Professional Documents
Culture Documents
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$110(&$" !
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~`
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 ~
#
G "
g g
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 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$110(&$" !
&
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
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
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
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
zsr }
bcq I
z
9
V WU
a Y `
9 Cq I
a Y ` X
q v A v H
T
( '
P G
P I
8 9
P I
P G
I P h
HC xB9 7
P I
( $ )' "
b c V
P h H G 9 I PC QB9 7 d
f PFB@7 IHC 9
o r o hu n~R$t } H V
1 1 # ' % # )4320)$(&$"!
g !D"
1 1 __ __ 16 9
H C F E C A I3GDB@
1 __ 4
1 C
t __ x 2
I "9 ) #0
u
8 U
&
a
&
#
a
C f
!7 w U!7 I
T
v A
T
U
7 f nzs or v
8 U
8 U
7 h q
f
Y U 8 U
7 C
7 w
q h
f
H 7 h v
a 7
7
U
` 7
v 2
C A
f
9 @7
rnn r `g
C
D"
1 1 # ' % # )4320)$(&$"!
"
` P I
d ! 9 d
f yFB9 7 IHC
o r Rn `ne$`d
"
& &
U U
'
q Cq
9
7 0 # q
7
v q f q
7 0
Iq
9 f
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
1 0.9
1
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
0.6
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
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~$v d {
1 1 # ' % # )4320)$(&$"!
T
T T
C 4 4 C 4
f
!
T w
T
4 9 T f
T
T
C
T
9
nk } e { r
P P
I (9 9
"
C A
f GB9 I A
o r d Rn n~ qn$ m qno
g
'
"
h "
w 9
f
( C
C T Q
T C
f
( & $ )'%"
10 10 ||aref||L2 / ||I||L2 10
7
f (
r n fhn } g
E FD C A ' 3 1 1
g 8
10
10
0.1
t 0.2 L = 20 L = 30 L = 40 L = 50 0.3
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
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
Solution calculated with approximation of convolution coefficients, 5 terms in the sum of exponentials, X=Y=1, T=150, dx=0.02, dy=0.02, dt=0.0002, J=50, K=50, V=0
0.1 10 0.05 30 0 0 40 5 10 15 20 25 30 35 x
Solution calculated with approximation of convolution coefficients, 20 terms in the sum of exponentials, X=Y=1, T=150, dx=0.02, dy=0.02, dt=0.0002, J=50, K=50, V=0
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
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
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
"$ " #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
$ " %#
( 0 (0
yxv rwu
( yv u ( ( ( (
67 3 (4( ((
((
89
l Q
w v u d
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
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
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
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
AST-UMW_CFD_Name -4
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
AST-UMW_CFD_Name -5
TM
Product Logo
Automatic Tet Mesh
Meshing Options
Automatic Hex Mesh Block Structured Mesh Multi-Domain Meshing
h Arbitrary Interfaces
AST-UMW_CFD_Name -6
TM
Product Logo
Calculation grids
AST-UMW_CFD_Name -7
TM
Product Logo
- 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
n (...)dA = n (...)dA
A j =1 Aj
n
nf
or
dA =
A k j =1
nf
Aj
dAk
Rate of change: R
Convection: C
Diffusion: D
Sources: S
AST-UMW_CFD_Name -10
TM Product Logo
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
AST-UMW_CFD_Name -11
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.
AST-UMW_CFD_Name -12
j = f j P + (1 f j ) P
)
( 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
AST-UMW_CFD_Name -13
TM Product Logo
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
AST-UMW_CFD_Name -14
TM Product Logo
AST-UMW_CFD_Name -15
-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
AST-UMW_CFD_Name -16
TM Product Logo
AST-UMW_CFD_Name -17
DU i P = + xi x j Dt
2 ui u j = 2 t Sij ij k 3
U U j 2 U i k ij ui u j + x j xi 3 xk
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 ?
AST-UMW_CFD_Name -18
TM Product Logo
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 ?
AST-UMW_CFD_Name -19
-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
AST-UMW_CFD_Name -20
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
-1
-1,5
AST-UMW_CFD_Name -21
Product Logo
Renault model
AST-UMW_CFD_Name -22
Product Logo
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
AST-UMW_CFD_Name -24R
DU i P = + Dt xi x j
U U j 2 U i k ij ui u j + x j xi 3 xk
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
AST-UMW_CFD_Name -25
C
TM Product Logo
U i k 2 2 C = ui u j / S x j S = 2 Sij Sij
- Equation above is checked by using DNS data.
AST-UMW_CFD_Name -26
TM Product Logo
-measurements -measurements
-calculations
-calculations
AST-UMW_CFD_Name -27
TM Product Logo
K- HTM
-the worst convergence rate achieved with RSM, see Figure above (the grid is orthogonal !). - Basara & Jakirlic (2003)
AST-UMW_CFD_Name -28
TM Product Logo
K-
AVL HTM
C = 0.02 0.2
-separation predicted with HTM and RSM but not with the k- model (Basara 2001).
AST-UMW_CFD_Name -29
Product Logo
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
AST-UMW_CFD_Name -30
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
AST-UMW_CFD_Name -31
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.
AST-UMW_CFD_Name -32
TM
Product Logo
Conclusions
- 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
AST-UMW_CFD_Name -33
10 10 Semilog Plot 10 10 10
6 3 0 3 6
500
Eigenvalues on scale 9
3 4
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
$ #
"!
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
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
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
scaling function
0.5
0 wavelet 0.5 1 0 1 2
0.2 0.4 0.6 0.8 1
scaling function
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 @
50
100
150
200
250
300
350
400
450
500
u u u s t s r h q 2 7p
fe i h feg fe d
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
3 4
!0 v 5
8 ' # # # # A C # y
4444444
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
$ # "!
10
4 5 6
$ 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
Motivation Evaluation of (fractional) norms Uniform condition numbers Fast fully iterative scheme Potential for adaptivity
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
vH,
I) 2 (I L2
H = DHL2 , L2
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
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
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 )
1.5
0.5
-1
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
line 1 line 1 line 1
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
Numerical examples - 2D Solution y and control u for Z = U = L2 and dierent choices for
line 1 line 1 line 1
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
0.001
0.0001
1e-05
1e-06
1e-07
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
Bioremediation Strategies
INJECTION RECOVERY
MACROSCALE
GROWTH MECHANISMS Attachment Detachment Reproduction Adsorption Desorption Filtration Interaction
MICROSCALE FLOW
MESOSCALE
INPUT Substrate Suspended Cells Oxygen
! 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
! Easier and cheaper to write a new version of the application using modern programming environments and liberal code re-use.
Modeling Process
PHYSICAL PROCESS
PHYSICAL MODEL
MATHEMATICAL MODEL
OUTPUT VISUALIZATION
DISCRETE MODEL
NUMERICAL MODEL
INPUTS
! Fault-tolerant algorithms
SCIRun Methodology
! 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)
! 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.
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).
Motivations
fC :=
f N + fW + f S + f E 4
f = 0
C Gaspar
C Gaspar
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 )
ak pk ( x j ) = 0 (k = 1,2,..., m)
C Gaspar
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
j ( r ) := (1 c j r ) 2 + j ( r ) := (1 c j r ) 4 (4c j r + 1) + .........
C Gaspar
C Gaspar
Remedies
domain decomposition preconditioning fast multipole evaluation methods compactly supported RBFs multi-level techniques direct multi-elliptic interpolation
C Gaspar
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 }
Theorem: The direct and the variational problems are equivalent and have a unique solution in W.
C Gaspar Advances in Numerical Algorithms, Graz, 2003 10
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 ||
|| 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
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
C Gaspar
17
C Gaspar
18
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
j L ( xk x j ) = f k
M
u ( x) := j ( xk x j ) with
j =1
L =
C Gaspar
20
u ( xk ) = f k
(k = 1,2,..., M )
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
C Gaspar Advances in Numerical Algorithms, Graz, 2003 21
Advantages: reduction of dimension reduction of mesh Drawbacks: large, dense and ill-conditioned matrices
C Gaspar
22
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
C Gaspar
24
C Gaspar
25
C Gaspar
26
C Gaspar
28
u * = v0 n
= w0
Theorem 3: || u u* ||
H 1 ()
C (|| u0 || 3/ 2 + || v0 || 1/ 2 + || w0 || 1/ 2 ) H H () H () () c
C Gaspar
29
(k = 1,..., N )
or
(k = 1,2,..., N )
30
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.
C Gaspar Advances in Numerical Algorithms, Graz, 2003 31
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
Advances in Numerical Algorithms, Graz, 2003
( k = 1,2,..., N )
32
C Gaspar
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
C Gaspar
34
Workshop on Advances in Numerical Algorithms Sept. 10-13, 2003 at the University of Graz
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
Solve
K h uh = f h
uh RNh
( Ps ) .
= P K P
(or K
=
s=1
PsT Ks Ps)
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
it 14 28 63
speedup
0 0 2 4 6 processors 8 10 12
0 0 2 4 6 processors 8 10 12
grad u grad v dx
= (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
1.5
1 0 5 10 15 20 25 DOFNUMBER 30 35 40
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
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
10
11
12
14
10
11
H i
j j
H
12
Decouplings of patches
H i
j j
H
13
(r)
.
h
H = i
i h =
h h r: (r)
i h
(r)
.
h iC
.
h iC
(r) H (r) i
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
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
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
i li
(r)
(r)
h i C ,
(r)
= i/
j .
j: (r) H j
weights:
(r) li (r) lt
s l
17
18
i h
(r)
(r)
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
(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
us = A s u Ms = AsMAT s
r =
s=1 P
AT r s s AT KFEM As s s
s=1
K =
21
Theorems/Proofs [Haase]
22
Solve
s=1
AT KAs u = f s
23
solve
s=1
Ks u = f
d
s=1
ds ; dH P ds
wH 0 pmg-e (KH , wH , dH , 1) w Pw
H P
; w
s=1
ws
Summary
Patch based AMG real automatic coarsening. Moving of shape functions results in intergrid transfer weights
(r) i
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
Sonderforschungsbereich 609
Outline: Abstract setting Error estimate Numerical realization Numerical examples Conclusions
Outline: Abstract setting Error estimate Numerical realization Numerical examples Conclusions
Discretization
Concept
Discretization
Concept
Discretization
Concept
Examples:
Discretization
Concept
= {u U ; a u b a.e. in },
Discretization
Concept
= {u U ; a u b a.e. in },
divy = 0}.
Discretization
Concept
yz
2 Z
2 U.
ad
Discretization
Concept
yz
2 Z
2 U.
ad
Discretization
Concept
yz
2 Z
2 U.
ad
1 S (Su z)
Discretization
Concept
yz
2 Z
2 U.
ad
1 S (Su z) 1 Sh(Shu z) h
u = P U h
ad
Discretization
Concept
yz
2 Z
2 U.
ad
1 S (Su z) 1 Sh(Shu z) h
u = P U h
ad
Ch2 z
U
Z,
and
U.
(S S ShSh)u
Ch2 u
Discretization
Concept
yz
2 Z
2 U.
ad
1 S (Su z) 1 Sh(Shu z) h
u = P U h
ad
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
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
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
Abstract setting, proof of error estimate Proof: Add necessary optimality conditions and utilize u, uh as test functions; u u h
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
Abstract setting, proof of error estimate Proof: Add necessary optimality conditions and utilize u, uh as test functions; u u h
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
Ch2{ u
+ z
Z}
u u h
U.
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
(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
Ch2{ u Implications:
+ z
Z}
u u h
U.
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
(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
Ch2{ u Implications:
+ z
Z}
u u h
U.
+C
h h
2
1
h2| ln h| 2
3 2
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
(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
Ch2{ u Implications:
+ z
Z}
u u h
U.
+C
h h
2
1
h2| ln h| 2
3 2
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
(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
Ch2{ u Implications:
+ z
Z}
u u h
U.
+C
h h
2
1
h2| ln h| 2
3 2
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
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
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
Decoupling of FE grid and active set. T active set contains at most 2 connected components.
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
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
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
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
Discretization
Numerical realization
Discretization
Numerical realization
1 uk dk = uk uk + Sh(Shuk z) = Sh(Shuk z) = .
Discretization
Numerical realization
1 uk dk = uk uk + Sh(Shuk z) = Sh(Shuk z) = .
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
Numerical realization
1 uk dk = uk uk + Sh(Shuk z) = Sh(Shuk z) = .
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. Drawback: Convergence can be guaranteed only for > ShSh .
Discretization
Numerical realization
1 uk dk = uk uk + Sh(Shuk z) = Sh(Shuk z) = .
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. 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
Numerical realization
Discretization
Numerical realization
(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
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
y(0) = y(1) = 0.
x Let ua = , e := 2 + x2 xmin ub, x and z(x) 2. Then the unique optimal solution
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
1.5
0.5
0.25
Discrete active set
0.5
Discrete active set (standard fem) Active set
0.75
Discretization
Numerical example
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
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
Discretization
Numerical example II
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 .
Discretization
Numerical example II
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
- 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
Numerical example II
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
- 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
Numerical example II
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
Numerical example II
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.
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!!!).
http://www.math.tu-dresden.de/hinze
joint work with: Manfred Kaltenbacher, Reinhard Lerch, Reinhard Simkovics, Department of Sensor Technology, University of Erlangen Stefan Reitzinger, University of Linz
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
Piezoelectric PDEs:
2d t2
= 0 = 0
in in
boundary conditions: NT Dn Dn = = = = 0 0 1 A qe 0 on on on on g e \ (g e)
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
dielectr. const.
2 kt =
33 33 fs 2 fp
tan
fp fs 2 fp
10
10 Impedanz
10
10
10
10
10 Frequenz (Hz)
10
. . . potential part in solution of piezo PDEs q e . . . surface charge at loaded electrode . . . Fourier transform .
10
10
10
10
frequency
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
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
25
30
35
40
45
50
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
= 0 = 0
S = |DIV T d|
E = |grad|
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
e
35 30 8 x 10
9
Identication of S
45 40
e
35 30 8 x 10
9
Identication of e
50 40
e
30 20 8 x 10
9
Identication of e and S
60 50
e
40 30 8 x 10
9
e
35 30 8 x 10
9
50 0 8 x 10
9
d dS
, and
d de
10
10
10
10
. . . 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
A) = JI nA = 0
in D on B
JI = = B=
1 |c| Ieds c
A ds = A ,
B nd
Instability
Linearized inverse problem:
F ()s = F ()
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
10
10
10
12
10
14
10
16
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
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 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
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
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)
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
details
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
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.
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
CT slice
MR slice
PET slice
SPECT slice
Multimodal visualization
3
Atlas-based segmentation: register each patient to a segmented atlas Atlas-based segmentation. De ning pathology and norms.
4
B. Dawant, 2001
(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
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(x, 1) =
j = 1, . . . , j u(x(, ), )d.
I(x, 1) =
x]
1 (I1 (x)). T
+[
x u]
=0[
x]
= I.
Excision: (s) =
s + (reg TV).
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
| I u + Iz |2dxdz I(x, 1) =
1(I1 (x))
0(I0 (x)),
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()
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
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
a) + a),
u + Iz )2 dxdz
xu T
11
xu
dx
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
` 7
I0
I0
(x
1)
I1
2)
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
K0
0 ()
I0 ()= I1()
where:
1 2 0 U (, , 1) det (
x) d, x) d,
c 0 i 0
() 2 U (, , ()) det ( 0
In words:
0()
13
v C (Q)
Q [(
I u)( I v) +
2
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
x +
1 0 u(x( , ), )d,
= 1, . . . , .
15
w(x( , ), )
)v(x(, ), )d d v(x( , ), )
)w(x(, ), )d d
x/ u2 without D 2u.
16
I0
=1
I1
=0
17
|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:
Object Excision
|I0() |I1()
I1()| I0()|
|I0() |I1()
I1()| I0()|
19
|I0() |I1()
I1()| I0()|
|I0() |I1()
I1()| I0()|
20
Strongly Rigid
I0()| I1()|
Weakly Rigid
I0()|
Weakly Rigid
22
- INSTITUT FR AN ONN
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
( v
w + vw)dx A1f
2 (H 1t ())
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
such that
I) 2 (I
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
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 ;
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
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
=: 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
|v| v
:= sup n1/ vn nI N
:= v + |v|
w +
close to
11
vN
< N s v
=s+1 2
target accuracy
work 1/s 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 <
Theorem
A scompressible ;
w
A bounded on
=s+1 2 <
1/s
w
13
satises
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
Gravitational lensing
Image 1
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:
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
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
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) =
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
Themes of Talk
Optimization-based multigrid
Distinct from (& superior to?) traditional multigrid Optimization models as subproblems
Good performance (even far from solution) Connection to other optimization methods
R. Michael Lewis, College of William & Mary Stephen G. Nash, George Mason University buckaroo@math.wm.edu snash@gmu.edu
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
R. Michael Lewis, College of William & Mary Stephen G. Nash, George Mason University buckaroo@math.wm.edu snash@gmu.edu
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.
buckaroo@math.wm.edu snash@gmu.edu
R. Michael Lewis, College of William & Mary Stephen G. Nash, George Mason University
compute e(2) = IHh (a(2) a(1)) line search: a(2) a(1) + e(2)
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,
R. Michael Lewis, College of William & Mary Stephen G. Nash, George Mason University buckaroo@math.wm.edu snash@gmu.edu
Alternatives to MG/Opt
Apply multigrid to PDE alone:
not always possible/sensible
Successive refinement:
interesting
R. Michael Lewis, College of William & Mary Stephen G. Nash, George Mason University buckaroo@math.wm.edu snash@gmu.edu
Justification #1:
R. Michael Lewis, College of William & Mary Stephen G. Nash, George Mason University
buckaroo@math.wm.edu snash@gmu.edu
Justification #2:
Convergence
lim f ( x) = 0
If we minimize f ( x)
lim f ( x) = 0
R. Michael Lewis, College of William & Mary Stephen G. Nash, George Mason University
buckaroo@math.wm.edu snash@gmu.edu
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
R. Michael Lewis, College of William & Mary Stephen G. Nash, George Mason University
buckaroo@math.wm.edu snash@gmu.edu
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)
Justification #3:
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]
R. Michael Lewis, College of William & Mary Stephen G. Nash, George Mason University
buckaroo@math.wm.edu snash@gmu.edu
Justification #4:
(I
H h
AI
h H
u (a ) u*
-4
-3
-2
-1
Solution at t=1
-4
-3
-2
-1
buckaroo@math.wm.edu snash@gmu.edu
R. Michael Lewis, College of William & Mary Stephen G. Nash, George Mason University
buckaroo@math.wm.edu snash@gmu.edu
Justification #5:
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
R. Michael Lewis, College of William & Mary Stephen G. Nash, George Mason University buckaroo@math.wm.edu snash@gmu.edu
Justification #6:
Computational Performance
R. Michael Lewis, College of William & Mary Stephen G. Nash, George Mason University
buckaroo@math.wm.edu snash@gmu.edu
Justification #7:
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
R. Michael Lewis, College of William & Mary Stephen G. Nash, George Mason University buckaroo@math.wm.edu snash@gmu.edu
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
R. Michael Lewis, College of William & Mary Stephen G. Nash, George Mason University buckaroo@math.wm.edu snash@gmu.edu
Questions?
R. Michael Lewis, College of William & Mary Stephen G. Nash, George Mason University
buckaroo@math.wm.edu snash@gmu.edu
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
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
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).
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!
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
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.
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)
Overview
High Performance Computers Example Application I: Material Science Example Application II: Biomedical Engineering Conclusions
12.09.2003
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
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
1G n n Merced
Year
Semiconductor Technology
1012 10 9 10 6 10 1
3
2017
kT
10 -3 10 -6 10 -9
2010
2020
Year
1024
1015
Atoms/Bit
10
Future Challenge
12.09.2003
11
12.09.2003
12
Diploma thesis: N. Threy Dissertation project: T. Pohl Cooperation mit Department of Material Science, C Krner KONWIHR-Project FreeWihr
Prof. Dr. Ulrich Rde Lehrstuhl fr Systemsimulation Universitt Erlangen-Nrnberg
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
12.09.2003
17
12.09.2003
18
Rising bubbles
12.09.2003
19
12.09.2003
20
12.09.2003
21
12.09.2003
22
12.09.2003
23
24
EEG Current
Future Practice
25
Source Localization
3D MRI data 512 x 512 x 256 voxels segment 4 compartments
12.09.2003
26
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
30
31
SR8000 Scalability
Parallel Multigrid for fast (forward) solution
Recall, SR 8000 has 1344 processors: could solve 20 billion unknowns in ~100 seconds!
Prof. Dr. Ulrich Rde Lehrstuhl fr Systemsimulation Universitt Erlangen-Nrnberg 12.09.2003 Graz, September 2003
32
Part IV Conclusions
12.09.2003
33
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,
Prof. Dr. Ulrich Rde Lehrstuhl fr Systemsimulation Universitt Erlangen-Nrnberg
12.09.2003
35
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
m f f! w y x x hy2uex g 2h d
C !
1f uw z x Y a
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
0 "
z %|x h z f 1he x x
" " 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
( 0
( )
'
X
tX d Yj
Joachim Schberl o FWF - Start Project hp-FEM Johannes Kepler University Linz, Austria
Joachim Schberl o
Introduction
Page
Poisson Equation:
u = f in
u = 0 on
Joachim Schberl o
Page
Poisson Equation:
u = f in
u = 0 on
A(u, v) =
v dx
and
f (v) =
f v dx
Joachim Schberl o
Page
Poisson Equation:
u = f in
u = 0 on
A(u, v) =
v dx
and
f (v) =
f v dx
Joachim Schberl o
Page
System PDEs
1 displacement u [H0,D ]d, strain operator (u) := 0.5( u + ( u)T ) Lam parameters , . e
Joachim Schberl o
Page
System PDEs
1 displacement u [H0,D ]d, strain operator (u) := 0.5( u + ( u)T ) Lam parameters , . e
t w
Joachim Schberl o
Page
System PDEs
Joachim Schberl o
Page
System PDEs
Linear elasticity: A(u, v) = Nearly incompressible materials: (u) : (v) + div u div v dx
Thin beam: t
Joachim Schberl o
Page
small parameter: symmetric bilinear form: Hilbert space: operator: with kernel: Assume:
uV
Joachim Schberl o
Page
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
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
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
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
Joachim Schberl o
Page
0.1
0.1
0.08
0.06
0.06
0.04
0.04
0.02
0.02
Joachim Schberl o
Page
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
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
w=
Eiwi
with
wi Vi :
Joachim Schberl o
Multigrid methods
Page
w=
Eiwi
with
wi Vi :
:= Cv, v
Vh Vh )
Joachim Schberl o
Multigrid methods
Page
w=
Eiwi
with
wi Vi :
:= Cv, v
Vh Vh )
inf
v= Ei vi vi Vi
vi
2 Ci
Joachim Schberl o
Multigrid methods
Page
Space splitting V =
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 =
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 =
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
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
Discrete Kernel:
w
wh h dx = 0.
w
Sub-space covering:
Vi
Joachim Schberl o
Multigrid methods
Page
11
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
inf
vh =EH vH + vi
vH
2 AH
vi
2 Ah
Ah requires:
AH )
(Vh, .
Ah )
Ah )
(VH , .
AH )
Joachim Schberl o
Multigrid methods
Page
12
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
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:
Rhv = RhF v
Joachim Schberl o
Multigrid methods
Page
15
Fortin operator
Error estimates are based on equivalent mixed formulations. Discrete LBB condition is usually veried by the Fortin operator F : V Vh:
Rhv = RhF v
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
Error estimates are based on equivalent mixed formulations. Discrete LBB condition is usually veried by the Fortin operator F : V Vh:
Rhv = RhF v
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
Error estimates are based on equivalent mixed formulations. Discrete LBB condition is usually veried by the Fortin operator F : V Vh:
Rhv = RhF v
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
Error estimates are based on equivalent mixed formulations. Discrete LBB condition is usually veried by the Fortin operator F : V Vh:
Rhv = RhF v
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)
Joachim Schberl o
Multigrid methods
Page
16
Numerical Experiments
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
Joachim Schberl o
Multigrid methods
Page
18
Thin 3D elements
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
A D 6 C 5 32 1 9 @ 7 8 7 4 B
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 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
` 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
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 !
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
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
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
' &
R C B
C
P I F
H D
'
2 $
C B C B
' 8
""
3 #
G F
E D
A @ ( 9
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
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
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, ...
Istituto per le Applicazioni del Calcolo "M. Picone"
computing optimal approximations of f ( x, y) by piece-wise smooth functions u(x, y ) (restrictions u i to the pieces R i differentiables)
Istituto per le Applicazioni del Calcolo "M. Picone"
Mumford-Shah Functional
Given the image f(x,y) let u be a real function defined on a domain and
( u) i
i=1
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
, ,
MSF Minimization
approximation of f(x,y) by u
(u - f )2 dxdy
/Su
u dxdy
( u) S = u i i =1
as short as
H1 (S u )
Istituto per le Applicazioni del Calcolo "M. Picone"
Parameters
, > 0 can be calibrated to eliminate
false edges , created by noise, and save the actual image;
is a scaling parameter,
effects;
u i D N 2 small squares
f ,S
N N grid,
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
Istituto per le Applicazioni del Calcolo "M. Picone"
-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
(ii)
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
Istituto per le Applicazioni del Calcolo "M. Picone"
-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
Euler Equations
minimizers of
( 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
Istituto per le Applicazioni del Calcolo "M. Picone"
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 )
Istituto per le Applicazioni del Calcolo "M. Picone"
Solution Algorithm
equation systems on the grid G and covering the domain :
M
L w =F
M ac M
* 0 w k = RelaxM w k ,L a ,F
Istituto per le Applicazioni del Calcolo "M. Picone"
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
Istituto per le Applicazioni del Calcolo "M. Picone"
{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
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
Istituto per le Applicazioni del Calcolo "M. Picone"
k=2
k=3
k=5
Lh a
Lh Circle 4x64 a 6 c
log( resz )
log(resu)
Istituto per le Applicazioni del Calcolo "M. Picone"
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
Istituto per le Applicazioni del Calcolo "M. Picone"
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
Istituto per le Applicazioni del Calcolo "M. Picone"
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
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)
(wa - I l+1 wa )
l+1
is an appropriate coarse-to-fine
In the FAS algorithm the idea is to store the full current approximation wa instead of storing just the correction
l
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
! " #
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 @ 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 @ 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
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
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
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
`U
v u @ z
@ C F GE
# !
I GH
z
t u a yB `U
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 `U X
n
f v wX h
@
gX
h k h d `U b d h
@ s P
t B
u `U
v g j h v wX
@
h ux
B
c B B po
d c
B
`U X U
B h u
po
B
X X
X U X
u h r
po po v
@ X ` s j s
`U c B r u ph
`U z
`U `U
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
`U v X t
v `U
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 `U 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
&
&
&
& $
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`
&
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
& @ )
'
) ' 7
&
&
'
&
&
@ '
( &
)
)
Y '
& ( 1 ) &
1 &
&
Stochastic vs. DE
Reaction: A+2X A X
3X
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
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 )
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
So, the magnetic eld H solves the following problem: Find H V Je,I : i
(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 ,
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 ,
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)
inf sup
(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 )}
(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)
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 ) .
I we have
I I
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
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 )
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
vh
qh
(26) NI,h
L2 (I )
2 C6 (1+C5 )1/2
. 2
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 }.
( 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
ch (NI , I )
:= =
NI I
NI
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
f Fint
+ +
f Fj
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
= = j
f Fj (curl vI,h
n)|f j meas(f )
curl vI,h n = 0 ,
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.
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
(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
(30)
.
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+
L2 (I )
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
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
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
y)
Ox F Ox
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
a(u, q)() = f () V
An Example Formulation of the . . . Optimization algorithm Discretization by Finite . . . Adaptive Mesh . . . Example 1 Example 2 Example 3 (Bypass . . . Conclusions
Iteration
Initial parameter: q0 k = 0, 1, 2 . . . Residuum: Update q :
Conclusions
rk = G (C c(qk )), k Hk q = rk
Gk = c (qk )
Hk :=
j(qk ) =
G Gk k
+ c(qk ) C, c (qk )
Gau-Newton algorithm
Hk := G Gk , k
Update methods
Hk := G Gk + Sk , k Globalization strategies:
Line search Trust-region techniques
Optimization algorithm Discretization by Finite . . . Adaptive Mesh . . . Example 1 Example 2 Example 3 (Bypass . . .
Conclusions
An Example Formulation of the . . . Optimization algorithm Discretization by Finite . . . Adaptive Mesh . . . Example 1 Example 2 Example 3 (Bypass . . . Conclusions
Unconstrained formulation
Go Back
Minimize
1 ch(qh) C 2
2 Z
Close Quit
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.
E(q) E(qh) = + P + R
Remarks
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
ihy) +
1 (xh)(u 2
ihu)
(xh)() := (f, ) a(uh, qh)() (xh)() := < Gh(G Gh)1 E(qh), C (uh)() > h au(uh, qh)(, yh)
Example 1 1 2
9
qu + 200u = 2 u = 0
in = (0, 1)2 , on .
q qh
4.09e-1 3.20e-2 8.10e-3 2.66e-3 8.01e-4 1.88e-4 2.24e-5
Ief f
1.60 0.91 0.97 1.03 1.05 1.04 1.04
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 . . . Optimization algorithm Discretization by Finite . . . Adaptive Mesh . . . Example 1 Example 2 Example 3 (Bypass . . . Conclusions
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
Formulation of the . . . Optimization algorithm Discretization by Finite . . . Adaptive Mesh . . . Example 1 Example 2 Example 3 (Bypass . . . Conclusions
An Example Formulation of the . . . Optimization algorithm Discretization by Finite . . . Adaptive Mesh . . . Example 1 Example 2 Example 3 (Bypass . . . Conclusions
0.01
0.001
0.0001
Go Back Close
1e-05 1000
Quit
10000
100000
Example 2 u u + f (u) = 0 in
An Example Formulation of the . . . Optimization algorithm Discretization by Finite . . . Adaptive Mesh . . . Example 1 Example 2 Example 3 (Bypass . . . Conclusions
Ox F
Go Back
Ox
Close Quit
Numerical parameter estimation for chemical models in reactive ows, Combustion Theory and Modelling, submitted, 2003
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
For T OL = 1% :
Global renement: N = 107585, e = 0.54% Local renement : N = 7081, e = 0.89%
Example 2 (IV)
An Example Formulation of the . . . Optimization algorithm Discretization by Finite . . . Adaptive Mesh . . . Example 1 Example 2 Example 3 (Bypass . . . Conclusions
v + v
v+
p = 0,
v = 0 in
Boundary Conditions:
v ? ? ?
1
= = = =
Calibration of PDE Models and Sensitivity Analysis with Adaptive Finite Elements: Application to CFD Problems, Journal of Comp. Physics, to submit, 2003
v + v
v+
p = 0,
v = 0 in
Boundary Conditions:
v nv p n nv p n nv p n
1
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
An Example Formulation of the . . .
0
S3
(1) in
S2
out
1 Minimize 2
under the constraint
(Ci(v) Ci)2
i=1
v + v
v+
p = 0,
v = 0 in , BC(q).
Ci(v) =
Si
vn
Example 3 (III)
An Example Formulation of the . . . Optimization algorithm Discretization by Finite . . . Adaptive Mesh . . . Example 1 Example 2 Example 3 (Bypass . . . Conclusions
Quantity of interest:
E(u) =
O
n (u) d ds,
Example 3 (IV)
An Example Formulation of the . . . Optimization algorithm Discretization by Finite . . . Adaptive Mesh . . . Example 1 Example 2 Example 3 (Bypass . . . Conclusions
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
Formulation of the . . . Optimization algorithm Discretization by Finite . . . Adaptive Mesh . . . Example 1 Example 2 Example 3 (Bypass . . . Conclusions
Numerical Examples
Efciency of the estimator Quality of the generated meshes
Go Back Close Quit
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
Introduction (Navier-Stokes / Oseen equations)
Institute of Computational Mathematics
Oseen equations
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Overview
Introduction (Navier-Stokes / Oseen equations)
Institute of Computational Mathematics
Oseen equations
Decoupled Approach
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Overview
Introduction (Navier-Stokes / Oseen equations)
Institute of Computational Mathematics
Oseen equations
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Overview
Introduction (Navier-Stokes / Oseen equations)
Institute of Computational Mathematics
Oseen equations
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Aim
Complex 3D domain
Institute of Computational Mathematics
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Aim
Complex 3D domain
Institute of Computational Mathematics
t u
Johannes Kepler University Linz
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
Institute of Computational Mathematics
t u
Johannes Kepler University Linz
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
Advances in Numerical Algorithms, Sept. 1013, 2003 p.3/31
Introduction q
Decoupled MG q
Aim
Complex 3D domain
Institute of Computational Mathematics
t u
Johannes Kepler University Linz
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
Advances in Numerical Algorithms, Sept. 1013, 2003 p.3/31
Convective Term
Linearization: Fixed point iteration ; Oseen
Institute of Computational Mathematics
equation
Johannes Kepler University Linz
u + (w
)u +
p=f
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Convective Term
Linearization: Fixed point iteration ; Oseen
Institute of Computational Mathematics
equation
Johannes Kepler University Linz
u + (w
Stabilization (SUPG):
)u +
p=f
replace v by v + h (w )v ;
((w
with h
Introduction q
= O(h).
Coupled AMG q Num. Results q
Advances in Numerical Algorithms, Sept. 1013, 2003 p.4/31
Decoupled MG q
C=
elts K
h2 ( p, q)0,K K
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
modied Taylor-Hood:
C=
elts K
h2 ( p, q)0,K K
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
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
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
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
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Pressure correction
Johannes Kepler University Linz
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
Pressure correction
Johannes Kepler University Linz
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
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
Institute of Computational Mathematics
Elman,Kay,Loghin, Silvester,Wathen
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Decoupled Approach BB
Black-Box-preconditioner
Institute of Computational Mathematics
Elman,Kay,Loghin, Silvester,Wathen
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Decoupled Approach BB
Black-Box-preconditioner
Institute of Computational Mathematics
Elman,Kay,Loghin, Silvester,Wathen
[ + (w
)]
[ + (w
)]s 1 s
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Decoupled Approach BB
Black-Box-preconditioner
Institute of Computational Mathematics
Elman,Kay,Loghin, Silvester,Wathen
[ + (w
)]
[ + (w
)]s 1 s
L pressure Laplacian,
A A,
Algebraic Multigrid
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Algebraic Multigrid
Complex geometry
Johannes Kepler University Linz
discretization
many unknowns.
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Algebraic Multigrid
Complex geometry
Johannes Kepler University Linz
discretization
many unknowns.
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Algebraic Multigrid
Complex geometry
Johannes Kepler University Linz
discretization
many unknowns.
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
unknowns.
Johannes Kepler University Linz
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
unknowns.
Johannes Kepler University Linz
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
unknowns.
Johannes Kepler University Linz
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.
Johannes Kepler University Linz
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.
Johannes Kepler University Linz
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:
Institute of Computational Mathematics
Prolongators
Johannes Kepler University Linz
l Pl+1
l Il+1
l Jl+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
(=
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Notation
Grid transfer:
Institute of Computational Mathematics
Prolongators
Johannes Kepler University Linz
l Pl+1
l Il+1
l Jl+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
(=
ne (u) mesh
u
2nd level
Coarsening
Renement
u p u. p
..
L-th level
u p
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Coarsening
u
2nd level
Coarsening
Coarsening
Renement
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
Coarsening
u
2nd level
Coarsening
Coarsening
Renement
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
Coarsening
u
2nd level
Coarsening
Coarsening
Renement
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
Coarsening
u
2nd level
Coarsening
Coarsening
Renement
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
Coarsening
u
2nd level
Coarsening
Coarsening
Renement
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
Coarsening
u
2nd level
Coarsening
Coarsening
Renement
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
Coarsening
u
2nd level
Coarsening
Coarsening
Renement
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
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
replacements
Convergence
Approximation Property Smoothing Property
Stability
Ql
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
replacements
Convergence
Approximation Property Smoothing Property
Stability
Ql
unstable behaviour:
Introduction q Decoupled MG q Coupled AMG q Num. Results q
Advances in Numerical Algorithms, Sept. 1013, 2003 p.14/31
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
P1-P1-stab Coarsening
Institute of Computational Mathematics
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
P1-P1-stab Coarsening
velocity and pressure live on the same nodes
Institute of Computational Mathematics Johannes Kepler University Linz
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
P1-P1-stab Coarsening
velocity and pressure live on the same nodes
Institute of Computational Mathematics Johannes Kepler University Linz
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
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
k,j
(akj )(vk vj )2
k,j
(a )(vk vj )2 . kj
vl
2 AD l .
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
k,j
(akj )(vk vj )2
k,j
(a )(vk vj )2 . kj
vl
2 AD l .
Bl (u, p; v, q) l v AD l + q Ml
AD l
+ p
Ml
Introduction q
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Use AMGe
Johannes Kepler University Linz
with element
agglomeration
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Use AMGe
Johannes Kepler University Linz
with element
agglomeration
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
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 velocity: identity on edges, energy minimization in the interior
Introduction q Decoupled MG q Coupled AMG q Num. Results q
Advances in Numerical Algorithms, Sept. 1013, 2003 p.20/31
nc P1 -P0 Stability
and
(1)
C vl1
vl1 Ul1 .
(2)
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
nc P1 -P0 Stability
and
(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.
Introduction q Decoupled MG q Coupled AMG q Num. Results q
Advances in Numerical Algorithms, Sept. 1013, 2003 p.21/31
Smoothing (1)
Use smoothers from GMG for saddle point systems, e.g.
Institute of Computational Mathematics
Vanka-Smoother:
Johannes Kepler University Linz
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 ,
Johannes Kepler University Linz
C + B A1 B T
A1 . . . -Jacobi).
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 ,
Johannes Kepler University Linz
C + B A1 B T
A1 . . . -Jacobi).
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 ,
Johannes Kepler University Linz
C + B A1 B T
A1 . . . -Jacobi).
residual
100
300
400
1e-4 1e-5 1e-6 residual 1e-7 1e-8 1e-9 1e-10 1e-11 0 1000 2000
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
Advances in Numerical Algorithms, Sept. 1013, 2003 p.24/31
Coupled AMG-SIMPLE
7 105 vars,
mids, prisms;
0 4000 8000 CPU-time [s] 12000 16000
= 0.0005
(geometries provided by AVL List GmbH) Introduction q Decoupled MG q Coupled AMG q Num. Results q
Advances in Numerical Algorithms, Sept. 1013, 2003 p.25/31
0.001
level 5
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
T0.1
1e-05
PSfrag replacements
1e-06
10
100
1000
10000
1/
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Standard aggl.
Institute of Computational Mathematics
ne coarse
Johannes Kepler University Linz
O(n)
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Standard aggl.
Institute of Computational Mathematics
Metis-renement
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Standard aggl.
Institute of Computational Mathematics
Metis-renement
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Standard aggl.
Institute of Computational Mathematics
Metis-renement
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Standard aggl.
Institute of Computational Mathematics
Metis-renement
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Standard aggl.
Institute of Computational Mathematics
Metis-renement
8e-6
Standard Metis
4e-6
T0.1
2e-6 1e-6 5
PSfrag replacements
6 refinement level 7 8
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
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
Introduction q
Decoupled MG q
Coupled AMG q
Num. Results q
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.
E. Weinmuller / ANA 2003 - Graz p.2/29
Therefore
C C[0, 1]
C C[0, 1]
C C[0, 1]
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
Remedy:
Remedy:
Remedy:
Problem class
z (t) =
M (t) t z(t)
Problem class
z (t) =
M (t) t z(t)
Problem class
z (t) =
M (t) t z(t)
Problem class
z (t) =
M (t) t z(t)
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
z (t) F (t, z(t)) = 0 plus BC
i,j
. . . ti,j . . . i+1
hi h
...
...
Collocation method
z (t) F (t, z(t)) = 0 plus BC
i,j
. . . ti,j . . . i+1
hi h
...
...
Collocation method
z (t) F (t, z(t)) = 0 plus BC
i,j
. . . ti,j . . . i+1
hi h
...
...
Collocation method
z (t) F (t, z(t)) = 0 plus BC
i,j
. . . ti,j . . . i+1
hi h
...
...
= O(h ),
E. Weinmuller / ANA 2003 - Graz p.13/29
Collocation method
z (t) F (t, z(t)) = 0 plus BC
i,j
. . . ti,j . . . i+1
hi h
...
...
= O(h ),
E. Weinmuller / ANA 2003 - Graz p.13/29
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
de Hoog and Weiss (1978): linear case, M has no eigenvalues with positive real parts
stage order: p z
= O(h ),
Convergence
=1
z (t) =
M +tC(t) z(t) t
+ f (t, z(t)) + BC
m
de Hoog and Weiss (1978): linear case, M has no eigenvalues with positive real parts
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
de Hoog and Weiss (1978): linear case, M has no eigenvalues with positive real parts
stage order: p z
= O(h ),
no superconvergence at i , in general.
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
de Hoog and Weiss (1978): linear case, M has no eigenvalues with positive real parts
stage order: p z
= O(h ),
no superconvergence at i , in general.
Auzinger, Koch, Weinmller (2002): nonlinear case, M has no eigenvalues with positive real parts Koch (2003): nonlinear case, spectrum of M arbitrary
E. Weinmuller / ANA 2003 - Graz p.14/29
Estimate QDeC/BEUL
z (t) F (t, z(t)) = 0 plus BC
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
z (t) F (t, z(t)) = 0 plus BC p (ti,j ) F (ti,j , p(ti,j )) = 0 plus BC
Estimate
p(ti,j ) z(ti,j ), i, j
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
p(ti,j ) z(ti,j ), i, j
Classical idea due to Zadunaisky, Frank, Stetter
F (ti,k , p(ti,k ).
F (ti,k , p(ti,k ).
Error estimate:
p(ti,j ) z(ti,j ) i,j i,j , i, j.
E. Weinmuller / ANA 2003 - Graz p.16/29
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) = .
Convergence
> 1 - Example
1 z (t) = 2 t
, 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) =
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
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
i,j fi,j , i = 1, . . . , N
i,j fi,j , i = 1, . . . , N
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.
Applications
Singularity of the second kind: Regular problems on semi-innite intervals:
z (t) = t f (t, z(t)), t [a, ), > 1.
Applications
Singularity of the second kind: Regular problems on semi-innite intervals:
z (t) = t f (t, z(t)), t [a, ), > 1.
Applications
Singularity of the second kind: Regular problems on semi-innite intervals:
z (t) = t f (t, z(t)), t [a, ), > 1.
Applications (2)
The nonlinear Schrdinger equation
u i + u + |u|2 u = 0, t t > 0.
Applications (2)
The nonlinear Schrdinger equation
u i + u + |u|2 u = 0, t t > 0.
Applications (2)
The nonlinear Schrdinger equation
u i + u + |u|2 u = 0, t t > 0.
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
Easy example
Harder example
Wrapped image and the unwrapped elevation model (Etna volcano region)
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 .
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.
(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.
P = 1.5
P = 1 and P=0
Another example
Original image Wrapped image
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).
D(q ) = E[ X q ( X ) 2 ] =
2
di
i =1 di 1
x ri
2 2
p ( x ) dx.
2000
1500
1000
500
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Gray Level
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
64 gray levels
2000
1500
1000
500
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Gray Level
D(q ) = E[ X q ( X ) 2 ] =
2
i =1 di 1
x ri
2 2
p ( x ) dx.
ri =
xp( x)dx
di
di 1
, di =
ri + ri+1 2
, i = 1, , n 1
p ( x)dx
di 1
ri k =
k ri k + ri+1
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 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
10
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
10
20
30
40
50
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
10
20
30
40
50
Local Minima
1.5
3 x 10 Multigrid Converged Solution
1.5
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
x 10
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
10
10
100
11
200
300
400
500
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
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
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.
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.
Image Binarization
Original Image
Nonuniform Illumination
ilted
Spherical
ilted
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