You are on page 1of 42

Introduction to Simulation - Lecture 23

Fast Methods for Integral Equations


Jacob White

Thanks to Deepak Ramaswamy, Michal Rewienski,


and Karen Veroy

Outline
Solving Discretized Integral Equations
Using Krylov Subspace Methods
Fast Matrix-Vector Products
Multipole Algorithms
Multipole Representation.
Basic Hierarchy
Algorithmic Improvements
Local Expansions
Adaptive Algorithms
Computational Results

Exterior Problem in Electrostatics


potential

+
v
-

2 = 0

Outside

is given on Surface
Dirichelet Problem

First Kind Integral Equation For Charge:


1
( x) =
( x ) dS
{

x
x

surface
1
424
3 Ch arg e
potential
Green's Density
Function

SMA-HPC 2003 MIT

Drag Force in a Microresonator

Courtesy of Werner Hemmert, Ph.D. Used with permission.

Resonator

Computed Forces
Bottom View

SMA-HPC 2003 MIT

Discretized Structure

Computed Forces
Top View

3-D Laplaces
Equation

Basis Function Approach


Piecewise Constant Basis

Integral Equation: ( x ) =

surface

Discretize Surface into


Panels

1
( x ) dS
x x
n

Represent ( x ) i i ( x )
{
i =1

Basis Functions

j ( x ) = 1 if x is on panel j
Panel j ( x ) = 0 otherwise
j
SMA-HPC 2003 MIT

3-D Laplaces
Equation
Put collocation points at
panel centroids

Basis Function Approach


Centroid Collocation

( )

xci = j

xci Collocation
point

j =1

G(x

panel j

ci

, x dS

14442444
3
Ai , j

A1,1 L L A1,n 1 ( xc1 )

M O

M M

= M
M
O M M M


An ,1 L L An ,n n xcn

( )

SMA-HPC 2003 MIT

Basis Function Approach

3-D Laplaces
Equation

Calculating Matrix Elements

xci Collocation

point

Ai , j =

panel j

1
dS
xci x

Panel j
One point
quadrature
Approximation
Four point
quadrature
Approximation
SMA-HPC 2003 MIT

Panel Area

Ai , j

xci xcentroid j
4

0.25* Area

j =1

xci x po int j

Ai , j

Basis Function Approach

3-D Laplaces
Equation

Calculating Self-Term

xci Collocation

point

Ai ,i =

panel i

1
dS
xci x

Panel i
One point
quadrature
Approximation

Ai ,i =

panel i
SMA-HPC 2003 MIT

Ai ,i

Panel Area

xci xci
1
424
3
0

1
dS is an integrable singularity
xci x

Basis Function Approach

3-D Laplaces
Equation

Calculating Self-Term
Tricks of the trade

xci Collocation

point

Panel i

Ai ,i =

panel i

Disk of radius R
surrounding
collocation point

Integrate in two Ai ,i =
disk
pieces
Disk Integral has
singularity but has
analytic formula
SMA-HPC 2003 MIT

disk

1
dS
xci x

1
1
dS +
dS

xci x
rest of panel xci x
R 2

1
dS =
xci x
0

1
rdrd = 2 R
r

Basis Function Approach

3-D Laplaces
Equation

Calculating Self-Term
Other Tricks of the trade

xci Collocation

point

Panel i

Ai ,i =

panel i

1
dS
xci x
1
424
3

Integrand is singular

1) If panel is a flat polygon, analytical formulas exist


2) Curve panels can be handled with projection

SMA-HPC 2003 MIT

Basis Function Approach

3-D Laplaces
Equation

Galerkin (test=basis)
n

( x ) ( x ) dS = ( x ) G ( x, x) ( x) dS dS
144
2443
1444442444443
i

bi

j =1

Ai , j

For piecewise constant Basis


n
1
dS dS
( x ) dS = j
14

panel
i
panel
j
x x
243 j =1
14444244443
bi
Ai , j
A1,1 L L A1,n 1 b1
M O

M

M = M
M
O M M M


L
L
A
A
n ,1
n,n
n bn
SMA-HPC 2003 MIT

3-D Laplaces
Equation

Basis Function Approach


Problem with dense matrix

Integral Equation Method Generate Huge


Dense Matrices
A1,1 L L A1,n 1 ( xc1 )

M O

M M

= M
M
O M M M


An ,1 L L An ,n n xcn

( )

Gaussian Elimination Much Too Slow!


SMA-HPC 2003 MIT

Solving Discretized
Integral Equations

compute Apk

( r ) ( Ap )
k T

k =

( Apk )

( Apk )

x k +1 = x k + k pk

The Generalized Conjugate


Residual Algorithm
The kth step of GCR

For discretized Integral


equations, A is dense
Determine optimal stepsize in
kth search direction

Update the solution


and the residual
r k +1 = r k k Apk
T
k
+
1
Compute the new
k
Ar
Ap
(
)
(
)
j
k +1
pk +1 = r
p j orthogonalized
T
j = 0 ( Ap ) ( Ap )
search direction
j
j
SMA-HPC 2003 MIT

Solving Discretized
Integral Equations

( r ) ( Ap )
k T

Vector inner products, O(n)

( Apk ) ( Apk )
x k +1 = x k + k pk
r k +1 = r k k Apk

pk +1 = r

k +1

Vector Adds, O(n)

Ar ) ( Ap )
(

p
( Ap ) ( Ap )
k

j =0

k +1 T

Complexity of GCR

Dense Matrix-vector
product costs O(n2)

compute Apk
k =

The Generalized Conjugate


Residual Algorithm

O(k) inner products,


total cost O(nk)

Algorithm is O(n2) for Integral Equations


even though # iters (k) is small!
SMA-HPC 2003 MIT

Solving Discretized
Integral Equations

The Generalized Conjugate


Residual Algorithm
Fast Matrix Vector Products

exactly compute Apk


Dense Matrix-vector product costs O(n2)

approximately compute Apk

Reduces Matrix-vector product costs to


O(n) or O(nlogn)

SMA-HPC 2003 MIT

SMA-HPC 2003 MIT

SMA-HPC 2003 MIT

SMA-HPC 2003 MIT

SMA-HPC 2003 MIT

SMA-HPC 2003 MIT

SMA-HPC 2003 MIT

SMA-HPC 2003 MIT

SMA-HPC 2003 MIT

SMA-HPC 2003 MIT

SMA-HPC 2003 MIT

SMA-HPC 2003 MIT

SMA-HPC 2003 MIT

SMA-HPC 2003 MIT

SMA-HPC 2003 MIT

SMA-HPC 2003 MIT

SMA-HPC 2003 MIT

SMA-HPC 2003 MIT

SMA-HPC 2003 MIT

1/(# panels)

SMA-HPC 2003 MIT

SMA-HPC 2003 MIT

SMA-HPC 2003 MIT

SMA-HPC 2003 MIT

SMA-HPC 2003 MIT

SMA-HPC 2003 MIT

SMA-HPC 2003 MIT

SMA-HPC 2003 MIT

Summary
Solving Discretized Integral Equations
GCR plus Fast Matrix-Vector Products
Multipole Algorithms
Multipole Representation.
Basic Hierarchy
Algorithmic Improvements
Local Expansions
Adaptive Algorithms
Computational Results
Precorrected-FFT Algorithms