT r s s AT KFEM As s s
s=1
K =
21

Repetition: Matrix patterns and its applications

The following operations can be performed in parallel (without communication): f = Ku Matrix M fullls the pattern condition: i, j : u = Mw f = MT r KH = MT K M [i] [j] = M[i,j] = 0
Theorems/Proofs [Haase]
22

Repetition: Parallel Multigrid : pmg(K, u, f, )

if == 1 then
P
Solve
s=1
AT KAs u = f s
else u Smooth(K, u, f, ) dfKu dH PT d wH 0 pmg (KH , wH , dH , 1) w P wH uu+w u Smooth(K, u, f, ) end if
23

(Parallel) Multigrid with new interpolation P: pmg-e(K, u, f, )

if == 1 then
P
solve
s=1
Ks u = f
else u Smooth(K, u, f, ) dfKu

P
d
s=1
ds ; dH P ds
wH 0 pmg-e (KH , wH , dH , 1) w Pw
H P
; w
s=1
ws
uu+w u Smooth(K, u, f, ) end if 24

Summary
Patch based AMG real automatic coarsening. Moving of shape functions results in intergrid transfer weights
(r) i
dier wrt. elements : P,
local support for coarse base functions, preserving of constant (null space), Galerkin aproach for coarse matrix K H with interpolation P . Slightly modied MG-algorithm necessary (P). Open : Relation: P K h P T
quality ?
P Kh P T
Implementation in progress.
25

R(20)-matrix
10
-0
10
-3
10
-6
10
-9
10
-12
10 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
-15
A new discretization concept in control constrained pde control and its numerical realization M. Hinze Technische Universitt Dresden a Institut f r Numerische Mathematik u hinze@math.tu-dresden.de www.math.tu-dresden.de/hinze
Technische Universitt Dresden a
Sonderforschungsbereich 609
Graz, September 10, 2003
Outline: Abstract setting Error estimate Numerical realization Numerical examples Conclusions
Outline: Abstract setting Error estimate Numerical realization Numerical examples Conclusions
There is no need to discretize controls
Discretization
Concept
Abstract setting S : U Y , U U , J (u) := J (Su, u), Y, U Hilbert spaces.

ad
u = argminuU J (y, u) = J (u) s.t. u U , where y = Su.

ad
Discretization
Concept

ad

ad
Sh : U Yh Y u = argminuU J (Shu, u) =: Jh(u) s.t. u U . h

ad
Discretization
Concept

ad

ad

ad
No discretization of the control space!!
Examples:
Discretization
Concept

ad

ad

ad
Examples: S = ()1, Sh FE solution operator, U = L2(), U

1 Y = H0 ().
ad
= {u U ; a u b a.e. in },
Discretization
Concept

ad

ad

ad
Examples: S = ()1, Sh FE solution operator, U = L2(), U

1 Y = H0 ().
ad
= {u U ; a u b a.e. in },
S = Stokes1, Sh Taylor-Hood FE semi discretization, U = L2([0, T ], L2()n), U = {u

ad
1 U ; a u b a.e. in (0, T ) }, Y = {y L2([0, T ], V ), yt L2([0, T ], V )}, V := {H0 ()n,
divy = 0}.
Discretization
Concept
Abstract setting, error estimate J (y, u) =

1 2
yz
2 Z
2 U.
Necessary optimality conditions: For v U (J (u), v u)U 0, where
ad
J (v) = v + S (Sv z).
Discretization
Concept

1 2
yz
2 Z
2 U.
Necessary optimality conditions: For v U (J (u), v u)U 0, where (Jh(u ), v u )U 0, where h h
ad
J (v) = v + S (Sv z).

Jh(v) = v + Sh(Shv z).
Discretization
Concept

1 2
yz
2 Z
2 U.
ad
J (v) = v + S (Sv z).

Characterization of continuous and discrete optimal controls: u = P U

ad
1 S (Su z)
Discretization
Concept

1 2
yz
2 Z
2 U.
ad
J (v) = v + S (Sv z).


ad
1 S (Su z) 1 Sh(Shu z) h
u = P U h
ad
Discretization
Concept

1 2
yz
2 Z
2 U.
ad
J (v) = v + S (Sv z).


ad
u = P U h
ad
Error estimate: Assume that Sh, Sh satisfy (S Sh)z U
Ch2 z
U
Z,
and
U.
(S S ShSh)u
Ch2 u
Discretization
Concept

1 2
yz
2 Z
2 U.
ad
J (v) = v + S (Sv z).


ad
u = P U h
ad
Error estimate: Assume that Sh, Sh satisfy (S Sh)z U
Ch2 z
U
Z,
and
U.
(S S ShSh)u
Ch2 u
Then u u h
U
Ch2{ u
+ z
Z }.
Discretization
Concept
Abstract setting, proof of error estimate Proof: Add necessary optimality conditions and utilize u, uh as test functions; u u h
2 U
Discretization
Concept
2 U
(S (Su z) Sh(Shu z), u u)U h h
Discretization
Concept
2 U
(S (Su z) Sh(Shu z), u u)U h h ((S S ShSh)u + (Sh S )z + ShSh(u u ), u u)U h h Sh (u u ) 2 h Y
Discretization
Concept
2 U
(S (Su z) Sh(Shu z), u u)U h h ((S S ShSh)u + (Sh S )z + ShSh(u u ), u u)U h h ((S S ShSh)u + (Sh S )z, u h Sh (u u ) 2 h Y u ) U
Discretization
Concept
2 U
Ch2{ u
+ z
Z}
u u h
U.
Discretization
Concept
2 U
Ch2{ u Implications:
+ z
Z}
u u h
U.
For nite element approximations of elliptic problems
Discretization
Concept
2 U
+ z
Z}
u u h
U.
For nite element approximations of elliptic problems u u h C

(S Sh)z + (S Sh )Su
+C
h h
2
1
h2| ln h| 2
3 2
n=1 for n=2 . n=3
Discretization
Concept
2 U
+ z
Z}
u u h
U.

(S Sh)z + (S Sh )Su
+C
h h
2
1
h2| ln h| 2
3 2
n=1 for n=2 . n=3
For = 0 we obtain some H 1 estimate:
Discretization
Concept
2 U
+ z
Z}
u u h
U.

(S Sh)z + (S Sh )Su
+C
h h
2
1
h2| ln h| 2
3 2
n=1 for n=2 . n=3
For = 0 we obtain some H 1 estimate: S(u u ) h

2 Y
Ch2 ( u
+ z
Z)
u u h
+ u u h
2 U
Discretization
Concept
Sketch
Imagine box constraints a u b and set p(h) := S(h)(S(h)u(h) z)
P(p/alpha) P(p_h/alpha)
Continuous setting
Active set
FE Discretization
Active set
Discretization
Concept
Sketch
Continuous setting
Active set
FE Discretization
Active set
Decoupling of FE grid and active set.
Discretization
Concept
Sketch
Continuous setting
Active set
FE Discretization
Active set
Decoupling of FE grid and active set. T active set contains at most 2 connected components.
Discretization
Concept
Sketch
Continuous setting
Active set
FE Discretization
Active set
Decoupling of FE grid and active set. T active set contains at most 2 connected components. Appropriate L estimates allow estimation of continuous and discrete active sets.
Discretization
Concept
Sketch
Continuous setting
Active set
FE Discretization
Active set
Decoupling of FE grid and active set. T active set contains at most 2 connected components. Appropriate L estimates allow estimation of continuous and discrete active sets. However: Is it possible to compute u numerically with reasonable numerical overhead? h
Discretization
Numerical realization
Projected Gradient Method
Discretization
u0 U , u0 := P[a,b]u0, k = 0, tol > 0.

dk := Jh(uk ) = uk + Sh(Shuk z)
= argmins0Jh(uk sdk ) uk+1 := P[a,b](uk dk ) If |uk+1 uk | tol, STOP k = k+1, goto 2.
Discretization
u0 U , u0 := P[a,b]u0, k = 0, tol > 0.

1 uk dk = uk uk + Sh(Shuk z) = Sh(Shuk z) = .
Discretization
u0 U , u0 := P[a,b]u0, k = 0, tol > 0.

For this choice of computation of P[a,b](uk dk ) only requires evalutation of the nite element
function Sh(Shuk z) on the computational grid.
Discretization
u0 U , u0 := P[a,b]u0, k = 0, tol > 0.

function Sh(Shuk z) on the computational grid. Drawback: Convergence can be guaranteed only for > ShSh .
Discretization
u0 U , u0 := P[a,b]u0, k = 0, tol > 0.

function Sh(Shuk z) on the computational grid. Drawback: Convergence can be guaranteed only for > ShSh .
Remedy (in the spirit of local convergence): primal-dual active set strategy, semi-smooth Newton method.
Discretization
Primal dual active set strategy Solve

( + ShSh)u + = Shz = r
(u, ; u) := max( + (u b), 0) + min( + (u a), 0) =
Discretization
Primal dual active set strategy Solve

( + ShSh)u + = Shz = r
(u, ; u) := max( + (u b), 0) + min( + (u a), 0) = Primal dual active set strategy: Initialize u0 = 0, 0 = r; set l = 1, Loop l
Aa := {l1 + (ul1 a) < 0} (= {r ShShul1 a < 0}, if = ), l Ab := {l1 + (ul1 b) > 0} (= {r S Shul1 b > 0}, if = ), l h
> 0 small.
Il := \ (Aa Ab). l l l 2, Aa = Aa , Ab = Ab , or (ul1, l1) l1 : u = ul1, = l , RETURN. l l1 l l1 Otherwise ul = a on Aa, ul = b on Ab, l = 0 on Il l l Solve for ul |I , l | a b
l
( + ShSh)ul + l = r
Al Al
l := l + 1.
Discretization
Numerical example
Consider min J (u) :=

y,u
1 2
0
|y z|2dx +
2
0
u2dx,
where u U := {v L2(0, 1); u ub} and y and u are related via the second order boundary
ad
value problem y = Su y + y = u + e in (0, 1),

2
y(0) = y(1) = 0.
x Let ua = , e := 2 + x2 xmin ub, x and z(x) 2. Then the unique optimal solution
of the control problem is given by y(x) = x x, and u(x) = min

2
x2 x
, ub ,
where the associated Lagrange multiplier of the equality constraint is given by p(x) = x2 x. Numerical discretization with linear nite elements.
Discretization
Numerical example
Comparison of discrete approaches

2.5 adjoint/alpha (discrete) control (discrete) adjoint/alpha (exact) control (exact) control (standard fem) 2
1.5
0.5
0.25
Discrete active set
0.5
Discrete active set (standard fem) Active set
0.75
Discretization
Numerical example
Comparison of discrete approaches

2.5 adjoint/alpha (discrete) control (discrete) adjoint/alpha (exact) control (exact) control (standard fem) 2
1.4
adjoint/alpha (discrete) control (discrete) adjoint/alpha (exact) control (exact) control (standard fem)
1.2
1
1.5
0.8
1
0.6
0.5
0.4
0.2
0 0 0.25
Discrete active set
0.5
Discrete active set (standard fem) Active set
0.75
0.117 0.156
0.25
Green: Standard approach (projection of nodal values onto admissible set). Red: Modied approach. Black: Exact solution.
Discretization
Numerical example II

y,u
1 2
0
|y z|2dx +
2
0
u2dx,
where u U := {v L2(); 0.2 u 0.2}, denotes the unit circle and y and u are related
ad
via Laplaces equation y = Su y = u in , y| = 0. Here z(x) = (1 |x|2)x1 and = 1.d 1.
Discretization

y,u
1 2
0
|y z|2dx +
2
0
u2dx,
where u U := {v L2(); 0.2 u 0.2}, denotes the unit circle and y and u are related
ad
via Laplaces equation y = Su y = u in , y| = 0. Here z(x) = (1 |x|2)x1 and = 1.d 1. Numerical discretization with linear nite elements. Given two grid sizes h1 = h2, the experimental order of convergence for some error functional E is given by EOC := ln E(h1) ln E(h2) ln h1 ln h2 .
Numerics by Ulrich Matthes.
Discretization
Experimental order of convergence Results for = 0.1. P refers to classical approach with piecewise linear, continuous controls. | |1
P EAS AS P
EuL2
E u
Eusem
E uH 1
EAS
AS -
1/1 2.18e-1 2.00e-1 8.66e-1 8.93e-1 1/2 5.54e-2 7.75e-2 4.78e-1 4.81e-1 1/4 1.16e-2 2.30e-2 2.21e-1 2.22e-1 1/8 3.02e-3 5.79e-3 1.15e-1 1.15e-1 1/16 7.66e-4 1.47e-3 6.09e-2 6.09e-2 1/32 1.93e-4 3.67e-4 2.97e-2 2.97e-2 1/64 4.82e-5 9.38e-5 1.41e-2 1.41e-2 1/128 1.17e-5 2.37e-5 6.40e-3 6.40e-3
1.97 2.25 1.94 1.98 1.99 2.00 2.04
- 5.05e-1
- 5.11e-1
1.37 0.86 1.75 1.11 1.99 0.94 1.98 0.92 2.00 1.03 1.97 1.07 1.98 1.14
0.89 5.05e-1 0.00 3.38e-1 0.60 1.12 5.05e-1 0.00 1.25e-1 1.43 0.95 2.60e-1 0.96 2.92e-2 2.10 0.92 1.16e-1 1.16 7.30e-3 2.00 1.03 4.98e-2 1.22 1.81e-3 2.01 1.07 1.88e-2 1.41 4.08e-4 2.15 1.14 6.98e-3 1.43 8.51e-5 2.26
Discretization
Experimental order of convergence Results for = 0.1. P refers to classical approach with piecewise linear, continuous controls. | |1
P EAS AS P
EuL2
E u
Eusem
E uH 1
EAS
AS -
1/1 2.18e-1 2.00e-1 8.66e-1 8.93e-1 1/2 5.54e-2 7.75e-2 4.78e-1 4.81e-1 1/4 1.16e-2 2.30e-2 2.21e-1 2.22e-1 1/8 3.02e-3 5.79e-3 1.15e-1 1.15e-1 1/16 7.66e-4 1.47e-3 6.09e-2 6.09e-2 1/32 1.93e-4 3.67e-4 2.97e-2 2.97e-2 1/64 4.82e-5 9.38e-5 1.41e-2 1.41e-2 1/128 1.17e-5 2.37e-5 6.40e-3 6.40e-3
1.97 2.25 1.94 1.98 1.99 2.00 2.04
- 5.05e-1
- 5.11e-1
1.37 0.86 1.75 1.11 1.99 0.94 1.98 0.92 2.00 1.03 1.97 1.07 1.98 1.14
0.89 5.05e-1 0.00 3.38e-1 0.60 1.12 5.05e-1 0.00 1.25e-1 1.43 0.95 2.60e-1 0.96 2.92e-2 2.10 0.92 1.16e-1 1.16 7.30e-3 2.00 1.03 4.98e-2 1.22 1.81e-3 2.01 1.07 1.88e-2 1.41 4.08e-4 2.15 1.14 6.98e-3 1.43 8.51e-5 2.26
Reference solution on grid with h = 1/256 Box constraints = 1/ in projected gradient method. Only 10 iterations on every renement level.
Discretization
Comparison of discrete approaches = 0.1, h = 1/8:

1
0.8
0.6
0.4
0.2
0.2
0.4
0.6
0.8
1 1
0.8
0.6
0.4
0.2
0.2
0.4
0.6
0.8
Red: Standard approach (projection of nodal values onto admissible set). Green and black: Modied approach. Cyan: Exact solution.
Discretization
Comparison of discrete approaches = 0.01, h = 1/4:

1
0.8
0.6
0.4
0.2
0.2
0.4
0.6
0.8
1 1
0.8
0.6
0.4
0.2
0.2
0.4
0.6
0.8
Red: Standard approach (projection of nodal values onto admissible set). Green and black: Modied approach. Cyan: Exact solution.
Conclusions and future work
New discretization concept for control constrained optimal control.
New discretization concept for control constrained optimal control. No explicit discretization of the control space.
New discretization concept for control constrained optimal control. No explicit discretization of the control space. Decoupling of active sets and FE grid.
New discretization concept for control constrained optimal control. No explicit discretization of the control space. Decoupling of active sets and FE grid. There is no need to discretize controls
New discretization concept for control constrained optimal control. No explicit discretization of the control space. Decoupling of active sets and FE grid. There is no need to discretize controls or if You do so: mimic on the discrete level the coupling between control and adjoint variables (First optimize, then discretize!!!).
Further informations on the topics
http://www.math.tu-dresden.de/hinze
Papers and movies are available
Material Parameter Identication in Piezoelectricity and Magnetics

Barbara Kaltenbacher
joint work with: Manfred Kaltenbacher, Reinhard Lerch, Reinhard Simkovics, Department of Sensor Technology, University of Erlangen Stefan Reitzinger, University of Linz
supported by the German Science Foundation DFG
Overview
Piezoelectricity Piezoelectric PDEs Material parameter identication Numerical results Temperature dependence, nonlinearity, hysteresis Nonlinear magnetics Instability Regularization by multilevel discretization Multigrid methods for unstable problems Conclusions and outlook
Piezoelectric Transducer
direct eect: indirect eect: applications: - ultrasound transducer (medical imaging and therapy) - force and accelleration sensor - actuator (injection valve in Common-Rail diesel eng.) - SAW (surface-acoustic-wave) sensor -...
Piezoelectric Eect
= cE S eT E D = eS + S E
S E d = (xx, yy , zz , yz , xz , xy )T . . . stress = (Sxx, Syy , Szz , Syz , Sxz , Sxy )T = DIV T d. . . strain = (Ex, Ey , Ez ) = grad. . . electr. eld
D = (Dx, Dy , Dz )T . . . electr. displacement = (dx, dy , dz ). . . mech. displacement . . . electr. potential
Piezoelectric PDEs:
2d t2
DIV cE DIV T d + eT grad div eDIV T d S grad
= 0 = 0
in in
boundary conditions: NT Dn Dn = = = = 0 0 1 A qe 0 on on on on g e \ (g e)
e . . . electrode q e . . . surface charge at electrode
g . . . grounding A. . . electrode area
Material Tensors
elasticity coecients cE 11 cE 12 cE 13 0 0 0 0 0 e31 cE 12 cE 11 cE 13 0 0 0 0 0 e31 cE 0 0 13 cE 0 0 13 cE 0 0 33 0 cE 0 44 0 0 cE 44 0 0 0 0 0 e15 0 e15 0 e33 0 0 0 0 0 0 0 0 0 0 0 0 0 e15 0 e15 0 1 E (c11 cE ) 0 0 12 2 0 S 0 11 0 0 S 11 0 0 0 e31 e31 e33 0 0 0 0 0 S 33
piezoelectric coupling coecients 10 dierent scalar coecients
dielectr. const.
State of Art (IEEE Standard, Eur. Norm)

test samples with special geometries model simplication to 1D case direct relation resonance frequencies coecients
example: thickness resonator thickness << radius 2 kt e2 33 2 = cE S 1k
t
2 kt =
33 33 fs 2 fp
tan
fp fs 2 fp
fs . . . lower critical frequency fp . . . upper critical frequency kt . . . thickness coupling factor
drawbacks: expensive measurements, inaccuracies!
Test sample scheme

(European Norm: Piezoelektrische Eigenschaften Keramischer Werkstoe und Bauteile 1998)
Identication by Simulation of Full System

10
0
10
10 Impedanz
10
Find material tensors cE , e, S from impedance measurements ()|e , Z() = e () q

4
10
10
10
10 Frequenz (Hz)
10
. . . potential part in solution of piezo PDEs q e . . . surface charge at loaded electrode . . . Fourier transform .
nonlinear operator equation F (cE , e, S ) = Newtons method
Sensitivities of measurements wrt. material parameters (1D case)
10
10
10
10
Z d/dc d/de d/d
frequency
Tests with thickness resonator

0.4 0.35 0.3 0.25 0.2 0.2 0.15 0.15 0.1 0.05 0 0 0.25 10 20 30 40 50 60 e 15 e 31 e
33
0.35 0.3 0.25
c 11 c 33 c12 c13 c44
0.1 0.05 0 0 0.4 0.35 0.3 0.25 0.2
10
20
30
40
50
60 11
33
0.2
0.15
0.1
0.15 0.1
0.05 0.05 0 0 10 20 30 40 50 60 0 0 10 20 30 40 50 60
relative error development: identication of all 10 coecients
inactive top outer electrode inner electrode inactive bottom
Stack actuator:
only part of the test samples can be produced
0
normed impedance
10
stack actuator
10
-1
reconstructions from available test sample measurements as starting values for material parameter identication for full stack actuator
10
-3
10
-2
measurement IEEE method inversion for

test sample full actuator
inversion for frequency [kHz]
25
30
35
40
45
50
Ultrasound array antenna material

(Motorola 3203HD)
frequency dependence of the material parameters
Further Identication Problems
temperature dependence nonlinear dependence hysteresis
Temperature dependence
external heating (e.g., sensors in cars) analogously to frequency dependence; cE = cE (), e = e(), S = S () heating by conversion of electrical and mechanical energy (large excitations over longer time intervals) coupling between mechanical strain S, electric eld E and temperature via = c E S eT E E D = eS + S E + pS = ET S + pST E + v
E pS v ... ... ... ...
temperature stress coe. pyroelectric const. thermic material const. entropy density
Nonlinear dependence
Large excitations (actuator applications):
2d t2
DIV cE (S)DIV T d + e(S, E)T grad div e(S, E)DIV T d S (E)grad
= 0 = 0
S = |DIV T d|
E = |grad|
innite dimensional problem, instability
First Results
1D case: dtt cE dx + e(|x|)T x e(|x|)dx S (|x|)x
x x
= 0 = 0
[Anderssen,Ringgaard,Bove,Albareda,Prez,Actuator 2000] e material parameters for Pz27 synthetic data computation in time domain
exact material parameters

45 40
e
35 30 8 x 10
9
7 0.02 E [kV/m] 280
Identication of S
45 40
e
35 30 8 x 10
9
7 0.02 E [kV/m] 280
Identication of e
50 40
e
30 20 8 x 10
9
7 0.02 E [kV/m] 280
Identication of e and S
60 50
e
40 30 8 x 10
9
7 0.02 E [kV/m] 280
Identication of S with 1% data noise

45 40
e
35 30 8 x 10
9
7 0.02 E [kV/m] 280
Identication of e with 1% data noise

100
50 0 8 x 10
9
7 0.02 E [kV/m] 280
Sensitivity of measurements wrt. S and e: mess mess

eigenvalues of sensitivity matrices
10
d dS
, and
d de
10
10
10
10
Nonlinear Magnetic B H Curves

B = (H)H H = (B)B
B (T) 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 0 1 2 H (A/m) 3 4 x 10 5
4
. . . permeability . . . reluctivity Identify from measurements of the magnetic ux for dierent currents I in an excitation coil. State of art: Epstein apparatus uniform magnetic eld B = nw , H = nw I At lt
Magnetic eld problem

(
A) = JI nA = 0
in D on B
JI = = B=
1 |c| Ieds c
impressed current density

ec
A ds = A ,
B nd
A magnetic vector potential = (| A|)
Inverse Problem Formulation

((| AI |) AI ) = JI I [I, I] (+boundary conditions) = AI ds c Nonlinear operator equation F () = F : Hpar Hmeas in D
Newtons method n+1 = n + sn F (n)sn F (n)
Instability
Linearized inverse problem:
F ()s = F ()
F () : Hpar Hmeas bounded and smoothing:

0.5 0.5 0.5 0.4 0.4 0.4 0.3 0.3 0.3 0.2 0.2 0.2 0.1 0.1 0.1 0 0 0 0.1 0.1 0.1
0.2
0.2
0.2
0.3
0.3
0.3
high frequency singular functions

0 5 10 15 20 25 30 35
0.4
0.4
0.4
0.5
10
15
20
25
30
35
0.5
10
15
20
25
30
35
0.5
10
10
10
correspond to small singular values of F ();

5 10 15 20 25 30
10
12
10
14
10
16
Unstable dependence of solution s on data

(noise level > || ||)
Regularization
regularization principles: Tikhonov method: (T T + I)x = T y , > 0 spectrum shift. Iterative solution: xn = (xn1, y ) (e.g. CG) early stopping. Projection onto nite dim. space: QhT xh = Qhy coarse discretization. regularization parameter choice: optimal balance between approximation and stability; convergence speed: in general arbitrarily slow!
Here: regularization: lin. Newton step equation: reg. by coarse discretization. multilevel methods Newton iteration: regularization by early stopping. choice of regularization parameter: discrepancy principle: maximize stability s.t. residual noise level h = max! n = min! s.t. s.t. ||F (n) + F (n)sh || C ||F (n ) || C
Multilevel methods for Parameter Id.

((| AI |) AI ) = JI AI ds = (I) c I= I 1 , . . . , I dI
Each iteration involves dI BVP solutions! meas keep dI = dim(Hh ) as small as possible. Adaptively for each iteration: start coarsely, sucessive renement. multilevel iteration speed up convergence by multigrid preconditioning (nonl.) full MG method
! ! e e e
3 1
! e e e
Multigrid Methods for Unstable Problems

Linear Problems F = y
two grid method B1

smoothing solution ne grid d
multigrid method Bl, l > 1: replace coarse grid solution by restriction prolongation m(l) applications of Bl1

d d
e e
e e
coarse grid
! e e
!e e e
! e e
!e e e
e e
! e e
!e e e
! e e
Problem: Smoothing Operator

magnetics example: j
10
8
10
10
10
12
well-posed problems Au = f (z.B. elliptic BVP) Sl A1 smoothing. unstable problems F a = y F smoothing but F 1 unbounded! Sl =?
10
14
10
16
5
0.5 0.4
10
0.5 0.4 0.3
15
20
25
0.5 0.4 0.3
30
uj
0.3
0.2
0.2
0.2
0.1
0.1
0.1
0.1
0.1
0.1
0.2
0.2
0.2
0.3
0.3
0.3
0.4
0.4
0.4
0.5
10
15
20
25
30
35
0.5
10
15
20
25
30
35
0.5
10
15
20
25
30
35
A Smoothing Operator for Unstable Problems

based on subspace decomposition [King92,Bramble&Leyk&Pasciak94]: Yl . . . discr. data space, Ql := P rojYl , Pl := P rojXl Sl : Y l Xl Sl := lPlF [I Ql1]
l h2p l
2
projection onto high frequency part
p . . . degree of instability (magnetics example: p = 2)

l [I Ql1]F
2
1,
l (Ql F )Ql
c>0
approximative inversion on high frequencies, cut o low frequencies ( coarse grid correction)
Convergence Results
h-independent contraction factors
||I BnQnT || c < 1 V-cycle convergence full MG iteration with discrepancy principle as a regularization method (stability, convergence) nonlinear multigrid methods
B H curve reconstructions
1.8
B (T)
1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 0 1 2 H (A/m) 3 4 x 10 5

4
0.6
0 1 2 3 4
5
B (T) 0 1 2 3 4
5
0.4
0.2
0 0
H (A/m)
2000
example with changing curvature, 1% data noise
details
Hysteresis: Preisach Model

P
D = eS + S E + P 0
p
Pr
1
Ec
-1
p(t) =
11
(, ), (e(t))dd
p(t) = e(t) =
P (t) P E(t) E
Conclusions and Outlook

extract material parameters from simple measurements by regularized Newton-multilevel methods using numerical eld simulations application in nonlinear magnetics and piezoelectricity general approach, applicable to other parameter identication problems hysteresis modelling optimization problems: ultrasound focussing, stack actuator design
Image Registration and Interpolation by Optical Flow with Maximal Rigidity

Stephen Keeling Wolfgang Ring Institut fr Mathematik u Karl-Franzens-Universitt Graz a Graz, Austria
Outline
Variational formulation to generalize rigid registration. Image similarity with scaling. Lagrangian and Eulerian regularization. Formulation of optimality system. Numerical solution of optimality system. Results and directions.
Image Registration Dened

De nition: Coordinate transformation identifying like points in images.
Top left: I0(). Top right: I1(). Bottom left: grid deformation, G(x()). Bottom right: morphing, I1(x()). Related processes: Image warping. Interpolation of images. Superposition of modalities.
2
L. Younes, 2000
Registration Among Imaging Modalities
CT slice
MR slice
PET slice
SPECT slice
Multimodal visualization
3
J. Pluim & T. Maintz, 2001
Registration Among Patients and Atlases
Atlas-based segmentation: register each patient to a segmented atlas Atlas-based segmentation. De ning pathology and norms.
4
B. Dawant, 2001
Registration within Temporal Sequences

Leading application: pathology tracking in mammography.
(a) before injection, (b) after injection, (c) di erence without registration. Image (a) after (d) rigid, (e) ane, and (f) non-rigid registration, with corresponding di erences from (b) in (g), (h), and (i).
5
D. Rueckert, 1998
The Coordinate System
at z = 1 I = I1
= slice Q = [0, 1]
at z = 0 I = I0
Physical coordinates: (x, y, z), x = (x, y). Curvilinear coordinates: (, , ), = (, ). Initially, (x, 0) = x. Otherwise, (x, z) = z. Point (x(, ), z) on trajectory with x(, 0) = . Displacement: d = x . Optical ow: u = (u, v) = x (, ). Trajectory tangent: (u, v, 1).
6
Proposed Formulation
Minimize: J(I,
0, 1 , u)
= +
1 0 1 0
| I u + Iz |2dxdz (| uT + u|2) + |uz |2dxdz

1 (I1 (x)).
subject to: I(x, 0) =

0 (I0 (x)),
I(x, 1) =
and possible landmark constraints: x(j , 1) = xj , where: x(, ) = +

0
j = 1, . . . , j u(x(, ), )d.
Similarity: | I u + Iz |2, and I(x, 0) = Rigidity: [

x u] T 0 (I0 (x)),
I(x, 1) =
x]
1 (I1 (x)). T
+[
x u]
=0[
x]
= I.
Excision: (s) =
s + (reg TV).
Depth regularization: |uz |2. Independent of image order.
Image Similarity
Simplest: |I0() subject to: I(x, 0) = I0(x), I(x, 1) = I1(x). Lagrangian to Eulerian:
1 0
I1(x(, 1))|2d =
1 0
|dI/d|2dd
|dI/d|2dd
x).
1 0
| I u + Iz |2dxdz
without 1/ det( With scaling:
| 0(I0()) subject to: I(x, 0) =
1(I1 (x(, 1)))| 1 0
| I u + Iz |2dxdz I(x, 1) =
1(I1 (x))
0(I0 (x)),
Alternatively di erential operator to match level curves: D I0() D I1(x(, 1))
Reciprocal Scaling
T
I1
c '
I0 I0 = S I1 = (1 + )S
x
T
I1
c '
0(I0 )
= I0
Scaling I0 not I1 x
E T
1(I1 )
= I1
'
S I0 Scaling I1 not I0 x
S
9
Lagrangian Regularization
Jlin(d) = + + Junl(x) = + + 4 2
1 0 1 0
[I0()

I1( + d(, 1))]2 d d +2

2
d +
dd
(d d )2 dd I1(x(, 1))]2 d x1

[I0()
1 0 1 0
1 + 1 + +2
2
x2 x2 x1
1
2
dd
x1
1
2
x2
dd
+ 0 (x x )2 dd Rigid versus collapse: x(, ) = R()( a) + a, d = x x(, ) = (1 2)( a) + a, d = x
= 16 ( )( + ) < 6 ( )( + ) = Jlin(d) Jlin(d) 3 8 x x Junl( ) = 0 < ( )( + ) = Junl( ). 15

10
Eulerian Regularization
Rigidity: [
x u] T
+[
x u]
=0[
x]
x]
= I.
Local u = x instead of global d = x J(I, u) = + With rigid eld: u(x, z) = x = W (x I = I0(R(z)T (x follows J(I, u) = 0. a),
Q(

displacements:
I u + Iz )2 dxdz x
xu
xu
|uz | dxdz
R (z) = W R(z) x I0() = I1( (, 1))
a) + a),
For the class: x(, ) = R()( a) + (a + b) u(x, z) = W (x a bz) + b need (|

x uz | 2
) and |uzz |2 instead of |uz |2.
For autonomous ows: J(I, u) = +

Q( xI
u + Iz )2 dxdz
xu T
11
xu
dx
Optimality Conditions for Intensity

d2 I ( , z ) [( I u + Iz )(u, 1)] = 2 + ( d I = 0, n u = 0, or dI u) = 0, Q d
I u + Iz = 0,
2)
(x
n u = 0 non-rigid. I u + Iz = 0 x2 x2 I1 I1 7 7 7
(x
1)
I1
(x
I0 x1 x1 Take I = 0 on . Lagrangian solution: I(x(, ), ) =

` 7
I0
I0
(x
1)
I1
2)
I0
I0()[1 I ()[1 0 or scaled with
U (, , 1)] + I1(x(, 1))U (, , 1), U (, , )], x(, ) ,

0 (I0 )
c 0 i 0
and
1(I1 ),
and:
0,
= U (, ) U (, , ) U (, ) U (, ) =
0
U(, 0) , 0) U(,
0
[0, ]
exp
u(x(, ), )d d ,
12
Optimality Conditions for Scaling

Lagrangian form of I inserted in J reveals optimal scaling:

I0 ()= I1 ()U()d I0()= U()d
K0
0 ()
I0 ()= I1()
U() d | I0()| c , K0 U() d I0 ()= | I0()| [0, 1]

c 0, i 0
where:
| I0| = 0 in { : I0() K0}, K0

I1(x(, 1)), I1() = 0, x(, ) , U() =

1 2 0 U (, , 1) det (
x) d, x) d,
c 0 i 0
() 2 U (, , ()) det ( 0
In words:
0()
is weighted average over {I0() = } of morphing I1
13
Optimality Conditions for Optical Flow

B(u, u, v) = F (v), where: B(w, u, v) = +
Q

v C (Q)
Q [(
I u)( I v) +
2

(uz v z )] dxdz vT + v dxdz
wT +
uT +
Q Iz
u :
F (v) =
I vdxdz
Theorem 1. Suppose I W 1,(Q) manifests suciently few symmetries that for every a R2 and c R: | I (a + cW x)|2dxdz > 0 Q unless a = 0 = c. Then with (s) = (x, z)s, 0 < 0 (x, z) 1 < , there exists a unique u H 1(Q) such that B( , u, v) = F (v), v C (Q). 1.0 0.5 0.0 0.0
T
Ie

e e
e e
e e
e e
eE
0.5
1.0
ue =
E
e e
z=1
ue = + 1 4
e e
z=0
z
1 4
Here TV cost is nite (and stationary), Gaussian not.

14
Optimality Conditions for Landmarks

Lagrangian functional: 1 L(u, ) = 2 J(u) + TE(u), where: E (u) = x( , 1) x = For: x(, ) (u; v) = u x( , ) u(x(, ), ) (u; v) + v(x( , ), ) d. u de ne the solution operator: Su,(, )Su,( , ) = Su,(, ), 0 1 Su,(, ) = I, [0, 1] Su,(, ) = u(x(, ), )Su,(, ) Su,(, ) = Su,(, ) u(x( , ), ) Stationarity conditions: L (u; v) = B(u, u, v) F (v) + TG(u, v) = 0, u v C (Q) L = E(u) = 0 where: 1 G(u, v) = 0 Su,(1, )v(x(, ), )d
0

x +
1 0 u(x( , ), )d,
= 1, . . . , .
15
Solving Optimality System with Landmarks

Theorem 2. Assume the conditions of Theorem 1. Suppose J(u) is minimized subject to E(u) = 0 at u W 1,(Q). Then there exists a R2j such that: B(u, v) = F (v) ()TG(u, v), v H 1(Q) E(u) = 0. Fixed Point Iteration: Given u, Solve u0: Solve uj : Solve : Update: B(u0, v) B(uj , v) 0 u = F (v) = Gj (u, v) = E(u0 + j uj ) = u 0 + j u j
Preferred: Augmented Lagrangian SQP... H (u, v, w) = + gives

2 1 0 Su, (1, ) 0 Su, (,
w(x( , ), )
)v(x(, ), )d d v(x( , ), )
1 0 Su, (1, ) 0 Su, (,
)w(x(, ), )d d
x/ u2 without D 2u.
16
Numerical Treatment of Intensity

Simple translation problem: I = I1
I = 0 I = 1

I = I0 with Eulerian discretization:
I0
=1
I1
=0
with Lagrangian discretization:
17
Numerical Treatment of Optical Flow

Lumped Finite Elements with Conjugate Gradient:
I0(), I1() I1(), I0()
|I0() |I1()
I1()| I0()|
L1-errors: 0.0094 on subdomains, 0.014 on full domains L-errors: 0.52 on subdomains, 1.0 on full domains Optical ow and morphed uniform grids:
Here nontrivial pixels reach . Extend images by zero to suciently large .

18
Object Excision
I0(), I1() I1(), I0()
|I0() |I1()
I1()| I0()|
I0(), I1() I1(), I0()
|I0() |I1()
I1()| I0()|
19
Strongly and Weakly Rigid Registration
I0(), I1() I1(), I0()
|I0() |I1()
I1()| I0()|
I0(), I1() I1(), I0()
|I0() |I1()
I1()| I0()|
20
MRI Application Images

I0(), I1(), |I0() I1()|
Strongly Rigid
I1(), I0(), |I1() I0(), I1(), |I0()
I0()| I1()|
Weakly Rigid
I1(), I0(), |I1()

21
I0()|
MRI Application Fields

Optical ow and morphed uniform grids: Strongly Rigid
Weakly Rigid
22
ELMS UNIVERSITT B WILH
- INSTITUT FR AN ONN
NDTE MATHEMATIK GEWA
Adaptive Wavelet Methods for LinearQuadratic Elliptic Control Problems

Angela Kunoth Institut fr Angewandte Mathematik, Universitt Bonn, Germany u a Supported by SFB 611
NISCHE FRIEDRICHRHEI
Wavelet Methods for Control Problems Motivation General Sobolev Norms in Minimization Functional Constraints in Form of Elliptic PDE Asymptotically Optimal Preconditioned Systems ; Fast Iterative Solvers Biorthogonal Wavelets ; Piecewise Polynomials ; Setup of Linear System Numerical Results: Carsten Burstedde Adaptive Schemes (joint work with Wolfgang Dahmen) Convergence Theory and Optimal Convergence Rates
1
; Modelling
Neumann Problem with Distributed Control minimize given y, f >0 subject to J(y, u) = y + y
y n 1 2
2 H 1s ()
2 (H 1t ())
=f +u =0
in on Av, w :=
I d R
A : H 1() (H 1()) ; Weak formulation minimize subject to J(y, u) = Ay

1 2
( v
w + vw)dx A1f
Nontrivial solution for y y y

2 H 1s ()
2 (H 1t ())
=f +u 0 s 1 smoothness parameter for state y 0 t 1 smoothness parameter for control u
s = t = 0 natural norms
2
Standard Discretizations A : H H isomorphism ; discretization ; nite dimensional problem ; fast iterative solvers ; ecient preconditioners In connection with control problem: Evaluation of fractional Sobolev norms New Paradigm (I) mapping property for A : Y Y (II) transformation into equivalent dimensional wellposed (III) convergent iteration for the dimensional (IV) adaptive application of operator A (V) convergence/complexity analysis ; set up control problem in wavelet coordinates
3
problem
problem
(Biorthogonal) Wavelets H Hilbert space with := { : I I}

H
H dual of H with , H Wavelets I (innite) index set I
(P1) Riesz basis property v H: (P2) locality v = vT :=

I I
such that
I) 2 (I
diam (supp ) 2||
|| resolution centered around 2||k
Representer for Control Problem in Wavelet Coordinates minimize subject to J(y, u) =

1 2
Ds (y
y ) A:
2
+ u 2
2
0 s, t 1 :=
2
Ay = f + Dt u
automorphism
Necessary and Sucient Conditions Lagr(y, u, p) := J(y, u) + p, Ay Ay = f + Dt u Lagr = 0 ; AT p = D2s (y y ) Q u = g u = Dt p Q: where Q := Dt AT D2sA1Dt + I g := DtAT D2s(y A1f )
5
(f + Dt u)
automorphism
symmetric positive denite
Convergent Iteration for the -Dimensional Problem Iterative solution of Mu = f ; ; un+1 = un + (f un+1 u un Mun) u M symmetric positive denite n = 0, 1, 2, . . . where := I 0 < M < 1
Ideal iteration ;
computable scheme for evaluation of Qun = (DtAT D2sA1Dt + I)un and g
Goal: Employ techniques by [Cohen, Dahmen, DeVore 99] on adaptive methods for elliptic systems
6
Adaptive Approximate Iterations Res [, Q, g, v] r determines for given > 0 a finitely supported r satisfying Coarse [, w] w determines for given > 0 a finitely supported w satisfying r
Qv
; w Realization: sort w into nonincreasing order nd smallest k such that sum of k largest coecients exceeds w 2 2 ; w Cost: 2N and N log N for sorting if N = #(supp w)
K := min{ : Main algorithm: # interior iterations is ( relaxation weight < 1 contraction number)
( + )
1 10 }
Adaptive Approximate Iterations Main Algorithm Solve [, Q, g] u (i) j = 0 u0 = 0 relaxation weight 0 := 1 c1(c1 f + y ) A 2 A < 1 contraction number
(ii) If j : stop and set u := uj Otherwise v0 := uj (ii.1) For n = 0, . . . , K and (ii.2) Apply Coarse [ 2 j , vK ] uj+1 5 1 compute Res [nj , Q, g, vn] rn vn+1 := vn + rn set j+1 := 1 j 2 j+1j go to (ii)
Theorem [Cohen, Dahmen, DeVore 01] For any > 0 Solve [, Q, g] u terminates after nitely many steps and
u
8
Routines for Realization of Res Apply [, A, v] w computes for given > 0 a finitely supported w satisfying w
Av
Solve [, A, f + u, y0, 0] y computes for given > 0, initial guess y0 for y with accuracy 0 a finitely supported y satisfying y y employs Resell [, A, f + u, y] r (i) Apply[ 1 , A, y] w 3 (ii) Coarse[ 1 , f ] f 3 Coarse[ 1 , u] u 3 (iii) set r := f + u w
Solve [, AT , y + y, p0, 0] p computes for given > 0 p p a finitely supported p satisfying

9
Realization of Res [nj , . . .] (n + 1)th iterate in (j + 1)th block := nj Res [, Q, g, v] r (i) u := n1j ( + n) y := c1 u + A
j+1,n+1,0 (ii) Coarse[4y, yj+1,n] y j+1,n+1,0 (iii) Solve[ 1 cA , A, f , uj+1,n, y ] y 2 4 (iv) Coarse[ u, pj+1,n] pj+1,n+1,0
=: yj+1,n+1
(v) Solve[ 1 , AT , y + y, pj+1,n+1,0] p 2 (vi) set r := p v
=: pj+1,n+1
10
Convergence and Complexity [Cohen, Dahmen, DeVore 99] Goal: Show that Solve realizes asymptotically the work/accuracy balance of best N term approximation Ideal Bench Mark Best N -Term Approximation v vN :=
# supp wN
min
; classify sparse sequences in 2 whose best N term approximation decays at certain rate ; Lorentz spaces
w
:= {v
w
: #{ I : |v| > } < } I
for 0 < < 2
|v| v
:= sup n1/ vn nI N
(v decreasing rearrangement) quasinorm for

w w
:= v + |v|
w +
for < + < 2
close to
11
Convergence and Complexity [Cohen, Dahmen, DeVore 99] Proposition v

w
vN
< N s v
=s+1 2
Ideal work/accuracy rate: Coarsening Lemma = v

w
target accuracy
work 1/s w
w nitely supported such that
output w of Coarse [4, w] satises # supp w < v w w

w
1/
w
1/s
< 5 < v
w
12
Convergence and Complexity Compressible Matrices [CDD 99] A scompressible: for any 0 < s < s there exists Aj with j 2j nonzero entries per row and column such that A v nitely supported ; Aj j 2sj jI 0 N
jI 0 N
j <
v[j] := v2j best 2j aproximations wj := Aj v[0] + Aj1v[1] + + A0v[j]
Theorem
A scompressible ;
w
A bounded on
for s < s and
=s+1 2 <
v nitely supported ; output w of Apply [, A, v] satises w

w
# supp w , #ops < 1/s v
1/s
w
13
Convergence and Complexity Main Theorem [Dahmen, K 02] Lemma w Res [, Q, g, v] w

w
satises
< max{ v w , g w } 1/s # supp w < 1/s max{ v w , g #ops # supp w
1/s
w
=s+1 2
Corollary Res has above properties for xed > 0 and any ( , 2) = for every > 0 Solve [, Q, g] u converges in nitely many steps u If u
w
u is nitely supported
1/s # supp u < 1/s u w #ops # supp u
14
for some ( , 2)
= u
w
<
Automatic detection of gravitational arcs in astronomical data using anisotropic diffusion and segmentation
Dipl.-Math. Frank Lenzen Department of Computer Science University of Innsbruck
Gravitational lensing
Image 1
background galaxy Observer gravitational lens Image 2
Gravitational lensing (scaled image)

3
Goal: Design of an algorithm for automatic detection of arcs resp. candidates Strategy: 1. Scaling the intensity values 2. Smoothing (Anisotropic Diffusion) 3. Segmentation 4. Selection of thin and elongated objects
4
Step 1 : Scaling
Step 1 - Scaling
Images are stored using the FITS format (allowing floating point intensity values).
Histogram:
Step 2: Smoothing
Step 2 - Smoothing
We use Anisotropic Diffusion as described in: Joachim Weickert: Anisotropic Diffusion in Image Processing, Teubner 1998 Advantages: - Reduction of noise - Edge enhancement - Closure of small gaps
Step 2 - Smoothing
For the Anisotropic Diffusion process the structure tensor is exploited:
v1, v2 will be used in the segmentation process !

9
Step 2 - Smoothing
Original image
Gaussian convolution
Anisotropic diffusion
10
Step 3: Segmentation
11
Step 3 - Segmentation
Each local maximum above a certain threshold cmin is expected to belong to an object. We use region-growing algorithm starting in a local maximum p0 using an adaptive threshold c( p0 ). The detection of local maxima is performed in a similar way as in standard watershed algorithms.
12
Step 3 - Segmentation
Determination of c( p0 ): - by edge detection along cross-section through p0 in direction v1 . - Two points of maximal descent with intensities c- and c+. Set c( p0 ) := max(c-,c+)
v1
p0
v2
Cross-section
13
Found objects
14
Step 4: Selection
15
Step 4 - Selection
Let R be an object consisting of pixels with coordinates xj= (x1j,x2j) und intensities Ij, j=1...n Interpretation as rigid body with density I(x,y) , mass m and mass center (mc1 , mc2) and mass moment:
16
Step 4 - Selection
Eigenvalues of M: Eccentricity: Thickness: Selection of thin and elongated Objects:
17
Results
18
Selected objects
19
Original (scaled) image
Detected objects
20
Computational effort
21
Conclusion
We designed an algorithm which
- uses the benefits of Anisotropic Diffusion for enhancing thin structures - performs a segmentation with adaptive edge detection based on the information gained by the diffusion process. - Provides good results (in comparison with another astronomical tool for object detection named Sextractor)
22
Credits
The work was done in cooperation with Prof. Sabine Schindler, Department of Astrophysics, IBK Prof. Otmar Scherzer, Department of Computer Science, IBK Paper: F. Lenzen, S. Schindler, O. Scherzer, "Automatic detection of arcs and arclets formed by gravitational lensing", Preprint Series of the Department of Computer Science, University of Innsbruck, Vol. 11, July 2003 Submitted to Astronomy & Astrophysics Supported by
23
Model Problems for the Multigrid Optimization of Systems Governed by PDEs
Robert Michael Lewis

College of William & Mary Williamsburg, Virginia, USA
Stephen G. Nash
George Mason University Fairfax, Virginia, USA
Copyright, 1996 Dale Carnegie & Associates, Inc.
Overview
Optimization problems governed by systems of differential equations minimize F ( a ) = f ( a , u ( a ))
a
subject to
S(a,u) =
S ( a , u ( a )) = 0 [other constraints] 0: system of PDEs
Design variables: State variables:
a
buckaroo@math.wm.edu snash@gmu.edu
R. Michael Lewis, College of William & Mary Stephen G. Nash, George Mason University
Applications
Engineering design Related applications:
VLSI design Shape optimization Traffic models Medical imaging
Related optimization problems:

Integral equations as constraints Systems of ODEs as constraints Variational problems (e.g., minimal surface)
R. Michael Lewis, College of William & Mary Stephen G. Nash, George Mason University buckaroo@math.wm.edu snash@gmu.edu
Themes of Talk
Optimization-based multigrid
Distinct from (& superior to?) traditional multigrid Optimization models as subproblems
Focus on algorithmic template

Ideas relevant to a class of algorithms Sample implementation
Focus on model problems

Illustrate essential ideas Analyze in depth
Some of the Justifications

Richer class of models Guarantees of convergence, progress Better operator properties than for PDEs alone:
Successful even for non-elliptic PDEs Reduced Hessian like an elliptic operator
Good performance (even far from solution) Connection to other optimization methods
The Algorithmic Template
General Idea
Optimization model is a family of problems
Discretization of functions Clustering of components Hierarchy of models
Linear vs nonlinear Approximate vs true
Better model more effort Use computations on a coarse grid to compute a search direction for a finer grid Use a line search on the finer grid
Multigrid for Optimization

MG/Opt Assume here that the differential equation Sh (a,u) = 0 is solved for uh (a) # of design variables a may be fixed Derived from full approximation scheme (Brandt, 1977) applied to optimality conditions f (a, u (a )) = 0
a h h
MG/Opt: Notation

h : fine grid, H : coarse grid IhH : downdate, IHh : update u : fine-grid vector, u : coarse-grid vector f : fine-grid objective function f : coarse-grid objective function g (a,u ): fine-grid gradient g(1)=g (a(1),u(1)), etc.
MG/Opt: Algorithmic Template

If coarsest grid, minimizea fh (a,uh (a)), else: partially minimize f h(a,u) to get a(1); a(1)=IhHa(1)
compute v = g (1) IhHg(1) Obtain a(2) from:
recursively minimize fH (a,uH(a)) vTa subject to bounds on the solution (to guarantee convergence)
compute e(2) = IHh (a(2) a(1)) line search: a(2) a(1) + e(2)
partially minimize f h(a,u) to get a(3)

Related Research
Existing Work
Multigrid methods
Mostly for linear/nonlinear equations, mainly elliptic PDEs
Related work:
Bramble, Brandt, Briggs, Gmez, Hackbusch, Heinkenschloss, Kelley, Mandel, McCormick, Polak, Sachs, Tseng, Zhu,
At this conference:
Petrova, Taasan, Schoeberl, Yavneh,
Alternatives to MG/Opt
Apply multigrid to PDE alone:
not always possible/sensible
Apply multigrid to optimality conditions:

Lose guarantees of convergence Not ideal for bounds, inequality constraints
Apply linear multigrid to Newton equations:

may give a better traditional method
Successive refinement:
interesting
Justification #1:
Richer Class of Models
Optimization Models are More Flexible

In this talk, the state equation is a PDE Could be an integral equation Could add additional constraints:
Bounds Inequalities
True generalization of multigrid
Justification #2:
Convergence
Analogy: Nonlinear equations vs. Optimization

If we solve f ( x) = 0
lim f ( x) = 0
If we minimize f ( x)
lim f ( x) = 0
Descent
eh is the search direction from MG/Opt If the additional bounds in the MG/Opt subproblem are appropriate, then eh is a descent direction The multigrid iteration will lead to improvement in the solution
Convergence
If (on each grid h )
underlying optimization algorithm is guaranteed to converge (to a stationary point) without multigrid strategy multigrid strategy produces a descent direction
Then
MG/Opt is guaranteed to converge (to a stationary point)
Related to trust-region approaches

Justification #3:
Suitability for HighPerformance Computing
User Requirements
Procedure to solve S (a,u ) = 0 for u given a Procedure to evaluate fh (a,u ) and a fh (a,u) for any grid h Procedures to implement downdate IhH and update IHh operators
Should satisfy IHh = (IhH)T <constant> [standard]
Flexibility of Algorithmic Template

Not dependent on a specific (single-grid) optimization algorithm [constrained, unconstrained, etc.] Isolates optimization algorithm, differential equation solver, grid generation Use state-of-the-art techniques for the individual components
Justification #4:
Ellipticity & Convexity
Properties of the Reduced Hessian

The reduced Hessian will be positive (semi) definite at the solution Multigrid works well for elliptic PDEs Connection between ellipticity and convexity Optimization methods work well on convex problems
Why Look at Model Problems?
Model Problem: Motivation

Goal: Gain better understanding of properties of MG/Opt by analyzing a specific, but simpler, model problem Has proven useful in analysis of Multigrid Allows detailed analysis (not possible in general nonlinear case) Model problem may be representative
Model Problem: Motivation (cont.)

Analysis of MG/Opt more complicated than for MG:
Downdated operator for Multigrid:
(I
H h
AI
h H
Downdated operator for MG/Opt:
( downdated differential operator ) h ( I hH AI H ) ( downdated differential operator )

T
A Sample Model Problem

Match a target function u*:
minimize F (a ) =
a 1 2
u (a ) u*
Where u (a) solves the 1-way wave eqn.:

ut cu x = 0 u ( x, 0) = a( x)
With 0 x 1, 0 t 1 Computations use c = constant = 1

Model Problem: Wave Eqn.

Hyperbolic equation Initial value moves without dissipation or dispersion Multigrid methods (applied to constraint alone) are not ideal: usual approach is to march forward in time
Initial Solution 2 1 0 -1 -2 -5 2 1 0 -1 -2 -5
-4
-3
-2
-1
Solution at t=1
-4
-3
-2
-1
Model Problem: Details

Consider an objective function of the form: 2 2 f (a, u (a)) = u u* + u x (u* ) x Assume that a(x) is discretized uniformly in x
Justification #5:
Analysis of Model Problem
Model Problem: Analysis of Continuous Problem

The reduced Hessian is
d 2 I + 2 dx
This is like the 1-dimensional Laplacian Ideal for multigrid Likely to cause difficulties for general-purpose large-scale optimization methods Analogous results for discretized model problem
Justification #6:
Computational Performance
Model Problem: Computations

# of design variables: n=1025 [1,051,650 total variables: n(n+1)] n=1025 it Optimization ls cg Successive refinement it ls cg it MG/Opt ls cg 99 100 967 23 100 956 10 20 57 25 26 216 12 24 65 25 26 225 14 28 79 25 26 214 16 32 89 25 26 220 18 36 112 19 20 145 232 242 1974 n=512 n=257 n=129 n=65 n=33
Choice of Comparative Algorithms

Why only compare MG/Opt with traditional optimization algorithms (for systems of equations)?
Inequality constraints may be present Optimality conditions may not be elliptic Hard to derive reduced Hessian/system (thus hard to identify a good preconditioner) No obvious relationship between original optimization model and reduced system
Justification #7:
Relation to other Optimization Methods
MG/Opt & Steepest Descent

Coarse-grid problem is a first-order approximation to the fine grid problem
Gradient of coarse-grid problem at aH = IhH [gradient of fine-grid problem at ah]
Analogous to the first-order approximation used to derive the steepest-descent method Additional bounds are needed since approximation is only useful within a limited region
Scaling of Search Direction

Let s ( ) = F (a + eh ) Well-scaled search direction: s '(1) 0 h Search direction of the form eh = I H eH If subproblems solved accurately, then:
s '(1) = e [ I ( Fh ) I FH ]eH + O( eH )
T H H h 2 h H 2 3
Search direction is Newton-like

Summary
Optimization-based multigrid provides flexibility and generality MG/Opt is an effective tool for discretized optimization problems MG/Opt is based on a first-order approximate model (descent, convergence) MG/Opt separates model, discretization, and optimization Variety of justifications (theoretical and computational) for using MG/Opt
Questions?
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
Computational Experience with large-scale Semidenite Programs

Franz Rendl Institut f r Mathematik u Universit t Klagenfurt a A - 9020 Klagenfurt, Austria franz.rendl@uni-klu.ac.at September 9, 2003
Semidenite Programs (SDP): C, A1, . . . , Am symmetric matrices of order n, b I m R sup C, X Ai, X = bi i X in short A(X) = b 0
Like Linear Programming, except X Rapid development in the last 10 years Applications: Control theory Robust Optimization Combinatorial Optimization
Unconstrained Optimization in binary variables: max xT Lx x {1, 1}n Global optimum difcult to nd How prove global optimality? Relaxations to get tractable problem Linearize: xT Lx = L, (xxT ) = L, X new variable is X: X = xxT obvious constraints on X: X diag(X) = e, rank(X) = 1 0 e all ones vector
but non-convex, so drop it
How solve Semidenite Programs? Most efcient way: Interior-point methods (Primal-Dual Path-following Variants) Features: Newtons method, so few iterations maintain X 0
dense linear algebra with matrices of order n solve linear system of order m System matrix positive denite, but dense and increasingly illconditioned
n seconds 200 1.84 400 8.92 600 24.10 800 51.45 1000 99.27 1500 314.99 2000 714.21
Table 1: Computation times (seconds) to solve the SDP on a PC (Pentium 4, 1.7 Ghz).
Representative Timings We solve max L, X : diag(X) = e, X 0
Matrices of order n, and n simple equations xii = 1
Combinatorial Cutting Planes: A simple observation:

x {1, 1}n, f =
1 1 1 |xT f | 1 0 . . . 0
Translated to X: xT f f T x = (xxT ), (f f T ) = X, f f T 1 Can be applied to any triangle i < j < k Nonzero elements of f also -1 This gives n 4 3 additional linear inequalities Further cutting planes could be added!

k = 500 k = 1000 k = 1500 k = 2000 k = 2500
n = 100 n = 200 n = 300 21 (19) 34 (20) 49 (19) 103 (21) 136 (22) 164 (21) 304 (24) 358 (24) 422 (24) 643 (25) 763 (26) 816 (24) 1090 (24) 1313 (26) 1360 (24)
Table 2: Computation times (seconds) on a PC (Pentium 4, 1.7 GHz) to compute the semidenite relaxation of Max-Cut for a graph with n nodes and k triangle inequalities. The number of interior point iterations is given in parentheses. n n 4n 2 3 50 1.225 78.400 100 4.950 646.800 200 19.900 5.253.600 500 124.750 82.834.000 Table 3: Triangle constraints as n increases
Computation times with k triangles added to original SDP The potential number of constraints grows fast ! Only n constraints determine off-diagonal part of X 2 Identifying good candidates for violated constraints ?
Bundle Method as Alternative: Maintain only part of constraints explicitely (x F ) Remaining constraints (Ax b) are dualized through Lagrangian
xF, Axb
max
cT x = max min L(x, )

xF 0
Lagrangian: L(x, ) = cT x + T (b Ax) Dual functional: f () := max L(x, )

xF
Minimizing f is equivalent to original problem f is convex, but non-smooth Bundle methods: minimize f using function and subgradient evaluation only Note: function evaluation means solving over x F . Which constraints should be kept explicitely (in F )? Here: F = {X : diag(X) = e, X 0})
Further simplication - Spectral Bundle: Instead of F = {X : diag(X) = e, X use F1 = {X : tr(X) = n, X max L, X : X F1 can be solved explicitely: Xopt = nvv T , where v is normalized eigenvalue of max(L). Spectral bundle method taylored for this purpose Working horse is fast eigenvalue computation of max 0}) 0})
problem spin5 spin6 spin7 spin8 spin9 spin10 G22 G27 G39
n cut initial bd gap (%) nal bd gap (%) time 125 108 125.3 15.99 111.1 2.87 0.73 216 182 211.8 16.37 186.9 2.69 3.16 343 304 347.9 14.44 308.4 1.45 10.55 512 454 524.3 15.48 463.1 2.00 33.44 729 630 744.7 15.27 666.7 5.83 42.75 1000 874 1032.6 18.15 921.4 5.42 106.1 2000 13220 14135.9 6.93 14045.8 6.25 278.06 2000 3186 4141.7 30.00 4048.4 27.05 406.66 2000 2314 2877.7 24.36 2672.7 15.50 533.36
Table 4: Results for Max-Cut. Initial bound (= no triangles). Final bound obtained after 100 function evaluations with the bundle method in case n 729, after 50 iterations in case 800 n 1000 and after 30 iterations for n = 2000.
Some computational results
Development of f during iterations:
Conclusions: Problems with huge number of constraints manageable Bundle methods converge slowly, Initial progress rapid, so good if inaccurate solutions are ok Can be applied to many other types of problem Joint work with I. Fischer (Klagenfurt) R. Sotirov (Dissertation in Klagenfurt, now McMaster Univ.) A. Wiegele (Klagenfurt, currently working on dissertation) I. Dukanoviv (Maribor, currently working on dissertation)
Numerical Simulation on High Performance Computers

Ulrich Rde
Lehrstuhl fr Systemsimulation Universitt Erlangen-Nrnberg www10.informatik.uni-erlangen.de Workshop on
Advances in Numerical Algorithms

Graz, Austria
September 10-13, 2003
Prof. Dr. Ulrich Rde Lehrstuhl fr Systemsimulation Universitt Erlangen-Nrnberg 12.09.2003 Graz, September 2003
Overview
High Performance Computers Example Application I: Material Science Example Application II: Biomedical Engineering Conclusions
Prof. Dr. Ulrich Rde Lehrstuhl fr Systemsimulation Universitt Erlangen-Nrnberg
12.09.2003
Graz, September 2003
Part I High Performance Computers
12.09.2003
12.09.2003
Prof. Dr. Ulrich Rde Lehrstuhl fr Systemsimulation H.W Meuer et al. 2002 Universitt Erlangen-Nrnberg
12.09.2003
Architecture example
8 Proc and 8 GB per node Performance: 1344 CPUs (168*8) 12 GFlop/node 2016 GFlop total Linpack: 1645 Gflop (82% of theoretical peak) Very sensitive to data structures
Hitachi SR 8000 at LRZ Munich, currently No. 29 in the Top 500

(No. 5 at time of installation in 2000)
12.09.2003
10.000.000.000 4G n 1.000.000.000 100.000.000 10.000.000 1.000.000 4M 1M n 256K n n 100.000 10.000 1K 1.000 100 1970 1975 1980 1985 1990 1995 2000 2005 n n 4004 n n DRAM Microprocessor (Intel) 80368 64K n n n n
Transistors/Die
Growth: 52% per year

64M n
1G n n Merced
n Pentium Pro Pentium 80468
Growth: 42% per year
Year
Moore's Law in Semiconductor Technology

(F. Hossfeld)
1021 1018 1015 1012 109
Semiconductor Technology
1012 10 9 10 6 10 1
3
106 103 1 1950 1960 1970 1980 1990 2000
2017
kT
10 -3 10 -6 10 -9
2010
2020
Year
Information Density & Energy Dissipation

(adapted by F. Hossfeld from C. P. Williams et al., 1998)
Energy/logic Operation [pico-Joules]
1024
1015
Atoms/Bit
Current Challenge: Parallelism on all levels and The Memory Wall

Parallel computing is easy, good (single) processor performance is difficult (B. Gropp, Argonne) There has been no significant progress in High Performance Computing over the past 5 years (H. Simon, NERSC) Instruction level parallelism Memory bandwidth and latency are the limiting factors Cache-aware algorithms (see DiME project) Conventional complexity measures (based on operation count) are becoming increasingly unrealistic
10
Future Challenge
Whats beyond Moores Law?
12.09.2003
11
Part II Example Application Material Science
12.09.2003
12
Example Application: Metal foams

Simulation with Lattice-Boltzmann methods
Diploma thesis: N. Threy Dissertation project: T. Pohl Cooperation mit Department of Material Science, C Krner KONWIHR-Project FreeWihr
12.09.2003
13
Metal Foams:
lightweight and stiff materials with high damping
Bubble growth, coalescence, collapse, drainage, rheology, etc. are still poorly understood. Simulation as a tool to better understand, control and optimize the process
12.09.2003
14
Lattice-Boltzmann Automaton
for computing fluid flow in a time varying geometry with free surfaces
12.09.2003
15
Stream-and-Collide
12.09.2003
16
The breaking dam
12.09.2003
17
Modelling of the two-phase flow
Reconstruction of curvature needed for Computing surface tension

12.09.2003
18
Rising bubbles
12.09.2003
19
Several rising bubbles
12.09.2003
20
Simulation and Experiment
Simulations and experiments: Diploma Thesis N. Threy Th
12.09.2003
21
LBM Status Quo

High resolution in space and time necessary O(n4) Enormous compute requirements requires parallelisation/supercomputer Low approximation order No local refinement Information transport/exchange only between next neighbors No multilevel structure
12.09.2003
22
Part III Example Application Biomedical Engineering
12.09.2003
23
Motivation: Bioelectric field computations

Reconstruction of electrical behaviour inside the head from measurements (Electroencephalography) Inverse problem important for Neurology and Neurosurgery
Neurosurgery at Erlangen University Head Hospital

Prof. Dr. Ulrich Rde Lehrstuhl fr Systemsimulation Universitt Erlangen-Nrnberg 12.09.2003
View through Operation microscope
24
EEG Current
Future Practice
Brain Mapping EEG Recording

25
Source Localization
3D MRI data 512 x 512 x 256 voxels segment 4 compartments
Dipole localization search algorithm = optimization
Localized Epileptic focus

Dissertation project: M. Mohr, Master thesis: S. Mhrs, Diploma thesis: H. Kstler hrs, Collaborators: Univ. of Utah (Chris Johnson), Ovidius Univ. Constanta (C. Popa) Univ. Univ. Popa) Bart Vanrumste (Gent, Univ. of Canterbury, New Zealand), G. Greiner, F. Fahlbusch Univ.
12.09.2003
26
Problem of inverse EEG/MEG

Direct Problem: Known: Sources (strength, position, orientation) Wanted: Potentials on the head surface Inverse Problem Known: Potentials on the head surface Wanted: Sources (strength, position, orientation)
12.09.2003
28
Challenges
Data acquisition MRI, CT, PET
segmentation, (co-) registration
EEG noise, filtering required eventually MEG and EEG data must be combined Forward problem: PDE with jumpy, anisotropic material parameters, high insecurity/noise in data and with singular source terms
dipole sources discontinuous coefficients anisotropic conductivities, from diffusion weighted MRI efficient forward solver multigrid Inverse problem - regularization (multi-) dipole model? Solution by optimization, which algorithm? Reduce large number of forward solutions reciprocity
Efficient Software is hard(!) to write

30
Computational effort for the bio-electric dipole localization inverse problem

Setup for optimization (using reciprocity) Compute 27 (#electrodes) forward solutions and store them (trivially parallel) Each forward solution based on 512 x 512 x 256 = 64 Mio. voxels/unknowns Is this a big problem?
Currently using (sequential) AMG from hypre package Better: Use geometric parallel multigrid
31
SR8000 Scalability
Parallel Multigrid for fast (forward) solution
Recall, SR 8000 has 1344 processors: could solve 20 billion unknowns in ~100 seconds!
32
Part IV Conclusions
12.09.2003
33
Why HPSim now

Simulation technology revolutionizes science and engineering Quantitative simulation of physical reality through simulation
Algorithmic modeling helps understand complex system behavior it is a new paradigm for model building and abstraction Predict Control Optimize
12.09.2003
34
Acknowledgements
Kompetenznetzwerk Technisch-Wissenschaftliches Hochleistungsrechnen (KONWIHR, Bayerische Forschungsstiftung)
Gridlib (grid data structures for Hitachi SR-8000) ParEXPDE (expression templates for PDE) FlowNoise (aero-acoustic modelling and simulation) FreeWihr (free surface flows for metal foams) Bayerisches Staatsministerium fr Umwelt- und Landesfragen, Project: Project: Natural Attenuation
Deutsche Forschungsgemeinschaft (DFG), Project DiME (data local iterative methods) More info: www10.informatik.uni-erlangen.de
Thanks to: B. Bergen, F. Hlsemann, C. Pflaum, C. Popa, M. Kowarschik, M. Mohr, J. Treibig, U. Fabricius, T. Pohl, C. Freundl, M. Kalmoun, H. Kstler,
12.09.2003
35
The purpose of computing is insight, not numbers is not yet in sight
12.09.2003
36
uyxw r u a Cfexg p zg d x x r {xw v {yxw g z r {y2ymf2Yw h uw ep w g w x x g z d {w w w u w } x w z d } o~x ea!| x 1 x yxw w { x f x w w w y z y v w s eutYr q p omk nl i g d jhfex yxw v w x w X t ` s W TucbHt YTr q phfecbaYW V i g` d ` X 7 A I U TS R 5 Q 9 3 A G 5 E 5 P 3 5 3 I 3 @ H7 G F7 3 E 5 B D 5 CA @ 9 86 5 4 3 5 B 3 7 ) $ 0 ) " & $ " 2$ 1('%#!
w w x w g 2|i ~2y z {eYx x q x z ui 21! e |Yfo g w w z x d x x w g x x %!1{ z x d x y w d e 2 f {w w hy2e z w w w t b ` i b t d cbfd ` r eCTt fw e e g z x d w y w hy2f{ w x d x x y z x z g x ~ { 2 efuw |x gCn x w p d i s t eTt Tt s X Tu TTr cb` HW b i t b s t W d g
x|y1 x hf ! uw z x d z g x x r i g x t1uwg( z h e1uw f d i y x i z d H nmx u{Cew w q yfd % r z d x r d r g n gf z 2fuw1 h hy2f ehfi x w x z w w x y w vx y v g fex v rt1yffyh ehyw2yg x z fy w x y g w |eCk x x x u2 uw w x hf hy2i p zw x w y g r z x d w i x i t j e | g x f z j(a fe~xgx ffa {H d g x i x g g x cbpum TTr bTY s Yfg Tjg ` g d s t W t g r` d s t W d
w y x g w x x w w d l hy2|ef2Y 2ee %u ( ( $ ''% " ''% " f z x z gg x x Cn z x d x z y %f!{o{ e e f '%
1fyh fy fC x k l %''% " g w g x xfuffl i` t TY uT` r` r b X r
d {w w | h {x h efw q xy z g z y y x x zw g 1 uw f! %eCx ! ug eeyg $0 u

m f f! w y x x hy2uex g 2h d
f wufh w x gw i fH t mYu (T# W ` d g r i g X t t
C !
1f uw z x Y a
w zw x y ehy2 ufe f(g n

z e%|x h q
tX b Yc V
b d i eTt s
i fr Y ` q ` d g ` s

b d i eTt s s jd X ' i fH t W b ` d g
i w x x 1ue1eef gfhh z {w w g w d 0 " 0 0 " " " " w w w x w z x d x d x z |x22w 21!e e f{ uw {eYx x q

$ 0 0 " " z x z gg x x Cn

bepuj|g TTr d i g d r gX dd s t W W j `d b t d r eCTt X YTY s t
0 "

z %|x h z f 1he x x

" " 0

0 0 00 0 0 0 | " 0 " 0 " " " |

" ""

0 | " |
0 0 e | " " | b ` d g d i i fH t Tt s
r x i x r x
$ %
! #
eh {| z ee { uw gf q w x w x d i g x
! "
|

t YcTr Tt s X b d i
xut( z %(fw u z g g x z x g z w d x d x d w g | 2ee { f{ hy2fw p

&
( 0
( )
'
ex x {w w x z d w d ruw v x w js x x y h {|y n 2 y2x w y d d z g w uT {{x fx
uyxw u {( w r x x fyxw gf2Y 2 w x x w w z y x d f f iw

&
w z g x d hy2e g e f2 w {u g w g x g x i gg x fi fuw g 1uwf z e~xu!x gw

&
X

tX d Yj
Multigrid Methods for Parameter Dependent Problems
Joachim Schberl o FWF - Start Project hp-FEM Johannes Kepler University Linz, Austria
Joachim Schberl o
Introduction
Page
Finite Element Method
Poisson Equation:
u = f in
u = 0 on
Joachim Schberl o
Parameter Dependent Problems
Page
Poisson Equation:
u = f in
u = 0 on
1 Hilbert space V = H0 () = {v L2 : v L2} Bilinear form A : V V R and linear form f : V R
A(u, v) =
v dx
and
f (v) =
f v dx
Variational problem: Find u V : A(u, v) = f (v) v V
Joachim Schberl o
Page
Poisson Equation:
u = f in
u = 0 on
1 Hilbert space V = H0 () = {v L2 : v L2} Bilinear form A : V V R and linear form f : V R
A(u, v) =
v dx
and
f (v) =
f v dx
Variational problem: Find u V : A(u, v) = f (v) v V
FEM: Finite element sub-space Vh V Find uh Vh : A(uh, vh) = f (vh) vh Vh
Joachim Schberl o
Page
System PDEs
Linear elasticity: A(u, v) = (u) : (v) + div u div v dx
1 displacement u [H0,D ]d, strain operator (u) := 0.5( u + ( u)T ) Lam parameters , . e
In principle the same
Joachim Schberl o
Page
System PDEs
1 displacement u [H0,D ]d, strain operator (u) := 0.5( u + ( u)T ) Lam parameters , . e
Timoshenko beam model: A(w, ; v, ) = + t2 (w )(v )
vertical displacement w, rotation , thickness t,
t w
In principle the same
Joachim Schberl o
Page
System PDEs
In principle the same but dependency on parameters
Joachim Schberl o
Page
System PDEs
Linear elasticity: A(u, v) = Nearly incompressible materials: (u) : (v) + div u div v dx
Thin beam: t
In principle the same but dependency on parameters
Joachim Schberl o
Page
Find u V : A(u, v) = f (v) with vV
1 A (u, v) = a(u, v) + c(u, v)
small parameter: symmetric bilinear form: Hilbert space: operator: with kernel: Assume:
(0, 1] a(u, u) 0 (Q, c(., .)) :V Q V0 := kern A1(u, u) u

2 V
uV
Joachim Schberl o
Page
FEM with Reduction Operators
The primal FEM Find uh Vh s.t.: 1 a(uh, vh) + c(uh, vh) = f (vh) vh Vh
often leads to bad results, known as locking phenomena. (One) explanation: Like penalty approximation to u = 0, but no FE functions fulll uh = 0, i.e. V0 Vh too small.
Joachim Schberl o
Page
FEM with Reduction Operators
The primal FEM Find uh Vh s.t.: 1 a(uh, vh) + c(uh, vh) = f (vh) vh Vh
often leads to bad results, known as locking phenomena. (One) explanation: Like penalty approximation to u = 0, but no FE functions fulll uh = 0, i.e. V0 Vh too small. Weaken the high energy term by reduction operator Rh: Find uh Vh s.t.: 1 a(uh, vh) + c(Rhuh, Rhvh) = f (vh) vh Vh
Large enough kernel Vh,0 = kernRh Vh Analysis: Equivalent mixed method
Joachim Schberl o
Page
Examples:
Nearly incompressible linear elasticity: FE space Vh...piecewise P2 on triangles Ah(uh, vh) = (uh) : (vh) + Rh[div uh] Rh[div vh] dx
Rh...L2-projection into piecewise constants
Joachim Schberl o
Page
Examples:
Nearly incompressible linear elasticity: FE space Vh...piecewise P2 on triangles Ah(uh, vh) = (uh) : (vh) + Rh[div uh] Rh[div vh] dx
Rh...L2-projection into piecewise constants Timoshenko beam: FE space Vh...piecewise linear for wh and h Ah(wh, h; vh, h) = hh dx + t2 (w Rhh)(v Rhh) dx
Rh...L2 projection into piecewise constants
Joachim Schberl o
Page
Numerical example: Timoshenko beam
Vertical load f = 1, compute w(1): Conforming FEM:

0.12 t=1e-1 t=1e-2 t=1e-3
With reduction operator:

0.12 t=1e-1 t=1e-2 t=1e-3
0.1
0.1
0.08 w(1) w(1)
0.08
0.06
0.06
0.04
0.04
0.02
0.02
0 1 10 100 Elements 1000 10000
0 1 10 Elements 100 1000
Joachim Schberl o
Page
Overlapping domain decomposition preconditioning
Split large problem into many small problems:
Joachim Schberl o
Multigrid methods
Page
Split large problem into many small problems: Preconditioning action w = C 1 d Block Jacobi preconditioner: C 1 =
i=1 T with Ei . . . N Ni embedding matrix and sub-space matrix Ai = Ei AEi m T EiA1Ei i
Joachim Schberl o
Multigrid methods
Page
Split large problem into many small problems: Preconditioning action w = C 1 d Block Jacobi preconditioner: C 1 =
i=1 m
Coarse grid problem:
T T EiA1Ei + EH A1EH i H
T with Ei . . . N Ni embedding matrix and sub-space matrix Ai = Ei AEi T Coarse to ne embedding matrix EH . . . N NH , coarse grid matrix AH = EH AEH .
Joachim Schberl o
Multigrid methods
Page
Additive Schwarz theory
Local spaces Vi, local forms Ci : Vi Vi R, embedding operators Ei : Vi Vh

Preconditioning action: C 1 : Vh Vh : d() w by
w=
Eiwi
with
wi Vi :
Ci(wi, vi) = d(Eivi) vi Vi
Joachim Schberl o
Multigrid methods
Page

w=
Eiwi
with
wi Vi :
To be a good preconditioner, the spectral estimate v is required ( v

2 C C
:= Cv, v
Vh Vh )
Joachim Schberl o
Multigrid methods
Page

w=
Eiwi
with
wi Vi :
To be a good preconditioner, the spectral estimate v is required ( v

2 C C
:= Cv, v
Vh Vh )
Additive Schwarz Lemma: v

2 C
inf
v= Ei vi vi Vi
vi
2 Ci
Lions, Nepomnyaschikh, Drya-Widlund, Xu, Zhang, Griebel-Oswald,...
Joachim Schberl o
Multigrid methods
Page
-Robust local preconditioner
Space splitting V =
Vi fullling the decomposition inequalities inf

uh = ui ui Vi
ui
2 V
c1(h) uh
2 V
u h Vh
inf
uh,0 = ui ui Vi Vh,0
ui
2 a
c2(h) uh,0
2 V
uh,0 Vh,0
Joachim Schberl o
Multigrid methods
Page
10
Space splitting V =

uh = ui ui Vi
ui
2 V
c1(h) uh
2 V
u h Vh
inf
ui
2 a
c2(h) uh,0
2 V
uh,0 Vh,0
Inverse inequality c(Rhwh, Rhvh) c3(h) wh Vh v V vh V Rhwh c h h inf sup
Joachim Schberl o
Multigrid methods
Page
10
Space splitting V =

uh = ui ui Vi
ui
2 V
c1(h) uh
2 V
u h Vh
inf
ui
2 a
c2(h) uh,0
2 V
uh,0 Vh,0
Inverse inequality c(Rhwh, Rhvh) c3(h) wh Vh v V vh V Rhwh c h h inf sup Then the (local) additive Schwarz preconditioner Dh fullls the -robust spectral estimates {c2(h) + c1(h)/c3(h)}1Dh Ah Dh
Related to Vassilevski-Wang, Cai-Goldstein-Pasciak, Arnold-Falk-Winther, Hiptmair, ... for H(div) and H(curl).
Joachim Schberl o Multigrid methods Page 10
Timoshenko beam splitting
Discrete Kernel:
w
wh h dx = 0.
w
Sub-space covering:
Vi

Joachim Schberl o
Multigrid methods
Page
11
Timoshenko beam splitting
Discrete Kernel:
w
wh h dx = 0.
w
Sub-space covering:
Vi

Point Jacobi:
1e+08 1e+07 1e+06 cond. num. cond. num. 100000 10000 1000 100 10 1 1 10 Elements 100 t=1e-0 t=1e-1 t=1e-2 t=1e-3
Block Jacobi:
1e+08 1e+07 1e+06 100000 10000 1000 100 10 1 1 10 Elements 100 1000 t=1e-0 t=1e-1 t=1e-2 t=1e-3
Joachim Schberl o
Multigrid methods
Page
11
Two-level preconditioner
2-level norm: vh Norm equivalence C

2 C
inf
vh =EH vH + vi
vH
2 AH
vi
2 Ah
Ah requires:
AH )
Continuous prolongation operator EH : (VH , .
(Vh, .
Ah )
Existence of continuous interpolation operator H : (Vh, .
Ah )
(VH , .
AH )
Joachim Schberl o
Multigrid methods
Page
12
-Robust two-Level preconditioner
Coarse grid bilinear form: A (uH , vH ) = a(uH , vH ) + 1c(RH uH , RH vH ) H VH,0 = kern RH Fine grid bilinear form: A (uh, vh) = a(uh, vh) + 1c(Rhuh, Rhvh) h Vh,0 = kern Rh Prolongation operator EH : VH Vh has to map EH : VH,0 Vh,0
Joachim Schberl o
Multigrid methods
Page
13
Example: Timoshenko beam
Coarse grid kernel function:
simply prolongated coarse grid kernel function:
locally adjusted to ne kernerl:
wH
wh
wh
Joachim Schberl o
Multigrid methods
Page
14
Fortin operator
Error estimates are based on equivalent mixed formulations. Discrete LBB condition is usually veried by the Fortin operator F : V Vh:
Continuous: Preserves weak constraints:
Rhv = RhF v
Joachim Schberl o
Multigrid methods
Page
15
Fortin operator
Rhv = RhF v
This is a robust interpolation operator from (V, F v h

2 A h
A )
to (Vh,
A ): h
F v h
2 a
+ 1 RhF v h
2 c
Joachim Schberl o
Multigrid methods
Page
15
Fortin operator
Rhv = RhF v

2 A h
A )
to (Vh,
A ): h
F v h v
2 V
2 a
+ 1 RhF v h
2 c
2 c
+ 1 Rhv
Joachim Schberl o
Multigrid methods
Page
15
Fortin operator
Rhv = RhF v

2 A h
A )
to (Vh,
A ): h
F v h v u
2 a
+ 1 RhF v h
2 c
2 1 Rhv 2 V + c 2 1 v 2 1 + c A
Joachim Schberl o
Multigrid methods
Page
15
Fortin operator
Rhv = RhF v

2 A h
A )
to (Vh,
A ): h
F v h v u u
2 a
+ 1 RhF v h
2 c
2 1 Rhv 2 V + c 2 1 v 2 1 + c A 2 A
Joachim Schberl o
Multigrid methods
Page
15
History
J. S.: Proceedings to EMG 96: Multigrid method with 2-level analysis for nearly incompressible materials and Timoshenko J. S.: Numer. Math. 99: Multigrid analysis for nearly incompressible materials J. S.: Thesis, 99: Multigrid method and analysis for Reissner Mindlin plates J. S. and W. Zulehner, 02,03: Iteration in mixed variables (Vanka smoother)
In preparation: J. S. and R. Stenberg: Multigrid for MITC and stabilized MITC
Joachim Schberl o
Multigrid methods
Page
16
Numerical Experiments
Two cubes, one nearly incompressible ( = 0.4999):
12288 tets, 28930 faces, 260370 unknowns
Joachim Schberl o
Multigrid methods
Page
17
Iterations numbers
Robust Multigrid (V-3-3): level unknowns = 0.3 1 0.5k 2 4.3k 20 3 33k 20 4 260k 21 Robust Smoother (3-3): level unknowns = 0.3 1 0.5k 2 4.3k 55 3 33k 98 Standard Multigrid (V-3-3): level unknowns = 0.3 1 0.5k 2 4.3k 62 3 33k 64 CG iteration, error reduction 1010
= 0.49 26 29 32 = 0.49 74 148 = 0.49 181 271
= 0.4999 30 36 42 = 0.4999 140 351 = 0.4999 1721 2000+
Joachim Schberl o
Multigrid methods
Page
18
Thin 3D elements
Reissner Mindlin reformulation for thin structures:
about 30 iterations of CG with multigrid preconditioning

Joachim Schberl o Multigrid methods Page 19
z | ~ t s } | y q t x r y {x r t z u q y q t x r w v u t s r q p o q t u x e v r d f w s qt n u y x u x t r w y x rt u s r x x x y e w t r y w m 8 l7 A k @ 6 j s w y v v w h i w w r q u t y w v q t e w y x t u q x x y u t er t h x v w q t r w r v w y x x u s d u s r t y w q t w r x d r u x y w u g x u u u t x y w u q t r w y x rt u x q w r t u q w u x wt y w x q x f qt d x e q s rt d u x y v y u s r x x x y y w t r y w v q p u x qt qt u x y v w t r w y v q t r w t t r v x s r y x r q x u x y v x y w u xt x r w y r u u x q w r t u r q xt x y x v w v ut s r q p fi hgfe d c @ 3 R S 7 6 b D 7 6 4 7C a ` Y XW D 8 V 49 R S 6 U9 R T BA 9 36 S9 7 R Q6 P I H G F E D 7 6 4 7C BA @9 4 8 76 54 32 1 0 )! ( ' & % $ # "!
q @ T D G 9 B 8 G 9 T 6 RG C 6 9 6 c B C B F 6 Q 9 C D b R 6 E V T 6 Q 9 6 C B P 6 9 E B F 6 Q 9 b D I9G 8 D H 6 X yD TG 6 Q 9 E B HH 6 y E B P E E 6 C 9 E H B c C D T 6 Q 9 b D E 6 VH B X 6 Q @ x i m l p o n m l g f g k j i hg f E B D 9 R 6 C C 6 b 6 C C 6 Q 9 C V bv H H B y C 6 F F V 6 Q 9 T D E 9 T G D F R 6 T 6 X 6 E 9 B E E 6 C 9 E H B c C D T 6 Q 9 6 C V E B 6 c 6 y P E E 6 8 D C F EG Q 9 U TG C V a @ yD C C B 6 U C BH 6 Q 9 b D T D G 9 8 6 C G R 6 Q 9 T G I9G 8 D H 6 X 9 T B 9 E T D 8 B 9 B H 6 T T B Q8 EG Q 9 Q U V D C Q 9 R 6 E E 6 C F EG 6 9 E B F 6 Q @ EHH B y RG UG C Q 9G y H 6 T T B Q8 H B 8G T D 8 B EG EG Q @ @ UG e TG T yD Q E EG 6 c 6 Q8 E 6 E D Q y 6 8G X 6 R B c D C b R 6 TG B 9 S D B 9 B R U TG E V IS R 6 9 B cG 9E 6 6 C B E C 6 9 6 c B C B F 6 Q @ d P T G HG B 9 6 R TG R 6 C 6 RG E T D 8 6 C B E T D G 9 B V 6 6 9 6 C 8 EG R C B 6 T G H T D T U T G 9H V E 6 C 6 Q 9 b D T DG 9 VH D E H B 8G C 6 c V T 6 Q 9 R T B c 6 9 E I E EG Q 9 b D T DG 9 B G 9 6 C 8 EG R 6 Q 9 C D b E R D Q 9 6 A @ 9T B 9E T D 8 B D 9 F V IH T D c 6 9E IE EG Q 9 IS R 6 T 6 R EG 6 C V E E 6 C F 6 Q 9 9 B Q 9 HH 6 y E B 6 9 D T D 9 IC B E E 6 8 6 T EG 9G IH H B T D G 9G R R @ c 6 H S D C F T DG 9 B 8 G 9 T 6 R G C 6 9 6 c B C B F U T G 9H V E 6 C 6 Q 9 b D 9 T 6 c 9 B 6 C 9 H B 8 G C 6 c V T 6 Q 9 D 9 6 U T 6 H H B Q8 6 H S B C 6 RG E T D 8 B E 6 E D F E T D G 9G R T D 8 I C B R T V D S E 9G Q 9G y C 6 Q 9 6 U D 9 c 6 9 E I E a 6 Q 9 b D I9G C B 6 TG H T D T 6 Q @ T DG 9 8 6 E 9 6 T 6 Q 9 TG R 6 S G C 8 E 6 R 6 8G X 6 R 6 Q 9 IS R 6 TG B 9 S D B 9 B R H B 9T 6 cG C 6 F 6 U TG E V IS E 6 VH BX 6 E 6 Q 9 R T D 9 EG 6 C 6 Q R 6 SG C 8 E 6 R 6 C V R 6 8 D C F T DG 9 B cG 9 E 6 C 6 9 6 c B C B F 6 Q 9 b D cG B 6 Q @ s r R T B P i P r P E C 6 9 6 c B C B F C V D b E 6 XH D X TG H 6 R D c 6H D Q y 6 Q @ x I C B R T V D S 6 Q 9 b D E 9 C B F 6 c D E T D R 6 E D F c G 6 S T B 8 E T D G 9 G R T D 8 I C B R T V D S y D 9 V D R T B y D T G E 6 R G E 6 v @ E H H B y T D x wv E T D G 9 G R T D 8 I C B R T V D S 6 Q 9 Q 9G y x v E a b D c 6 9 E I E 6 Q 9 6 XB Q 6 y E 6 9 E B F 8G c B C 6 8 6 Q 9 b D E yD 6 Q 9 b D T DG 9 FG C 8 E 6 R H B 8G 9 B c 6 Q 9 B c 6 Q 9 C D b P E V Q @ s r 9G cGH U TG RGH E B R T B i C D 9 8 B b U TG RGH E HH B y B E C 6 9 6 c B C B F C B H B 8 E H B T DG 9G R R B D y9 6 R VH 8 TG E T D G 9G R T D 8 6 E 6 Q @ I C B R T V D S 6 Q 9 T D R 6 9 8 6 C G R D 8 6 C B I9G 8 D H Y 6 X H B G 9T 6 U T B 9 6 Q 9 R T B E E 6 C 9E H BG 9T 6 U T B 9 6 Q 9 9 B Q 9 6 c V E E B 6 y 6 C 6 @ IH 6 XG 9 8 6 F E 6 C I C B R T V D S 6 Q 9 D 9 E C D 9 8 6 X 9 T 6 U T B 9 9G T V R T B H B c C D T 9G T V 6 Q 9 6 C B d R T B g 6 C 6 Q y x wv t u h d ep t sr q g epi h g f ed
c C D b 6 Q 9 E B Q P a ` T GP 9 B Q 9 T D G 9 G R T D 8 I C B R T V D S 6 F I9 Y C 6 G XB W B TG U T G 9H V E 6 C P U T G RG H E I S R 6 SG C 8 E 6 R E G EG Q 9 P IH H B 8 G F D 8 E D C 8 B A @ 9 8 6 7 6 ' % " ! $ # " ( ' " ( 5 3 4 3 & 2
1 0 !' ) # " ! % ( ' 0 0 0 & % ! $ # " !
s v s t ~ w s s t x x w u v w { x v s s t x z y u t ~ y w z t u w { x ~ t z x w v u w { x u w y x s t x x y t w { x w v w u t x u w v u t s v t { x w y s x t z w w u x u x ~ t s t x z w { x w u w u x { x w w u z u w ~ t y w z t u s t x y x t v w s u x s t z u w s z s t s z s t s w y v w x s t x v w z { z s t x s ~ w x w t w { x v s z w v t y w { x ~ t s t x x w u v w { t u x w u v s x t v w u u s t x z t z u w y s w { w w u x z y u t s v w u w y w { x ~ t w s w u w v w { x w x y t u v z t { v s z x w v w s w v w u w w u x w { x ~ t w s w u w v w { x x s x s t t x z s t w u x z y u t s w { x w s w v z w v t y w { x { t { x z s t x w v v u v s x v s s t x x w w { x } v w x u x v zz y u t s z xs w v s w w v s w t x v w y w u xs w y w u w y zz ~ s t x z w w s w u w v w { x u t ~ s t x w v v u v s x w { x w u w u w { }
s y
y wz t u s t x y x t v w s u x s t u w sz s t s w { x ~ t s t x z t w { x v w s y u w x w v s w { x w u u w x w y u x w u u t w { x s t w { x x u t s w x w u x w { x v s u v s t w { x t x u t x w z y u t s x s w { x w u z w x w w u v s w u w { } xs t u w w x w u x z y u t s w { x w x y t v s s t x w y t u ~ v s vz w w { x x w s w } u w x w y u w { x ~ t x w s w u w w u t ~ z t w s xz y w u x z y u t s w { x ~ t w z v w u w y w { w } u w w u t s t x w u v } t | s w u t s x s t x s w y w u w y w { x w x t s w
y u t ~ w { x s s t x v s t u v s t w { x v s y w x w { x s u y t s t x w z xs w u w v z x u w { x zz w x u } w } w v w v z w w u w u w { x v s vz w x t z w w { x w s w v y w x { x u w x w y u u t ~ ~ t x w u w w u t x s t w u w u v s s t x v s t u v s t x t z w z x u w t u w w t y w } u v s t }t | x t w { x s x s w y w w z t { } w { x s tz x tz w }t | s x s x s t w { x s ~ w x t z w w { x u t ~ s t x v s t u v s t x w z { u w t y w } u v s t } t | s w { x s s t x v s t u v s t w y w } z z } v u w { x t x s v s t w u u t u v s t w { x ~ t x u w { x s s t x s ~ x t v w z u w s w w { x { x } y w x w { x w t y w } { { } s t s t w u s t z t w { x u t u w x s w { x u w v s t w } w w v w { x w v s w x w { x ~ t }t | w { x z w v t y t x u w v u t s r
b H G 9 8 9 7 F
8 9 7 Ea 9 2 7
& ! ` ! ! # # 3 I Y X W ! ! 9 7 ! # ! # 6 F ( D A @ V 9 8 B A @ V 7 ! 8 TA @ E DA @ 2 R U 8 H S 9 R 8 H G 9 8 7 F 8 9 8 7 6 54 3
7 P Q F P G2 3 I 8 B A @ E D A C B A @ 2 F 6 @ E DAC BA @ 2
2 ! ! " 1 1 ( 0
! ) ! ( ! ) ! ! ) ( " ! ! ) ! ' ( " " & % ! & " ' " !
& % % ! ! " " & % $ & % $ # ! " " !

l ptp e i
u r e u s iv e g p s r q uh l
s r x s r t r p e u d v s r x g g p p e u d v g s vx u r y x fv t w w r p t r g p q x t u r y p e d g s v x u r q x w w r r p t sx p u r x q p t w w r i e x e g x iux t r f v y d v s x g g x e d u t r w s v x u r x y x u w v s r s r u t r w n u s x v j r s r n r v u s x i p x e d v u s x v w s v x u r u s p y p w y x s r g us p y p t q sx t p u p y r t r w s r g us p y p t q sx y v t d p w f v q p p s r q ux p u r u g e ux i g p r x t r e d v g p w iu e u v t p ev i g f v p s r u f y x g g p u r t p ux y e ux t v r p e u e f v e u r u r e u g p e t p g v p s p u r t p ux e u p e u u r s vx u r f r e p s r p y g p qx s h
sx y
z u | { u ~y { v ~ u w w { z ~ v z y z
wz y z y x w v u t y z ~ | { | z x | z | { y ~ u }| { z y x w v u t
s x r sx p eq u pg g gx v e u p y n q p o n r f u w p q s v q m l y e ux t v r sx iv v d p e u sx r u v p i h w v v t p s sx p e u sx s vx u r t p ux s v u i p k p sx g p s v u g f j sx y t v dt p w i w v v p ug p s sx s v w g p t t v q p e u w f g r p t p i d h eq r v t w w r fv gq r p e u sx e ux i v e u p y r d v p w y r fp s r gx t p s sx s r p x ev t w v u g r e p s v p f re u s p ugx g s v q r sx r u v v u t p t v sx p t f p q v t w s vx u f v g p f re t p u p y r t r w eq r p t v d p t p e i h s vx u r f y t v d y p
v t w u r e u v u p f x s eq p u sx y y r t v t w qx u r t r f rx us p f p g r s vx u r qx w w r p e d u g sx y
g r y p vt w s vx ur x yx u w v p e u p u r ug pt s r q p i h g fe d
V " P ) % V ' c " $ P % c `' " 0% Q $ ( ( " P % " $ % P " $ c b P # $ W P $ & % " $ % g T S R " ' " 0 ' ) $ ( $ " %' `' ' ( 0 P " $ & W P " $ ( ( " T S R Q ) g EP " ' ) " b " " $ P I y x w v u X p i % 0 ' ) $ ( % " d % 0 % ( " ' ) % ) $ V $ & $ ( $ ' " $ " 0% # % V ( 0 P V # " "W " $ " '% 0% R X 0 ' ) $ ( $ " c $ & & % V' $ & ` $ " P 0% ' ) V `' " 0% Q $ ( ( " `' % 0 ' ) $ ( % " d % 0 % ( ' c W' ' ( 0 " Q % % V ' ) " P " $ ( ( " V " ' V 0% " " 0 P % c b % P " 0 0 " P E ' f $ V $ ( " `% P ' W & % " V ( 0 P ' # % " b % $ " ( 0 P " t X P E " $ P g $ ( e % c # % I I s R " 0 $ & $ ( ) W" % " V ( 0 P Y X $ V # f % P% ( ' V $ ' % ` 0 % $ # ' " # % ' V $ Y X p i % V P % $ " 0 ' ) $ ( " ' % " r X P " $ ( ( " T S R P V $ % % c $ W' $ " % ' P " Q % % c ( P " W b % P % V P % ( ` & % " P % ' ( ( " % c # % ' " I X q p i X & Ph ` $ P # $ W P # % `' $ V % $ W% # P% c b 0 ' ) $ ( % " d% 0% ( % " $ g P ' ' " $ W $ & 0 T S R & " % $ " W P f% P ( $ " 0 T S R P V e X P " & % d% " ( 0 " c c &% 0 T S R P V $ `' ' " % $ " ( " 0 % ' ' " P c b % " $ P W% P " ' ' " & ' $ E ( V P V $ % 0 ' ) $ ( P % R X ( ` ` $ a & " ' V 0 $ & " ( V ` ) `'' " P% ( ` " 0 % Q $ ( ( " $ " $ % % V ) `' P " Q ( V '% V ) % " % U P V $ Y X 0 T S R P V $ " '' " P % W ) " V P% 0 ' " " " % U P V $ '' " P % " % U $ " ' 0 T S R & Q P I H G @ 7 9 7F 3 7
& ' $ E & % " ) "
A D 6 C 5 32 1 9 @ 7 8 7 4 B
" 0 ' ) $ ( ' " % #% $ & " % # " $ # " !

e j ~ j j d ~ d m j ~ m i j h g g ~g i m m j d e j m e d j i d j m j j j d j ~ e i d m j j m j ~ i j ~ j j e i sr s r
m m j d f m f e d m d j j j f ~ d j m e f j ~ j j g j m j m j j g g i j d j f i d e h g d j m j f j m j j d m j j m m g f } m d ~ d ~ e } ~ j m d m j j | g j j j f i j ~ j m e i j d h g m ~ j j ~ d j m j e m e ~ g ~ m ~ m e d e g g g g f ~ f | g m f d j j ~ j e } d ~ j | g m m d f m j m f j j f i g e j ~ m m i f i ~ g m e m e m e m j f d j j m l e m f m ~ j m d ~ } m m |
v xp t r s x y { z y r x w v u ts r q p o
g m f l f e j k e j e i h g f e d
21 76 5 76 5 ) 0
4
3

)
(

21 ) 0
' & % " #
"

h G 9 @ T U T C y T F E9 T W T U 9 9 C D C R C G s I C Q E D C 9 C T c ` h bX R Q 9 E Rl T E 9 R F D E Q I U @ C R Q l h QU y R U EQ I U @ C C G E T F H 9 T C QI Q EI @ q C S R U T S U C G E C T U r y RU r c CTC
` Y bX
f k f f k f f f ` X ` X jf gf h ` X iX ` `
` X X h i h `r X ` rr r r r X
h 9 Q ` bX a ` d Y X q C I A F T G E Q P y C E U Q D F 9 9 U x ~ y } | x { z x u y v x u w v u t C G s f i i h II U E U RF yRC C y EF R 9C F y y RU R Q T U C RQ I 9Q R F Q E D R @ H C Q E D C B A F C G E c U E U y R C Q S C T U ` oY X R Q r q 9 C @ I U C G E C D R Q p ` oY X H F R F Q E U I @ q T F H C T E D U T E 9 A U CT F q g U E 9 @ B 9 Q ` d Y X R F Q E D R @ H C Q E D C B A F C G E E U G E C E F n T F E D C y U F I C E C T k9 Q y C G E i D 9Q y R U Q T E U q 9 9 C R Q E 9 E R C q CI C C EQ R W C G E 9 Q c C T C m`Y r r YX CTC G P h ` bX
h l f k xw s f r h l k ut s f f
h i jf e ` X i r h i gf h ` X e FE EDC BA @9
` dY X
h r ` r X RQ q
f EQ Q E D @ y R F D y C Q E C T D 9 Q y hC G E G E Q P T F E D C C I A U Q T U C C T H C G E y R U CI A U Q T U C E U E9 y C Q E C T D 9Q y C G E G EQ P T F E D C C G E HQ E R C yQ C P C T C G P q CI A F T S RQ q q U T S F T T U C RQI R F R I U R FQ 9 R C qQ y C EQ R W S RQ PFII F H C G E E C S C P T C E C a q U T U R F Q E U Q E C T D 9 Q y U G EQ P R Q U q F y C G E H F R F Q E U Q E C T D 9 Q y E R C q CI C C EQ R W U T C E H y C W 9Q E U 9 I 9 @ FQ A F 9Q E R C q C TQ @ C T 9 Q G E C 9 U D T @ F R CI A U Q E R C T C Q y I 9 @ F @ R Q E R F D C D Q PE C A F E y C 9 F a @ 9 C T U 9 E R Q U T E 9 R F D E Q I U @ C R Q C G E y R U R F Q E D R @ H C Q E D C B A F C G E c 9 q C I A F Tp S RQ q q U T S F T T U C RQI R F R H F 9 R FQ E UI @ q T F H I U T C R C S R C @I U W x wR Cf Q S y Cu t s fU 9 Q s y R U EQ Q E D @ y R F D C G E T F H 9 EQ qQ I C Q EQ 9 F R C Q S Q T F Q T U C T U yRU c CTC xw s f v f v ut s f p i hg f e d c ` b Y X a` Y Y X 9 C W 9 Q E U 9 V 9 ERQ U T E 9 R F D S RQ PFII F H C G E F E E D C B A @ 9
` E R Q U T E 9 R F D F A E Q Q E D @ y R F DX ` Y X r ` E R Q U T E 9 R F D 9 9 U qX
Xk
j X
b iD g g g f f f f f f f hD f a f f f f f f f f f f f WD f f f ' !f f f f f f D f f f f f f f f f f f f H D f f f f f e ED e e
VR G G G d' G T G R G G G G G G G HW G G G G G H E ' WH H HH EH G G G E HE G
C B c 4 X ` 4 7 @ d # 8 @ 4 @ 4 7 7 4 8 4 ` 4 u w i q u w t x g e 4 4 X X ' !
d 4 B c 8 4 `4 @ ' # 8 4 B ` 9 B @ 8 7 8 7 `9 8 4 $ @ B 2 d b a ! ` 9 4 v s x h r f g x q r g wf 2 d 8 9 @ 9 B 4 A 8 @ B 4 b a ! ` 9 d b a 4 7 ` 8 @ 8 7 9 B @ ` 9 4 B ` 9 4 8 @ 4 B c 4 4 @ @ 8 @ ` 5 4 9 4 8 @ 5 B 5 8 @ B X 1 U # ! bV R
T #! a R
v s x h x q v sr y h x g w h v f t f u e t s r qf p g i h g f e ` 9 X X 4 7 ` 8 @ 8 7 4 B 4 9 B 4 ` 9 A 7 7 4 8 @ 5 2 d 4 4 U V C ` B c b a 4 @ B 8 4 4 ` 8 T S R # U T7 T S R # 4 8 4 Y X X X 4 7 4 4 @ 8 7 & 1 G ! 0 ' ! W D Y X
G !
G ! ( ' & ! D U V U T T S R # $ T S R # ' D $ Q P H I ! H D G ! % & $ F E D ! E D
B 4 4 8 9 @ C 9 B 4 A 8 @ 9 7 8 4 7 2 6 6 54 3 2
1 0 ' ) $ ( ' & % $ # " !
q u w m r n z n x r s o n rx w r u n o o n v q t s o r t n n q s m t w x r t x r q x u q x m r | t p u n m q x t x r q x u n o s q u w m r n z u t s r n o x u n o n m l u n x s n w q s t s t w x r w q n m r o w v q u w m r n z r n o x t x u n q w w o t n n n s m q z n w o r t x w n u u s q v w t w x r w q n m r o w v z m r x o w s s l s n m r t w | n | u n q s q u w m r n z n x r s o n r p u n w q n o s q t w x r s n q n w r o nx s u t s q n w r t n m l q rtx w n u u s q v w t wx r s r z w n m r tx n qx o s q t wx r s n o s n tx v w q z n r q y q n rx t n u t x o s x z x n rx t n u t x y s x yr qx y { x o r s z n m r n r w t n u u t s r x o q q n m r r x z w n l t w l w t z w o v {x o r s z n m ~ qx n ux q u t s m rm x o n m r o w v yo rt n u n x u w z n m r u t s } ~ y u n s n o lw t qx v w y o rt n n m
z n r q y q t x l w w v n m r v w n u x q u t s m r m x o n m r q x u t s { x o r s z n m r q x n o n m l ~
~ y u n r w t n u | ~ v w q r t n z n o t x n m r o w v z n r q y q o s n t x t x l w w v n m r r n n l | v w l w o u t w n q n m r t x t x r r x r q ~ } ~
| y n z s t | v w q lw o m r u t s m r n m r z w o v u t s o w v q rt n z n o tx n m r n r s tx zx n n l lw o w r n l w o n t w r q q x ~ u t s {x o r s z s t w s x u s q x | t w x r s n s x r t n r w x o r n n n m r v w {x o r s z q q n t x r q n m r q x ~ r s m r n r w ~ ~ ~ ~ ~ ~
q t wx q q n o { n tx lw w v n m r u t s y t n x n o s v w q n x o r t n w o n t w t t x t x s z n o n m q n x o r s z s t w s x u n o s s x u ~ u t s s x u ~ } | { x o r s z y r x r t n u x n m r o w v q u t s r q p n o n m l
% D " ( 4 32 7 7 6 5 9 9
" 4 3 3 2 " " ! E 7 7 6 D " & B C C ! ! " ! % ( " 3 3 & A ( B ! ( % @ " (

776
" 4 32 ( 43 3 2 ( % 1 " 0 ! & ! & ) ( " " ( " ' % & % $ ! # " !

I e e i Pd P I i I V a i I X T W R V Y U a X X T U g a q W b R T b G Y U ` S U a b a G S R G g G S a T R hX V R qG S q R U I V ` hq I WbR i XG R I i UTbSUR H I i q a I I P
I d i d d e d i d V a i I R G W T g U R T b G V b a b b G X S b T U U ` q W W T W S b TG X b R U S a b aG S RV ` qV R q T h S U a X b aG S RV ` g U a g T X ` aG U R I XG T U I W b R a UG I
I e d i p p p P p I i p V a i I X g T V Y a U S b G a T V W W R X H bG V a X U a X W a h S T g T G S R U T SG a X X RV q I S b R Sb T U T X I W b R i S U T V T I i b R I e e e I d e d I i
V T W a g X Sb T U U ` q W W T a b aG S R q G S X ` q TV T W T I
a i e p I V a i I X b a G S R ` T V V T R g T h S U a I WbR i R ` I iGUR g g I
rl n r r r
t u w v w u s u t u y x wv uy d d y x t y ux w s t s txv w f y f y w x w x t u y x t u w u } t y v u x v x w v u x y x w v u t v s t s t v u t u x w w e u x v w w f t w u x t w u u v x w x y t d w y x t u w u t w ~ } { | {z x w v u t x t t y y wf u yv xv s t f xt x u uw y x yt t y yv w
v n r u r t s rq po n ml k
j d i xv y us y s t v x f w s t h x t d xv ux y s x w y s v t u w w f g w u f e y x w v u t u v v s d t w u y s w y v u t u v y t y s u v t y x w v u t u v t y s s x y u u uv y x wv u t s
r X S q R S b a q p i h X T g e ed f e ed c b a G S ` Y G U S X G W V R G U T S R Q I P I H G F
1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0
10
20
30
40
50
60
70
80
90 100 0
10
20
30
40
50
60
70
80
90
100
cb

"
P SB
$ E
P
a ) # 7 "
% $ % ` @ B
V

Y X A
P

3 ) 0 3 W
B A V 3 3
U %

T P

% " & $ P S B 3
7 & & & % " % % " % B
' &
R C B
C

P I F
H D
'
2 $
C B C B
' 8
""
3 #
G F
E D
A @ ( 9
& & % 3 ' 8 7
6 5 4

3 1 2 1
0 '
) ( '
% & & % $ " # "
!

Multigrid Computation for Variational Image Segmentation Problems: Euler equations and approximation
Rosa Maria Spitaleri
Istituto per le Applicazioni del Calcolo-CNR Viale del Policlinico 137, 00161 Rome, Italy e-mail: spitaleri@iac.rm.cnr.it
Advances in Numerical Algorithms Graz, September 10-13, 2003

Istituto per le Applicazioni del Calcolo "M. Picone"
Variational Image Segmentation and Computational Approach

minimization of the Mumford-Shah functional definition of a sequence of -convergent functionals solution of associated Euler equations finite difference approximation nonlinear system solution multigrid computation geometric and synthetic images visualization : computed solution (reconstructed image and edge ), convergence histories
Segmentation Problem
appropriate decomposition of the domain of a function f(x,y) (computer vision ) striking a plane domain at the point with coordinates (x, y) the function f(x,y) is called image
f(x,y) is the strength of the light signal
Discontinuity Causes
light reflected off surfaces S i of solid objects O i , seen from P , the camera or eye point, will strike the domain (retina or film) in various open subsets R i which could have common boundaries (edges of the objects in foreground), surfaces with different orientation (edges of a cube), discontinuity in illumination (shadows), textured, partially transparent, highly-reflecting objects, ...
the segmentation problem consists in computing a decomposition of
D = R ... R 1 n such that

the image f ( x, y) varies smoothly and/or slowly within each R i the image f ( x, y) varies discontinuously and /or rapidly across most of the boundary between different R i
computing optimal approximations of f ( x, y) by piece-wise smooth functions u(x, y ) (restrictions u i to the pieces R i differentiables)
Mumford-Shah Functional
Given the image f(x,y) let u be a real function defined on a domain and
(u) (u) D = {R , ..., R } a decomposition of n 1 n (u) = R i Su , such that

where S u = and
i =1
( u) i
i=1
( u) (u) i the boundary of R i
MSF Definition
the MSF is defined in the following form:
E(u) = (u - f ) dxdy +
2
where H1 (S
/ S
u
2
u dxdy + H ( Su )
1
= ( u / x ) + ( u / y ) u ) the Hausdorff measure of S u , > 0 assigned parameters.
, ,
MSF Minimization
approximation of f(x,y) by u
(u - f )2 dxdy
(u) u smooth on each R i 2
/Su
u dxdy
the boundary possible
( u) S = u i i =1
as short as
H1 (S u )
Parameters
, > 0 can be calibrated to eliminate
false edges , created by noise, and save the actual image;
is a scaling parameter,
controls the noise
effects;
defines the threshold to detect the edge
Interest and Expectation

droping any of the three terms inf E = 0 : u = 0, S u = , s without the first: s without the second: u = f , S = , u
s without the third: u = mean
u i D N 2 small squares
f ,S
N N grid,
(u,S u ) is a cartoon of the actual image f(x,y) :

s a new image in which the edges are drawn sharply and precisely and the objects are drawn smoothly without texture, s idealization of a complicated image, representing essentially the same scene
Variational Convergence
the problem of minimizing E has been conjectured to be well posed (open problem), these functionals have minimizers in the spaces of Special functions of Bounded Variation (SBV), the minimization problem is difficult for the presence of the set of discontinuity contours S u as unknown variational convergence to solve minimization of functional depending on discontinuities: approximation of a variational problem by a sequence of more tractable problems
-Convergence
the sequence of functionals {F (u)} on a metric space U is -convergent to the functional F(u) if, u0 U : (i) sequence {uk} converging to u0
k
lim inf F (u k ) F(u 0 )
(ii)
a sequence {uk} converging to u0 such

that
k
lim sup F (u k ) F(u 0 )
Properties
variational property- {F (u)} sequence of functionals on U -convergent to F(u) k {u* } of minimizers of {F (u)} -if a sequence k converges, then the limit is a minimizer of F(u) and {Fk (u* )} converges to the minimal value of k F(u) k the -convergence is a variational convergence -convergence stability under continuous perturbations- Let G be a continuous functional k F Fk + G F+ G G F
-Convergent Functionals
sequence of -convergent functionals
E ( u, z) =
k
(u - f )
2
dxdy +
u dxdy +
z 2 k(1-z)2 ) dxdy + 4 k
? E k (u, z) = (u - f)2 dxdy+ E k (u,z) or

k E stability property: E
Discontinuity Curves by Control Function

the function z(x,y ) controls the gradient of u and has values ranging between 0 and 1, z * is close to 0 in a the minimizer k neighbourhood of the set S u , which shrinks as k , and close to 1 in the continuity regions u* thus is permitted to become the gradient of k arbitrarily large along S u (jumps in the solution) z * converge to a function equal the minimizers k to 0 along S u and 1 everywhere else : the -limit E(u) does not depend on z
Euler Equations
minimizers of
E k (u, z) are the solutions of the
following coupled Euler equations
( z 2 u ) + ( z 2 u ) = 1 (u f ) x y x y
2 k 2 (1 z) z = k z u 4
Neumann boundary conditions local minimum of the associated functional more accurate approximation as k increases
Finite Difference Approximation

the discretization of with grid spacing h andthe h finite difference operator L a ( complete c approximation of u and z )
1 + 4 h2 k z 4 h2
4 z 2u z u 2 ? h 2 2 ux + u y + 4 + 2 h
1 z z x ux + z y uy = f 2 2h k2 4 ? k2 = 0 4 h2 z 4
? w = 1 4 (w(x + h,y) + w(x h,y ) + w(x,y + h) + w(x,y h)) wx = w(x + h,y) w(x h,y) wy = w(x, y + h) w(x,y h) where w = (u,z )
Solution Algorithm
equation systems on the grid G and covering the domain :
M
, with mesh size h
L w =F
M ac M
onegrid ( l = M , l is the grid level ) computation of the solution
* 0 w k = RelaxM w k ,L a ,F
Gauss-Seidel relaxation rotated lexicographical ordering: a grid point (x1 ,y 1 ) precedes another point (x 2 , y 2 ) if and only if
x1 < x 2 or
system associated to the functional E (u, z) for a fixed value of the index k observations: the discretization step of the finite difference method should decrease as k increases optimal choice of the parameters and is a delicate problem
0 w k initial guess for GS-relaxation applied to each k
{x1 = x 2 , y1 < y 2 }
Image Segmentation
f ( x, y ) a given image, and S u = {( x, y ) : z ( x, y ) = 0 }
geometrical and realistic problems:
one or more squares and circles, a vase
u* and computed results: smoothed image k z * (discontinuity contours) control function k

convergence histories: residuals, norms, logarithm values, iteration numbers result visualization
Experimental Evaluation
experimental choice of the parameters: = 3. , = 0.05 image resolution: 64x64, 128x128, 256x256, 101x101 brightness measurements : 256 levels 0 0 initial guess: uk equal to the input image, z k equal to 1 everywhere even small values of k can be used
k=2
k=3
k=5
Lh a
Lh Circle 4x64 a 6 c
log( resz )
log(resu)
k and h link
S u discontinuity set, we can define z(x,y)= ((x, y)) in Bk = { x,y ): (x, y) < k }, ( where (x,y ) is the distance of (x, y) from S u and kt k (t) =1 e 2 (convergence in Bk ) z k (x,y ) between 1 and 0 in [ 5 l,5l ] , l = 2 k : we have mh =10 l = 20/k where m is the node number in this interval for a given h by p = hk we can control the gap approximation
Conclusion
We have defined a multigrid finite difference method able to improve numerical solution of Euler equations in variational image segmentation Application to segmentation problems shows the capabilities of the method in computing solutions and providing satisfactory convergence histories Future research deals with improving performances of multigrid computation
Multigrid Computation for Variational Image Segmentation Problems: Multigrid approach

Rosa Maria Spitaleri Istituto per le Applicazioni del Calcolo-CNR
Viale del Policlinico 137, 00161 Rome, Italy e-mail: spitaleri@ iac.rm.cnr.it
Advances in Numerical Algorithms, GRAZ, September 10- 13, 2003
A few research topics

numerical simulation (PDE approximation) linear and non-linear multigrid computation applications: grid generation shallow water equations image segmentation numerical results visualization interactivity
Multigrid Computation
In the multigrid computation a system of finite-difference or finite-element equations is solved on a given grid by continuous interactions with a hierarchy of coarser grids Main components of the multigrid computation:
relaxation for error smoothing restriction (fine-to-coarser transfer operator) prolongation (coarse-to fine transfer operator) MG cycling ( cycling by grid change)
Let G , G , ... , G be a sequence of grids, all approximating the domain , with mesh sizes h0, h1, ... , hM such that h0 > h1 >... > hM. Let us assume the system Lw = F in , w = on and the difference equations along with the associated boundary conditions L w =F
M M M
on G ,
w = on
M M M
Full Approximation Storage ( FAS ) algorithm for non linear problems The general form of the correction equation is Lw = F ,
l l l l+1 l l l
w =
l l l
where -L wa
l+1 l+1 l+1
F = L (I l+1 wa )+I l+1 (F

l l l l+1 l l+1
l+1
) ),
= (I l+1 wa )+I l+1 (
l+1
- wa
l+1
wa is the computed approximation on the l finer grid and I l+1 an appropriate fine-to-coarse transfer operator (restriction)
In the fine grid correction step ( wa ) new= ( wa ) old + Il where Il

l+1 l+1 l+1 l+1
(wa - I l+1 wa )
l+1
is an appropriate coarse-to-fine
transfer operator (prolongation)
In the FAS algorithm the idea is to store the full current approximation wa instead of storing just the correction
l
FAS two-grid method

One iteration step of the (l+1,l) two-grid method, l +1 l +1 which computes w new from w old , is composed of the following steps: Pre-smoothing l +1 ? compute w by applying 1 (0) sweeps of a given (non-linear) relaxation method with starting l +1 guess w old
l +1 ?l +1 = Relax 1 (w old ; Ll + 1 , Fl +1 , l +1 , l +1 ) w
Post-smoothing
compute w by applying 2 (0) sweeps of a given (non-linear) relaxation method with l +1 starting guess w ?
w
l +1 new
l +1 new
? = Relax (w ;L ,F , , )
l+1 l+1
l +1
l +1
l +1
The non-linear relaxation procedure, suitable for error smoothing, is applied 1 times before and 2 times after the coarse grid correction step
Coarse-grid correction
?l +1 = F l+1 Ll +1 wl +1 ? computation of the residual f ?l = I ll+1 f l+1 ? f restriction of the residual restriction of the current l l l +1 w = l+1 w ? ? approximation computation of the l l ?l + Ll wl ? exact solution of L w = f l l l V = w w computation of the correction ? l+1 l l +1 interpolation of the correction V = I l V computation of the correct l +1 l+1 l +1 w = w +V ? ? approximation
Multigrid Components
coarse grid construction (hierarchy of grids): meshsize doubling relaxation for error smoothing ( Relax , Relax ): Gauss-Seidel relaxation-non collective relaxation ordering : rotated lexicographical multigrid cycling : V, W cycles
1
2
restriction ( fine-to-coarse-grid transfer l l operator )( I l+1 , l+1 ) : straight injection, weighted --> nine-point prolongation (coarse-to-fine-grid transfer l+1 operator ) ( I l ) : piecewise linear interpolation, weighted-->nine-point
Finer grid
1/2 1/4 1/4
Coarser grid
1/2 1/4
1/4
References
R. M. Spitaleri, A Scientific Computing Environment for Differential Modeling, Mathematics and Computer in Simulation, 62, 4: (2003) R.M. Spitaleri, R. March, D. Arena, A Multigrid Finite Difference Method for the Solution of Euler Equations of the Variational Image Segmentation Applied Numerical Mathematics , 39: 181-189 (2001) R. M. Spitaleri, Numerical Insights in the Solution of Euler Equations of the Variational Image Segmentation, Non Linear Analysis , 47: 3333-3344 (2001) Spitaleri R.M., Full-FAS Multigrid Grid Generation Algorithms, Applied Numerical Mathematics 32, 4: 483-494 (2000) R.M Spitaleri, R. March and D. Arena, Finite Difference Solution of Euler Equations Arising in Variational Image Segmentation, Numerical Algorithms, 21: 353-365 (1999) R.M. Spitaleri (Ed.) Special issue on Applied Scientifc Computing-Grid Generation, Approximated Solutions and Visualization, Applied Numerical Mathematics, (2003to appear) R.M. Spitaleri (Ed.) Special Issue on Applied Numerical Computing: Grid Generation and Solution Methods for Advanced Simulations, Applied Numerical Mathematics 46, 3-4, ( 2003)
Organization 2003
MASCOT 03
IMACS/ISGG Worshop 3rd Meeting on Applied Scientific Computing and Tools, Grid generation, Approximation, Simulation and Visualization FORTE VILLAGE RESORT, Sardinia, October 3-4, 2003
in TCN-CAE International Conference on CAE and Computational Technologies for Industry, October 2-5, 2003
EWM2003-Annual Conference of EWM-European Women in Mathematics, Luminy, France, November 3-7, 2003 Session Numerical Methods
@ A9 R SC ( 1 TB A4 C ( 0 E 1 V WU a bY` D dc efg yx Y f r q g dh pi qrf e h cs t Ys u vh w X B4 15 C B H 5 1 4 I PC 5' 5 4 (G ( 78 %& 0 FE 12 5 64 3 $ B4 DA4 C 4 (0 )' " @Q

! " #
H @C
e f
du
g j hi
c d
| x @ @ 5 I0 @ @ ( H0 2 ( 5 64
l opn q rm j @0 x 1 H 4 8 C I0 x H0 B5 @ @ 0 4 1 I ( x H x 1 5 4 E x 1 ( 7 4 8 @C (5 @ 5 8 { D 1I R SC ( 8C ~ @C (5 10 E @ w x 0 w t { @ 4 D@ C 0 ( I0 x I0 z r x { | } ~ t ~ u 6t m
v x yw
d s g m m
0 D 20 1C 5 H 8 C 1PC 5I E 1PC I 04 @ 1 ( 12 @ I ( 1( D B 4 B 8 B D 1( B 4 8 ( H H ( 0 1 R B 8 ( H 0 B 8 12 @ I ( ( ( 8 R @ ( 8 C 8 ( 5 @ ( I0 7 @ H( 4 H B5 7 7 D 57 7 I0 8 ( @ H( ( H 5 @ 8 1 4C G H B5 @ 8X 20 H( 1PC I5 B5 8 ( @ D 04 0 8 H 07 5 8X 1 1 4C G H 5 B64 ( BC D X @ |
0 @ 0 I 8 1 ( 4 H B 0 8 4 l 5 64
q j qg j m
I D8 0 BC @
@C 8 @C (5 1 (
58 B4 IPC T A4 C
x 2 @ 4 @ 4 ~ B B ~ | ~ V 4 @ | 5 1 @ ( 7 0 D 5E B5 I DB 5 1 2 E x 0 4 @C #| 4 (G 4 x x B (7 @ ( x z 57 4 B @0 (7 1 1 w C w T 0 E t @ 0 w t w } 5 B64 ~ } ~ } 1 } 5 T 4C (G 5 H I0 @ 1C 4 5 x H( 7 (7 C 1( 4 (G B4 ~ T | R SC ( @ (G (G 1C 5 4 4 G 4 7 1 H 5 4C @ T 0
@0 ( T
( 4
d B4 @C t V t w gj ~ t w 5 @ 8 7 S t d m m i ~ ~ j i m k m t
e f g j hi
4 (G 1 B4 ( @C D I0 @ 5 64 I H ~ 1 4C 4 ( @ | 5 B 5 4 15 8 8 8 D 5 T 4C 0 E H 8 1 @C ( B G R5 15 C 04 C @5 x x 1 ( H @C 8 @C (5 @C 58 B4 IPC T A4 C C w 4 (G 8 @C (5 10 E 1 ( 0 8 @C 4 t @ 7 ~ T ( 4 R5 SC 7 ( @ 8 @ 4 T8 0 E 4 G D B I ( { 7 C x @0 (G ( @ 4 1 2 R SC ( @ ~ 5 x I5 E 7 @ 0 (G 7 @ @ @
0 5
( (7 @ ( 4 ~ (G 8C @C (5 10 E 1 w H
0 E B I f I @ ( x 0 58 ( D 1 5 04 C @ x @ R5 15 C ( 1( B5 4 (G 8 x T R R5 C B5 ( 4 04 1 0 x (7 4 (G 7 C 0 15 1 8 8 ( I 4 1C U( 7 I E 5 5 G B ~ 5 4 4 G 8 1( 8 I E G C 15 x 8 ~ 0 E D B H D 5 ( 4 0 @ x 1 D2 ~ 1( 4 x 0 E @ @ 7 0 4C C ( 4 I G 4 T8 (C @ 58 0 ( B D8 04 C I0 (G B4 C (G H 1 (5 D 4 5 8 T A4 C C (
E 3 0 1)
0 1 @ T8 @ ( 5 0 E D H 4 C 8 (C I0 R B 12 ( B 04 H 5 @ ( 1 8 x @ | x 7 I ~ B @C 7 @C 2 C @ 5 T C 04 ( H 12 D8 B4 0 3 ( 0 R @
('
@C 58 I ( H 1 ( 8 4C
0 0 1(
04 @
@ A9
5 SY) E 43 2 !
1 1( B0 @

0 4 9 @
G
5C C 2 `H @ B C2 C2 0 70 53 @ U B dc ) f ge h 7 G
i
5 786 9 @A43 9

4C
a3
2

H 5 0 ~ 4C
7 C
B ( B0 G
B
D EG F) &
G
@
1( 5 4 C @ 4
CH 3H b 4 Q P
2 I 1H

( 8
5 4 C 4
|
I0 1 (G

@
( 0 E @ 1I I5
4C
(G 5 H ( D
5 R ST9 U R9
@ Q f 2
R V82 ) 3 E 3 4 E 2 R W3 2 X RE R
(5 0 T 8
B D8 x @
1( B0
( D B I I0
@ I @
E 04
2
@C @ 0 E 04 C D2 2
7 @C
@0
5G
C BC (7 1( 1( 1 H 58 1C U( 25 D B 1 @ E I U( 2 57 @ ( 5 8 C ( @ B C
"
x$
0 D 20 1C 5 H 8 C 1PC 5I E 1PC I 04 @ 1 ( 12 @ I ( 1( D B 4 B 8 B D 1( B 4 8 ( H H ( 0 1 R B 8 ( H 0 B 8 12 @ I ( ( ( 8 R @ ( 8 C 8 ( 5 @ ( I0 7 @ H( 4 H B5 7 7 D 57 7 I0 8 ( @ H( ( H 5 @ 8 1 4C G H B5 @ 8X 20 H( 1PC I5 B5 8 ( @ D 04 0 8 H 07 5 8X 1 1 4C G H 5 B64 ( BC D X @ |
0 @ 0 I 8 1 ( 4 H B 0 8 4 l 5 64
q j qg j m
I D8 0 BC @
@C 8 @C (5 1 (
58 B4 IPC T A4 C
0 3 8 R ( { 5 H @ E 9
r Fpq ts
~ @ 7
5 ~
( 1
( 4 B(
0G
vu x Fw
B( 15 5 @ 20 04 7 ~ 4 x 8 8 7 E 0
@C ( x
~

0B 4 V
B4 | (

(

0 @ @ 0 x 8 B( @
(B 4 ~

x x
14 (C 0B 4 V B 5 4C C B0 8 I5 B T 88 D ( 18 B ( 4 @C (5 T1 8 I0 R 12 @ ( 4 @ @ ( 0 E 0 (G B D8 04 C ( I 4 7 (7 4 H 5 G 1 4C G 20 (G 8 E 3 8 0B 4 3 9 E (
(G 5 20 1 4C G H
x | l n zn n |k } l | l m mx l | z h w l n |k l z x h l h m l m n | w |k ~ h n |k l | ~ |k z { m l n mn t xg z { z { n zn | n n x q r x v Tu x yw t 8s t z l z l l ~ } m |
ln 8 jk { n lmn lmn y p qo ep f l l jk n| x lx | ts vu x Fw
8 h h h ig r Fpq
H( 0 4
@ A9 I5 G 57 8 (C (B
@C
@C
5 @ ( 4C 1 7 C H( @ 1 0 B4
| ~}
@
(7 BC D B 2 H( @ ( 4 7 5 2 1 5 15
l
@
5 C
h j l z 8 x z x { | m zn l n zn n m m| l m n l n n zn z l n zn l x { z | l l n { x n z | { z | n { x n m x x | x n n l |k m zm x n { x m l x |k n l m m 8 l n
h ig j zx l | l n m z zx n x m n m m| | | n |
z l

0.2 0.4 0.6 0.8 1.2 0 1 0.2
o o
0
F
symmetry conditions
o
0.5

g2=250
o
1
F
0.2 0 0.2
1.5
0.2 0.4 0.6 0.8 0 1 0.2
8 1 1
0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
x n z n l mn n m zn h h g h g h | x m zm l l m x h l n | h m h g s h w w l n |k n | x n | z ~ zn n z l z | l x n | x z n { m n { | m| n |k x | l x x n zx n | n | l m l | l ~ zm l m n nk m m n | { l |k | l x n | x z | ~ g o t t h h g { n l zx n n k n ln n
0.2 1
0.4
0.6
0.8
1.2
0.2 0.5 0
0.5 1 1.5 2 2.5 3 3.5
o o o o
s g o g g o g g g g g o g g g g g g o g s

g g o
g g
g g g g o g g g o g
1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0.2
=104 5 =10 6 =10 =107

0.1 0
d

h j l z 8 x z x { | m zn l n zn n m m| l m n l n n zn z l n zn l x { z | l l n { x n z | { z | n { x n m x x | x n n l |k m zm x n { x m l x |k n l m m 8 l n
h ig j zx l | l n m z zx n x m n m m| | | n |
z l
d X
c B@
Q R
d h

e f c
6 75

B
X U B X
z
gX U
Bd
w
8

U I
$h U h h s P X t u ` v wX c c B B h X X X v ux h U @ c h U @ d S
~
B
yB
xB
@
@ A9
!
# $" B C D
z {
S
B
@
U VT
|
h I B
z { z
@ W X Y
' (& ) 0
2 01

U B p qh H B
zx
5 B
{ u ~ w
@ 3

@ p h r r
'2
Ù
v u @ z
@ C F GE
# !
I GH
z
t u a yB Ù
r
I @ v wX
ts
b 5 C
w
v ux h P
x
S B c
z {
f B @ h B c s ` c
X c U

B B

h i
B @ ep s l d U
@
U

h X W
v wX j

t j B f i
$h
e p h f
d ea
me n @ v wT X v
Xj B
c r

` j U
fT X

f h
@
h
j h

j X g Xj B h gX
B Ù X
n
f v wX h
@
gX

h k h d Ù b d h
@ s P
t B
u Ù

v g j h v wX
@
h ux

B

c B B po

d c
B
Ù X U

B h u
po
B
X X
X U X

u h r

po po v

@ X ` s j s

Ù c B r u ph
Ù z
Ù Ù
v v v
y p
X
fx d X
H g u pqh
jX
@ v T T v v T @ @ B B B

U P h h `
X
t @

s u ph I
U `
X X
d { |u P P b P b b yB

Ù v X t
v Ù
x gX s
u pqh

B X @ y p @ h t

u P zu h u I
h h
w u I
B
B

P S S
c e f d c d B q
~ po po

B c

c @ d B
g X h | @ @ v wX i h @ v vP t u t c d @ B B t r h X H U c h U
@
g jX
X U X
B h

h t
X U

X h U

X X h B
@
H u pqh

U X h c

r B d @
U p qh

d @ B
X
t
h h d v

@ } v v
X Ù v wX B h v xh @ @ B

r zu
p ` s

u u pqh t t d
d
r s
v w

h c B X u I B
c h S B
@ B c d B Bd @ @ @ Bd d ea h c @ B @ U U @ @ @ @ c @ @ Bd Bd c
X t c

B c B@ B c

c B@

B B
@
B
@ @
B B
d @
B d
@
@
B
B S
@ ~

@ ~
8
d @ B c B d B @ d @
B
d c
@ 8
B
d @
B B d

@ d B
@ @
@ @
B d B
B
B B
B
d B B

B d B

d V c d ea d wa c U B v I d S | v T v I B B B B d @ d h
@ c ~

` `
X 9 s
v v b h d a h { h X t t I

( s
&

1 & ) & 7 8

D EBC
%

7 q

&
1 ) P
d ) &
3 7
'
IH
GF
&

( 2 & 3 d q Y `

&
Q R SB

&
&
2 &

) &
a 1

X T ) 3 V WU
) 9
3
(
'
7 1

5 64 0 e i p
d
" Y ` a b @

'

5 &
&
) &

@
' &
)
&
'
! 1

g hf
3 q

'
" ) Y

Y `
` a b a
( 0 g

! " X 7 )
' &
&
&
Y 1
)
a b
'
' 2
u vt
w $ 0

w & $ x )
&
w
&
w 0

u

)
&
x ) 1 ) 4 )
& &( &
&
&
&
) 1 ' ) 5

'
) 5

)
& ) 1

1 )
) 1

7
& )
)

&
&
&
X r
y &
& (
) (
( ) &
X &

@ X X X X X
r f
a b
r X
&
&

r X `

r X
X f
&
r X
e
'
&
&
'
) 2

0.15
0.05
0.1
0.2 0 0.05
0.1
&
0.5
i
(
0 0
0.2
0.4
0.6 0.8 1
o e d
i q 3

k l q

k l
p
k l
k l
dm 3

h i j `
i
s
3 q
3 q
i k l
s
k l q
k l
k l n d
k l on

h i j
t t

&
&
u& ' )& 1 & 1

&
& $
i d
&
i 2

' )

| {

7

&
%
5 &
k l` p p y zw y zw y zw y zw e eo o y w zx` qd n k k l` k q y zw
k l`
&
&v & 2 &

2 1
d o k o q y wd o o o y w n q y w k k o o
d ` p k q p y w
k dp y w
k q y w `
&
'
)

&
p k p y zw e p k p p y w q
k p o y w o
k d p q y zw m
k l` ` y w zx y w zx` d
&v 1 p k

1 )& p k p q y w m e ` y w p y we k o k o q y w y w`

k o
n o
&
` d
&
&
2 & &

7 &

) ) 2
&
x
9 ) u }~ '

)
1 3
'
( f & 2 &
) & '

' x
"
@ '
&

z
@ & '
&
5 &
) & 5
'
(
1 ' )

1 X
'
7 ) &

&

&
) ) 2

` s ) &
&
3 & 1 &

'
& Y

` ) )
'
& & u

'

(
s
'
' t &
'
&

w &

w $ @ & y

w $ ) &
& '
'
1

'
'
&
'
29

)

& )

&
& @

'
'

& 2
&4
& ( ( )

&

&
7 7 )

'& &

&
)
& ( &

) ( &
&

w $ 1
w ) ) 29
' & 1 &

& @ )

'
) ' 7
&
&
'
&

&
@ '
( &

)
)
Y '
& ( 1 ) &

1 &
&

Differential Equations Models Do not Capture Some Essential Features
Stochastic vs. DE
Reaction: A+2X A X
3X
Multiple steady states! ODE Stochastic model
Mixed Finite Element Approximation of Eddy Current Problems

Ana Alonso Rodr guez Ralf Hiptmair Alberto Valli
Abstract Finite element approximations of eddy current problems that are entirely based on the magnetic eld H are haunted by the need to enforce the algebraic constraint curl H = 0 in non-conducting regions. As an alternative to techniques employing combinatorial Seifert (cutting) surfaces in order to introduce a scalar magnetic potential, we propose mixed multi-eld formulations, which enforce the constraint in the variational formulation. In light of the fact that the computation of cutting surfaces is expensive, the mixed nite element approximation is a viable option despite the increased number of unknowns.
Introduction
The governing equations of electromagnetic elds and currents E, H, B, D, J are Maxwells equations completed by constitutive laws in order to model the eldmatter interaction. In what follows we shall restrict ourselves to the Maxwell model of memoryless linear materials with Ohms law (see [9]): t D + curl H t B + curl E = J = Je + E , = 0, D B = E, = H . (1)
Here is the magnetic permeability, the dielectric tensor, and stands for conductivity. and are assumed to be symmetric and uniformly positive denite 3 3-matrices, whereas is supposed to be symmetric and uniformly positive denite inside the conducting region C , but vanishes in the air region I . All the material parameters are functions of the spatial variable x only. Under these circumstances, if the source current Je is of the form Je (t, x) = Re[Je (x) exp(it)], where Je is a complex-valued vector eld and = 0 is a xed angular frequency, the elds E, H, B, D also have this harmonic dependence on
Dipartimento di Matematica, Universit` di Milano, via Saldini 50, I-20133 Milano, a alonso@mat.unimi.it Seminar for Applied Mathematics, ETH Zrich, CH-8092 Zrich, hiptu u mair@sam.math.ethz.ch Dipartimento di Matematica, Universit` di Trento, I-38050 Povo(Trento), a valli@science.unitn.it
time. The Maxwell model with Ohms law then assumes the following strong form. iD + curl H = Je + E , iB + curl E = 0 , D = E, B = H . (2)
The unknows now are the complex amplitudes E, H, B, D, independent of time. In many situations it is possible to consider simpler quasi-static models that still oer a suciently accurate description of electromagnetic phenomena. The most popular among these simplied models is the so-called eddy current model, which consists in neglecting the term iD in (2) [12, 5]. Then compliance with Amp`res law entails e div Je,I = 0 in I ,
j
Je,I n dS = 0 ,
j = 1, . . . , p ,
(3)
where j , j = 1, . . . , p , are the connected components of the boundary of C . The latter is denoted by := C . Here and in the sequel we denote by vL the restriction of a vector eld v to L , L = I, C. We introduce an articial computational domain R3 , which is a box containing the conductors and their immediate neighborhood, big enough so that one can assume a zero eld beyond. As before we write C for the conductor region and I := \ C . For the sake of simplicity, we assume that I is connected. On homogeneous boundary conditions for either H or E are imposed: throughout we will demand H n = 0 on , modelling a container made by an innitely permeable magnetic material (see [9], p. 232; [?], p. 408). However, with simple modications our results also hold for the boundary condition H n = 0 on . The boundary condition H n = 0 on implies another compatibility condition for Je,I , namely Je,I n = 0 on . (4)
Obviously, we cannot expect a solution for E to be unique, because it can be altered by any gradient supported in I and will still satisfy the equations. However, imposing the constraints div( EI ) = 0 in I , EI n = 0 on , EI n dS = 0, j = 1, . . . , p 1,
j
(5)
(that are implied by (2)) will restore uniqueness of the solution for E. 2
The complete eddy current model we consider in the sequel is then: curl H = Je + E , div( EI ) = 0 in , En=0,
j I
iH + curl E = 0 in , EI n dS = 0 , Hn=0 j = 1, . . . , p 1 , (6)
on .
The existence and uniqueness of a solution of problem (6) has been proved in [4], assuming that the entries of the matrices and belong to L (), those of belong to L (I ), and nally that the source current Je belongs to (L2 ())3 and satises (3) and (4). Dropping the displacement current converts Amp`res law into the purely e algebraic constraint curl H = Je,I in I . This raises problems not encountered with the full Maxwells equations. This paper will be devoted to how to deal with these problems in the context of a variational formulation based on the magnetic eld H. We will focus on approaches that forgo the direct option to incorporate the constraint into the trial space. Instead it is enforced by means of augmented variational equations. Adding extra equations may seem wasteful and, indeed, it is, because the resulting formulations will, after a nite element Galerkin discretization, feature many additional degrees of freedom. However, this is the price to pay for avoiding the cumbersome topological preprocessing, that is the construction of cuts [15], that is indispensable in the case of the direct option. Hence, these augmented formulations can become relevant for practical computations. Here we are going to present a couple of possibilities to take into account the seemingly simple constraint curl H = Je,I in I . Each variant will come with its own issues of stability and uniqueness. A brief outline of the paper is as follows: in the next section we introduce notation and function spaces needed for the remainder of the article. Then we review the well-known H-based variational formulation of the eddy current problem. From these basic equations we derive augmented mixed formulations in the fourth section. In the fth section their nite element Galerkin discretization will be discussed. Finally, in Sect. 6 we give a-priori error estimates.
Basic concepts
As usual, we denote by H s () or H s (), s R, the Sobolev space of order s of real or complex measurable functions dened on or , respectively. 1 If we denote by H0, () the subspace of H 1 () consisting of those 1 1 functions satisfying | = 0. As usual H0 () := H0, (). The space H(curl; ) (respectively, H(div; )) denotes the set of real or complex vector valued functions v (L2 ())3 such that curl v (L2 ())3 (respectively, div v L2 ()). If , by H0, (curl; ) we denote the subspace of H(curl; ) of those functions v satisfying (v n)| = 0. We set H0 (curl; ) := H0, (curl; ). H 0 (curl; ) denotes the subspace of curl-free
0 functions of H(curl; ) and H0, (curl; ) = H0, (curl; )H 0 (curl; ). Analogously H0, (div; ) stands for the subspace of H(div; ) containing functions v satisfying (v n)| = 0. As above, we set H0 (div; ) := H0, (div; ). Moreover, H 0 (div; ) denotes the subspace of divergence-free functions of H(div; ) 0 and H0, (div; ) := H0, (div; ) H 0 (div; ). Finally, H s (curl; ) denotes the space of vector functions v (H s ())3 such that curl v (H s ())3 . Topology enters our considerations through the space of harmonic vector elds 0 0 H := H0, (curl; I ) H0, (div; I ) .
(7)
Moreover, for the sake of brevity, we introduce the space of admissible electric elds W I := {NI (L2 (I ))3 | NI satises (5)} , and the space of unique vector potentials
0 Y I := H0, (curl; I ) H0, (div; I ) H .
(8)
(9)
They owe their name to following result, which will be useful in the sequel. It is essentially contained in [1, 13]. For a constructive proof see [3]. Theorem 2.1 For each vI W I there is a unique qI Y I such that curl qI = vI and qI
L2 (I )
C1
vI
L2 (I )
The H-based variational formulation
Basically, two dierent variational formulations of (6) exist, either based on the electric eld E or the magnetic eld H [8]. They correspond to the primal and dual formulation of a second order elliptic problem. However, the algebraic constraint on curl H manifests itself in a entirely dierent way in the two formulations. Therefore we restrict ourselves to the H-based approach. The generic form of the H-based variational formulation involves the Hilbert space of complex-valued vector functions V 0 := {v H0 (curl; ) | curl vI = 0 in I } , endowed with the natural norm ||v||2 0 := V
|v|2 +
C
| curl vC |2
(recall that vI and vC denote the restrictions of v to I and C , respectively). We will also need the ane space V Je,I := {v H0 (curl; ) | curl vI = Je,I in I } = H + V 0 , 4
where H is a function in H0 (curl; ) such that curl H = Je,I in I . The I magnetic eld we are looking for belongs to V Je,I . Moreover for each v V 0 0=
(iH + curl E) v = i
H v +
C
EC curl vC .
Using the strong form of Amp`res law in the conductor, namely EC = e 1 (curl HC Je,C ), we arrive at 0 = i
H v +
C
1 (curl HC Je,C ) curl vC .
So, the magnetic eld H solves the following problem: Find H V Je,I : i
H v + C 1 curl HC curl vC = C 1 Je,C curl vC v V 0 .
(10)
The existence and uniqueness of a solution of (10) follows from the LaxMilgram lemma, since, under our assumptions on the material coecients, the bilinear form is trivially coercive on V 0 . Next, we have to recover the electric eld in . In C , from Amp`res law we have e EC = 1 (curl HC Je,C ) , while in I there exists a unique EI H(curl; I ) such that curl EI = iHI , EI n = 0 on ,
j
(11)
div( EI ) = 0
in I ,
(12) (13) (14)
EI n dS = 0 j = 1, . . . , p 1 , on .
EI nI = EC nC
Here, nL denotes the unit outward normal vector on L , L = I, C. We refer to [4] for more details. Then (H, E) with H solution of (10) and E dened as E= EC EI in C in I
is the unique solution of (6). Remark 3.1. We note that a nite element method based on (10) would have to deal with the constrained space V Je,I . The direct way to deal with the constraint in V 0 makes use of scalar magnetic potentials by representing
1 V 0 |I = grad H (I ) H
(see [3, 7]). It would be a perfect solution, unless we had to construct a basis of H in order to continue with discretization. Such a basis is readily available, once we have cuts at our disposal, i.e. a collection of surface in I that cut any non-bounding cycle [9, 18]. Finding these cuts for arbitrary shape of C seems to be a challenging problem [15]. 5
Mixed formulations
The main idea is to reformulate (10) as a saddle point problem in nonconstrained vector spaces by introducing Lagrange multipliers. Let us dene the bilinear form in H0 (curl; ) a(w, v) := i
w v +
C
1 curl wC curl vC .
We can introduce a Lagrange multiplier AI (L2 (I ))3 and consider the saddle point problem Find (H, AI ) H0 (curl; ) (L2 (I ))3 : a(H, v) + I curl vI AI (15) = C 1 Je,C curl vC v H0 (curl; ) curl HI NI = I Je,I NI NI (L2 (I ))3 . I This problem does not have a unique solution as it is possible to add any function 0 of H0, (curl; I ) to AI . However if (H, AI ) is a solution of (15) then H is the solution of (10), and AI satises curl AI = iHI = curl EI and AI nI = 1 (Je,C curl HC ) nC = EC nC = EI nI on . Thus, in order to restore uniqueness of solution it is natural to look for AI in the constrained space W I dened in (8). Then it is obvious that AI = EI . From Theorem 2.1 it is easily veried that W I is equal to the range space 1 curl H0, (curl; I ). Thus, we consider the two-eld formulation: Find (H, AI ) H0 (curl; ) W I : a(H, v) + I curl vI AI (16) = C 1 Je,C curl vC v H0 (curl; ) curl HI NI = I Je,I NI NI W I . I Theorem 4.1 A unique solution of (16) exists. Proof. We can appeal to Theorem 2.1 and the general theory of variational saddle point problems [10]. 2 Again, the space W I involves some constraints. So we introduce another Lagrange multiplier to impose these: we consider the space
1 H (I ) := { H 1 (I ) | |j is constant j = 1, . . . , p 1, |p = 0},
and it is easily veried by integration by parts that NI W I if and only if 1 NI (L2 (I ))3 and I NI I = 0 for all I H (I ). Eventually we
confront the following problem: 1 Find (H, AI , I ) in H0 (curl; ) (L2 (I ))3 H (I ) : a(H, v) + curl vI AI I = C 1 Je,C curl vC v H0 (curl; )
I
(17)
curl HI NI + I NI = I Je,I NI AI I = 0
NI (L2 (I ))3
1 I H (I ) .
We note that if (H, AI , I ) is a solution of (17) then I = 0 (just take NI = I in (17)), and (H, AI ) is solution of (16). Introducing the bilinear forms b(, ) : H0, (curl; I ) (L2 (I ))3 C , and
1 c(, ) : (L2 (I ))3 H (I ) C ,
b(vI , NI ) :=
I
curl vI NI ,
c(NI , I ) :=
I
NI
I ,
and the linear operators F (v) :=

C
1 Je,C curl vC ,
v H0 (curl; )
and G(NI ) :=
I
Je,I NI ,
NI (L2 (I ))3 ,
problem (17) can be rewritten as 2 I 3 1 I Find (H, AI , I ) in H0 (curl; ) (L ( )) H ( ) : = F (v) v H0 (curl; ) a(H, v) + b(vI , AI ) b(HI , NI ) + c(NI , I ) = G(NI ) c(AI , I ) = 0 NI (L2 (I ))3
1 I H (I ) .
In order to prove that (17) has a unique solution, we can use the following result, which is Lemma 4.1 in [11] extended to complex Hilbert spaces. Lemma 4.2 Let X, Q, M be three complex Hilbert spaces and a : X X C, b : X Q C, c : Q M C be three continuous bilinear forms, i.e. there exist three positive constants c1 , c2 , c3 such that |a(v, w)| c1 v X w X , |b(v, N)| c2 v X N Q , |c(N, )| c3 N Q M for all v, w X, N Q 7
and M . Given f X , g Q , l M , let us consider the saddle point problem Find (H, A, ) in X Q M : a(H, v) + b(v, A) = f, v vX (18) b(H, N) + c(N, ) = g, N NQ c(A, ) = l, M . Let Q0 Q and X 0 X be two subspaces as follows: Q0 = {N Q | c(N, ) = 0 X 0 = {v X | b(v, N) = 0 Assume that a(, ) is X 0 -coercive, i.e., |a(v, v)| v
2 X
M }, N Q0 }.
v X 0,
(19)
and that the following infsup conditions hold

NQ0 vX
inf sup
|b(v, N)| , v X N Q |c(N, )| , N Q M
(20)
M NQ
inf sup
(21)
for some positive constants , , . Then problem (18) has a unique solution. Now we are in a position to prove the following result: Theorem 4.3 Problem (17) has a unique solution. Proof. In order to verify the assumptions of Lemma 4.2, rst recall that the spaces W I and V 0 can also be characterized as W I = {NI (L2 (I ))3 |
I
NI
I = 0
1 I H (I )}
and V 0 = {v H0 (curl; ) | curl vI NI = 0 NI W I } ,

I 1
(in the latter case, just take NI = curl vI ). Since the bilinear form a(, ) is coercive on the space V 0 , we need only show that the two inf-sup conditions are satised, more precisely, that there exist two positive constants and such that sup
vH0 (curl;) I
curl vI NI NI v H(curl;) 8
L2 (I )
(22)
for all NI W I , and sup

NI (L2 (I ))3 I
NI
NI
L2 (I )
H 1 (I )
(23)
1 for all I H (I ). Poincares inequality gives us a constant C2 > 0 such that I H 1 (I ) 1 C2 I L2 (I ) for all I H (I ). Moreover, since is assumed to be uniformly positive denite, there exist two positive constants and such that for all NI (L2 (I ))3
NI
2 L2 (I )
NI NI
NI
2 L2 (I ) .
1 Hence, given I H (I ) and choosing NI =
I we have
I I
supNI (L2 (I ))3
NI I
L2 (I )
NI
L2 (I )
L2 (I )
C2
H 1 (I ) .
Concerning (22), by Theorem 2.1 for all NI W I there exists qI Y I such that NI = 1 curl qI . Let q H(curl; ) be a continuous extension of qI into C ; hence, q H(curl;) C3 qI H(curl;I ) . By the stability estimate of Theorem 2.1 we can infer that q Thus, supvH0 (curl )
I
2 H(curl;)
2 C3 qI
2 H(curl;I )
2 2 C3 (1 + C1 ) curl qI
2 L2 (I ) .
curl vI NI
H(curl;)
curl qI NI
H(curl;)
1 2 (1+C1 )1/2 C3
curl qI NI I curl qI L2 (I )
1 2 (1+C1 )1/2 C3
NI NI
L2 (I )
NI
2 (1+C1 )1/2 C3
NI
L2 (I )
. 2
Finite element discretization
We are aiming for Galerkin nite element discretization of both the two-eld problem (16) and the three-eld formulation (17). In both cases we want to verify the assumptions of the theory of discrete saddle point problem [10, Chap. 2]. We assume that , C , I are Lipschitz polyhedra and consider a family of regular tetrahedral meshes {Th }h of such that each element K Th is 9
contained either in C or in I . We denote TC,h , TI,h the restriction of Th to C and I , respectively. The parameter h will also provide the meshwidth of Th . We employ Ndlecs curl-conforming edge elements of the lowest order to e e approximate the magnetic eld: let Vh be the nite elements space dened by the complex valued functions Vh := {vh H(curl; ) | vh (x)|K = aK + bK x K Th } , where aK and bK are two constant vectors in C3 . It is known that any function vh Vh is uniquely determined by the following degrees of freedom [16, Sect. 3.2] Me (v) = {
e
v ds | e is an edge of Th },
where is the unit vector along the edge e. These edge moments make sense for any v (H s ())3 with curl v (Lp ())3 with s > 1/2 and p > 2 (see [6, Lemma 4.7] and [16, Lemma 3.13]). Moreover the following interpolation error estimate holds (see [2, 11] and [16, Theor. 3.14]). Lemma 5.1 Denoting by h w Vh the interpolant of w, for 1/2 < s 1, we have h w w
L2 (K)
C4 hs ( wI K
H s (K)
+ curl wI
H s (K) )
w H s (curl; K),
where hK is the diameter of K Th . The homogeneoous boundary conditions on are incorporated by setting degrees of freedom on to zero. Thus we end up with the spaces
I Xh := Vh H0 (curl; ) and Xh := {vh|I | vh Xh } .
For additional information about edge elements the reader is referred to [16, Chap. 3], [2], and [14, Chap. III, Sect. 5.3].
5.1
Two-eld formulation
The challenge is the approximations of the constrained space W I . However, we can take the cue from the representation in Theorem 2.1 and lift it into the discrete setting. More precisely, we choose
I Wh := 1 I curl Xh
as trial space for W I . Note that this is a conforming discretization in the sense I that Wh W I . This results in the following discrete two-eld problem. I Find (Hh , AI,h ) in Xh Wh : a(H , v ) + curl vI,h AI,h h h I (24) = C 1 Je,C curl vC,h vh Xh , I curl HI,h NI,h = I Je,I NI,h NI,h Wh . I 10
Theorem 5.2 Problem (24) has a unique solution. Proof. As in the case of the continuous problem, it is straightforward that the bilinear form a(, ) is continuous in Xh and coercive in
0 Xh := {vh Xh | I I 0 since in particular 1 curl vI,h Wh , so that vh Xh implies curl vh|I = 0. To prove a uniform discrete inf-sup-condition we rely on the following lemma. It is a variant of a discrete Poincar-Friedrichs inequality, see [16, Theorem 4.7] e for a proof. I,0 I Lemma 5.3 Let Xh := {vI,h Xh | curl vI,h = 0} and I,0 I (Xh ) := {pI,h Xh | I,0 pI,h vI,h = 0 vI,h Xh } . I curl vI,h NI,h = 0 NI,h Wh } ,
(25)
Then there exists a positive constant C5 , independent of h, such that for all I,0 pI,h (Xh ) pI,h
L2 (I )
C5 curl pI,h
L2 (I )
I I I By the denition of Wh , for each NI,h Wh there exists qI,h Xh such that I,0 I,0 1 NI,h = curl qI,h . By projecting on (Xh ) , we nd a unique qI,h (Xh ) with the same property. Let qh be some uniform discrete extension of qI,h to C (the existence of such an extension has been proved in [2]). Then qh Xh and
qh
H(curl;)
C6 qI,h
H(curl;I )
2 C6 (1 + C5 )1/2 curl qI,h
L2 (I )
2 Since the constant C6 (1 + C5 )1/2 is independent of h, qh is a suitable candidate in the discrete inf-sup condition: |
I
supvh Xh
curl vI,h NI,h |

H0 (curl;)
vh
curl qI,h NI,h |

H0 (curl;)
qh
(26) NI,h
L2 (I )
2 C6 (1+C5 )1/2
. 2
All assumptions of [10, Chap. II, Theor. 1.1] are satised.
I Remark 5.1. The only way to implement the space Wh is to rely on its 1 very denition, that is we obtain its elements as curl of edge element functions. Yet, these will no longer be unique. The bottom line is that in a practical implementation of the two-eld method we will face a singular system of linear equations. As its kernel is well separated, conjugate gradient type iterative solvers will perform well.
11
5.2
Three-eld formulation
Apart from H we have to approximate AI (namely, EI ) and I in (17). Let Pk denote the standard space of complex polynomials of total degree less than or equal to k with respect to the real variable x. To discretize AI L2 (I ) we choose piecewise constant vector functions in the space QI := {NI,h (L2 (I ))3 | NI,h|K (P0 )3 K TI,h } . h
1 In order to approximate the Lagrangian multiplier I H (I ) it would be natural to rely on piecewise linear Lagrangian nite elements. However, it turns out that this space is too small to guarantee uniform stability of the discretized mixed formulation. We have to switch to a larger space for the approximation of the Lagrangian multiplier; it will be the nonconforming Crouzeix-Raviart elements, dened as follows: I Uh := {I,h L2 (I ) | I,h|K P1 K TI,h and I,h is continuous at the centroid of any face f common to two elements in Th }.
Then the discrete I will belong to

I I Mh := {I,h Uh | I,h (p) is equal for all centroids p of faces of j , j = 1, . . . , p 1, and I,h (p) = 0 for all centroids p of faces of p } . I Note that, since functions in Uh are no longer continuous, they are no longer 1 I in H ( ). Therefore we must dene a modied bilinear form ch : (L2 (I ))3 1 I 1 I 1 I (H (I ) + Mh ) C and a norm on H (I ) + Mh . For each I H (I ) + Mh we denote I the function in (L2 (I ))3 dened as
( I )|K :=
1 Note that if I H (I ), then form
(I|K ) K TI,h . I = I
I 1 NI (L2 (I ))3 , I H (I ) + Mh
I . Similarly, we dene the bilinear
ch (NI , I )
:= =
NI I
NI
I 1 and the norm in H (I ) + Mh
2 h
:=
K K
| I |2 =
2 L2 (I )
Then, the nite elements approximation of (17) can be formulated as follows: I I Find (Hh , AI,h , I,h ) in Xh Qh Mh : a(Hh , vh ) + b(vI,h , AI,h ) = F (vh ) vh Xh (27) b(HI,h , NI,h ) + ch (NI,h , I,h ) = G(NI,h ) NI,h QI h I ch (AI,h , I,h ) = 0 I,h Mh . 12
To show that this problem has a unique solution we need the following lemma, (see [17]).
I Lemma 5.4 We have the L2 (I )-orthogonal decomposition QI = curl Xh h I Mh . I Proof. The proof has two parts. In the rst part we show that for all vI,h Xh I and I,h Mh we have the orthogonality I curl vI,h I,h = 0. In the I I second part we establish that dim(QI ) = dim(curl Xh ) + dim( Mh ). h I I For any vI,h Xh and I,h Mh integration by parts yields I
curl vI,h
I,h
= = =
curl vI,h
I,h
curl vI,h n I,h

f
f Fint
curl vI,h n [I,h ]f

f F p j=1 f
+ +
curl vI,h n I,h

f
f Fj
curl vI,h n I,h ,
where Fint is the set of internal faces of the triangulation TI,h , F and Fj denote the set of faces of TI,h on and j , respectively, and [I,h ]f denotes the jump of I,h across the face f . Note that, for all f Fint , (curl vI,h n)|f is constant and f [I,h ]f = 0 since [I,h ]f is a linear function and it is equal zero in the centroid of f . Moreover (curl vI,h n)|f = 0 for all f F , and, using that for all faces f Fj one has f I,h = j meas(f ), for all j = 1, . . . , p , we nally nd
f Fj f
curl vI,h n I,h
= = j
f Fj (curl vI,h
n)|f j meas(f )
curl vI,h n = 0 ,
hence curl vI,h

I
I,h = 0 .
Let us introduce the Raviart-Thomas nite element space [10, Chap. III] RTh := {vh H(div; I ) | vh (x)|K = aK + bK x K TI,h }, where aK is a constant complex vector and bK is a complex number, and the subspaces RT0, := RTh H0, (div, I ),
0 0 RT0, := RTh H0, (div, I ).
By arguments from discrete cohomology, it can be proved (see [9]) that, as vector spaces on C,
I 0 dim(curl Xh ) = dim(RT0, (I )) (p 1).
13
Let us denote by #K the number of tetrahedra of TI,h , by #F the total number of faces of TI,h and by #F , #F , the number of faces of TI,h on and by respectively. It is not dicult to prove that:
0 dim(RT0, ) = dim(RT0, (I )) dim(div(RT0, (I ))) = (#F #F ) #K, I I dim(Mh ) = (#F #F ) + (p 1) = dim( (Mh )), I dim(Qh ) = 3#K.
Since 4#K = 2#F (#F + #F ) then

I I dim(curl Xh ) + dim( Mh ) = [(#F #F ) #K (p 1)] + [(#F #F ) + (p 1)] = 2#F (#F + #F ) #K = 4#K #K = dim(QI ). h I Since, trivially, curl Xh QI and h I Mh QI , the proof is nished. h
Using Lemma 5.3, we can now prove the main result of this section. Theorem 5.5 Given a triangulation Th of , assume that the entries of the matrix are piecewise constants in I . Then Problem (27) has a unique solution. Proof. Conditions (19) and (20) follow as in the proof of Theorem 5.2, provided that we show that the space {vh Xh | b(vI,h , NI,h ) = 0 where QI,0 QI is dened as follows h h QI,0 := {NI,h QI | ch (NI,h , I,h ) = 0 h h
I I,h Mh },
NI,h QI,0 } , h
0 coincides with the space Xh dened in (25). In fact, it is enough to prove that I,0 I Qh = Wh . Since the entries of the matrix are piecewise constants, for each I,0 NI,h Qh we have NI,h QI . Therefore, using Lemma 5.4, we obtain that h I I NI,h curl Xh , hence QI,0 1 curl Xh . The converse is straightforward, h proceeding as in the proof of Lemma 5.4. I Concerning the infsup condition (21) note that for all I,h Mh one has I,h QI , hence from the denition of the norm h h
supNI,h QI h
|ch (NI,h ,I,h )| NI,h L2 (I ) I,h
|ch ( I,h ,I,h )| I,h L2 (I ) I,h

I
(28)
I,h I,h I,h
2 L2 (I )
14
2 Remark 5.2. Note that Je,I = curl Ke,I for some Ke,I H0, (curl; I ). If h Ke,I is welldened, we can dene Gh (NI ) := I curl(h Ke,I ) NI . If in problem (27) we replace G with Gh , it is easily showed that the new I,h is equal to zero.
Error estimates
Given the discrete inf-sup-conditions established in Sect. 5.1, the quasioptimality of the discrete solution of the two-eld problem is standard [10, Chap. 2]. Here, we are only concerned with the discrete three-eld problem (27). We denote by c1 and c2 the continuity constants of the bilinear forms a(, ) and b(, ) respectively, by the coercivity constant of a(, ) in V 0 and by and two positive constants, independent of h, such that inf sup vh |b(vI,h , NI,h )| , NI,h L2 (I ) (29)
NI,h QI,0 vh Xh h
H(curl,)
and
I I,h Mh NI,h Qh
inf
sup
|ch (NI,h , I,h )| . I,h h NI,h L2 (I )

2 C6 (1+C5 )1/2
(30)
.
From (26) and (28) we can take =
and =
1 Theorem 6.1 Let (H, AI , I ) H0 (curl; ) (L2 (I ))3 H (I ) be the soI I lution of Problem (17) and (Hh , AI,h , I,h ) Xh Qh Mh the solution of Problem (27). Then the following error estimates hold:
H Hh
H(curl;)
1+
c1
c2
1+
c2
vh Xh
inf
H vh
H(curl;)
(31)
AI AI,h
L2 (I )
1+
inf NI,h QI,0 AI NI,h

h
L2 (I )
(32) + c1 H Hh I I,h = I,h

H(curl;)
c2 H Hh
H(curl;)
(33)
Proof. The proof follows the lines of the proofs in [10, Ch. 2]. For all vh , vh Xh I and NI,h Qh a(Hh vh , vh ) + b(vI,h , AI,h NI,h ) = F (vh ) a(vh , vh ) b(vI,h , NI,h ) = a(H vh , vh ) + b(vI,h , AI NI,h ) .
15
0 Note that if vh Xh then curl vI,h = 0 in I , therefore a(Hh vh , vh ) = a(H G vh , vh ). If we take vh Xh := {vh Xh | b(vI,h , NI,h ) = G(NI,h ) NI,h 0 QI,0 }, then Hh vh Xh and we nd h a(Hh vh , Hh vh ) = a(H vh , Hh vh ). 0 Since Xh V 0 , from coerciveness we conclude
H Hh
H(curl;)
H vh
H(curl;) H vh
+ Hh vh
H(curl;)
1+
c1
(34)
H(curl;) G vh Xh .
Moreover, from the infsup condition (29), for all vh Xh there exists a unique 0 zh (Xh ) such that b(zI,h , NI,h ) = b(HI vI,h , NI,h ) for all NI,h QI,0 and h zh
H(curl;)
c2 H vh
H(curl;) .
Setting vh := zh + vh , for all NI,h QI,0 we have h b(vI,h , NI,h ) = b(HI , NI,h ) = b(HI , NI,h ) + c(NI,h , I ) = G(NI,h ) , G hence vh Xh . Furthermore, Hvh
H(curl;)
Hvh
H(curl;) +
zh
H(curl;)
1+
c2
Hvh
H(curl;)
and (31) follows from (34). To obtain (32) we use the infsup condition (29). For each NI,h QI,0 we h nd 1 |b(vI,h , AI,h NI,h )| AI,h NI,h L2 (I ) sup . vh Xh vh H(curl,) On the other hand b(vI,h , AI,h NI,h ) = F (vh ) a(Hh , vh ) b(vI,h , NI,h ) = a(H Hh , vh ) + b(vI,h , AI NI,h ), then AI,h NI,h
L2 (I )
c1 H Hh
H(curl;)
c2 AI NI,h
L2 (I ) ,
which yields (32). To obtain (33) we use the infsup condition (30) that in particular gives I,h
h
sup
NI,h Qh
|ch (NI,h , I,h )| . NI,h L2 (I )
16
On the other hand ch (NI,h , I,h ) = G(NI,h ) b(HI,h , NI,h ) = b(HI , NI,h ) + c(NI,h , I ) b(HI,h , NI,h ) = b(HI HI,h , NI,h ) , then I,h
h
c2 H Hh
H(curl,) .
2
I Remark 6.1. It is worthy to note that QI,0 = 1 curl Xh and that there h I exists qI H0, (curl; ) such that AI = curl qI . Hence
NI,h QI,0 h
inf
AI NI,h
L2 (I )
I qI,h Xh I qI,h Xh
inf
(curl qI curl qI,h )

H(curl;I )
L2 (I )
inf
qI qI,h
References
[1] A. Alonso and A. Valli. Some remarks on the characterization of the space of tangential traces of H(rot; ) and the construction of an extension operator. Manuscripta mathematica, 89:159178, 1996. [2] A. Alonso and A. Valli. An optimal domain decomposition preconditioner for low-frequency time-harmonic Maxwell equations. Math. Comp., 68(226):607631, April 1999. [3] A. Alonso Rodr guez, P. Fernandes, and A. Valli. The time-harmonic eddycurrent problem in general domains: solvability via scalar potentials. In C. Carstensen, S. Funken, W. Hackbusch, R.H.W. Hoppe, and P. Monk, editors, Computational Electromagnetics, volume 28 of Lecture Notes in Computational Science and Engineering, pages 143163. Springer, Berlin, 2003. [4] A. Alonso Rodr guez, P. Fernandes, and A. Valli. Weak and strong formulations for the time-harmonic eddy-current problem in general domains. Euro. J. Appl. Math., 14:387406, 2003. [5] H. Ammari, A. Bua, and J.-C. Ndlec. A justication of eddy currents e e model for the Maxwell equations. SIAM J. Appl. Math., 60(5):18051823, 2000.
17
[6] C. Amrouche, C. Bernardi, M. Dauge, and V. Girault. Vector potentials in threedimensional nonsmooth domains. Math. Meth. Appl. Sci., 21(9):823 864, 1998. [7] A. Bermdez, R. Rodr u guez, and P. Salgado. A nite element method with Lagrange multipliers for low-frequency harmonic Maxwell equations. SIAM J. Numer. Anal., 40(5):18231849, 2002. [8] A. Bossavit. Two dual formulations of the 3D eddycurrents problem. COMPEL, 4(2):103116, 1985. [9] A. Bossavit. Computational Electromagnetism. Variational Formulation, Complementarity, Edge Elements. Academic Press, San Diego, 1998. [10] F. Brezzi and M. Fortin. Springer, New York, 1991. Mixed and Hybrid Finite Element Methods.
[11] Z. Chen, Q. Du, and J. Zou. Finite element methods with matching and non-matching meshes for Maxwell equations with discontinuous coecients. SIAM J. Numer. Anal., 37(5):15421570, 2000. [12] H.K. Dirks. Quasistationary elds for microelectronic applications. Electrical Engineering, 79:145155, 1996. [13] P. Fernandes and G. Gilardi. Magnetostatic and electrostatic problems in inhomogeneous anisotropic media with irregular boundary and mixed boundary conditions. Math. Models Meth. Appl. Sci., M3 AS, 7(7):957991, 1997. [14] V. Girault and P.A. Raviart. Finite Element Methods for NavierStokes Equations. Springer, Berlin, 1986. [15] P.W. Gross and P.R. Kotiuga. Finite element-based algorithms to make cuts for magnetic scalar potentials: topological constraints and computational complexity. In F.L. Teixeira, editor, Geometric Methods for Computational Electromagnetics, volume 32 of PIER, pages 207245. EMW Publishing, Cambridge, 2001. [16] R. Hiptmair. Finite elements in computational electromagnetism. Acta Numerica, pages 237339, 2002. [17] P. Monk. A mixed method for approximating Maxwells equations. SIAM J. Numer. Anal., 28:16101634, 1991. [18] P. P. Silvester and R. L. Ferrari. Finite Elements for Electrical Engineers. Cambridge University Press, Cambridge, 1990. [19] T. Tarhasaari and L. Kettunen. Topological approach to computational electromagnetism. In F.L. Teixeira, editor, Geometric Methods for Computational Electromagnetics, volume 32 of PIER, pages 189206. EMW Publishing, Cambridge, 2001.
18
A posteriori error estimation for nite element discretization of parameter identication problems
Boris Vexler University of Heidelberg (joint work with Roland Becker) Overview:
An Example Formulation of the . . . Optimization algorithm Discretization by Finite . . . Adaptive Mesh . . . Example 1 Example 2 Example 3 (Bypass . . . Conclusions
An Example Formulation of the problem Solution algorithm A posteriori error estimation Numerical examples
Go Back Close Quit
An Example u u + f (u) = 0 in
E A exp( 1y )y(c
An Example Formulation of the . . . Optimization algorithm
Arrhenius law: f (u) =
y)
Discretization by Finite . . . Adaptive Mesh . . . Example 1 Example 2 Example 3 (Bypass . . . Conclusions
Ox F Ox
Go Back Close Quit
An Example Formulation of the . . .
An Example (II) Discretization of the state equation leads to an error in parameter A Ah

Goals:
Optimization algorithm Discretization by Finite . . . Adaptive Mesh . . . Example 1 Example 2 Example 3 (Bypass . . . Conclusions
Quantitative control of this error Minimization of the cost to achieve a given tolerance
Go Back
|A Ah| T OL
Close Quit
Formulation of the problem using Least Squares method 1 Minimize J(u) := C(u) C 2
under the constraint
An Example Formulation of the . . . Optimization algorithm Discretization by Finite . . . Adaptive Mesh . . . Example 1
2 Z
Example 2 Example 3 (Bypass . . . Conclusions
a(u, q)() = f () V
u V , q Q = R np C : V Z = Rnm is an observation operator C Z denotes the measurements
Go Back Close Quit
Solution operator S:QV a(S(q), q)() = f () V
Reduced observation operator c(q) := C(S(q))
Unconstrained least squares formulation 1 Minimize j(q) := c(q) C 2

2 Z
Go Back Close Quit
Optimization algorithm First-order necessary optimality condition j (q) = 0 G(c(q) C) = 0, G = c (q)

An Example Formulation of the . . . Optimization algorithm Discretization by Finite . . . Adaptive Mesh . . . Example 1 Example 2 Example 3 (Bypass . . .
Iteration
Initial parameter: q0 k = 0, 1, 2 . . . Residuum: Update q :
Conclusions
rk = G (C c(qk )), k Hk q = rk
Gk = c (qk )
Go Back Close Quit
Next parameter: qk+1 = qk + q
Optimization algorithm (II)

An Example
Choice of the matrix Hk :

Full Newton algorithm
2
Formulation of the . . . Optimization algorithm Discretization by Finite . . . Adaptive Mesh . . . Example 1
Hk :=
j(qk ) =
G Gk k
+ c(qk ) C, c (qk )
Example 2 Example 3 (Bypass . . . Conclusions
Gau-Newton algorithm
Hk := G Gk , k
Update methods
Hk := G Gk + Sk , k Globalization strategies:
Line search Trust-region techniques
with an update formula for Sk
Go Back Close Quit
Derivatives of the reduced observation operator
Optimization algorithm Discretization by Finite . . . Adaptive Mesh . . . Example 1 Example 2 Example 3 (Bypass . . .
Gij := ci(q) = Ci (u)(wj ) qj

with u = S(q) and wj V is the solution of the help problem:
Conclusions
au(u, q)(wj , ) = aqj (u, q)(1, ) V.
Go Back Close Quit
Discretization by Finite Elements Constrained formulation 1 Minimize C(uh) C 2

2 Z
a(uh, qh)(h) = f (h) h Vh
Unconstrained formulation
Go Back
Minimize
1 ch(qh) C 2
2 Z
Close Quit
Adaptive Mesh Renement Algorithm
An Example Formulation of the . . . Optimization algorithm Discretization by Finite . . .
1. Choose an initial mesh Th0 and set k = 0 2. Construct the nite element space Vhk 3. Compute uhk Vhk , qhk Q solving PIP on Vh 4. Evaluate the a posteriori error estimator hk 5. If hk T OL quit 6. Rene Thk Thk+1 using information from hk 7. Increment k and go to 2.
Adaptive Mesh . . . Example 1 Example 2 Example 3 (Bypass . . . Conclusions
Go Back Close Quit
A posteriori error estimation(I)
An Example Formulation of the . . . Optimization algorithm Discretization by Finite . . . Adaptive Mesh . . .
E(q) E(qh) = + P + R
Remarks
Example 1 Example 2 Example 3 (Bypass . . . Conclusions
R is a cubic remainder term due to linearization |P | C e C(u) C

1
R. Becker and B. Vexler
Go Back Close Quit
A posteriori error estimation for nite element discretization of parameter identication problems, Numerische Mathematik, to appear, 2003 2nd Leslie Fox Prize, Cambridge 2003
An Example
A posteriori error estimation (II) =

We solve:
1 (xh)(y 2
Formulation of the . . . Optimization algorithm Discretization by Finite . . . Adaptive Mesh . . .
ihy) +
1 (xh)(u 2
ihu)
Example 1 Example 2 Example 3 (Bypass . . . Conclusions
au(u, q)(, y) = < G(GG)1 E(q), C (u)() >

and
(xh)() := (f, ) a(uh, qh)() (xh)() := < Gh(G Gh)1 E(qh), C (uh)() > h au(uh, qh)(, yh)
Go Back Close Quit
Example 1 1 2
9
An Example Formulation of the . . . Optimization algorithm Discretization by Finite . . .
(u(i) Ci)2 inf

i=1
Adaptive Mesh . . . Example 1 Example 2
qu + 200u = 2 u = 0
in = (0, 1)2 , on .
Example 3 (Bypass . . . Conclusions
Go Back Close Quit
Example 1 (II) Efciency of the error estimator N

1089 1521 2209 3297 6041 17025 47761
Optimization algorithm Discretization by Finite . . . Adaptive Mesh . . . Example 1 Example 2
q qh
4.09e-1 3.20e-2 8.10e-3 2.66e-3 8.01e-4 1.88e-4 2.24e-5
2.56e-1 3.52e-2 8.35e-3 2.58e-3 7.62e-4 1.81e-4 2.15e-5
Ief f
1.60 0.91 0.97 1.03 1.05 1.04 1.04
Example 3 (Bypass . . . Conclusions
Go Back Close Quit
An Example
Example 1 (III) Comparison with an energy estimator
Formulation of the . . . Optimization algorithm Discretization by Finite . . . Adaptive Mesh . . . Example 1 Example 2 Example 3 (Bypass . . . Conclusions
Go Back Close Quit
Example 1 (IV) Comparison with an energy estimator

global energy our 0.1
0.01
0.001
0.0001
Go Back Close
1e-05 1000
Quit
10000
100000
An Example
Example 1 (V)
Comparison with a heuristic approach
Go Back Close Quit
Example 1 (VI) Comparison with a heuristic approach

global r=0.1 our 0.1
0.01
0.001
0.0001
Go Back Close
1e-05 1000
Quit
10000
100000
Example 2 u u + f (u) = 0 in
E Arrhenius law: f (u) = A exp( 1y )y(c y)
Ox F
Go Back
Ox
Close Quit
R. Becker, M. Braack and B. Vexler
Numerical parameter estimation for chemical models in reactive ows, Combustion Theory and Modelling, submitted, 2003
Example 2 (II) Initial parameters : q = (4, 0.15)

Optimal parameters: q = (6.9, 0.07)
Go Back Close Quit
Example 2 (III)
0.01
An Example
global local
Formulation of the . . . Optimization algorithm Discretization by Finite . . . Adaptive Mesh . . . Example 1 Example 2 Example 3 (Bypass . . .
0.001
Conclusions
0.0001
10000
100000
Go Back Close Quit
For T OL = 1% :
Global renement: N = 107585, e = 0.54% Local renement : N = 7081, e = 0.89%
Example 2 (IV)
Go Back Close Quit
Example 3 (Bypass Simulation)

(2) in 0 (1) in out
v + v
v+
p = 0,
v = 0 in
Boundary Conditions:
v ? ? ?
1
= = = =
0 on 0 (1) ? on in (2) ? on in ? on out
Go Back Close Quit
Calibration of PDE Models and Sensitivity Analysis with Adaptive Finite Elements: Application to CFD Problems, Journal of Comp. Physics, to submit, 2003
Example 3 (Bypass Simulation)

(2) in 0 (1) in out
v + v
v+
p = 0,
v = 0 in
Boundary Conditions:
v nv p n nv p n nv p n
1
= 0 on 0 (1) = q1 n on in (2) = q2 n on in = 0 on out
Go Back Close Quit
Calibration of PDE Models and Sensitivity Analysis with Adaptive Finite Elements: Application to CFD Problems, Journal of Comp. Physics, to submit, 2003
Example 3 (II)
(2) in
S1
0
S3
Optimization algorithm Discretization by Finite . . .
(1) in
S2
out
Adaptive Mesh . . . Example 1 Example 2 Example 3 (Bypass . . . Conclusions
1 Minimize 2
(Ci(v) Ci)2
i=1
v + v
v+
p = 0,
v = 0 in , BC(q).
Go Back Close Quit
Ci(v) =
Si
vn
Example 3 (III)
Quantity of interest:
E(u) =
O
n (u) d ds,
1 = ( v + ( v)T ) pI, d = (0, 1)T 2

Error E(u) E(uh ) = ?
Go Back Close Quit
due to the discretization due to the error in parameter q qh
Example 3 (IV)
uniform local enegry 0.1
0.01
Go Back Close
0.001
Quit
1000 10000 100000
An Example
Conclusions Parameter identication problem Discretization by nite elements A posteriori error estimation
Error estimator for the error in parameter Error estimator for the quantity of interest An auxiliary equation
Numerical Examples
Efciency of the estimator Quality of the generated meshes
Go Back Close Quit
Coupled Algebraic Multigrid Navier-Stokes Methods for thep Equations

Markus Wabro Joint work with Walter Zulehner Institute of Computational Mathematics Johannes Kepler University Linz
www.numa.uni-linz.ac.at
This work has been supported by the Austrian Science Foundation Fonds zur F orderung der wissenschaftlichen Forschung (FWF) under the grant P14953 Robust Algebraic Multigrid Methods and their Parallelization
Advances in Numerical Algorithms, Sept. 1013, 2003 p.1/31
Overview
Introduction (Navier-Stokes / Oseen equations)
Institute of Computational Mathematics Johannes Kepler University Linz
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Overview
(Algebraic) multigrid methods for the solution of the

Oseen equations
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Overview

Oseen equations
Decoupled Approach
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Overview

Oseen equations
Decoupled Approach Coupled Approach

Coarse-level construction Smoothing
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Overview

Oseen equations
Decoupled Approach Coupled Approach

Coarse-level construction Smoothing Numerical results
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Aim
Complex 3D domain
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Aim
Complex 3D domain
the incompressible Navier-Stokes equations
t u
u + (u
)u +
p = f,
div u = 0,
+ boundary/initial conditions,
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Aim
Complex 3D domain
t u
u + (u
)u +
p = f,
div u = 0,
+ boundary/initial conditions, Finite Element Discretization
f u A(u) B T = ; p B C g
Coupled AMG q Num. Results q
Introduction q
Decoupled MG q
Aim
Complex 3D domain
t u
u + (u
)u +
p = f,
div u = 0,
+ boundary/initial conditions, Finite Element Discretization
f u A(u) B T = ; p B C g
We want an efcient Solver.

Introduction q Decoupled MG q
Convective Term
Linearization: Fixed point iteration ; Oseen
equation
u + (w
)u +
p=f
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Convective Term
Linearization: Fixed point iteration ; Oseen
equation
u + (w
Stabilization (SUPG):
)u +
p=f
replace v by v + h (w )v ;
((w
with h
Introduction q
)u, v)+ h (w u, w v), +(amongst others)
= O(h).
Decoupled MG q
Used Mixed Finite Elements

P1 -P1 -stab:
C=
elts K
h2 ( p, q)0,K K
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q

P1 -P1 -stab:
modied Taylor-Hood:
C=
elts K
h2 ( p, q)0,K K
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q

P1 -P1 -stab:
modied Taylor-Hood:
C=
elts K
h2 ( p, q)0,K K
nc P1 -P0 :
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
How should we solve the linear saddle point problem

f u A(w) B T = g p B C
(indenite, in general unsymmetric)?
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
How should we solve the linear saddle point problem

f u A(w) B T = g p B C
(indenite, in general unsymmetric)? Efciently!
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
MGM for Saddle-Point-Problems
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q

Iterative Decoupling
Pressure correction
methods: SIMPLE, Uzawa Precond. Krylov space methods (e.g. GMRES, BiCGstab) MG for the elliptic systems.
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q

Iterative Decoupling
Coupled MG for the whole system. Needs appropriate MG components (smoother,
Pressure correction
methods: SIMPLE, Uzawa Precond. Krylov space methods (e.g. GMRES, BiCGstab) MG for the elliptic systems.
inter-grid-transfer).
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Decoupled Approach SIMPLE

u A = f B T pk ,
uk+1 = u D1 B T p, pk+1 = pk + p,
S p = B u Cpk g,
damping .
with D diagonal of A, A A, S C + BD 1 B T ,
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Decoupled Approach BB
Black-Box-preconditioner
Elman,Kay,Loghin, Silvester,Wathen
Find a good preconditioner (not necessarily solver),

combine it with Krylov space method (e.g. BiCGstab)
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q

combine it with Krylov space method (e.g. BiCGstab) I 0 A1 0 I BT 1 = , K 0 S 1 0 I 0 I
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q

combine it with Krylov space method (e.g. BiCGstab) I 0 A1 0 I BT 1 = , K 0 S 1 0 I 0 I
[ + (w
)]
[ + (w
)]s 1 s
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q

combine it with Krylov space method (e.g. BiCGstab) I 0 A1 0 I BT 1 = , K 0 Q1 Ap L1 0 I 0 I
[ + (w
)]
[ + (w
)]s 1 s
Ap pressure conv.-diff. operator, Q pressure mass matrix (lumping),

Introduction q Decoupled MG q Coupled AMG q Num. Results q
L pressure Laplacian,
A A,
Algebraic Multigrid
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Algebraic Multigrid
Complex geometry
discretization
many unknowns.
No further renement possible.
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Algebraic Multigrid
Complex geometry
discretization
many unknowns.
Algebraic Multigrid. Starting from the nest levels

generate the coarse levels (almost) only from matrix information.
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Algebraic Multigrid
Complex geometry
discretization
many unknowns.
Algebraic Multigrid. Starting from the nest levels

generate the coarse levels (almost) only from matrix information. O.K. for elliptic problems, application to the decoupled approach is straightforward. What about coupled AMG?
Coupled Approach Strategy
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q

Separate pressure and velocity-component
unknowns.
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q

unknowns.
Use interaction of u and p indicated by nite element.
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q

unknowns.
Use interaction of u and p indicated by nite element. Try to translate as much as possible from geometric MG for saddle point problems and from elliptic AMG.
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q

unknowns.
Use interaction of u and p indicated by nite element. Try to translate as much as possible from geometric MG for saddle point problems and from elliptic AMG. Black box coupled AMG?
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q

unknowns.
Use interaction of u and p indicated by nite element. Try to translate as much as possible from geometric MG for saddle point problems and from elliptic AMG. Black box coupled AMG?
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Notation
Grid transfer:
Prolongators
l Pl+1
l Il+1
l Jl+1
l l l l Il+1 = diag(Il+1 , Il+1 , Il+1 ),
l Il+1 : Ul+1 Ul ,
l Jl+1 : Ql+1 Ql ,
Ul := (Rnl )d Ql := Rml
l T Pl+1 ).
Restrictors
Kl =
T Al Bl Bl Cl
Pll+1
(=
Coarse level systems:

l = Pl1 Kl1 Pll1
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Notation
Grid transfer:
Prolongators
l Pl+1
l Il+1
l Jl+1
l l l l Il+1 = diag(Il+1 , Il+1 , Il+1 ),
l Il+1 : Ul+1 Ul ,
l Jl+1 : Ql+1 Ql ,
Ul := (Rnl )d Ql := Rml
l T Pl+1 ).
Restrictors
Kl =
T Al Bl Bl Cl
Pll+1
(=
Coarse level systems:

l = Pl1 Kl1 Pll1
With two exceptions (h-dependence in conv. stab. and in element stab.)!

Mod. Taylor-Hood Coarsening

1st level
ne (u) mesh
u
2nd level
Coarsening
Renement
coarse (p) mesh
u p u. p
..
L-th level
u p
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
PSfragMod. Taylor-Hood replacements

u
Coarsening
u
2nd level
Coarsening
Coarsening
Renement
coarse (p) mesh
1st level
p ne (u) mesh
u p u. p
..
L-th level
Renement ne (u) mesh coarse (p) mesh
u p
1st level
2nd level
..
L-th level
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q

u
Coarsening
u
2nd level
Coarsening
Coarsening
Renement
coarse (p) mesh
1st level
p ne (u) mesh
u p u. p
..
L-th level
u p
1st level
2nd level
..
L-th level
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q

u
Coarsening
u
2nd level
Coarsening
Coarsening
Renement
coarse (p) mesh
1st level
p ne (u) mesh
u p u. p
..
L-th level
u p
1st level
2nd level
..
L-th level
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q

u
Coarsening
u
2nd level
Coarsening
Coarsening
Renement
coarse (p) mesh
1st level
p ne (u) mesh
u p u. p
..
L-th level
u p
1st level
2nd level
..
L-th level
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q

u
Coarsening
u
2nd level
Coarsening
Coarsening
Renement
coarse (p) mesh
1st level
p ne (u) mesh
u p u. p
..
L-th level
u p
1st level
2nd level
..
L-th level
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q

u
Coarsening
u
2nd level
Coarsening
Coarsening
Renement
coarse (p) mesh
1st level
p ne (u) mesh
u p u. p
..
L-th level
u p
1st level
2nd level
..
L-th level
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q

u
Coarsening
u
2nd level
Coarsening
Coarsening
Renement
coarse (p) mesh
1st level
p ne (u) mesh
u p u. p
..
L-th level
u p
1st level
2nd level
..
L-th level
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
replacements
Convergence
Approximation Property Smoothing Property
Stability
Standard AMG Standard GMG Techniques
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
replacements
Convergence
Standard AMG Standard GMG LBB-Condition e.g. Techniques
Stability
vBlT p inf ? sup qQl vUl v Ul q
Ql
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
replacements
Convergence
Standard AMG Standard GMG LBB-Condition e.g. Techniques
Stability
vBlT p inf ? sup qQl vUl v Ul q
Ql
unstable behaviour:
Mod. Taylor-Hood Stability
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Mod. Taylor-Hood Stability

no analytic results
only heuristics (enough velocity unknowns)

Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
P1-P1-stab Coarsening
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
velocity and pressure live on the same nodes
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
velocity and pressure live on the same nodes
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
P1-P1 -stab Stability (1)

Geometric case: Franca, Stenberg (1991) using idea of Verfurth (1984):

sup
0=vVh
(div v, p) p v 1
1 2
h2 K
elts K
2 0,K
p Qh
Stability
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q

AMG case: Assume
ADl is of essentially positive type i.e. for all v l Ul

k,j
(akj )(vk vj )2
k,j
(a )(vk vj )2 . kj
For all vl Ul we can nd a l+1 vl Ul+1 such that l

l vl Il+1 l+1 vl l
2 Dl
vl
2 AD l .
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q

AMG case: Assume
ADl is of essentially positive type i.e. for all v l Ul

k,j
(akj )(vk vj )2
k,j
(a )(vk vj )2 . kj
For all vl Ul we can nd a l+1 vl Ul+1 such that l

l vl Il+1 l+1 vl l
2 Dl
vl
2 AD l .
Then for all levels l and all (u, p) Ul Ql

sup
0=vUl 0=qQl
Bl (u, p; v, q) l v AD l + q Ml
AD l
+ p
Ml
Introduction q
with Bl (u, p; v, q) = uT ADl v + pT Bl v + uT BlT q pT Cl q .

Decoupled MG q Coupled AMG q Num. Results q
nc P1 -P0 Coarsening (1)
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Use AMGe
Brezina,Cleary,Falgout,Henson, Jones,Manteufel,McCormick,Ruge Jones, Vassilevski
with element
agglomeration
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Use AMGe
Brezina,Cleary,Falgout,Henson, Jones,Manteufel,McCormick,Ruge Jones, Vassilevski
with element
agglomeration
adapted to our situation.
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
agglomeration identify unknowns build prolongation
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
pressure: identity prolongation
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
pressure: identity prolongation velocity: identity on edges, energy minimization in the interior
nc P1 -P0 Stability
Lemma: Assume existence of lin.Op. l : Ul1 Ul l1 with

b(l vl1 , ql ) = b(vl1 , Jll1 ql ) l1

and
ql Ql , vl1 Ul1 l vl1 l1

A
(1)
C vl1
vl1 Ul1 .
(2)
Then inf-sup in Ul1 Ql1 implies inf-sup in Ul Ql .
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
nc P1 -P0 Stability
Lemma: Assume existence of lin.Op. l : Ul1 Ul l1 with

b(l vl1 , ql ) = b(vl1 , Jll1 ql ) l1

and
ql Ql , vl1 Ul1 l vl1 l1

A
(1)
C vl1
vl1 Ul1 .
(2)
Then inf-sup in Ul1 Ql1 implies inf-sup in Ul Ql . (1), (2) can be shown using purely geometric arguments.
Smoothing (1)
Use smoothers from GMG for saddle point systems, e.g.
Vanka-Smoother:
Solve smaller local problems. (1 pressure unknown + connected velocities) Combine solutions via (multiplicative) Schwarz method. Zulehner/Schoberl: theoretical basis for GMG (additive variant).
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Smoothing (2)
Braess-Sarazin-Smoother: (symm. inexact Uzawa) u A( j+1 uj ) = f Auj B T pj ,
A(uj+1 uj+1 ) = B T (pj+1 pj ), where A A S C + B A1 B T

(AMG for
S(pj+1 pj ) = B uj+1 Cpj g,

(Gauss-Seidel)
C + B A1 B T
A1 . . . -Jacobi).
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Smoothing (2)

(AMG for

(Gauss-Seidel)
C + B A1 B T
A1 . . . -Jacobi).
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Smoothing (2)

(AMG for

(Gauss-Seidel)
C + B A1 B T
A1 . . . -Jacobi).
Problem: Must keep matrices for inner (S ) AMG

needs more memory.
Numerical Results P1 isoP2 -P1

1 1e-1 1e-2 BiCGStab,BBPre AMG-SIMPLE Coupled,MSM
2D valve, 105 vars,
residual
1e-3 1e-4 1e-5 1e-6
100
200 CPU-time [s]
300
400
1e-4 1e-5 1e-6 residual 1e-7 1e-8 1e-9 1e-10 1e-11 0 1000 2000
AMG-SIMPLE Coupled,Braess,1-shift Coupled,Braess,2-shift
3D valves, 7 105 vars,
3000
4000
5000
6000
7000
CPU-time[s]
(geometries provided by AVL List GmbH) Introduction q Decoupled MG q Coupled AMG q Num. Results q
Numerical Results P1-P1-stab
0.0001 1e-05 1e-06 residual 1e-07 1e-08 1e-09 1e-10
Coupled AMG-SIMPLE
7 105 vars,
mids, prisms;
0 4000 8000 CPU-time [s] 12000 16000
mesh: tets, hexas, pyra-
= 0.0005
(geometries provided by AVL List GmbH) Introduction q Decoupled MG q Coupled AMG q Num. Results q
Numerical Results h-Dependency

Stokes, driven cavity (unstructured grid)
0.01
0.001
BiCGStab,BBPre Coupled,Braess Coupled,MSM AMG-SIMPLE
level 5
unknowns 10,571 52,446 166,691 665,155 2,657,411
T0.1
0.0001
6 7 8
5 6 7 refinement level 8 9
1e-05
placements
1e-06
1e-07
T0.1 :=
CPU time [min.] for the red. of the norm of res. by 0.1 number of unknowns
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Numerical Results -Dependency

2D valve, Taylor-Hood, 103,351 unknowns
0.0001
BiCGStab,BBPre Coupled,Braess Coupled,MSM AMG-SIMPLE
T0.1
1e-05
PSfrag replacements
1e-06
10
100
1000
10000
1/
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
nc Numerical Results P1 -P0
Standard aggl.
Metis-renement use metis to nd an initial partition (coarsest level), rene recoursively.
ne coarse
agglomerate coarser elements
O(n)
O(n log n)!
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Standard aggl.
Metis-renement
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Standard aggl.
Metis-renement
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Standard aggl.
Metis-renement
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Standard aggl.
Metis-renement
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Standard aggl.
Metis-renement
8e-6
Standard Metis
4e-6
T0.1
2e-6 1e-6 5
PSfrag replacements
6 refinement level 7 8
(level 8: 1,180,930 unknowns)

Summary(1) (nearly the last slide)

We have developed coupled AMG methods for
saddle point problems originating in the Navier

Stokes equations, for
the mod. Taylor-Hood element the P1 -P1 stab element (stable!) the P1nc -P0 element (AMGe, stable!)
We have compared them to methods using a decoupled approach.
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Summary(2) (last slide but one)

BUT
We have absolutely no Black-Box method!

We have no rigorous prove for most aspects (e.g. convergence)!
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
The Last Slide
Thank you for your attention!
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Numerical Solution of Singular Boundary Value Problems in ODEs

Ewa B. Weinmuller
joint with Winfried Auzinger and Othmar Koch
e.weinmueller@tuwien.ac.at. Vienna University of Technology
E. Weinmuller / ANA 2003 - Graz p.1/29
Contents
Introductory remarks.
Contents
Introductory remarks. Problem class.
Contents
Introductory remarks. Problem class. Singularity of the rst kind: Numerical solution by collocation. Error estimate via QDeC/BEUL. MATLAB 6 Code sbvp1.0.
Contents
Introductory remarks. Problem class. Singularity of the rst kind: Numerical solution by collocation. Error estimate via QDeC/BEUL. MATLAB 6 Code sbvp1.0. Singularity of the second kind: Numerical solution by collocation. Error estimate via QDeC/BEUL. Error estimate via QDeC/BOX. Error estimate based on mesh halving.
Contents
Introductory remarks. Problem class. Singularity of the rst kind: Numerical solution by collocation. Error estimate via QDeC/BEUL. MATLAB 6 Code sbvp1.0. Singularity of the second kind: Numerical solution by collocation. Error estimate via QDeC/BEUL. Error estimate via QDeC/BOX. Error estimate based on mesh halving. Conclusions and future work.
Introduction - sol. structure

Case study: Consider linear system
M z (t) = z(t), t (0, 1], z(t) R3 . t

M z (t) = z(t), t (0, 1], z(t) R3 . t
Does a continuous solution z C[0, 1] exist?

M z (t) = z(t), t (0, 1], z(t) R3 . t
Does a continuous solution z C[0, 1] exist? Decouple the system: D = E 1 M E = diag(+ , 0, )

D (t) = (t), (t) := E 1 z(t) t

M z (t) = z(t), t (0, 1], z(t) R3 . t

D (t) = (t), (t) := E 1 z(t) t 1 (t) = c1 t+ C[0, 1]

M z (t) = z(t), t (0, 1], z(t) R3 . t

D (t) = (t), (t) := E 1 z(t) t 1 (t) = c1 t+ C[0, 1] 2 (t) = c2 C[0, 1]

M z (t) = z(t), t (0, 1], z(t) R3 . t

D (t) = (t), (t) := E 1 z(t) t 1 (t) = c1 t+ C[0, 1] 2 (t) = c2 C[0, 1] 3 (t) = c3 t C(0, 1]
Introduction - sol. structure (2)

Positive eigenvalue of M :
1 (t) = c1 t+

Zero eigenvalue of M :
2 (t) = c2

Negative eigenvalues of M :
3 (t) = c3 t

c1 t + z(t) = E(t) = E c2 = c1 t+ E1 + c2 E2 + c3 t E3 c3 t
General solution of the original problem:

Projections: Q ... onto the subspace of R3 spanned by E3 and P := I Q.

Projections: Q ... onto the subspace of R3 spanned by E3 and P := I Q. Note that

Qz(t) = c3 t E3 and P z(t) = c1 t+ E1 + c2 E2 .

Projections: Q ... onto the subspace of R3 spanned by E3 and P := I Q. Note that

Qz(t) = c3 t E3 and P z(t) = c1 t+ E1 + c2 E2 . z C Qz(0) = 0 and P z(1) = P , R3 arbitrary.
Therefore

Generalizations: (F. de Hoog and R. Weiss, 1976)
z (t) =
M t z(t) + f (t),
f C[0, 1], z(t) Rn

z (t) =
M t z(t) + f (t),
f C[0, 1], z(t) Rn
general spectrum of M (complex eigenvalues and Jordan canonical form)

z (t) =
M t z(t) + f (t),
f C[0, 1], z(t) Rn

z (t) =
M +tC(t) z(t) + f (t), t
C C[0, 1]

z (t) =
M t z(t) + f (t),
f C[0, 1], z(t) Rn

z (t) = z (t) =
M +tC(t) z(t) + f (t), t
C C[0, 1]
M +tC(t) z(t) + f (t, z(t)) t

z (t) =
M t z(t) + f (t),
f C[0, 1], z(t) Rn

z (t) = z (t) =
M +tC(t) z(t) + f (t), t
C C[0, 1]
M +tC(t) z(t) + f (t, z(t)) t
Structure of the boundary conditions:

Qz(0) = 0 z C[0, 1] P z(1) = P Ba z(0) + Bb z(1) =
Introduction - num. solution

M z (t) = z(t), t (0, 1] + BC t
Introduction - num. solution

M z (t) = z(t), t (0, 1] + BC t
Apply box scheme on the equidistant mesh: stepsize h = 1/N , meshpoints ti , i = 0, 1, . . . , N : Let i = 1, 2, . . . , N, then
i i1 M i + i1 h [i ] := = 0 + BC h ti1/2 2
Introduction - num. sol. (2)

Classical approach: Stability + Consistency = Convergence!

Classical approach: Stability + Consistency = Convergence! Write the scheme as
h [] = 0, h : R3(N +1) R3(N +1) .

h [] = 0, h : R3(N +1) R3(N +1) .
Stability (means that h is one to one mapping):

const. h [] h []

h [] = 0, h : R3(N +1) R3(N +1) .
Stability (means that h is one to one mapping):

const. h [] h []
Consistency: For z sufciently smooth

Rh [z(ti )] := h [z(ti )] = O(h2 ).

Consequently,
h [] = 0, h [z] = O(h2 )

Consequently,
h [] = 0, h [z] = O(h2 ) z const. h [z]h [] = const. h [z] = O(h2 )

Consequently,
h [] = 0, h [z] = O(h2 ) z const. h [z]h [] = const. h [z] = O(h2 ) 2 h Difculty: Rh [z(ti )] = O ti1/2 = O(h).

Consequently,
Remedy:

Consequently,
Remedy:
Invert h and estimate

Consequently,
Remedy:
Invert h and estimate

z = 1 Rh [z] ! h
Problem class
z (t) =
M (t) t z(t)
+ f (t, z(t)), t (0, 1],
Qz(0) = 0, Baz(0) + Bbz(1) = .
Problem class
z (t) =
M (t) t z(t)
+ f (t, z(t)), t (0, 1],
Qz(0) = 0, Baz(0) + Bbz(1) = .

= 1 . . . singularity of the rst kind ( < 1 weak)
Problem class
z (t) =
M (t) t z(t)
+ f (t, z(t)), t (0, 1],
Qz(0) = 0, Baz(0) + Bbz(1) = .

= 1 . . . singularity of the rst kind ( < 1 weak) > 1 . . . singularity of the second kind (essential)
Problem class
z (t) =
M (t) t z(t)
+ f (t, z(t)), t (0, 1],
Qz(0) = 0, Baz(0) + Bbz(1) = .

= 1 . . . singularity of the rst kind ( < 1 weak) > 1 . . . singularity of the second kind (essential)
Assume: There exists a solution z . The problem is well-posed. Solution z is appropriately smooth.
Collocation method
z (t) F (t, z(t)) = 0 plus BC
Collocation method
i,j
. . . ti,j . . . i+1
hi h
...
...
Collocation method
i,j
. . . ti,j . . . i+1
hi h
...
...
Globally continuous, piecewise polynomial function p(t) of maximal degree m satisfying

p (ti,j ) F (ti,j , p(ti,j )) = 0 plus BC.
Collocation method
i,j
. . . ti,j . . . i+1
hi h
...
...

Convergence regular case:
Collocation method
i,j
. . . ti,j . . . i+1
hi h
...
...

Convergence regular case: stage order: p z
= O(h ),
Collocation method
i,j
. . . ti,j . . . i+1
hi h
...
...

Convergence regular case: stage order: p z
= O(h ),
superconvergence at i .
Convergence
=1
z (t) =
M +tC(t) z(t) t
+ f (t, z(t)) + BC
Convergence
=1
z (t) =
M +tC(t) z(t) t
+ f (t, z(t)) + BC
de Hoog and Weiss (1978): linear case, M has no eigenvalues with positive real parts
Convergence
=1
z (t) =
M +tC(t) z(t) t
+ f (t, z(t)) + BC
m
stage order: p z
= O(h ),
Convergence
=1
z (t) =
M +tC(t) z(t) t
+ f (t, z(t)) + BC
m
stage order: p z
= O(h ),
no superconvergence at i , in general.
Convergence
=1
z (t) =
M +tC(t) z(t) t
+ f (t, z(t)) + BC
m
stage order: p z
= O(h ),
Auzinger, Koch, Weinmller (2002): nonlinear case, M has no eigenvalues with positive real parts
Convergence
=1
z (t) =
M +tC(t) z(t) t
+ f (t, z(t)) + BC
m
stage order: p z
= O(h ),
Auzinger, Koch, Weinmller (2002): nonlinear case, M has no eigenvalues with positive real parts Koch (2003): nonlinear case, spectrum of M arbitrary
Estimate QDeC/BEUL
Estimate QDeC/BEUL
z (t) F (t, z(t)) = 0 plus BC p (ti,j ) F (ti,j , p(ti,j )) = 0 plus BC
Estimate QDeC/BEUL
Estimate
p(ti,j ) z(ti,j ), i, j
Estimate QDeC/BEUL
Estimate
p(ti,j ) z(ti,j ), i, j
Classical idea due to Zadunaisky, Frank, Stetter
(1975-1978). Modication: QDeC, Auzinger, Koch, Weinmller (2001-2002).

Estimate QDeC/BEUL (2)

Auxiliary scheme for OP: BEUL
i,j i,j1 = F (ti,j , i,j ) + BC. i,j

i,j i,j1 = F (ti,j , i,j ) + BC. i,j
Neighboring scheme: BEUL

i,j i,j1 = F (ti,j , i,j ) + di,j + BC, i,j di,j :=
p(ti,j )p(ti,j1 ) i,j
m+1 k=1 j,k
F (ti,k , p(ti,k ).

i,j i,j1 = F (ti,j , i,j ) + BC. i,j
Neighboring scheme: BEUL

i,j i,j1 = F (ti,j , i,j ) + di,j + BC, i,j di,j :=
p(ti,j )p(ti,j1 ) i,j
m+1 k=1 j,k
F (ti,k , p(ti,k ).
Error estimate:
p(ti,j ) z(ti,j ) i,j i,j , i, j.
QDeC/BEUL - ASCOR
Consider
M (t) z (t) = z(t) + f (t, z(t)), t (0, 1], t Qz(0) = 0, Ba z(0) + Bb z(1) = .
QDeC/BEUL - ASCOR
Consider
M (t) z (t) = z(t) + f (t, z(t)), t (0, 1], t Qz(0) = 0, Ba z(0) + Bb z(1) = .
Then, if h is sufciently small,

(p(ti,j ) z(ti,j )) (i,j i,j ) = O(h
m+1
Auzinger, Koch, Weinmller (2002), Koch (2003).
MATLAB Code sbvp1.0

Auzinger, Kneisl, Koch, Weinmller (2002) http://www.math.tuwien.ac.at/ewa/ General purpose code written in M ATLAB.
MATLAB Code sbvp1.0

Auzinger, Kneisl, Koch, Weinmller (2002) http://www.math.tuwien.ac.at/ewa/ General purpose code written in M ATLAB. Suitable for a wide range of tolerances variable order method.
MATLAB Code sbvp1.0

Auzinger, Kneisl, Koch, Weinmller (2002) http://www.math.tuwien.ac.at/ewa/ General purpose code written in M ATLAB. Suitable for a wide range of tolerances variable order method. Efcient estimate of the global error.
MATLAB Code sbvp1.0

Auzinger, Kneisl, Koch, Weinmller (2002) http://www.math.tuwien.ac.at/ewa/ General purpose code written in M ATLAB. Suitable for a wide range of tolerances variable order method. Efcient estimate of the global error. Global error estimate made available to the user.
MATLAB Code sbvp1.0

Auzinger, Kneisl, Koch, Weinmller (2002) http://www.math.tuwien.ac.at/ewa/ General purpose code written in M ATLAB. Suitable for a wide range of tolerances variable order method. Efcient estimate of the global error. Global error estimate made available to the user. Global error control as basis for the mesh adaptation strategy reecting smoothness of the solution only.
Convergence

> 1 - Example
1 z (t) = 2 t
z2 (t) z3 (t) z4 (t) 1 ez1 (t)/2
, t (0, 1], 0 0 1 0 0 0 0 1 0 0 0 1
1 0 0 0
0 1 0 0
0 0 0 0
0 0 0 0
z(0) +
0 0 0 0
0 0 0 0
z(1) =
Equidistant coll.-points, m=4

Convergence of collocation
h 1/10 1/20 1/40 1/80 1/160 errgrid 2.06e17 4.26e18 1.65e19 6.22e21 3.84e22 2.27 4.69 4.73 4.02 pgrid errmesh 2.06e17 4.26e18 1.65e19 6.22e21 3.74e22 2.27 4.69 4.73 4.06 pmesh
Reference solution determined with h = 1/320.

QDeC/BEUL
Test problem: z (t) = t1 z(t) + et 3 Exact solution z(t) = et . Matrix of the BEUL scheme
h 1/10 1/20 1/40 1/80 1/160 cond 6.99e+05 3.28e+09 1.55e+15 9.60e+23 5.10e+37
=3
et t3 ,
z(1) = e
norm 1.00e+03 8.00e+03 6.40e+04 5.12e+05 4.10e+06
norminv 6.99e+02 4.09e+05 2.42e+10 1.88e+18 1.25e+31

Estimate QDeC/BOX
Neighboring scheme:
i,j i,j1 i,j + i,j1 ) + di,j + BC. = F (ti,j 1 , 2 i,j 2
h 1/2 1/4 1/8 1/16 1/32 1/64 1/128 errgrid 8.80e06 5.41e07 3.02e08 1.82e09 1.11e10 6.80e12 4.23e13 errgrid est 7.32e06 4.54e07 1.54e08 6.09e10 2.81e11 1.20e12 5.11e14 4.01 4.88 4.66 4.43 4.55 4.56 pgrid est errmesh 8.80e06 4.65e07 2.77e08 1.71e09 1.07e10 6.67e12 4.19e13 errmesh est 7.32e06 4.54e07 1.54e08 6.09e10 2.81e11 1.20e12 5.11e14 4.01 4.88 4.66 4.43 4.55 4.56 pmesh est
BEUL versus BOX

Test problem: z (t) = t1 z(t) + f (t), z(1) = e 3 Exact solution z(t) = et .
BEUL versus BOX

Test problem: z (t) = t1 z(t) + f (t), z(1) = e 3 Exact solution z(t) = et . Apply BEUL scheme
i,j1 = i,j 1 3 ti,j
O(h2 )
i,j fi,j , i = 1, . . . , N
BEUL versus BOX

Test problem: z (t) = t1 z(t) + f (t), z(1) = e 3 Exact solution z(t) = et . Apply BEUL scheme
i,j1 = i,j 1 3 ti,j
O(h2 )
i,j fi,j , i = 1, . . . , N
Apply BOX scheme

i,j1 = 2t3 i,j1/2 i,j 2t3 i,j1/2 + i,j
O(1)
i,j + fi,j1/2 , i = 1, . . . , N
Mesh halving
Collocation on mesh h: ph . Collocation on mesh h/2: ph/2 . Error structure: p(t) z(t) = e(t)hm + O(hm+ ), > 0 2m Error estimate: E(t) := 12m (ph/2 (t) ph (t))
h 1/2 1/4 1/8 1/16 1/32 1/64 1/128 errgrid 8.80e06 5.41e07 3.02e08 1.82e09 1.11e10 6.80e12 4.23e13 errgrid est 2.90e07 1.16e08 3.75e10 1.61e11 6.93e13 3.61e14 7.11e15 4.65 4.95 4.54 4.54 4.26 4.34 pgrid est errmesh 8.80e06 4.65e07 2.77e08 1.71e09 1.07e10 6.67e12 4.19e13 errmesh est 1.61e07 1.16e08 3.75e10 1.61e11 6.93e13 3.61e14 5.44e15 3.80 4.95 4.54 4.54 4.26 2.73 pmesh est
Summary
(1) Singularity of the rst kind Convergence of collocation: Theory ASCOR of the estimate QDeC/BEUL: Proof MATLAB Code sbvp1.0
Summary
(1) Singularity of the rst kind Convergence of collocation: Theory ASCOR of the estimate QDeC/BEUL: Proof MATLAB Code sbvp1.0 (2) Singularity of the second kind, experiments: Convergence of symmetric collocation schemes Error estimate QDeC/BEUL FAILS! Error estimate QDeC/BOX Error estimate based on mesh halving Higher order terms : O(hm+ ), < 1.
Future work
(1) Theory: Singularity of the second kind. Convergence of the collocation scheme. ASCOR of the estimate QDeC/BOX. ASCOR of the estimate based on mesh halving. Grid adaptation strategy.
Future work
(1) Theory: Singularity of the second kind. Convergence of the collocation scheme. ASCOR of the estimate QDeC/BOX. ASCOR of the estimate based on mesh halving. Grid adaptation strategy. (2) Software: Upgrade of the code sbvp1.0. Eigenvalue problems.
Numerical Solution of Singular Boundary Value Problems in ODEs

Ewa B. Weinmller
joint with Winfried Auzinger and Othmar Koch
e.weinmueller@tuwien.ac.at Vienna University of Technology
Applications
Singularity of the second kind: Regular problems on semi-innite intervals:
z (t) = t f (t, z(t)), t [a, ), > 1.
Applications
z (t) = t f (t, z(t)), t [a, ), > 1.
Extended Frank-Kamenitskii equation (convection-explosion theory, ame and wave propagation)

u 2m u = (1)m+1 2m + eu , t x x R, t > 0.
Applications
z (t) = t f (t, z(t)), t [a, ), > 1.
Extended Frank-Kamenitskii equation (convection-explosion theory, ame and wave propagation)

u 2m u = (1)m+1 2m + eu , t x x R, t > 0.
Self-similar blow-up solutions computed from

(1)m+1 z (2m) (t) tz (t) + ez(t) 1 = 0, t > 0.
Applications (2)
The nonlinear Schrdinger equation
u i + u + |u|2 u = 0, t t > 0.
Applications (2)
u i + u + |u|2 u = 0, t t > 0.
Associated nonlinear eigenvalue problem for ODE

d1 z (t)+ z (t)z(t)+ia(tz(t)) +|z(t)|2 z(t) = 0, t t > 0.
Applications (2)
u i + u + |u|2 u = 0, t t > 0.
Associated nonlinear eigenvalue problem for ODE

d1 z (t)+ z (t)z(t)+ia(tz(t)) +|z(t)|2 z(t) = 0, t t > 0.
Nonlinear optics, plasma physics.
Two-Dimensional Phase Unwrapping

Joint works (one each) with Gregory Dardyk Reuven Bar-Yehuda Iddit Shalem
Phase Unwrapping Application

Used in coherent signal processing applications, such as SAR, MRI. All coherent signal applications are based on a single signal property known as phase. The actual phase cannot be extracted from the physical signal All we can get is the wrapped phase ,i.e. the phase values forced into the interval (-,] by a modulo 2 operation.
SAR interferometry
Two images of the same scene are acquired from slightly different locations The images are coregistered into spatial alignment with one another The pixel-wise phase differences are proportional to terrain elevations
The Aim of Unwrapping

(x) real phase, unknown (x) = (x) + 2k(x) known wrapped phase, where k(x) is the unknown integer function that forces < < The aim of unwrapping reconstruct from The task is easy if
There are no discontinuities in the wrapped phase The resolution is sufficient
Easy example
Harder example
Wrapped image and the unwrapped elevation model (Etna volcano region)
Classical Unwrapping Methods

Path-following methods Minimum-norm methods
Path-following algorithms
Perform integration of the discrete gradients (wrapped differences) along paths Problem: need to pick paths that avoid problematic areas where the data is inconsistent
Minimum-norm algorithms
More global approach Minimize the distance between and the discrete gradient estimated from the values of the wrapped function .
Minimum Lp-norm functional

We require the grid-based solution (i,j) to minimize the discrete functional
h=
M 1 N
x i, j i +1, j i, j i =0 j =0
+ iy j , i, j +1 i, j i =0 j =0
M N 1
where xi,j and yi,j are the wrapped differences of , and p is the norm parameter.
Wrapped difference approximation

We assume that if (i,j) and (k,l) are adjacent points, then |i,j k,l|< Then, xi,j=W(i+1,j i,j) and yi,j=W(i,j+1 i,j), where W is the wrapping operator defined by W(fi,j) = fi,j +2ki,j with integer ki,j chosen so that W(fi,j)(-,]
The linear case

For p=2 this results in the discrete Poisson equation:
(i+1, j 2i, j +i1, j) + (i, j+1 2i, j i, j1) = (ix, j ix1, j) + (iy, j iy, j1)
This equation can be easily solved by Multigrid methods as well as others. The reconstruction is often unsatisfactory because of the exaggerated effect of outlying values.
The nonlinear case

Better reconstructions may be found using p<2 However, for such values of p the equations are nonlinear, often with nearly discontinuous coefficients For p<1 the problem is no longer convex. But p -> 0 generally yields the best results. This problem is NPhard.
Some numerical results

Wrapped image P=2
P = 1.5
P = 1 and P=0
Another example
Original image Wrapped image
Wrapped, view from above
P=2
Noisy data
P=1.5
P=0
Summary
We can obtain reconstructions efficiently even with small p using (semi)-classical nonlinear multigrid; (with G. Dardyk). But p<1 requires a continuation from p=1. For p -> 0, we have a factor-2 approximation algorithm, based on the local-ratio technique; (with R. Bar-Yehuda). Currently working on a discrete multi-scale algorithm; (with I. Shalem).
A Multi-level Approach to Quantization

Yair Koren, Irad Yavneh, Alon Spira Department of Computer Science Technion, Haifa 32000 Israel
One dimensional (scalar) quantization

Consider the image I consisting of G representation (gray) levels. We would like to represent I with n < G representation levels as best as possible. More formally, given a signal X (image, voice, etc.), with probability density function (histogram) p(x), we would like an approximation q(x) of X, which minimizes the distortion:
D(q ) = E[ X q ( X ) 2 ] =
2
di
i =1 di 1
x ri
2 2
p ( x ) dx.
Here, all x [di 1 , di ) are represented by ri .
Example (Lena, 512x512)

3000 2500
2000
1500
1000
500
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Gray Level
Left Lena (gray level image), Right Lenas histogram, p(x).
Lena, 8 gray levels

3000
2500
2000
1500
1000
500
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Gray Level
Representing Lena with less levels
128 gray levels
64 gray levels
Lena, 4 gray levels

3000 2500
2000
1500
1000
500
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Gray Level
Nave vs. Optimal Quantization
Lena, 8 levels, Left optimal, Right - Naive
The Lloyd Max Iterative Process

We wish to minimize
n di
D(q ) = E[ X q ( X ) 2 ] =
2
i =1 di 1
x ri
2 2
p ( x ) dx.
Differentiating w.r.t. r and d yields the Lloyd-Max equations:

di
ri =
xp( x)dx
di
di 1
, di =
ri + ri+1 2
, i = 1, , n 1
p ( x)dx
di 1
Max and Lloyd proposed a simple iterative process:

Given some initial guess, d 0, iterate for k = 1, 2, until some convergence criterion is satisfied:
d ik 1 dik1 1 dik 1 dik1 1
ri k =
xp ( x)dx , dik = p ( x)dx
k ri k + ri+1

We can rewrite the Lloyd-Max equations in terms of d alone:
di +1 di xp ( x ) dx xp ( x ) dx 1 di1 di + di+1 d i = di , i = 1, , n 1. 2 p ( x ) dx d p ( x ) dx d i 1 i
This is a generally a nonlinear system.

However, for the simple case, p = 1, L-M reduces to
dik = 1 k 1 di 1 + 2di + di +1 ) , i = 1, , n 1. ( 4
This is nothing but a damped Jacobi relaxation with damping factor 1/2 for the discrete Laplace equation. Evidently, multigrid acceleration is likely to help. We employ a nonlinear multigrid algorithm, using the Lloyd Max process for relaxation (with overrelaxation 4/3), and a nonlinear interpolation which retains the order of d.
Numerical Tests
We compare three algorithms: 1. Lloyd-Max, starting with a uniform representation 2. Our multigrid algorithm, starting similarly 3. LBG (Linze et al., 1980): Sequential refinement (coarse-to-fine). In all the algorithms, the basic iteration is Lloyd-Max.
P(x)=x
1 0.8 0.6 0.4 0.2 0 10 10
15
Converged Solution
10 10 10
RMS of residual vs. iterations
10
Multigrid LloydMax LBG 0

5
20
0.2
0.4
0.6
0.8
10
20
30
40
50
Residual reduction factor vs. iterations 1 0.8 0.6 0.4 0.2 0 0 10 20 30 4 3.9 3.8 3.7 3.6 LloydMax Multigrid LBG 40 50 3.5 3.4 3.3
x 10
Distortion vs. iterations
Multigrid LloydMax LBG
10
20
30
40
50
Step Function and local minima

8 6 4 2 0 Multigrid Converged Solution 8 6 4 2 0 LloydMax Converged Solution
0.2
0.4
0.6
0.8
0
3
0.2
0.4
0.6
0.8
8 6 4 2 0
LBG Converged Solution 1.75 1.7 1.65 1.6 1.55 1.5 1.45 0 0.2 0.4 0.6 0.8 1 1.4
x 10
Multigrid LloydMax LBG
10
20
30
40
50
Local Minima
1.5
3 x 10 Multigrid Converged Solution
1.5
3 x 10 LloydMax Converged Solution
0.5
0.5
0
3
0.2
0.4
0.6
0.8
0
8
0.2
0.4
0.6
0.8
1.5
x 10
LBG Converged Solution 4.2 4
x 10
Distortion vs. iterations Multigrid LloydMax LBG
3.8 3.6
0.5
3.4 3.2 3
0.2
0.4
0.6
0.8
20
40
60
80
100
Convergence Criterion
1.4 1.2 1 0.8 0.6 0.4 0.2 0 0 0.2 0.4 0.6 0.8 1 10
6 6 x 10 LloydMax Converged Solution
10
RMS of residual vs. iterations
10
10
100
11
200
300
400
500
1.1 1 0.9 0.8 0.7
Residual Reduction Factor vs. Iterations 2 1.9 1.8 1.7 1.6 1.5 1.4 0 100 200 300 400 500
x 10
100
200
300
400
500
Discrete Vector Quantization

The 1D problem is used mainly as a preliminary study towards higher-dimensional problems, viz., vector quantization (e.g., for color images).
Also, the p histogram is discrete in practice, and usually quite sparse and patchy and there are many different solutions (local minima). Standard multigrid methods do not seem appropriate.
Decision regions (Voronoi cells) and representation levels (centers of mass) for P(x,y)1
Equal height contours of P(x,y) = x*y
Decision regions for P(x,y)=x*y

Let G denote the number of possible representationlevels (D-tuples), P the number of such levels for which p does not vanish, and R the number of quantized representation levels. Typically,
A Lloyd Max iteration costs at least O(P) operations. As it doesnt seem possible to usefully coarsen p, coarselevel iterations will be equally expensive, resulting in O(P log(R)) complexity for the multigrid cycle.

Sketch of algorithm (V Cycle):
For j = 0,1, , log ( R ) 1, log ( R ) , log ( R ) 1, ,1, 0, Relax ( j )

Sketch of Relaxation algorithm:
Partition the R variables into, say, 2log( R ) j aggregates of 2 j representation levels each. Then sweep over the aggregates, changing each in turn so as to (approximately) minimize the fine-level functional (Gauss-Seidel style).
Conclusions
The multi-level approach is very promising for the problem of quantization. In 1D and (semi-) continuous p we get Much faster convergence. Often better minima. Sounder convergence criterion. The real dividends are expected for vector quantization (as in color images). This is a significantly harder and more important problem. Research on this is in progress, led by Yair Koren.
A Multigrid Approach to Binarization
Ron Kimmel and Irad Yavneh
Image Binarization
Original Image
Nonuniform Illumination
ilted
Spherical
Nave (threshold) binarization
ilted
Nave (threshold) binarization
Spherical
Yanowitz-Bruckstein Binarization
Isolate the locations of edge centers, for example, the set of points,
s = {( x , y ) : I > T }
for some threshold T. Use the values I(x,y), for (x,y) in s, as constraints for a threshold surface, u, which elsewhere satisfies the equation
u = 0.
For this we use our version of a multigrid algorithm with matrix-dependent prolongations.
Edges
ilted
Spherical
Results
ilted
Spherical

(Slide) Advances in NA2003

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

(Slide) Advances in NA2003

Uploaded by

Copyright:

Available Formats

 vn ~ VP~Rq$s n r r r n } ~RqRp`eRfd } snb$eRpEfR } r d n o r d r  r t d n `n } 3ihf~rR8 z } R~R `n~~ { n dr r gz E`n~  kn$ r u r d l o r Rn~ } R~r`fzn RVn`ne$`d r r n `n } zs`  Rqn h`s no snrj

u t r r  I~rfRf H nR ~R~ { u n ' r n d r ~yr zs`&fF wn  $`n~gz { n

ef } r } Rnsr } j o n n r  r } pzs } ~ 1| n dr n qp }  Pe } } VRn p~R  q { o o r

i~ r r ReRkoR~r r eg xB@7 I EC A9 r ot `nr } i ovn } } n  d rd r wFsR&`qnq

d f zh`' e sr } sfhrf  PPi~Pi n nd t t

rv  nfdn~g }  64 I 1B9 08 } r n 44 ( C ) 7 WWY ( ' Ax%&$ f PH CDAB9 $ !$ # Ax f yH DB9 $ W X C I C I C A w  h f d 7 ` VWWW  7

UTS Q yH DB@7 `nhr } i I C A9 r ot r d l `n~ } GR~r`fzn R 2  v {

"      )    Gg  " g

h H G XV T R YWUSH I H 9 h f yH ( 9  7 I C I ` h IPQ( @7 d d % C 9 f u~ n`s8$r } n ~Re$~ A r d r n d r r

r drn d n`s8$r } n ~Re$~ { r r

$r } n ~s`nh`' e~R } r d r   sR VvF z`n``r } si~qnzd on r dr n V s hd~riqR`d Rn s 2 { r o

gz sReRp`hf ur d r r n n r i tn `g` PF1  R k Ri$$p dRq

8 9   " 4 2 0 ( & $ "   76531)'%#!  v    I d C Q T d A d 7 PH T G 9 A I  9 7 h  7 9 T

dR`sq `n~heR`d } nhe$ r t

C I " !7 I "9 8#9 9 #) f #9 ! 7 I "

o r n nn } `gIsrd  knpo

V 2  nz~ d or v T w @" H v ` h Y h U ! 7 9 h " q A

  TUS Q &C  "  U 7 9 % $ f #9 U ! 7 f  " I " I # a  f 9 U B9  k$ nRid C 7 n d

q 7 0 f 0f gz `n``r } iPH 7 r dr n gI r d l o o `n } rRfsr`fzn $epGn5  Re

y s i sVRS`n~qR`gvzzsr r o n r d o n d u d  $Pe$`rno V eR`d VRS`n~$tRd { o n r  qn$ nRiIeRd { o

 ` ns vd s kBA 9 h d ~R } r  n  r '  n s`n(e ~$ } r ~r

r dr n f gz in`ssr } y zqR`Sf d r d l o f `n~ } VRfsrzn R0nP  R 2 (

I &`nq } eU~r$ { 9 no r n T T 7 h 9 7 $ f 7 V eRd {  #9 )  T  f I 9 I "  n r or t hu~R } r H n~  nzs  qRir {

Initial Function at t=0, X=1, Y=1, dx=0.05, dy=0.05, V=0

0.8 0.7 0.6 psi(x,y,0)

0.5 0.4 0.3 0.2 0.1 0

0.8 |psi(x,y)|, t=T

1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 15 0 10 5 0 y 15 x 10 5

DTBC vs. exp.ansatz (L=10,20,30), 40000 time steps

gz nr } r dr n ' n qpo rn(e R } `g V R(sg~$vs { on r t d

& 5I 9 n   v`n } q on r d & I q 9 % zv~R } Cv`n } q n  r o n r d r r r t s`n~ } ~P  H$s`nno ~~RRd { f A

o n r R RVhvqn v( } vn eRd I q  T 9 f R`R Rn } Rn n~  qn$ ~ id { rr n o  o r

" g ! g  g  5 " #    "  

relative L2Norm of reflected part for different number of coefficients

n sRz(  8` p } jy ~  kn$ g  qnH$ q ( t ~Re

1 0.9 0.8 0.7 |psi(x,y)|, t=T 0.6

0.5 0.4 0.3 0.2 0.1 20 0 0 30 10 20 30 40 40 x

0.014 0.012 0.01 0.008 0.006 0.004 20 0.002 30 0 0 5 10 15 20 25 30 35 x 40 50 50 40 45 y 10

z } Rlp sr$Fpg  ~ g } f A V t r r n `n } `g f ( fhn } g r n

2fIp fAf2 t y t u e w y t yw b u t r vfxAvRsc q fihfda p b g e c b  $W  0Y(X5

s tqrS TpRQP gPfQS 7 QRP edb c b`PYU&i4ehX W cVUTSRQP B ( ( c a ( ( ( 00( (0 0 ( H 34 I@ 4(CC( F

p gg A2 q h c t 8 11 1 1 ))))) 02) ) ) ) @ ) )h ( ( ( ( c t 3F 4( ( 4C C )) )) F 6 4C C @ DC( EAC( B G 4( ( ( 3

F 4C C pYXP4T3 Q ( ( h Pp@ h (( " $"      '&%#!

edjwkghcRb|{d{c  kYh2t jgw yfd z t t r w g c u ew b 8 ml f g t At 2t SjA0jx w y e t w y t gw e `

 !   "     

` XWH F H aF YQIw GE v s dz yU b g 2 s T rh2 p t z cw g e t g R PH F SQIw GE zD B d s k C A @ l k w d i v i z i { s o H o d v 9r i p o h d 8s { z 7 s { tYA i ~ t 6 z ge g  ! % ('& #$#  $ $% # 5 3 42 1 0 )

 !   "     

 t i h  { e i     ie i b { ' i  { f i     '% i e w g c b z t t r o w v k d i p o Asrhd|vRsc f c n "'&$%#!# # h% ! % ('$ "      '  $ s

100 0 90 Transmission Loss 10*log_10 |p|^2 80 70 60 50 40 30

10 12 Range r [km] Example 1

Workshop Advances in Numerical Algorithms Advanced Simulation Technologies

Graz, Austria, September 10-13, 2003

Turbulence modelling from the perspective of the commercial CFD

Dr. B. Basara AVL List GmbH

Some real-life examples

TM Product Logo Turbulence models

-known as Reynolds stresses are modelled

- robust ? - accurate ? - where is the optimum?

hUser Defined Boundary Layer hEntirely Conform

h User Defined Boundary Layer h Entirely Hex h Entirely Conform

h Semi-Automatic h Entirely Conform

Finite Volume Method

V (...)dV = A n (...)dA or V x k dV = A dAk

vn ~ VP~Rq$s n r r r n } ~RqRp`eRfd } snb$eRpEfR } r d n o r d r r t d n `n } 3ihf~rR8 z } R~R `n~~ { n dr r gz E`n~ kn$ r u r d l o r Rn~ } R~r`fzn RVn`ne$`d r r n `n } zs` Rqn h`s no snrj

u t r r I~rfRf H nR ~R~ { u n ' r n d r ~yr zs`&fF wn $`n~gz { n

ef } r } Rnsr } j o n n r r } pzs } ~ 1| n dr n qp } Pe } } VRn p~R q { o o r

i~ r r ReRkoR~r r eg xB@7 I EC A9 r ot `nr } i ovn } } n d rd r wFsR&`qnq

d f zh`' e sr } sfhrf PPi~Pi n nd t t

rv nfdn~g } 64 I 1B9 08 } r n 44 ( C ) 7 WWY ( ' Ax%&$ f PH CDAB9 $ !$ # Ax f yH DB9 $ W X C I C I C A w h f d 7 ` VWWW 7

UTS Q yH DB@7 `nhr } i I C A9 r ot r d l `n~ } GR~r`fzn R 2 v {

" ) Gg " g

h H G XV T R YWUSH I H 9 h f yH ( 9 7 I C I ` h IPQ( @7 d d % C 9 f u~ n`s8$r } n ~Re$~ A r d r n d r r

$r } n ~s`nh`' e~R } r d r sR VvF z`n``r } si~qnzd on r dr n V s hd~riqR`d Rn s 2 { r o

gz sReRp`hf ur d r r n n r i tn `g` PF1 R k Ri$$p dRq

8 9 " 4 2 0 ( & $ " 76531)'%#! v I d C Q T d A d 7 PH T G 9 A I 9 7 h 7 9 T

C I " !7 I "9 8#9 9 #) f #9 ! 7 I "

o r n nn } `gIsrd knpo

V 2 nz~ d or v T w @" H v ` h Y h U ! 7 9 h " q A

TUS Q &C " U 7 9 % $ f #9 U ! 7 f " I " I # a f 9 U B9 k$ nRid C 7 n d

q 7 0 f 0f gz `n``r } iPH 7 r dr n gI r d l o o `n } rRfsr`fzn $epGn5 Re

y s i sVRS`n~qR`gvzzsr r o n r d o n d u d $Pe$`rno V eR`d VRS`n~$tRd { o n r qn$ nRiIeRd { o

` ns vd s kBA 9 h d ~R } r n r ' n s`n(e ~$ } r ~r

r dr n f gz in`ssr } y zqR`Sf d r d l o f `n~ } VRfsrzn R0nP R 2 (

I &`nq } eU~r$ { 9 no r n T T 7 h 9 7 $ f 7 V eRd { #9 ) T f I 9 I " n r or t hu~R } r H n~ nzs qRir {

gz nr } r dr n ' n qpo rn(e R } `g V R(sg~$vs { on r t d

& 5I 9 n v`n } q on r d & I q 9 % zv~R } Cv`n } q n r o n r d r r r t s`n~ } ~P H$s`nno ~~RRd { f A

o n r R RVhvqn v( } vn eRd I q T 9 f R`R Rn } Rn n~ qn$ ~ id { rr n o o r

" g ! g g 5 " # "

n sRz( 8` p } jy ~ kn$ g qnH$ q ( t ~Re

z } Rlp sr$Fpg ~ g } f A V t r r n `n } `g f ( fhn } g r n

2fIp fAf2 t y t u e w y t yw b u t r vfxAvRsc q fihfda p b g e c b $W 0Y(X5

F 4C C pYXP4T3 Q ( ( h Pp@ h (( " $" '&%#!

edjwkghcRb|{d{c kYh2t jgw yfd z t t r w g c u ew b 8 ml f g t At 2t SjA0jx w y e t w y t gw e `

! "

` XWH F H aF YQIw GE v s dz yU b g 2 s T rh2 p t z cw g e t g R PH F SQIw GE zD B d s k C A @ l k w d i v i z i { s o H o d v 9r i p o h d 8s { z 7 s { tYA i ~ t 6 z ge g ! % ('& #$# $ $% # 5 3 42 1 0 )

! "

t i h { e i ie i b { ' i { f i '% i e w g c b z t t r o w v k d i p o Asrhd|vRsc f c n "'&$%#!# # h% ! % ('$ " ' $ s

} u o t o y o ~ } | lu { p t z yq u x wv i v l nq u o t s r m q p o n m l k j i h g f e d y x w u t t sr a` U` q V p T i W dh fTTaU U` W` UgTUY V` fv Y`RWWa setdsWr` UWQSTTcY ` a b a`Y Q T X W V Q R UU TS R Q P I @ H G @ B @ B A @ F E D C @ B @ B A @ 9 87 65 4 3 2 1 1 0

! ! ) ( ' $ ! ! ' & %$ # " ! ! !

' & % ! # $ # " !

H F % & !& ( & ) 1 f 5 04 3 @ 7 5 0 8 y x w `Y v v 1

`Y 8 X4 ) ) 8 6 6 W ) c F b ' $ D @ 8 9 a 82 V ' G T U $ %$ & T

x w v ut s r q # " ' I dV p i h XV Pg` Y fT P eV P dX P`` Y W X c W P` b SQa ` YT SV XT P WV U PT S R PQ PIH ! 6 5 4444444 ) 6 F ' G F A C E D AC # " B A @ 9 # 8 7 6 # 5 %

3 3 3 1 2 1 0 ) ' ( % ( % & ( % ! ' % %& % " ! $ #

9 8 7 1 0 ) ( ' & 5 6 5 3 2 1 0 ) ( ' & % 4