c gasparadvances in numerical algorithms, graz, 2003 1 fast interpolation techniques and meshless...
TRANSCRIPT
1Advances in Numerical Algorithms, Graz, 2003
C Gaspar
Fast interpolation techniques and Fast interpolation techniques and meshless methodsmeshless methods
Csaba GáspárSzéchenyi István University,
Department of Mathematics
Győr, Hungary
C Gaspar Advances in Numerical Algorithms, Graz, 2003
2
• surface approximation
• completion of incomplete data sets
• construction of meshfree methods
4: ESWN
Cffff
f
0f
MotivationsMotivations
C Gaspar Advances in Numerical Algorithms, Graz, 2003
3
The scattered data interpolation problem inThe scattered data interpolation problem in 2R
Vectorial problems: some additional conditions are required
0 div u
0rot u
),...,1()(
such that :function smooth)tly (sufficien a Find
.,..., values theand
in spaced)ly (irregular ,..., points Given the :
2
1
21
Nkuxu
u
uu
xxProblem
kk
N
N
RR
R
C Gaspar Advances in Numerical Algorithms, Graz, 2003
4
Method of radial basis functionsMethod of radial basis functions
where the basis functions depend on ||x|| only.The coefficients are a priori unknown and can be determined by solving the interpolation equations:
The form of the interpolation function:
N
jjjj xxxu
1)(:)(
m
jjj xpa
1)(
),...,2,1()( Nkuxu kk
and the orthogonality conditions:
),...,2,1(0)(1
mkxpaN
jjkk
C Gaspar Advances in Numerical Algorithms, Graz, 2003
5
Some popular RBFsSome popular RBFs
Method of Multiquadrics (Hardy):
Thin Plate Splines (Duchon):(biharmonic fundamental solution)
Traditional DRM (Partridge, Brebbia):
22:)( jj crr
rrrj log:)( 2
Gaussian functions:22
:)(rc
jjer
rrj 1:)(
Compactly supported RBFsCompactly supported RBFs (Wendland):
.........
)14()1(:)(
)1(:)(
4
2
rcrcr
rcr
jjj
jj
Globally supported RBFsGlobally supported RBFs
C Gaspar Advances in Numerical Algorithms, Graz, 2003
6
Why do we like the Why do we like the RBFs soRBFs so much? much?
Main advantages:
• excellent approximation properties
• easy to implement
But what is the But what is the price to payprice to pay??
We have to face that:
• RBFs generally lead to a system with large, dense and ill-conditioned matrix
• memory requirement: O(N 2)
• computational cost: O(N 3)
C Gaspar Advances in Numerical Algorithms, Graz, 2003
7
• domain decomposition• preconditioning• fast multipole evaluation methods• compactly supported RBFs• multi-level techniques• direct multi-elliptic interpolationdirect multi-elliptic interpolation
RemediesRemedies
C Gaspar Advances in Numerical Algorithms, Graz, 2003
8
Methods based on iterated elliptic operatorsMethods based on iterated elliptic operators
Observation: Thin plate splines are fundamental solutions of the biharmonic operator
Idea: Let us define the interpolation function in such a form:
)conditionstion (interpola :)( kk uxu )conditions(boundary 0:| u
• Second-order operators: unusable!• Fourth-order operators:
biharmonic operator; bi-Helmholtz-operator; Laplace-Helmholtz-operator
• Higher order operators:triharmonic operator; tetraharmonic operator; iterated Helmholtz-operator, etc.
},...,,{\in 0 21 NxxxLu
C Gaspar Advances in Numerical Algorithms, Graz, 2003
9
””Harmonic interpolation”: a counterexampleHarmonic interpolation”: a counterexample
),...,2,1(),(:),(
0:|
)},(),...,.(),,{(\in 0
0
2211
Nkyxfyxu
u
yxyxyxu
kkkk
NN
Numerical features:•ill-posed problem•poor interpolation surface•logarithmic singularities logarithmic singularities at the interpolation pointsat the interpolation points
yxyxf sinsin:),( ),1,0()1,0( : 0
C Gaspar Advances in Numerical Algorithms, Graz, 2003
10
Biharmonic interpolationBiharmonic interpolation
}0)()(:)({: 120 NxwxwHwW
)(20 Hu
),...,2,1()()(
},...,,{\in 0
0
21
Nkxuxu
xxxu
kk
N
Equivalent formulations: Let us introduce the closed subspace:
Classical formulation: For a given function , find a function such that
Direct problem: For a given function , find a functionsuch that
)(200 Hu
},...,,{\in 0)( 210 Nxxxvu
Variational problem: For a given function , find a function such that
)(200 Hu
)(200 Hu
Wv
Wv Wwwvu L 0),( )(0 2
Theorem: The direct and the variational problems are equivalent and have a unique solution in W.
C Gaspar Advances in Numerical Algorithms, Graz, 2003
11
Representation of the biharmonic interpolationRepresentation of the biharmonic interpolation
Theorem: u is uniquely represented in the form:
where w is an everywhere biharmonic function, and is the biharmonic fundamental solution:
N
jjj xxxwxu
1)()()(
rrr log)( 2
...but the coefficients can be calculated from: without solving any system of equations!without solving any system of equations!
This is an RBF-like interpolation...This is an RBF-like interpolation...
N
jjj xxxu
1)()(
C Gaspar Advances in Numerical Algorithms, Graz, 2003
12
ApproximationApproximation
||||infsup:},...,{ 1
yxhNxxyx
Separation distance:
Theorem: The following estimations hold:
)(02)(0
)(02
1)(0
20
10
202
||||||||
||||||||
HH
HL
uhCuu
uhCuu
C Gaspar Advances in Numerical Algorithms, Graz, 2003
13
Bi-Helmholtz interpolationBi-Helmholtz interpolation
)(20 Hu
),...,2,1()()(
},...,,{\in 0)(
0
2122
Nkxuxu
xxxuIc
kk
N
Classical formulation: For a given function , find a function such that
Direct problem: For a given function , find a functionsuch that
)(200 Hu
},...,,{\in 0)()( 21022
NxxxvuIc
Variational problem: For a given function , find a function such that
)(200 Hu
)(200 Hu
Wv
Wv Wwwvu L 0),( )(0 2
The direct and the variational problems are equivalent, and have a unique solution in W. The representation and approximation theorems are still valid.
Now the fundamental solution is Now the fundamental solution is rapidly decreasingrapidly decreasing ("almost compactly supported").("almost compactly supported").
Representation formula:
Coefficients can be computed directly:
)()(,)()( 11
crrKrxxxuN
jjj
N
jjj xxxu
1)()(
C Gaspar Advances in Numerical Algorithms, Graz, 2003
14
Further generalizationsFurther generalizations
• The use of even higher order differential operators• The use of different operators simultaneously
In each case, the representation automatically produces an RBF-likeIn each case, the representation automatically produces an RBF-likeinterpolation based on the fundamental solution of the applied operator.interpolation based on the fundamental solution of the applied operator.
0uTriharmonic operator:
Mixed Laplace-Helmholtz operators:
0)( 2 fIc
0)( 2 fIc
0)( 22 fIc
C Gaspar Advances in Numerical Algorithms, Graz, 2003
15
The main computational problem:The main computational problem:
How to solve How to solve the multi-elliptic interpolation equation?!the multi-elliptic interpolation equation?!
C Gaspar Advances in Numerical Algorithms, Graz, 2003
16
Numerical techniquesNumerical techniques
Main features:• No algebraic interpolation equationNo algebraic interpolation equation has to be solved.• Instead, the problem is converted to an auxiliary differential equationauxiliary differential equation.• The domain is not predefinedThe domain is not predefined• The boundary conditions are not strictly prescribedThe boundary conditions are not strictly prescribed
Numerical solution techniques:• Traditional FDM or FEM ?• FFT ?• MGR ?• quadtree-based multigrid techniquesquadtree-based multigrid techniques
C Gaspar Advances in Numerical Algorithms, Graz, 2003
17
Quadtree cell systemsQuadtree cell systems
• Non-uniform cell subdivision generated by a finite set of points• Schemes based on
finite volumes• Multigrid solution
technique• Computational costComputational cost
can be reduced to can be reduced to
OO((NN log log NN))
Demo application
C Gaspar Advances in Numerical Algorithms, Graz, 2003
18
BBiharmonic iharmonic interpolation on a QT cell systeminterpolation on a QT cell system::an ean examplexample
: ( , ) ( , ), ( , ): sin sin 0 1 0 1 2 20f x y x y
The QT-cell system The interpolation functionRelative L2-errors (%):
N \ lmax 7 8 9 10 32 17.314 7.759 8.305 9.202 64 4.159 4.026 4.529 2.692128 0.770 1.332 0.691 0.581256 0.213 0.154 0.143 0.148512 0.095 0.062 0.049 0.041
C Gaspar Advances in Numerical Algorithms, Graz, 2003
19
Bi-Helmholtz interpolation: an exampleBi-Helmholtz interpolation: an example
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
))4
1(2(sin))
4
1(2(sin:),(),1,0()1,0(: 22
0 yxyxu
c = 10
c = 50
C Gaspar Advances in Numerical Algorithms, Graz, 2003
20
Construction of meshless methods.Construction of meshless methods.General approachesGeneral approaches
Original problem: )(in fLu + boundary conditions
Method 1. The solution u is directly interpolated:
Method 2. First, the function f is interpolated:
M
jjj xxxu
1)(:)(
M
jjj xxxf
1)(:)(
M
jkjkj fxxL
1)(
LxxxuM
jjkj with)(:)(
1
C Gaspar Advances in Numerical Algorithms, Graz, 2003
21
Construction of meshless methodsConstruction of meshless methodsby direct multi-elliptic interpolationby direct multi-elliptic interpolation
Model problem: )(in fu + boundary conditions
Method 1. The solution u is directly interpolated by tetraharmonic interpolation:
Method 2Method 2. First, the function . First, the function ff is interpolated by biharmonic interpolation. is interpolated by biharmonic interpolation. Then, Then, using the same QT cell systemusing the same QT cell system, the model problem is, the model problem is solved, solved, without boundary conditionswithout boundary conditions..
),...,1()(
),...,2,1()(
},...,,{\in 0 214
NMkuxu
Mkfxu
xxxu
kk
kk
N
...A bit uncomfortable......A bit uncomfortable...
This procedure mimics the computation of Newtonian potentialThis procedure mimics the computation of Newtonian potentialReduces the domain problem to a Reduces the domain problem to a boundary problemboundary problem
}0)(...)(
,0)(...)(:)({:
1
140
NM
M
xuxu
xuxuHwW
C Gaspar Advances in Numerical Algorithms, Graz, 2003
22
How to solve the boundary problem?How to solve the boundary problem?
Traditional tool: the Boundary Element Method
Direct Multi-Elliptic Boundary InterpolationDirect Multi-Elliptic Boundary Interpolation
DrawbacksDrawbacks:• large, dense and ill-conditioned matrices
AdvantagesAdvantages:• reduction of dimension• reduction of mesh
C Gaspar Advances in Numerical Algorithms, Graz, 2003
23
Boundary interpolation 1. The basic ideaBoundary interpolation 1. The basic idea
Model problem: Laplace equation with Dirichlet boundary condition
The interpolation points are located at the boundary only, and...The interpolation points are located at the boundary only, and...
The use of the mmixed Laplace-Helmholtz operatorixed Laplace-Helmholtz operator:
0u'Harmonic' interpolation:
0uBiharmonic interpolation:
'Quasi-harmonic' interpolation:
0)( 2 uIc
0)( 2 uIc
02
1( ) ( ( || ||) log( || ||))
2x K c x c x
c Fundamental solution:
C Gaspar Advances in Numerical Algorithms, Graz, 2003
24
Boundary interpolation 2. Boundary interpolation 2. Application to the Laplace equationApplication to the Laplace equation
Key issue: the proper choice of the scaling constant c
The original problem: 0* 0, * |u u u
The associated boundary interpolation problem:
002|,|,0
1v
n
uuuu
cu
Theorem 1: 1/ 2 1/ 22 ( ) 0 0( ) ( )|| * || (|| || || || )L H H
Cu u u v
c
1 12
0 0( ) ( ) ( )|| * || (|| || || || )
H H L
Cu u u v
c Theorem 2:
With more regular data:
C Gaspar Advances in Numerical Algorithms, Graz, 2003
25
Dirichlet boundary conditionDirichlet boundary conditionExamplesExamples
%)718.1(500c
%)149.0(250c
%)102.26(0c
C Gaspar Advances in Numerical Algorithms, Graz, 2003
26
Neumann boundary conditionNeumann boundary conditionThe naive approachThe naive approach
The original problem:
The associated boundary interpolation problem:
),...,2,1()()(
},...,{\in ,0)(
0
122
Nkxvxn
u
xxuIc
kk
N
0|*
,0* vn
uu
)(30 Hu
The method fails...The method fails...
uu approximates approximates vv00 but fails to approximate but fails to approximate uu* !* !
C Gaspar Advances in Numerical Algorithms, Graz, 2003
27
Neumann boundary conditionNeumann boundary conditionAn improved approachAn improved approach
Idea: Locate additional Dirichlet points along Neumann boundary in outward normal direction; Define Dirichlet data to enforce Neumann conditions
0:~ vduu
)(:~0 n
uvduu
or:
This procedure mimics theThis procedure mimics themethod of fundamental solutionsmethod of fundamental solutions
C Gaspar Advances in Numerical Algorithms, Graz, 2003
28
Neumann boundary conditionNeumann boundary conditionExamplesExamples
025.0d
050.0d
100.0d
C Gaspar Advances in Numerical Algorithms, Graz, 2003
29
Boundary interpolation 3.Boundary interpolation 3.Application to the biharmonic equationApplication to the biharmonic equation
The original problem: 0 0*
* 0, * | ,u
u u u vn
The associated boundary interpolation problem:
2
0 0 02 2
10, | , ,
u uu u u u v w
nc n
Theorem 3: 1 3/ 2 1/ 2 1/ 20 0 0( ) ( ) ( ) ( )|| * || (|| || || || || || )
H H H H
Cu u u v w
c
Further applications:• 2D Stokes flow
C Gaspar Advances in Numerical Algorithms, Graz, 2003
30
Vectorial interpolation problemsVectorial interpolation problems
),...,1()(such that :function a Find
.,..., values theand in ,..., points Given the :22
21
21
Nkx
xxProblem
kk
NN
uuRRu
RuuR
Componentwise interpolation works well...
... but often additional conditions are required:
0 div
or ,0rot
u
u
Use the potential/stream function approach
RBF-method failsRBF-method fails
),...,2,1()(,)( 21 NkuxDvxD kkkk
C Gaspar Advances in Numerical Algorithms, Graz, 2003
31
Tri-Helmholtz-interpolation for vectorial Tri-Helmholtz-interpolation for vectorial problemsproblems
Classical formulation: For a given functionfind a function such that
),...,2,1()()(),()( 21 NkxuxDxvxD kkkk
Variational formulation: DefineFor a given function find a function such that
},...,1,)( grad :)({: 30 NkxwHwW k 0
)()(),( 22 HHvuu
)(300 H
)(30 H
)(30 H
WwwIcIcc
wIcIc
L
L
0)(),)((
)(grad),)((grad
)(
20
22
)(
20
2
2
2
Theorem: The variational problem has a unique solution in W.
The vectorial interpolation function: (grad ) is divergence-free.
yxyx 2sin2sin),(
C Gaspar Advances in Numerical Algorithms, Graz, 2003
32
Representation and RBF-like solutionRepresentation and RBF-like solution
Representation formula: If c > 0 then is represented in the form:
where is the fundamental solution of the operator :
The theory assures existence and uniqueness also for the The theory assures existence and uniqueness also for the RBF-like interpolationRBF-like interpolation:
32 )( Ic
N
jjj
N
jjj xxDxxDx
12
11 )()()(
),...,2,1()()(
)()(
122
112
112
111
NkuxxDxxD
vxxDxxD
k
N
jjkj
N
jjkj
k
N
jjkj
N
jjkj
))(2)((16
1)( 10
224
rcrcKrcKcrc
r
Componentwise TPSExactVectorial tri-Helmholtz
C Gaspar Advances in Numerical Algorithms, Graz, 2003
33
Summary and conclusionsSummary and conclusions
• Interpolation based on iterated elliptic equationsInterpolation based on iterated elliptic equations• Boundary interpolation by singularly perturbed Boundary interpolation by singularly perturbed equationsequations• Special RBF- and RBF-like methods are Special RBF- and RBF-like methods are automaticallyautomatically generatedgenerated
• No explicit use of RBFs is neededNo explicit use of RBFs is needed• No large and dense algebraic equations are to be solvedNo large and dense algebraic equations are to be solved• Quadtree-based multigrid methods are usedQuadtree-based multigrid methods are used
34Advances in Numerical Algorithms, Graz, 2003
C Gaspar
ThankThank you you for your for your attention!attention!