Methode D Elements Frontieres

Methode delements de fronti`ere
Marc Bonnet
UMA (Dept. of Appl. Math.), POems, UMR 7231 CNRS-INRIA-ENSTA
32, boulevard Victor, 75739 PARIS cedex 15, France
mbonnet@ensta.fr
Ecole doctorale MODES Methodes numeriques avancees, 28 mars 2013

http://uma.ensta.fr/
mbonnet/enseignement.html
Marc Bonnet (POems, ENSTA) Methode delements de fronti`ere 1 / 183
Outline
1. Review of boundary integral equation formulations
Electrostatics
Laplace
Elastostatics
Frequency-domain wave equations
2. Review of classical BEM concepts
3. The GMRES iterative solver
4. The fast multipole method (FMM) for the Laplace equation
Multipole expansion of 1/r
The single-level fast multipole method
The multi-level fast multipole method
5. The fast multipole method (FMM) for elastostatics
6. The fast multipole method for elastodynamics
7. Other acceleration methods
Exponential representation of 1/r
FMM using equivalent sources
Clustering and low-rank approximations
Kernel-independent acceleration via kernel interpolation
Adaptive cross approximation
8. Preconditioning
Review of boundary integral equation formulations
Electrostatics
Laplace
Elastostatics
8. Preconditioning
Review of boundary integral equation formulations Electrostatics
Electrostatics
Laplace
Elastostatics
8. Preconditioning
Electrostatics
Well-known, and simple, physical setting
Allows to introduce important concepts of integral equation formulations

with a clear physical meaning
Said concepts will generalize to other settings (elasticity, electromagnetics...)
Also helpful later for a physical understanding of the fast multipole method
(FMM)
Electrostatics: discrete charged particles
Coulomb interaction force:

F
12
=
q
1
q
2
4
1
r
2
r
12
q
1
r
q
2
r
12
Electrostatic eld:
F
12
= q
2
E
12
, E
12
=
q
1
4
1
r
2
r
12
Electrostatic potential:
E
12
=
2
V, V =
q
1
4
1
r
(with : permittivity of the medium (material constant))
Electrostatics: continuous charge distributions
Continuous charge distribution: dq = dV :

E(x) =
1
4
_
V
()
r
2
r dV
r = x, r = |r|, r = r/r
Gauss theorem:
div E =

Poisson equation (Gauss theorem with E = V):

V +

= 0
Electrostatics: continuous charge distributions
Proof of Gauss theorem:
_
V
div E dV =
_
V
EndS
=
1
4
_
V
_
_
V
1
r
2
(rn(x)) dS
x
_
() dV
=
1
4
_
V
()() dV
where () is the solid angle of the (closed) surface V from origin :

() = 4 ( V), () = 0 ( R
3

V)
Hence:
_
V
div E dV =
_
V
dV
Since this is true for any domain V, one has div E = /, i.e. the Gauss theorem
holds.
Electrostatic volume potential
The electrostatic volume potential results from the superposition of electric elds
generated by elementary charges dV distributed in a volume V:
1[, V](x) =
1
4
_
V
()
r
dV
As already seen, 1[, V] satises the Poisson equation:

1[, V] +

= 0 (x V), 1[, V] = 0 (x R
3

V)
Properties of electrostatic volume potentials
The volume integral is weakly singular (i.e. singular, but integrable) for
x V, so that 1[, V] is well-dened inside the charged domain V: one has
dV
x
= r
2
dr d = O(r
2
)
1[, V] is continuous everywhere, and in particular across the boundary V.

Electrostatic single-layer potential
The electrostatic single-layer potential results from the superposition of electric
elds generated by elementary charges dS distributed on a surface S:
o[, S](x) =
1
4
_
S
()
r
dS
o[, S] is harmonic outside of S:

o[, S] = 0 (x R
3
S)
Properties of electrostatic single-layer potentials
The surface integral is weakly singular (i.e. singular, but integrable) for x S,
so that o[, S] is well-dened on the charged surface V: one has
dS
x
= r dr d = O(r )
o[, S] is continuous everywhere, and in particular across S.

Electrostatic double-layer potential
The electrostatic double-layer potential is the limiting case of the superposition of
two single-layer potentials of (i) arbitrary close supports S
, (ii) opposite charge

density, (iii) nite dipolar moment q:
T[q, S](x) = lim
h0
o[
h
, S
+
](x) +o[
h
, S
](x)
with
lim
h0
[h
h
](x) = q(x)
(note: q is analogous to a concentrated moment)
n
x
+
x
x
x
+
= x +
h
2
n
x
= x
h
2
n
S
+
S
Performing the limit h 0, one nds:

T[q, S](x) =
1
4
_
S
1
r
2
rn() q() dS

Electrostatic double-layer potential
T[q, S](x) =
1
4
_
S
1
r
2
rn() q() dS
Properties of electrostatic double-layer potentials
T[q, S] is harmonic outside of S:

T[q, S] = 0 (x R
3
S)
The surface integral is weakly singular (i.e. singular, but integrable) for x S,
so that T[q, S] is well-dened on the charged surface V.
This is not obvious and stems from
dS
x
= r dr d = O(r ) and
1
r
2
rn = O(
1
r
)
T[q, S] is discontinuous across S, with

T[q, S](x
+
) T[q, S](x
) = q(x)
Electrostatic potentials: comments
(a) Electrostatic potentials have a clear physical meaning as the potential elds
associated with volume, surface or dipolar charge distributions.
(b) Electrostatic potentials, as mathematical constructs, dene harmonic elds for
arbitrary choices of supports V, S and densities , q.
(c) As we will shortly see, any harmonic function can be expressed in terms of
such potentials.
Review of boundary integral equation formulations Laplace
Electrostatics
Laplace
Elastostatics
8. Preconditioning
Physical problems governed by the Laplace or Poisson equation
Poisson equation:
u + b = 0 (on ) + unspecied well-posed BCs
(Laplace equation if b =0)
Electrostatics (u: electrostatic potential);
Potential uid ow (u: velocity potential);
Thermal equilibrium (u: temperature);
Torsion (u: warping function dened on 2-D section of shaft)

...
Reciprocity identity and integral representation
u + b = 0 (in ) + unspecied well-posed BCs
Integral representation of u based on two ingredients:
(i) Reciprocity identity:
_
(uv vu) dV =
_
(uv
,n
vu
,n
) dV (u
,n
un)
(ii) Fundamental solution:
G(x, ) + ( x) = 0 (in O)
Choosing v =G(x, ) in (i) yields the integral representation formula:
u(x) =
_
G(x, )b() dV
+
_
_
G(x, )u
,n
() G
,n
(x, )u()
_
dS
(x )
Full-space fundamental solution (O=R
3
):
G(x, ) =
1
4r
(r =| x|)
Volume and single-layer potentials
The single-layer potential o solves the Laplace equation in R

3
, where
o[, ](x) :=
_
G(x, )() dS
The volume potential 1 solves the Poisson equation, where

1[b, ](x) :=
_
G(x, )b() dV
Continuity across (with [[f ]] := f

+
f

):
[[1(x)]] = 0, [[o(x)]] = 0, (x )
Double-layer potentials
The following potential solve the Laplace equation in R

3
:
T[, ](x) :=
_
G
,n
(x, )() dS
(double-layer potential)
Jump relation:
[[T(x)]] = (x) (x )
Boundary traces of double-layer potential:

T[, ](x
) = (c(x) 1)(x) +T[, ](x),

T[, ](x
+
) = T[, ](x
) + (x)
with the denitions ((x): solid angle of seen from x ):
T[, ](x) :=
_
G
,n
(x, )() dS
, c(x) = (x)/(4) (x )
x
(x)
When is smooth at x (the usual case), (x) =2 and c(x) =1/2
Integral representation as sum of potentials
Expressing the integral representation formula in terms of potentials yields
u(x) = 1[b, ](x) +o[u
,n
, ](x) T[u, ](x)
Any harmonic function is representable as a single-layer, or a double-layer
potential, or a linear combination of both.
Any solution of the Poisson equation is representable as a sum of the volume
potential of density b and arbitrary surface potentials.
Singular integral equations
Boundary integral equations are obtained as the limiting situation when
z z of integral representations:
u(x) = lim
zx
u(z)
= lim
zx
_
1[b, ](z) +o[u
,n
, ](z) T[u, ](z)
_
= 1[b, ](x) +o[u
,n
, ](x) T[u, ](x) + (1c(x))u(x)
to obtain:
c(x)u(x) +T[u, ](x) o[u
,n
, ](x) = 1[b, ](x) (x )
i.e. (in expanded form)
c(x)u(x) +
_
_
G
,n
(x, )u() G(x, )u
,n
()
_
dS
=
_
G(x, )b() dV
(x )
Outline of the boundary integral equation method
c(x)u(x) +T[u, ](x) o[u
,n
, ](x) = 1[b, ](x) (x )
1. Insert given boundary data;
Solve for the remaining boundary unknown
2. Then, invoke integral representation for evaluation of u (and related
quantities, e.g. u) at interior points
Example ( = S
u
S
q
, with u = u
D
on S
u
and u
,n
= q
D
on S
q
):
1. Solve for u[
S
q
and u
,n
[
S
u
the integral equation system
c(x)u(x) +T[u, S
q
](x) o[u
,n
, S
u
](x)
= T[u
D
, S
u
](x) +o[q
D
, S
q
](x) +1[b, ](x) (x S
q
)
T[u, S
q
](x) o[u
,n
, S
u
](x)
= c(x)u
D
(x) T[u
D
, S
u
](x) +o[q
D
, S
q
](x) +1[b, ](x) (x S
u
)
2. Integral representation:
u(x) = 1[b, ](x) +o[u
,n
, ](x) T[u, ](x)
Fundamental solution, Greens function
u + b = 0 (in ) u =u
D
(on S
u
), u
,n
=q
D
(on S
q
)
Fundamental solution (again):

G(x, ) + ( x) = 0 (in O)
i.e. any eld induced in by a unit point source placed at x
Greens function: fundamental solution with homogeneous BCs on :
((x, ) + ( x) = 0 (in )
1(x, ) := ((x, )n() = 0 (on S
q
),
((x, ) = 0 (on S
u
)
Fundamental solution boundary unknowns u (on S

q
) and u
,n
(on S
u
)
(boundary integral equation needed)
Greens function explicit integral representation formula (boundary

integral equation not needed)
u(x) =
_
((x, )b() dV
+
_
S
q
((x, )q
D
() dS
_
S
u
1(x, )u
D
() dS
(x )
Greens functions (Laplace): half-space
Method of images:
((x, ) = G(x, ) + G(x
sym
, ) Neumann BC on free surface
((x, ) = G(x, ) G(x
sym
, ) Dirichlet BC on free surface
x
sym
x
Greens functions (Laplace): sphere
2 2
x = x
_
(a / )
x
a
((x, ) =
1
4
_
1
r

a
1
r
_
(Dirichlet BC on surface)
Indirect boundary integral equations:
Seek solutions of Laplace equation using potentials:

u(x) = o[, ](x) +T[, ](x) (x )
Dirichlet problem (u =u
D
on ) using double-layer potential:
c(x)(x) +T[, ](x) = u
D
(x) (x )
Neumann problem (u
,n
=p
D
on ) using single-layer potential:
c(x)(x)
n(x)
o[, ](x) = p
D
(x) (x )
Seek solutions of Poisson equation using potentials:

u(x) = 1[b, ](x) +o[, ](x) +T[, ](x) (x )
Review of boundary integral equation formulations Elastostatics
Electrostatics
Laplace
Elastostatics
8. Preconditioning
Additive decomposition of strain into elastic and initial (e.g. thermal, plastic,
visco-plastic) parts (assuming innitesimal strain):
=
E
+
I
where = C:
E
Constitutive equation (C: fourth-order tensor of elastic moduli):

= C: (
I
)
For isotropic elasticity (: shear modulus, : Poisson ratio):

(
ijk
=
_
2
1 2
ij
k
+ (
ik
j
+
i
jk
)
_
Governing eld equation for unknown displacement eld u():

C
ijab
(u
a,bj

I
ab,j
) + b
i
= 0 ( )
Elastostatic fundamental solution
Kelvin fundamental solution: unit point force applied at x R
3
along
k-direction in unbounded elastic medium, i.e.:
C
ijab
U
k
a,bj
+ ( x)
ik
= 0 ( R
3
)
U
k
i
(x, ) =
1
16(1 )
1
r
_
r
i
r
k
+ (3 4)
ik
k
ij
(x, ) =
1
8(1 )
1
r
2
_
3r
i
r
k
r
j
+ (1 2)(
ik
r
j
+
jk
r
i

ij
r
k
)
x, R
3
T
k
i
(x, ) =
k
ij
(x, )n
j
()
( r = [ x[ , r
i
= (
i
x
i
)/r )
Governing eld equation for unknown displacement eld u():
_
_
C
ijab
(u
a,bj

I
ab,j
) + b
i
_
U
k
i
(x, ) dV
= 0 (1)
Governing eld equation for fundamental solution:
_
_
C
ijab
U
k
a,bj
+ ( x)
ik
_
u
i
() dV
= 0 (2)
Performing (1)-(2) and invoking the divergence identity, one obtains the integral
representation formula of the displacement:
u
k
(x) =
_
_
U
k
i
(x, )t
i
() T
k
i
(x, )u
i
()
_
dS
+
_
_
U
k
i
(x, )b
i
() +
k
ij
(x, )
I
ij
()
_
dV
Partially unknown contribution of (BC + unknown trace)
Known contribution of (if

I
is known beforehand)
Displacement boundary integral equation
Limiting process as x z in integral representation:
u
k
(x) =
_
_
U
k
i
(x, )t
i
() T
k
i
(x, )u
i
()
_
dS
x
+
_
_
U
k
i
(x, )b
i
() +
k
ij
(x, )
I
ij
()
_
dV
x
n
s
Non-integrable singularity of T
k
(x, ):
Limit to the boundary approach
Direct approach using exclusion neighbourhood of z
Indirect regularization approach

Displacement boundary integral equation
Integral equation, singular form:
1
2
u
k
(x) + P.V.
_
T
k
i
(x, )u
i
() dS
x

_
U
k
i
(x, )t
i
() dS
x
=
_
_
U
k
i
(x, )b
i
() +
k
ij
(x, )
I
ij
()
_
dV
Integral equation, regularized form:

_
_
T
k
i
(x, )[u
i
() u
i
(x)] U
k
i
(x, )t
i
()
_
dS
x
=
_
_
U
k
i
(x, )b
i
() +
k
ij
(x, )
I
ij
()
_
dV
Both forms require u C

0,
(otherwise process x z breaks down)
Numerical implementation based on (well-documented) singular element

integration methods.
Boundary-only formulations in the absence of body forces and initial strains.
Treatments (double / multiple reciprocity methods) sometimes allow to

convert domain integrals with b,
I
into boundary integrals.
Volume, single-layer and double-layer elastic potentials
Volume potential (prescribed body forces):

1
b
k
[b, ](x) =
_
U
k
i
(x, )b
i
() dV
(displacement eld created in R

3
by b given on )
Volume potential (prescribed initial strains):

1
k
[
I
, ](x) =
_
k
ij
(x, )
I
ij
() dV
(displacement eld created in R

3
by
I
given on )
Single-layer elastic potential:

o[, ]
k
(x) =
_
U
k
i
(x, )
i
() dS
x
Double-layer elastic potential:

T[, ]
k
(x) =
_
T
k
i
(x, )
i
() dS
x
Greens tensor (elasticity): half-space (Mindlin solution)
U
k
i
(x, ) =
1
16(1 )
_
r

ik
+
1
r

ik
+
r
,i
r
,k
r
+
r
,i
r
,k
r
+
2x
3
y
3
r
3
[
ik
3r
,i
r
,k
]
+

r (1 +r
,3
)
_
ik

r
,i
r
,k
1 +r
,3
__
U
k
3
(x, ) =
1
16(1 )
_
r
,3
r
,k
r
+
r
,k
r
_
r
,3

2x
3
r
_
6x
3
y
3
r
3
r
,3
r
,k
+
r
,k
r (1 +r
,3
)
_
U
3
i
(x, ) =
1
16(1 )
_
r
,i
r
,3
r
+
r
,k
r
_
r
,3

2x
3
r
_
+
6x
3
y
3
r
3
r
,i
r
,3

r
,i
r (1 +r
,3
)
_
U
3
3
(x, ) =
1
16(1 )
_
r
+
1 +
r
+
r
2
,3
r
+
r
3
r
3
r

2x
3
y
3
r
3
[1 3r
,3
r
,3
]
_
x
sym
r
_
x
r
( = 3 4, = 4(1 )(1 2))

Greens tensors (elasticity)
Full-space (Kelvin):
exact solutions (within full-space idealization):
Response to arbitrary eigenstrain distribution:
u
k
(x) =
_
k
ij
(x, )
I
ij
() dV
Elastic ellipsoidal inhomogeneity (Eshelby problem), using

equivalent-inclusion approach
Half-space with free surface (Mindlin, 1936), Boussinesq as special case:

exact solutions (within half-space idealization):
Soil mechanics and geotechnics;
Contact mechanics (Hertz solution, Galin formulae)
Two perfectly-bonded half spaces (Rongved, 1955) closed form
Elastic layer between two parallel planar free surfaces (Benitez and Rosakis
1987) integral transform
Review of boundary integral equation formulations Frequency-domain wave equations
Electrostatics
Laplace
Elastostatics
8. Preconditioning
u + k
2
u = 0 (in ) + unspecied well-posed BCs
Time-harmonic problems (u(, t) =Re[u()e
it
]); wavenumber k = /c.
e.g. linear acoustics (u: pressure, (i)
1
u: velocity)
Reciprocity identity:
_
(uv vu) dV =
_
(uv
,n
vu
,n
) dV (u
,n
un)
Fundamental solution (full space):

G(x, ) + k
2
G(x, ) + ( x) = 0 (in O)
G(x, ) =
e
ikx
4| x|
(O=R
3
)
Integral representation:
u(x) =
_
_
G(x, )u
,n
() G
,n
(x, )u()
_
dS
(x )
Singular integral equation
Dynamic (Helmholtz) and static (Laplace) fundamental solutions have same

singularity:
G(x, ) =
e
ikx
4| x|
=
1
4| x|
+ O(1) ( x)
Limit to the boundary (or other) approach yields the same free term c(x) as
with Laplace problems
Singular integral equation:

c(x)u(x) +
_
_
G
,n
(x, )u() G(x, )u
,n
()
_
dS
= 0 (x )
Singular integrals: invoke methods for handling static singular kernels:

_
G
,n
(x, )u() dS
=
_
_
G
,n
(x, ) G
,n
(x, ; =0)
_
u() dS
+
_
G
,n
(x, ; =0)u() dS

Half-space problems
Method of images still applicable:
((x, ) = G(x, ) + G(x
sym
, ) Neumann BC on free surface
((x, ) = G(x, ) G(x
sym
, ) Dirichlet BC on free surface
x
sym
x
Unbounded media
Wave equations (scalar, elastic, Maxwell...) frequently employed for media
idealized as unbounded

R
S
R
R
Limiting case as R of
c(x)u(x) +
_
_
S
R
+
_
_
_
G
,n
(x, )u() G(x, )u
,n
()
_
dS
= 0 (x )
Unbounded media, radiation conditions
Integral equation
c(x)u(x) +
_
_
G
,n
(x, )u() G(x, )u
,n
()
_
dS
= 0 (x )
valid if u satises a radiation condition at innity:
Integral form:
lim
R
_
S
R
_
G
,n
(x, )u() G(x, )u
,n
()
_
dS
= 0
Local form, sucient, known as Sommerfeld condition:

ux iku = o(|x|
1
) |x|
(Sommerfeld is known to imply u = o(1), i.e. decay of u at innity)
The radiation condition is satised by G(x, ), and consequently also by
The fundamental solution
Volume, single-layer, double-layer potentials
Integral representation formula
Scattering of incident waves by hard obstacles
Governing equations (hard obstacle[s])

u + k
2
u = 0 in
un = 0 on (no normal velocity, i.e. hard obstacle)
Known incident wave (or free-eld) u

F
; radiation conditions not assumed
(e.g. plane wave):
u
F
+ k
2
u
F
= 0 (in R
3
)
Decomposition:
u = u
F
+ v = incident + scattered, + radiation conditions for v

u
F
Scattering of incident waves by hard obstacles
Scattered eld veries integral equation (by virtue of radiation cnds):

c(x)v(x) +
_
_
G
,n
(x, )v() G(x, )v
,n
()
_
dS
= 0 (a)
Free-eld veries integral equation for interior of scatterer:

[c(x) 1]u
F
(x) +
_
_
G
,n
(x, )u
F
() G(x, )u
F
,n
()
_
dS
= 0 (b)
Simplied integral equation formulation (a)+(b) in terms of total eld:

c(x)u(x) +
_
G
,n
(x, )u() dS
= u
F
(x)
Fictitious eigenfrequencies
BIE formulations for exterior problems break down when is an eigenfrequency
for a certain interior problem.
Example:
Direct BIE formulation for exterior Neumann problem (scattering by rigid

obstacle):
c(x)u(x) +
_
G
,n
(x, )u() dS
= u
F
(x) (a)
Indirect BIE formulation (using a double-layer potential representation) for

interior homogeneous Dirichlet problem (using same normal as (a)):
c(x)(x) +
_
G
,n
(x, )() dS
= 0 (b)
(b) has non-trivial solutions if is a Dirichlet eigenvalue.
Therefore, so does (a) as the governing integral operator is the same

Remedies include:
(i) enforcing an extra set of integral identities at interior points;
(ii) combining (with complex coecients) two BIE formulations having dierent
eigenvalues (see treatment in Pyl, Clouteau, Degrande 2004)
Scattering of incident waves by penetrable obstacles
Penetrable inclusion (
(), c
() and with denitions =/
, =c/c
):
( + k
2
)u = 0 (in

B) ( +
2
k
2
)u
= 0 (in B)
u = u
, u
,n
= u
,n
(on B)
Domain integral equation of Lippman-Schwinger type (proof: combine

reciprocity identities written on B for u
and on R
3
B for u u
F
):
u(x) +
_
B
_
( 1)G(x, )u() + (1
2
)k
2
G(x, )u()
dV
= u
F
(x)
c ( ), ( )
* *
x x
u
F
B
Linear elastodynamics
Governing eld equation:
div (C: [u]) +
2
u + b = 0 (in )
Fundamental solution (full space):
U
k
i
(x, ) =
1
k
2
S
_
(
qs
ik

qk
is
)

x
q
s
G(|x|; k
S
) +

x
i
k
G(|x|; k
P
)
_
G(z; k) =
e
ikz
4z
, k
2
S
=

2
, k
2
P
=
12
2(1)
k
2
S
Radiation conditions for unbounded media, local form:
[u
P
]x ic
P
= o(|x|
1
)
[u
S
]x ic
S
= o(|x|
1
)
_
|x|
Review of classical BEM concepts
Electrostatics
Laplace
Elastostatics
8. Preconditioning
Interpolation
Partition into elements (possibly curvilinear and with curvilinear edges):

=
N
e
e=1
E
e
Isoparametric representation (most commonly used) of and unknown :

x =
n(e)
q=1
N
q
(a)x
q
(x) =
n(e)
q=1
N
q
(a)
q
(a = (a
1
, a
2
)
e
)
( )
a
1
2
3
1
2
a
e
E
_
a
_
a
e
x
x
x
Connectivity table:
Q(e, q) global number of q-th node of E
e
(1 e N
e
, 1 q n(e))
Isoparametric interpolation: N = N
N
(for scalar problems).
(with N
N
: number of nodes and N: number of unknown nodal DOFs)
Collocation BEM
Sample integral equation (Laplace + Dirichlet, direct or single-layer formulation):
_
G(x, )() dS
= b(x) (x )
Principle: enforce integral equation at the N

N
nodes x = x
1
, . . . , x
N
N
.
Leads to linear system of equations

A = b (AR
NN
, b R
N
)
where
A
PQ
=
eI (Q)
_
e
G(x
P
, (a))N
Q
(a)J(a) da (1 P, Q N)
b
P
= b(x
P
)
Matrix A square, fully-populated, invertible, non-symmetric obtained by

assembly of element matrices
A
e
(x
P
) R
1,n(e)
=
__
e
G(x
P
, (a))N
q
(a)J(a) da
_
1qn(e)
Evaluation of element integrals
A
e
(x
P
) R
1,n(e)
=
__
e
G(x
P
, (a))N
q
(a)J(a) da
_
1qn(e)
If x
P
,E
e
(nonsingular element integral): Gaussian quadrature
A
e
(x
P
)
G
g=1
w
g
G(x
P
, (a
g
))N
q
(a
g
)J(a
g
)
If x
P
E
e
(singular element integral): specialized treatment:
Weakly singular integrals (O(r

1
) kernel in 3-D) removed by suitable
transformaation of parametric coordinates a
Strongly singular integrals (O(r

2
) in 3-D) either
(i) recast into weakly singular integrals using regularization techniques
(ii) evaluated directly as Cauchy principal values
For simple element shapes and interpolations (e.g. 3-noded isoparametric

triangle), analytic singular integration available
Galerkin BEM
Example (simplest): solve Dirichlet problem for Laplace equation using single-layer
potential
H
1/2
(), nd H
1/2
(), B(, ) = b,
with B(, ) =
_
()G(x, ) (x) dS
dS
x
Leads to linear system of equations

A = b (AR
NN
, b R
N
)
where
A
PQ
=
I (P)
eI (Q)
_
N
P
(a
)
_
_
e
G(x(a
), (a))N
Q
(a)J(a) da
_
J(a
) da
(1 P, Q N)
b
P
=
I (P)
_
N
P
(a
)b(x(a
))J(a
) da
Matrix A square, fully-populated, invertible, symmetric

Limitations of traditional BEM
CPU for the main steps of traditional BEMs:
(a) Set-up of A: CPU = O(N
2
);
(b) Solution using direct solver (usually LU factorization): CPU = (N
3
);
(c) Evaluation of integral representations at M points: CPU = O(N M).
Besides:
(d) O(N
2
) memory needed for storing A.
= Problem size N at most O(10
4
)
Reasons fors (a)-(d):
G(x, ) non-zero for all (x, );
Element matrices A
e
(x
P
) recomputed for each new collocation point x
P
.
Overcoming the limitations of traditional BEM
Two issues:
1. To accelerate the BEM (i.e. to reduce its O(N
3
) complexity)
2. To increase permitted problem sizes.
Main ideas:
(i) Iterative solution of BEM matrix equation
= CPU = O(N
2
N
I
), with usually N
I
/N 0;
(ii) Acceleration of matrix-vector product A for given density .
= complexity lower than O(N
2
).
Several strategies available for developing fast BEMs
The Fast Multipole Method (FMM) is the most developed to date.
The GMRES iterative solver
Electrostatics
Laplace
Elastostatics
8. Preconditioning
The GMRES algorithm
Linear system
Au = b
_
_
_
AR
NN
or C
NN
, A invertible
u R
N
ou C
N
,
b R
N
ou C
N
Generalized Minimal RESiduals (GMRES): principle

u
(k)
= arg min
uu
(0)
+V
k
|bAu|
2
, V
k
= Vectv
1
, . . . , v
k
to be specied
Explicit form of minimisation:
u
(k)
= u
(0)
+
k
j =1
(k)
j
v
j
with
(k)
(
(k)
1
, . . . ,
(k)
k
) = arg min
1
,...,
k
_
_
_
_
r
(0)
j =1
j
Av
j
_
_
_
_
2
Iteration k: basis (v
1
, . . . , v
k1
) augmented with a new vector v
k
, hence
V
k1
V
k
.
Minimization problem size increases with k: restart when k >m, GMRES(m)
The GMRES algorithm
If k =N, one must have u

(N)
= u (hence convergence within N iterations
In practice: (i) N
iter
N is desired, (ii) exact convergence not necessary.
r
(k)
|b Au
(k)
| r
(0)
(: tolerance)
Construction of subspace V
k
= Vectv
1
, . . . , v
k
using Krylov vectors:
V
k
= Vect w
1
, w
2
= Aw
1
, . . . , w
k
= Aw
k1
with w
1
= r
(0)
Sequence (v
1
, . . . , v
k
) constructed using orthonormalization of Krylov
vectors (w
1
, . . . , w
k
):
v
T
v
k
= 0
|v
k
| = 1
Vectw
1
, . . . , w
k
= Vectv
1
, . . . , v
k
= V
k
for all k 1
Main contribution to computational cost: evaluation of matrix-vector
products Aw for given w.
Preconditioning
Left preconditioning:
M
1
Au = M
1
b
Improved convergence (i.e. less iterations) if M
1
A better conditioned than
A
Krylov sequence associated with matrix M

1
A:
r
(0)
= M
1
b M
1
Au
(0)
= w
1
i.e. Mw
1
= b Au
(0)
w
k+1
= M
1
Aw
k
i.e. Mw
k+1
= Aw
k
(k 0)
Modied convergence criterion:

|M
1
bM
1
Au
(k)
| |M
1
b|
Many approaches available for denir preconditioning matrices M:

Diagonal preconditioneur M
ij
= A
ij
ij
;
Incomplete LU factorization of A;
Sparse approximate inverses;
Multigrid approaches;
Preconditioners exploiting specic features of the problems, e.g.
single-inclusion case for many-inclusion problems.
The fast multipole method (FMM) for the Laplace equation
Electrostatics
Laplace
Elastostatics
8. Preconditioning
Origins of the FMM: fast computation of potentials
(x
i
) = C
N
j =1
q
j
|
j
x
i
|
(1 i N
x
)
C =(4
0
)
1
, electric charges q
j
(electrostatic); C =(, masses q
j
(gravitation)
Straightforward computation: CPU=O(N

x
N
);
Reason: inuence coecient |

j
x
i
|
1
depends on both x
i
and
j
;
Fast summation (Greengard, 1985): CPU=O(N

x
+N
)
Iterative solution of integral equation
Model problem:
nd ,
_
G(x, )() dS
= b(x), i.e. o[, ](x) = b(x) (x )

Krylov vector: A discretized version of
o[, ](x) :=
_
G(x, )() dS
Integral operator o: a generalization to innite-dimensional function spaces (here

H
1/2
()) of the concept of matrix.
Using traditional BEM: CPU = O(N

2
) for each evaluation of A;
Aim of the Fast Multipole Method: evaluation of A at CPU cost lower

than O(N
2
).
FMM: main ideas
o[, ](x) :=
_
G(x, )() dS
Main idea: seek to reuse element integrations (w.r.t. ) when collocation

point x is changed;
Method: express the fundamental solution as a series of products:

G(x, ) =
n=0
g
n
(x)h
n
()
and truncate the series at suitable level p:
G(x, ) =
p
n=0
g
n
(x)h
n
() +
G
(p)
Consequence:
o[, ](x) =
p
n=0
g
n
(x)
_
h
n
()() dS
+ (p)
The p integrations are independent on x and are reusable as x is changed.
FMM: main ideas
How to express G(x, ) as a sum (series) of products? Taylor expansion
about origins x
0
and
0
:
G(x, ) =
m0
1
m!
[
m
G](x,
0
) (
0
)
m
=
n0
m0
1
m!n!
[
n
x
G](x
0
,
0
) (xx
0
)
n
(
0
)
m
0
x
0
x
r
r
0
For Laplace kernel 1/r : sophisticated version of Taylor expansion leads to
multipole expansion (see next).
The fast multipole method (FMM) for the Laplace equation Multipole expansion of 1/r
8. Preconditioning
The Multipole expansion of 1/r is given (see proof later) by:
1
| x|
=
+
n=0
n
m=n
R
n,m
(xx
0
)
+
=0
n
=n
(1)
n
S
n+n
,m+m
(
0
x
0
)R
n
,m
(
0
)
R
n,m
(z) =
1
(n + m)!
P
m
n
(cos )e
im
n
S
n,m
(z) = (n m)!P
m
n
(cos )e
im
1
n+1
with z =
_
sin(cos e
1
+ sin e
2
)
_
+ cose
3
Conditions for convergence of the multipole expansion:

|xx
0
| < | x
0
| and |
0
| < |x
0
|
0
x
0
x
r
r
0
Multipole expansion of 1/r : computation of solid harmonics R
n,m
,
S
n,m
Solid harmonics R
n,m
and S
n,m
evaluated using Cartesian coordinates using
recursion formulae derived from those for Legendre polynomials:
The R
n,m
(z) are computed recursively by setting R
0,0
(z) = 1 and using
R
n+1,n+1
(z) =
z
1
+ iz
2
2(n + 1)
R
n,n
(z)
((n + 1)
2
m
2
)R
n+1,m
(z) (2n + 1)z
3
R
n,m
(z) +[z[
2
R
n1,m
(z) = 0 (n m)
The S
n,m
(z) are computed recursively by setting S
0,0
(z) = 1/ [z[ and using
S
n+1,n+1
(z) =
(2n + 1)(z
1
+ iz
2
)
[z[
2
R
n,n
(z)
[z[
2
S
n+1,m
(z) (2n + 1)z
3
S
n,m
(z) + (n
2
m
2
)S
n1,m
(z) = 0 (n m)
Finally, R
n,m
(z) and S
n,m
(z) for negative values of m are computed via the
identities
R
n,m
(z) = (1)
m
R
n,m
(z) S
n,m
(z) = (1)
m
S
n,m
(z)
Multipole expansion of 1/r : computation of solid harmonics R
n,m
,
S
n,m
Derivatives of the R
n,m
:
z
1
R
n,m
(z) =
1
2
(R
n1,m1
R
n1,m+1
)(z)
z
2
R
n,m
(z) =
i
2
(R
n1,m1
+ R
n1,m+1
)(z)
z
3
R
n,m
(z) = R
n1,m
Multipole expansion of 1/r : proof (1/4)
0
x
r
G(x, ) =
1
4| x|
x = (
0
) (x
0
)
Spherical coordinates:
(
0
) =
_
sin(cos e
1
+ sin e
2
) + cose
3
(x
0
) = R
_
sin (cos e
1
+ sin e
2
) + cos e
3
1
| x|
= (R
2
2R cos +
2
)
1/2
cos =
1
R
(
0
)(x
0
) = sin sin cos( ) + cos cos
1
| x|
= (R
2
2R cos +
2
)
1/2
=
1
R
(1 2zt + t
2
)
1/2
_
z =cos , t =

R
_
Since (12zt +t
2
)
1/2
is the generating function of the Legendre polynomials,
i.e.:
(1 2zt + t
2
)
1/2
=
n0
P
n
(z)t
n
(t < 1)
one has:
1
| x|
=
+
n=0
P
n
(cos )
R
n+1

n
( < R)
0
x
r
1
| x|
=
+
n=0
P
n
(cos )
R
n+1

n
( < R)
To recast as a series of products g(, , )h(R, , ): addition theorem for
Legendre polynomials:
P
n
(cos ) =
n
m=n
(n m)!
(n + m)!
_
P
m
n
(cos )e
im
__
P
m
n
(cos )e
im
_
1
| x|
=
+
n=0
n
m=n
R
n,m
(
0
)S
n,m
(x
0
)
R
n,m
(z) =
1
(n + m)!
P
m
n
(cos )e
im
n
S
n,m
(z) = (n m)!P
m
n
(cos )e
im
1
R
n+1
The series is convergent if |
0
| < |x
0
|.
1
| x|
=
+
n=0
n
m=n
R
n,m
(
0
)S
n,m
(x
0
)
Application to evaluation of potentials:
(x
i
) = C
N
j =1
q
j
|
j
x
i
|
= C
+
n=0
n
m=n
M
n,m
(
0
)S
n,m
(x
i
0
)
_
with multipole moments dened by
M
n,m
(
0
) =
N
j =1
q
j
R
n,m
(
j
0
)
Truncation of series to n <p (with error control available, see later):
Evaluation of N
N
x
inuence coecients |
j
x
i
|
1
replaced with
evaluation of p
2
N
x
products M
n,m
(
0
)S
n,m
(x
i
0
)
Multipole expansion of 1/r : proof
Insertion of local origin x
0
into x
i
0
:
Note identity
S
n,m
(z) = (1)
n
_

z
1
+ i

z
2
_
m
nm
z
nm
3
1
|z|
(m > 0) (a)
= (1)
n+m
_

z
1
i

z
2
_
m
n+m
z
n+m
3
1
|z|
(m < 0) (b)
Invoke multipole expansion of 1/|z| with z = (x

i
x
0
) (
0
x
0
):
1
|x
i
0
|
=
+
=0
n
=n
R
n
,m
(x
i
x
0
)S
n
,m
(
0
x
0
)
Use representation (a,b) for S

n
,m
(
0
x
0
);
Exploit harmonicity of 1/|

0
x
0
| via
_

0
1
+ i

0
2
__

0
1
i

0
2
_
1
|
0
x
0
|
+

2
0
3
2
1
|
0
x
0
|
= 0
Reorder and reorganize resulting formula

S
n,m
(x
i
0
) =
+
=0
n
=n
R
n
,m
(x
i
x
0
)S
n+n
,m+m
(
0
x
0
)
Multipole expansion: truncation error
Assume one can nd R > 0 and > 1 such that
_
|xx
0
| < R et | x
0
| > R
_
et
_
|
0
| < R et |x
0
| > R
_
An upper bound of the error arising from truncating the multipole expansion at
order p is:
1
| x|

p
n=0
n
m=n
R
n,m
(xx
0
)
p
=0
n
=n
(1)
n
S
n+n
,m+m
(
0
x
0
)R
n
,m
(
0
)

1
R( 1)
p+1
0
x
0
x
r
r
0
The truncation error is scale-independent
The fast multipole method (FMM) for the Laplace equation The single-level fast multipole method
8. Preconditioning
Single-level fast multipole method
Boundary of interest enclosed in
a cubic grid
d
Convergence of multipole
expansion guaranteed if x and
lie in non-adjacent cells, with

3
Cx
C
(A(Cx))
C
(A(Cx))
d
Single-level FMM
Matrix-vector product integral operator evaluation

o[, ](x) =
_
G(x, )() dS
Split into near and far contributions:

_
A(C
x
)
_
C
/ A(C
x
)
_
C
o[, ](x) = o[, ]

near
(x) +o[, ]
far
(x)
Single-level FMM
Far contribution o[, ]
far
(x):
o[, ]
far
(x) =
/ A(C
x
)
_
C
G(x, )() dS
(x (
x
)
Introduce (truncated) multipole expansion of G(x, ), with
0
and x
0
chosen as
centres of cells (
and (
x
:
o[, ]
far
(x)
/ A(C
x
)
1
4
p
n=0
n
m=n
R
n,m
(xx
0
)
p
=0
n
=n
(1)
n
S
n+n
,m+m
(
0
x
0
)
_
C
R
n
,m
(
0
)() dS
=
1
4
p
n=0
n
m=n
R
n,m
(xx
0
)
/ A(C
x
)
p
=0
n
=n
(1)
n
S
n+n
,m+m
(
0
x
0
) M
n
,m
((
)
=
p
n=0
n
m=n
R
n,m
(xx
0
)L
n,m
((
x
)
Single-level FMM
C
x
Single-level FMM: complexity
1. Compute and store multipole moments M
n,m
((
) for each integration cell:

CPU = O(p
2
N
B
(N/N
B
)) = O(p
2
N).
2. For each collocation cell (
x
:
(a) Compute local coecients L
n,m
(C
x
, C
) (M2L):
CPU = O(p
2
p
2
N
B
) = O(p
4
N
B
);
L
n,m
(C
x
) =
/ A(C
x
)
p
=0
n
=n
(1)
n
S
n+n
,m+m
(
0
x
0
) M
n
,m
(C
)
(b) Far contribution S[, ]
far
(x) to integral operator, for all x C
x
:
CPU = O(p
2
(N/N
B
))
S[, ]
far
(x) =
+
n=0
n
m=n
R
n,m
(xx
0
)L
n,m
(C
x
)
(c) Near contribution S[, ]
near
(x) to integral operator, using standard BEM
procedures: CPU = O
_
|A(C
x
)| (N/N
B
) (N/N
B
)
_
= O(|A
_
C
x
)| N
2
/N
2
B
_
Single-level FMM: optimal complexity
Total CPU time for one evaluation of o[, ]:
CPU = Ap
2
N + N
B
_
Bp
4
N
B
+ Cp
2
(N/N
B
) + D [/((
x
)[ N
2
/N
2
B
_
= (A + C)p
2
N + Bp
4
N
2
B
+ DN
2
/N
B
Optimal choice: N
B
= O(N
2/3
), yielding CPU / GMRES iteration = O(N
4/3
)
Single-level FMM (Laplace and other elliptic PDEs): CPU = O(N

4/3
);
To further improve complexity: multi-level FMM

The fast multipole method (FMM) for the Laplace equation The multi-level fast multipole method
8. Preconditioning
Multi-level FMM
1
1
Multi-level FMM: initialization of multipole moments
24
19
20 21
27
31
33
35
36
37
40
41
42
43
18
22 28
32 39
23
26
30
38
25
34
44
29
9 11 12 13 14 16 15 10 6 7 8 17
3 5
1
2 4
20 21 22 26 27 28 29 30 31 32 33 34 35 36 37 38 40 41 42 43 39 44 19 18 23 24 25
Computation of multipole moments, leaf level cells ( =

=3 here):
M
n,m
((
(
;
(
)
0
) =
_
C
(
R
n,m
(
(
)
0
)() dS

Multi-level FMM: upward pass
7
11
15
17 14
13
12
8
16
10 9 6
3 5
1
2 4
16 6 7 8 9 10 11 12 13 14 15 17
Computation of multipole moments, parent cells ( =3 =2 here):
M
n,m
((
(1)
;
(1)
0
) =
C
()
Children(C
(1)
)
M
n,m
((
()
;
(1)
0
) (M2M)
Upward pass needs translation of origin
()
0

(1)
0
in
M
n,m
((
()
;
()
0
)
Multi-level FMM: M2M translation formula for upward pass
M
n,m
((
(1)
;
(1)
0
) =
C
()
Children(C
(1)
)
_
C
(
R
n,m
(
(1)
0
)() dS
M2M translation identity:

R
n,m
(
(1)
0
) =
=0
n
=n
R
n
,m
(
()
0

(1)
0
)R
nn
,mm
(
()
0
)
M
n,m
((
()
;
(1)
0
) =
=0
n
=n
R
n
,m
(
()
0

(1)
0
)M
nn
,mm
((
()
;
()
0
)
Multi-level FMM: M2M translation formula for upward pass
Proof:
write |x|
1
in two ways (inserting either
(1)
0
or
()
0
as pole):
1
|x|
=
+
n=0
n
m=n
R
n,m
(
()
0
)S
n,m
(x
()
0
) (a)
=
+
n=0
n
m=n
R
n,m
(
(1)
0
)S
n,m
(x
(1)
0
) (b)
Invoke identity
S
n,m
(x
(1)
0
) = (1)
n
+
=0
n
=n
R
n
,m
(
(1)
0

()
0
)S
n+n
,m+m
(x
()
0
)
Redene summation indices ((n, m) (n+n
, m+m
)), reorder summations
Identify cofactors of S
n,m
(x
(1)
0
) in (a) and (b)
R
n,m
(
(1)
0
) =
=0
n
=n
R
n
,m
(
()
0

(1)
0
)R
nn
,mm
(
()
0
)
Multi-level FMM: denition of interaction list
Denition of interaction list J(():
(
J(() (
, /(() but Father((
) /(Father(())
(
J(()
(
/(()
=2 =3
Note: (
J(() (
, /(() for level =2.

Multi-level FMM: denition of interaction list
General case, for generic cell (:
C
(
J(()
(
/(()
J(() contains up to
6
3
3
3
= 189 cells (3-D);
6
2
3
2
= 27 cells (2-D)
Multi-level FMM: M2L translation, upper level =2
M2L translation formula between disjoint same-level cells:
L
n,m
((
()
x
) =
/ A(C
()
x
)
p
=0
n
=n
(1)
n
S
n+n
,m+m
(
()
0
x
()
0
) M
n
,m
((
()
)
x
0
0
(2)
(2)
Multi-level FMM: M2L translation, upper level =2
17 16
15
14 13
12
11
10 9 6
8
7
Collocation cell (
(0)
x
: no action taken
Collocation cell (
(1)
x
: no action taken
Collocation cell (
(2)
x
: initialize M2L
L
n,m
((
(2)
x
) =
C
(2)
/ A(C
(2)
x
)
p
=0
n
=n
(1)
n
S
n+n
,m+m
(
(2)
0
x
(2)
0
) M
n
,m
((
(2)
)
To apply M2L to /((
(2)
x
): subdivision
Multi-level FMM: M2L translation, downward pass
23
24 25
9 11 12 13 14 16 15 10 6 7 17
3 5
1
2 4
20 21 22 26 27 28 29 30 31 32 33 34 35 36 37 38 40 41 42 43 39 44 19 18 23 24 25
8
L
n,m
((
()
x
; x
()
0
) = L
n,m
((
(1)
x
; x
()
0
)
+
C
()
I(C
()
x
)
p
=0
n
=n
(1)
n
S
n+n
,m+m
(
()
0
x
()
0
) M
n
,m
((
()
)
Downward pass entails translation of origin x
(1)
0
x
()
0
Multi-level FMM: M2L translation formula for downward pass
L
n,m
((
(1)
x
; x
()
0
) =
=0
n
=n
R
n
n,m
m
(x
()
0
x
(1)
0
)L
n
,m
((
(1)
x
; x
(1)
0
)
Proof: one must have
n=0
n
m=n
R
n,m
(xx
()
0
)L
n,m
((
(1)
x
; x
()
0
) (a)
=
n=0
n
m=n
R
n,m
(xx
(1)
0
)L
n,m
((
(1)
x
; x
(1)
0
) (b)
Then, insert identity
R
n,m
(xx
(1)
0
) =
=0
n
=n
R
n
,m
(xx
()
0
)R
nn
,mm
(x
()
0
x
(1)
0
)
into (b) and identify cofactors of R
n,m
(xx
()
0
) in (a) and (b)
Multi-level FMM: overall complexity
Fixed number M of DOFs per leaf cell
=

= O(log N) and O(N/M) leaf cells;
Each non-empty level- cell has (on average) 4 non-empty children cells
= on average, N
()
=O(4
N) DOFs per level- cell

(i) Evaluation of multipole moments in leaf cells:
CPU = O(p
2
M (N/M)) = O(p
2
N)
(ii) Upward pass (M2M):
CPU = O(p
4
(N/M)[1 + 4
1
. . . + 4
3
] = O(p
4
N/M)
(iii) Transfer (M2L) at each level 2
from interaction list of each cell:

CPU = O(p
4
(N/M)[1 + 4
1
. . . + 4
2
] = O(p
4
N/M);
(iv) Downward pass (L2L) at each level:
CPU = O(p
4
(N/M)[1 + 4
1
. . . + 4
3
] = O(p
4
N/M)
(v) Evaluation of local expansions at leaf cells:
CPU = O(p
2
N/M)
(vi) Evaluation of near contributions o[, ]
near
(x) using standard BEM
techniques:
CPU = O((N/M) M [/(()[M) = O([/((
x
)[MN)
Overall complexity: CPU = O(N)
The fast multipole method (FMM) for elastostatics
Electrostatics
Laplace
Elastostatics
8. Preconditioning
Elastostatics
Reformulation of Kelvin solution in terms of 1/r :
U
k
i
(x, ) =
1
16(1)
_
(34)
ik
1
r
+ (
i
x
i
)

x
k
1
r
_
(a)
T
k
i
(x, ) =
1
8(1)
_
(12)
_
n
k
()

x
i
1
r
n
i
()

x
k
1
r
+ 2(1)
ik
n
j
()

x
j
1
r
+ (
j
x
j
)n
j
()

2
x
i
x
k
1
r
_
(b)
Substitute multipole expansion of 1/r into (a) and (b):
1
| x|
=
+
n=0
n
m=n
R
n,m
(xx
0
)
+
=0
n
=n
(1)
n
S
n+n
,m+m
(
0
x
0
)R
n
,m
(
0
)
(again)
Elastostatics: multipole expansion of Kelvin solution
U
k
i
(x, ) =
1
16(1 )
+
=0
n
=n
_
F
n
,m
ki
(x
0
)R
n
,m
(
0
)
+ G
n
,m
k
(x
0
)(
i
i 0
)R
n
,m
(
0
)
_
F
n
,m
ki
(x
0
) =
+
n=0
n
m=n
(1)
n
S
n+n
,m+m
(
0
x
0
)
_
(3 4)
ik
R
n,m
(xx
0
) + (
i 0
x
i 0
x
i
x
i 0
)

x
k
R
n,m
(xx
0
)
_
G
n
,m
k
(x
0
) =
+
n=0
n
m=n
(1)
n
S
n+n
,m+m
(
0
x
0
)

x
k
R
n,m
(xx
0
)
A similar formula (not shown) can be established for the multipole representation
of T
k
i
(x, )
Elastostatics: multipole moments
M
t
n
,m
;i
((
;
0
) =
_
C
R
n
,m
(
0
) t
i
() dS
M
t
n
,m
((
;
0
) =
_
C
R
n
,m
(
0
) (
i
i 0
)t
i
() dS
M
u
n
,m
;ki
((
;
0
) =
_
C
R
n
,m
(
0
) n
k
()u
i
() dS
M
u
n
,m
;k
((
;
0
) =
_
C
R
n
,m
(
0
) n
k
()(
i
i 0
)u
i
() dS

M2M, M2L and L2L formulae
M2M, M2L and L2L formulae are derived using those for 1/r . For example, the
elastostatic M2M formulae are:
M
t
n,m;i
((
()
;
(1)
0
) =
=0
n
=n
R
n
,m
(
()
0

(1)
0
)M
t
nn
,mm
;i
((
()
;
()
0
)
M
t
n,m
((
()
;
(1)
0
) =
=0
n
=n
_
R
n
,m
(
()
0

(1)
0
)M
nn
,mm
((
()
;
()
0
)
+ (
()
i 0

(1)
i 0
)M
t
nn
,mm
;i
((
()
;
()
0
)
M
u
n,m;ki
((
()
;
(1)
0
) =
=0
n
=n
R
n
,m
(
()
0

(1)
0
)M
nn
,mm
;ki
((
()
;
()
0
)
M
u
n,m;k
((
()
;
(1)
0
) =
=0
n
=n
_
R
n
,m
(
()
0

(1)
0
)M
nn
,mm
;k
((
()
;
()
0
)
+ (
()
i 0

(1)
i 0
)M
t
nn
,mm
;ki
((
()
;
()
0
)
Numerical example: uniform thermal strain in ellipsoidal region
x
y
z
a
1
a
2
a
3
FEM
BEM
e
x
z
y
FEM
BEM
Mesh Nodes Elements Oct-tree DOFs
BEM FEM Max level Leaves N
BEM
N
F
N
BEM
+ N
F
1 267 346 979 3 42 1050 276 1326
2 822 1038 3153 3 100 3126 903 4029
3 1362 1540 5563 3 103 4632 1770 6402
4 2274 2418 9626 4 301 7266 3189 10455
5 5881 5200 26602 4 422 15612 9837 25449
6 12868 9402 61770 5 1175 28218 24495 52713
7 20258 12842 100200 6 1403 38538 41505 80043
Numerical example: uniform thermal strain in ellipsoidal region
-0.096
-0.095
-0.094
-0.093
-0.092
-0.091
0 1 2 3 4 5 6 7 8 9 10
s
Sxx
Exact
Mesh 3
Mesh 5
Mesh 7
(a)
-0.077
-0.076
-0.075
-0.074
-0.073
-0.072
0 1 2 3 4 5 6 7 8 9 10
s
Syy
Exact
Mesh 3
Mesh 5
Mesh 7
(b)
-0.057
-0.056
-0.055
-0.054
0 1 2 3 4 5 6 7 8 9 10
s
Szz
Exact
Mesh 3
Mesh 5
Mesh 7
(c)
Mesh Precond. (s) FMM Iters Total time
BEM FEM (s/iter) n (s)
1 10 <1 1 37 43
2 36 <1 2 37 154
3 50 <1 3 37 202
4 64 3 6 36 277
5 169 18 11 37 721
6 349 101 19 38 1425
7 512 279 42 38 2913
1.E+00
1.E+01
1.E+02
1.E+03
1.E+04
1.E+03 1.E+04 1.E+05
DOFs
Total Time (sec.)
FMM-CBEM
Traditional CBEM
Cross over point
Numerical example: many-inclusion problem
Nodes Elements Oct-tree DOFs
BEM FEM Max. level Leaves N
BEM
N
F
N
BEM
+ N
F
93227 122880 326493 5 7176 374784 92289 467073
Precond. (s) Time (s) Iters Total time
BEM FEM Upw. Downw. Direct Cycle n (s)
6609 19 47 48 84 180 147 39656
Numerical example: Pian Telessio dam
Dam
Rock
Water
1050
1010
448
Dam
Rock
Water
1050
1010
448
Mesh Nodes Elements Oct-tree DOFs
BEM FEM Levels Leaves N
BEM
N
F
Total
1 25443 21684 (T3) 73569 (T4) 9 8953 38118 43797 81915
2 23433 7726 (T6) 10307 (T10) 8 3548 50490 46773 97263
3 51978 15296 (T6) 14462 (T10) 8 6786 96636 64152 160788
A 406035 279742 1218105
Mesh Precond. (s) Time (s) Iters Total time
BEs FEs Upw. Down. Direct Cycle n (s)
1 186 26 23 27 24 76 83 7916
2 328 114 11 21 23 57 82 5818
3 1215 223 23 36 102 165 85 17775
A 3749
-4.00E-03
-3.00E-03
-2.00E-03
-1.00E-03
0.00E+00
1.00E-03
2.00E-03
3.00E-03
4.00E-03
5.00E-03
0 100 200 300 400 500
s (m)
Ux (m)
ABAQUS
Mesh 1
Mesh 2
Mesh 3
(a)
-1.60E-02
-1.40E-02
-1.20E-02
-1.00E-02
-8.00E-03
-6.00E-03
-4.00E-03
-2.00E-03
0.00E+00
0 100 200 300 400 500
s (m)
Uy (m)
ABAQUS
Mesh 1
Mesh 2
Mesh 3
y
z
x
s
(b)
-1.80E-03
-1.60E-03
-1.40E-03
-1.20E-03
-1.00E-03
-8.00E-04
-6.00E-04
-4.00E-04
-2.00E-04
0 100 200 300 400 500
s (m)
Uz (m)
ABAQUS
Mesh 1
Mesh 2
Mesh 3
(c)
-4.0E+05
-3.5E+05
-3.0E+05
-2.5E+05
-2.0E+05
-1.5E+05
-1.0E+05
0 5 10 15 20 25 30
s (m)
Sxx (N/m
2
)
ABAQUS
Mesh 1
Mesh 2
Mesh 3
-1.4E+06
-1.2E+06
-1.0E+06
-8.0E+05
-6.0E+05
-4.0E+05
-2.0E+05
0.0E+00
2.0E+05
0 5 10 15 20 25 30
s (m)
Syy (N/m
2
)
ABAQUS
Mesh 1
Mesh 2
Mesh 3
a=151 m
s
x
y
z
-2.0E+06
-1.6E+06
-1.2E+06
-8.0E+05
-4.0E+05
0.0E+00
4.0E+05
8.0E+05
0 5 10 15 20 25 30
s (m)
Szz (N/m
2
)
ABAQUS
Mesh 1
Mesh 2
Mesh 3
Exemple, calcul damortissements dans les MEMS
Figure 5. More realistic model of one-fourth of the MEMS: geometry and mesh.
Frangi A., Spinola G., Vigna B., IJNME 68:10311051
Exemple, calcul damortissements dans les MEMS
Table II. Comparison between meshes of increasing renement.
Mesh employed Mesh 1 Mesh 2 Mesh 3
Number of unknowns 125 058 272 364 548 388
Global force (N) 1.80 10
4
2.01 10
4
2.12 10
4
Figure 6. Convergence of the GMRES solver and of the force computed.
Frangi A., Spinola G., Vigna B., IJNME 68:10311051
Exemple, homogeneisation numerique
Fig. 10 Contour plot of surface stresses
for a model with 216 randomly distributed

and oriented short bers
Fig. 9 A BEM mesh used for the short ber inclusion with 456
elements
Liu Y.J., Nishimura N. et al., ASME J. Appl. Mech. 72:115128 (2005)
Fig. 11 A RVE containing 2197 short bers with the total DOF3 018 678
Fig. 14 A RVE containing 5832 long bers with the total DOF10 532 592
Fig. 15 Estimated effective Youngs moduli in the x-direction for the composite model with
up to 5832 long rigid bers ber volume fraction3.85%
The fast multipole method for elastodynamics
Electrostatics
Laplace
Elastostatics
8. Preconditioning
Avec: Jean-Fran cois Semblat (Civil Engineering Research Lab [LCPC], Paris)
Nicolas Nemitz (doctoral thesis, 2002-2006)
Stephanie Chaillat (doctoral thesis, 2005)
Eva Grasso (doctoral thesis, 2008)
Tekoing Lim (doctoral thesis, Atomic Energy Commission [CEA], 2007)
Motivation
Modelling of elastic wave propagation in large/unbounded domains
Soil-structure interaction
Site eects
Computational forward solution method for inverse problems

Pros and cons of BEMs for elastic waves
FEM, FDM, DG...
Domain mesh
Approx. radiation conditions
Sparse matrix
BEM
Surface mesh (i.e. reduced dimensionality)
Exact radiation conditions
Fully-populated matrix
BEM adequate for large (unbounded) media with simple (linear) properties.
Fully-populated BEM inuence matrix is a priori a severe limiting factor
Standard BEM (3-D elastodynamics, frequency domain)
Governing integral equation for boundary displacements and tractions
c
ik
(x)u
i
(x) =
_
_
t
i
(x)U
k
i
(x, ; ) u
i
(x)T
k
i
(x, ; )
dS
(x )
Full-space elastodynamic fundamental solutions
U
k
i
(x, ; ) =
1
4k
2
S
_
(
qs
ik

qk
is
)

x
q
y
s
G
S
(x ) +

x
i
y
k
G
P
(x )
_
T
k
i
(x, ; ) = C
ijh
U
k
h
(x, ; )n
j
()
G
(z) =
exp(ik
z)
z
(fund. sol. Helmholtz eqn., = P, S)
BEM discretization = fully-populated system of linear equations.
Computational limitations of standard BEM
Solution of fully-populated matrix equation
Direct solvers (LU factorisation, . . .) :
Pros: robust, accurate;
Cons: O(N
2
) memory and O(N
3
) CPU
Iterative solvers (GMRES, . . .) :
Pros: O(N
iter
N
2
) CPU;
Cons: O(N
2
) memory; N
iter
may be large
(N: number of BE DOFs)
Limitations of standard BEM
High memory cost

problem size limit N = O(10
4
) (PC, single-proc.)
Limited geometric complexity, (piecewise) heterogeneity, frequency range

Fast multipole accelerated BEM
Fast Multipole Method (FMM):
Based on iterative linear equation solvers (GMRES)
Fast method for evaluating matrix-vector products, i.e. discretized versions of

e.g.
_
t
i
(x)U
k
i
(x, ; ) dS
(for given solution candidate t)

A few milestones
Laplace: Rokhlin (1985)
Electrostat.: Greengard (1988)
Electromag.: Chew (1994), Darve (2000), Sylvand (2002)...
Elastodyn. freq. domain: Fujiwara (2000)
Elastodyn., time domain: Nishimura (2002)
BEM-FEM : Margonari, Bonnet (2004), Gaul et al...
Eective prop. of composite mater.: Nishimura, Liu (2005)

Decomposition of Helmholtz fundamental solution
Multipole expansion formula (diagonal form, Epton and Dembart 1995)
0
x
0
x
r
r
0
exp(ik| x|)
| x|
=
ik
4
lim
L+
_
sS
e
iks.
(
L
(s; r
0
; k)e
iks.x
ds
Transfer function
(
L
(s; r
0
; k) =
L
p=0
(2p + 1)i
p
h
(1)
p
(k|r
0
|)P
p
_
cos(s, r
0
)
_
Single-level FMM
Boundary of interest enclosed in
cubic grid
d
Convergence of multipole
expansion assured if x and lie
in non-adjacent cells
Cx
C
(A(Cx))
C
(A(Cx))
d
Single-level FMM
Matrix-vector product evaluation of integral operator
Must compute e.g.:

[/t](x) :=
_
t
i
(x)U
k
i
(x, ; ) dS
Split integrals into near and FM contributions:

_
A(C
x
)
_
C
/ A(C
x
)
_
C
[/t](x) = [/t]
near
(x) + [/t]
far
(x)
Single-level FMM algorithm: principle
C
1

2
3

4
C
x
x
0
x
1
x
2
x
3
x
4
_
t
i
(x)U
k
i
(x, ; ) dS

exp(ik| x|)
| x|
=
ik
4
lim
L+
_
sS
e
iks.
(
L
(s; r
0
; k)e
iks.x
ds
compute multipole moments for each cell (
and quadrature point
Transfer (M2L) from (
to non-adjacent (
x
Evaluate FM contribution to matrix-vector product
Add near contribution to matrix-vector product (computed using standard

BEM techniques)
Complexity of single-level elastodynamic FMM: O(N
3/2
) per GMRES iteration
Multi-level FMM
level =0
level =1
level =2
level =3
.
.
. level =

(leaf)
highest level for which FMM is applicable.
computation organization based on
recursive subdivision (oc-tree)
Complexity of multi-level elastodynamic FMM: O(N log
2
N)
Computational issues: truncation of transfer function
Transfer function
(
L
(s; r
0
; k) =
L
p=0
(2p + 1)i
p
h
(1)
p
(k|r
0
|)P
p
_
cos(s, r
0
)
_
Choice of truncation parameter:
L too small: convergence not reached for (

L
(s; r
0
; k);
L too large: divergence of h

(1)
p
Empirical formula used
L=
3k
S
d + C
log
10
(
3k
S
d + )
(see Darve 2000 and Sylvand 2002 for Maxwell eqns)
Computational issues: adjustment of constant C
d
L =
3k
S
d + C
log
10
(
3k
S
d + )
0 5 10 15 20
C
10
-4
10
-2
10
0
10
2
R
M
S

s
o
l
u
t
i
o
n

e
r
r
o
r

o
n

c
a
v
i
t
y

s
u
r
f
a
c
e
erreur
20
30
40
50
60
70
C
P
U

/

i
t
e
r
.
CPU
C
= 7.5 (consistent with Sylvand 2002, Maxwell eqns)

Computational issues: number of levels
d
min
2d
min
4d
min
L =
3k
S
d + C
log
10
(
3k
S
d + )
# of levels d
min
k
S
/2 rel. err. / BEM CPU / iter (s)
4 1.32 1.1 .10
5
367
5 0.66 4.7 .10
4
134
6 0.33 3.7 .10
3
104
7 0.17 5.1 .10
2
200
8 0.083 1.7 .10
1
380
Choice of leaf cell size ( choice of # of levels):
inuence on CPU
inuence on accuracy
d
min
0.3
S
Computational issues: discretization relative to wavelength
10 20
points per S wavelength
10
-3
10
-2
R
M
S

s
o
l
u
t
i
o
n

e
r
r
o
r

o
n

c
a
v
i
t
y

s
u
r
f
a
c
e
erreur
0
50
100
150
200
250
300
C
P
U

/

i
t
e
r
.
CPU
10 points per S wavelength
Numerical verication of theoretical complexity (CPU), Helmholtz
Example: sphere under uniform normal velocity. Mesh renement, mesh
density / wavelength kept xed
N. Nemitz, M. Bonnet, Eng. Anal. Bound. Elem (2008)
Numerical verication of theoretical complexity (CPU),
elastodynamics
1e+02 1e+03 1e+04 1e+05 1e+06
N
1e-02
1e+00
1e+02
1e+04
C
P
U

/
i
t
e
r

(
s
)
BEM
O(N
2
)
single-level FMM
O(N
3/2
)
multi-level FMM
O(N log
2
N)
Chaillat S., Bonnet M., Semblat J.F., Comp. Meth. Appl. Mech. Engng. (2008)
Example (Helmholtz)
Identication of a dual hard scatterer
1
/a
2
/a
6 4 2 0 2 4 6
6
4
2
0
2
4
6
12
10
8
6
4
2
x 10
4
N. Nemitz, M. Bonnet, Eng. Anal. Bound. Elem (2008).
Example (Helmholtz)
FMM: numerical parameters
Element and DOF count (FM-BEM):
Cube size Cube Obstacle Total (
Elements DDLs Elements DDLs Elements DDLs
16a 76800 38402 336 170 77136 38572 100
3
32a 307200 153602 336 170 307536 153772 150
3
CPU timing (single-CPU PC) and GMRES iteration count
Cube size u
true
on S
true
u on S u on S T on (
16a 1444s (N
iter
=435) 969s (N
iter
=282) 1163s (N
iter
=342) 852s
32a 6461s (N
iter
=439) 5615s (N
iter
=388) 6818s (N
iter
=476) 1860s
N. Nemitz, M. Bonnet, Eng. Anal. Bound. Elem (2008).
Example: scattering of a plane P wave by a spherical cavity
O
R
z
x
y
plane wave
Example: scattering of a plane P wave by a spherical cavity
=0
1 1.5 2 2.5 3
r/R
-2
-1
0
1
r
a
d
i
a
l

d
i
s
p
l
a
c
e
m
e
n
t
k
P
a/ = 1 analytical
k
P
a/ = 1 numerical
k
P
a/ = 4 analytical
k
P
a/ = 4 numerical
= 0
=/4
1 1.5 2 2.5 3
r/R
-0.5
0
0.5
1
r
a
d
i
a
l

d
i
s
p
l
a
c
e
m
e
n
t
k
P
a/ = 1 analytical
k
P
a/ = 1 numerical
k
P
a/ = 4 analytical
k
P
a/ = 4 numerical
= /4
=/2
1 1.5 2 2.5 3
r/R
-0.4
-0.2
0
0.2
0.4
r
a
d
i
a
l

d
i
s
p
l
a
c
e
m
e
n
t
k
P
a/ = 1 analytical
k
P
a/ = 1 numerical
k
P
a/ = 4 analytical
k
P
a/ = 4 numerical
= /2
=3/4
1 1.5 2 2.5 3
r/R
-0.5
0
0.5
1
1.5
2
r
a
d
i
a
l

d
i
s
p
l
a
c
e
m
e
n
t
k
P
a/ = 1 analytical
k
P
a/ = 1 numerical
k
P
a/ = 4 analytical
k
P
a/ = 4 numerical
= 3/4
z
x
y
N = 122, 886
Example: scattering of a plane P wave by a hemispheric canyon
y
z
plane P wave
a
D = 3a
free surface
elastic half-space
A
B C
(N = 23382)
Simplied conguration for topographic site eect
Low frequency: comparison with other published results
Higher frequency: FMM

Comparaison with earlier results, k
P
a = 0.25 (low frequency)
x, y
z
a
D = 3a
A
B C
s
0 1 2 3
s / a
0
1
2
3
4
d
i
s
p
l
a
c
e
m
e
n
t

m
o
d
u
l
u
s
|
u
y
|
(present FMM)
|
u
y
|
(Sanchez-Sesma)
|
u
y
|
(Reinoso et al.)
|
u
z
|
(present FMM)
|
u
z
|
(Sanchez-Sesma)
|
u
z
|
(Reinoso et al.)
Results for higher frequency k
P
a = 5
N = 287 946 (86 iter., 210 s / iter, single-proc. 3 GHz PC)
x, y
z
a
D = 3a
A
B C
s
0 1 2 3
s / a
0
1
2
3
d
i
s
p
l
a
c
e
m
e
n
t

m
o
d
u
l
u
s
|
u
y
|
(present FMM)
|
u
z
|
(present FMM)
Study of convergence of GMRES
k
P
a =0.25 k
P
a =0.5 k
P
a =0.75 k
P
a =1.5 k
P
a =5 k
P
a =10
D =3a 7 (23382) 10 (23382) 12 (23382) 19 (23382) 86 (287946) >280 (1145700)
D =5a 7 (61875) 10 (61875) 15 (61875) 28 (61875) 159 (774180)
D =7a 8 (77565) 13 (77565) 17 (77565) 43 (77565)
D =20a 14 (98844) 39 (98844) 43 (98844)
Example: scattering of a plane P wave by a semi-elliptical canyon
B
C
D
x
y
z
a
b
a
y
z
plane P wave
a
D = 6a
free surface
elastic half-space
D E B A
C
Simplied conguration for topographic site eect
Low frequency: comparison with other published results

S
a = 0.5 (low frequency), N = 25 788
-3 -2 -1 0 1 2 3
s / R
0.5
1
1.5
2
2.5
3
3.5
d
i
s
p
l
a
c
e
m
e
n
t

m
o
d
u
l
u
s
|
u
y
|
(present FMM)
|
u
y
|
(Reinoso et al.)
|
u
z
|
(present FMM)
|
u
z
|
(Reinoso et al.)
Results for higher frequency k
s
a = 2
0.012
0.630
1.250
1.870
2.500
x
y
0.28
0.97
1.66
2.35
3.05
N = 353 232 (32 iter., 140 s / iter, single-proc. 3 GHz PC)
Example: scattering of a plane P wave by an alluvial hemispheric
valley
x, y
z
plane P wave
R
D = 5R
free surface
elastic half-space
A
B C
(N = 17409)

2
=0.3
1
,
2
=0.6
1
,
1
=0.25,
2
=0.3
Low frequency: comparison with Sanchez-Sesma (1983) and Delavaud (2007)

valley
P
R = 0.5
Sanchez-Sesma, 1983 (semi-analytical)
Delavaud, 2007 (spectral nite element method)

x, y
z
R
D = 5R
A
B C
s
0 1 2
x/R
0
1
2
3
4
5
6
d
i
s
p
l
a
c
e
m
e
n
t

m
o
d
u
l
u
s
|
u
y
|
(present FMM)
|
u
y
|
(Sanchez-Sesma 1983)
|
u
y
|
(Delavaud 2007, SEM)
|
u
z
|
(present FMM)
|
u
z
|
|
u
z
|
valley
P
R = 0.7:
Sanchez-Sesma, 1983 (semi-analytical)
Delavaud, 2007 (spectral nite element method)

x, y
z
R
D = 5R
A
B C
s
0 2
x/R
0
1
2
3
4
5
6
d
i
s
p
l
a
c
e
m
e
n
t

m
o
d
u
l
u
s
|
u
y
|
(present FMM)
|
u
y
|
|
u
y
|
|
u
z
|
(present FMM)
|
u
z
|
|
u
z
|
valley
Results for a higher frequency k
P
R = 1
N = 84 882 (76 iterations, single-proc. 3 GHz PC)
x, y
z
R
D = 5R
A
B C
s
0 1 2
x/R
0
1
2
3
4
5
6
d
i
s
p
l
a
c
e
m
e
n
t

m
o
d
u
l
u
s
|
u
y
|
(present FMM)
|
u
z
|
(present FMM)
Other acceleration methods
Electrostatics
Laplace
Elastostatics
8. Preconditioning
Other acceleration methods Exponential representation of 1/r
8. Preconditioning
Partial Fourier transform along (

1
,
2
) coordinates (
1
,
2
transformed
variable):
G(x, ) + ( x) = 0 =

G
,33

2
G +
e
i.x
4
2
(
3
x
3
) = 0
Solve analytically for

G(
1
,
2
,
3
; x) (case (
3
>x
3
)):
G(
1
,
2
,
3
; x) =
1
8
2
exp
_
(
3
x
3
) + i
_
1
(
1
x
1
) +
2
(
2
x
2
)
_
Exponential representation of 1/r (case

3
>x
3
):
1
| x|
=
1
2
_
+
0
_
2
0
e
[ (
3
x
3
)i( cos (
1
x
1
)+sin (
2
x
2
) ) ]
dd
Similar (but distinct) formula available for the case

3
<x
3
.
Cheng, H., Greengard, L., Rokhlin, V. A fast adaptive multipole

algorithm in three dimensions. J. Comp. Phys., 155:468498 (1999).
1
| x|
=
1
2
_
+
0
_
2
0
e
[ (
3
x
3
)i( cos (
1
x
1
)+sin (
2
x
2
) ) ]
dd (
3
>x
3
)
Division of interaction list into 6 (3-D) or 4 (2-D) sublists:
U U U
U U U U U U
U U U U U U
D
D D D D D D
D D
E E W
C
Uplist (U)
Downlist (D)
Eastlist (E)
Westlist (W)
Northlist (N), for 3-D
Southlist (S), for 3-D

Exponential representation of 1/r : numerical quadrature
1
| x|

s
k=1
w
k
M
k
M
k
j =1
e
k
[ (
3
x
3
)i( cos
j ,k
(
1
x
1
)+sin
j ,k
(
2
x
2
) ) ]
<
Outer numerical quadrature: points

k
and weights w
k
given by Cheng,
Rokhlin, Yarvin (1999);
Inner numerical quadrature: M

k
equally-spaced angles
j ,k
=2j /M
k
, with
M
k
also given by Cheng, Rokhlin, Yarvin (1999).
Comparison with traditional multipole expansion:

M
1
+ . . . + M
s
= O(p
2
Cheng, H., Rokhlin, V., Yarvin, N. Nonlinear optimization, quadrature

and interpolation. SIAM J. Optim., 9:901923 (1999).
Exponential representation of 1/r and diagonal translations
1
| x|

s
k=1
w
k
M
k
M
k
j =1
e
k
[ (
3
x
3
)i( cos
j ,k
(
1
x
1
)+sin
j ,k
(
2
x
2
) ) ]
<
Insert poles x, , dene multipole moments and translation operations.
Multipole moments:
M(k, j ) =
_
C
k
[ (
3
3,0
)i( cos
j ,k
(
1
10,
)+sin
j ,k
(
2
2,0
) ) ]
() dS
M2M, M2L, L2L translations are diagonal, e.g.:

L(k, j ) = M(k, j ) e
k
[ (
3,0
x
3,0
)i( cos
j ,k
(
1,0
x
1,0
)+sin
j ,k
(
2,0
x
2,0
) ) ]
Speeds up M2M, M2L, L2L operations
(O(p
2
N) instead of O(p
4
N) using traditional FMM)
Summation w.r.t. (k, i ) performed on local expansions, at the very end (after
upward, M2L and downward phases)
Exponential expansions available for other kernels, e.g. Helmholtz

Useful for FMM for low-frequency wave problems
Other acceleration methods FMM using equivalent sources
8. Preconditioning
Main idea: express elds at remote points in terms of equivalent density, e.g.:
_
C
G
,n
(x, )u() dS
=
_
S
d
G(x, z)(z) dS
z
for some (x ,/((
))
x
0
0
S
d
Ying, L., Biros, G., Zorin, D. A kernel-independent adaptive fast

multipole in two and three dimensions. J. Comp. Phys., 196:591626 (2004).
Main idea: express elds at remote points in terms of equivalent density, e.g.:
_
C
G
,n
(x, )u() dS
=
_
S
d
G(x, z)(z) dS
z
for some (x S
c
)
x
0
0
S
d
c
S
Solve the above (Fredholm, 1st kind) integral equation for
Truncation parameter = discretization of

FMM using equivalent sources: M2M translations
S
(l)
d
S
(l1)
d
S
(l1)
c
Find
(1)
,
_
S
(1)
d
G(x, z)
(1)
(z) dS
z
=
_
S
()
d
G(x, z)
()
(z) dS
z
(x S
(1)
c
)
FMM using equivalent sources: L2L translations
(l)
d
S
S
(l1)
d
(l)
S
c
Find
()
,
_
S
()
d
G(x, z)
()
(z) dS
z
=
_
S
(1)
d
G(x, z)
(1)
(z) dS
z
(x S
()
c
)
Kernel-independent acceleration method;
Truncation parameter p = discretization of (scale-independent for kernels

associated with elliptic problems);
Found by Ying, Biros, Zorin (2004) to have O(p

2
N) complexity / iteration;
Requires solving 1st kind integral equations (ill-conditioned integral operator)

Other acceleration methods Clustering and low-rank approximations
8. Preconditioning
DOF clustering
Spatially local DOFs (e.g. nodal values on a BEM mesh);
Recursive subdivision of set of DOFs into subsets, e.g. by bisection

Example (1-D BE mesh, e.g for a 2-D crack problem):
15 14 13 9 7 3 12 4 6 8 12 5
{7,8,9,10,11} {5,6} {1,2,3,4}
{1,...,6} {7,...,15}
{1,...,15}=I
{12,13,14,15}
{1,2} {3,4} {5,6} {} {7,8,9} {12,13} {14,15} {10,11}
1011
Block clustering of inuence matrix
Let
1
,
2
denote the geometrical support of two clusters (i.e. of two subsets
of the DOF index list /). For instances,
1
,
2
are cubic cells.
1
,
2
are admissible if
Min
_
Diam(
1
), Diam(
2
)
_
Dist(
1
,
2
)
2
Dist( , )
2
Diam( )
1
Diam( )
Block clustering of (BEM) inuence matrix into n blocks:

(I
1
J
1
) . . . (I
n
J
n
) = //, I
k
, J
k
generated by DOF clustering of /
Such block clustering is not unique for a given index set /.
Hierarchical, recursive block clustering of (BEM) inuence matrix into n

blocks:
Block // is not admissible.
If block (I J) is admissible, do nothing
Else, create sub-blocks of (I J) using children sublists of (I , J)
Block clustering of inuence matrix: example
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
15 14 13 9 7 3 12 4 6 8 12 5
{7,8,9,10,11} {5,6} {1,2,3,4}
{1,...,6} {7,...,15}
{1,...,15}=I
{12,13,14,15}
{1,2} {3,4} {5,6} {} {7,8,9} {12,13} {14,15} {10,11}
1011
Concept of panel clustering in the BEM (Hackbusch and Nowak, 1989)
Concept of 1-matrices (Hackbusch, 1999)
Acceleration via low-rank approximation of blocks
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Idea: use low-rank approximations of blocks:
A(I , J)
r
k=1
a
k
b
T
k
Ideally, r [I [, [J[
Matrix-vector product:
A(I , J)u(I )
r
k=1
_
b
T
k
u(I )
_
u(I )
Block clustering + low-rank approximation of blocks = acceleration of matrix
operations
SVD and block rank
Any block A(I , J) of size mn admits a singular value decomposition

(SVD):
A(I , J) = USV
T
=
R
k=1
s
k
u
k
v
T
k
s
1
s
2
. . . s
R
>0, u
T
k
u
= v
T
k
v
=
k
R Min(m, n) is the (numerical) rank of A(I , J)
A(I , J) has (approximate) low rank r if s

k
is suciently small for k >r
Computing complete SVD of A(I , J) needs O(mn) memory + O(mn

2
) CPU
not acceptable; other strategies required
FMM (analytic decomposition of kernel required)
Kernel interpolation (analytic decomposition of kernel not required)
Algebraic treatment of matrix blocks: adaptive cross approximation
Wavelet transformation of basis functions (not addressed here)
FMM as block clustering with low-rank approximation of blocks
The multi-level Fast Multipole Method features block clustering (through
hierarchical octree of cubic cells)
and low-rank approximation through truncated multipole expansion
1
| x|
=
p
n=0
n
m=n
R
n,m
(xx
0
)
p
=0
n
=n
(1)
n
S
n+n
,m+m
(
0
x
0
)R
n
,m
(
0
)
Other acceleration methods Kernel-independent acceleration via kernel interpolation
8. Preconditioning
In many cases:
Fundamental solution available (not necessary ion closed form);

Availability of high-order derivatives problematic at best
Taylor-based expansion impractical
Analytic expansion (e.g. multipole, exponential) not available

FMM treatment impossible
Idea: polynomial interpolation of G(x, ) in product of two non-adjacent cells
(x(
x
, (
)
x
0
0
G(x, )
P
p=1
Q
q=1
P
p
(x)G(x
p
,
q
)Q
q
()
x
p
: interpolation nodes in (
x
(Cartesian product of 1-D set of nodes);
P
p
(x): interpolation polynomials (e.g. Cartesian product of 1-D Lagrange polyn.);
q
: interpolation nodes in (
(Cartesian product of 1-D set of nodes);

Q
q
(x): interpolation polynomials (e.g. Cartesian product of 1-D Lagrange polyn.);
x
p
C
C
x
q
Evaluation of
o[, ]
far
(x) =
A(C
x
)
_
C
G(x, )() dS
Multipole moments:
M
q
(
0
) =
_
C
Q
q
()() dS
M2L translation:
L
p
(x
0
) =
Q
q=1
G(x
p
,
q
)M
q
(
0
)
M2M (upward) translations by expressing the Q

(1)
q
(;
(1)
0
) in terms of
the Q
()
q
(;
()
0
) (e.g. Taylor expansion for polynomials)
L2L (downward) translations similarly

Other acceleration methods Adaptive cross approximation
8. Preconditioning
Aim: nd recursively a low-rank approximation of a mn block A of the form:
A = R + S, S =
r
k=1
u
k
v
T
k
, |R|
F
To be used in conjunction with a 1-matrix block clustering.
Bebendorf, M., Rjasanov, S.. Adaptive low-rank approximation of

collocation matrices. Computing, 70:124 (2003).
Adaptive cross approximation: partially-pivoted ACA
1. Initialization: S = 0, / = , r =0
2. Recursion:
(a) k = Min{j , j K}, K = K{k}; STOP if |K| =n,
(b) Row generation: a = A
k
,
(c) Row of residual, pivot column: R
k
= a
r
i =1
v
i
(u
i
)
k
, =Argmax|R
k
|,
(d) Test: if Max|R
k
| = 0, go to 2(a),
(e) Column generation: a = A
,
(f) Column of residual, pivot row: R
= a
r
i =1
u
i
(v
i
)
, k =Argmax|R
k
|,
(g) New vectors: u
m+1
= (R
k
)
1
R
, v
m+1
= R
k
,
(h) Stopping criterion: u
r +1
F
v
r +1
F
S
F
,
(i) New approximation of block: S = S + u
r +1
v
T
r +1
,
(j) Recursion: r = r +1, go to 2(b)
Algorithm requires O(r

2
(m+n)) operations,
Complete ACA requires O(N

1+
) operations for any >0 if kernel

asymptotically smooth
[Kurz, Rain, Rjasanov, 2006]
Example: crack propagation analysis
N 45000,
CPU 4500s
RAM= 1.5GB
Kolk, K., Weber, W., Kuhn, G. Investigation of 3D crack propagation

problems via fast BEM formulations. Comp. Mech., 37:3240 (2005).
Preconditioning
Electrostatics
Laplace
Elastostatics
8. Preconditioning
Preconditioning
Diagonal preconditioneur M
ij
= A
ij
ij
;
Sparse matrix of near contributions in FMM;
Incomplete LU factorization of A;
Sparse approximate inverses;
Multigrid approaches;
Fast BEM solution method (e.g. FMM, ACA) with low truncation;
Preconditioners exploiting specic features of the problems, e.g.
single-inclusion case for many-inclusion problems.
Preconditioning
Sparse approximate inverse (SPAI)
Denition:
[A
] = arg min
[E]
| [I ] [E][A] |
2
F
[E] R
NN
sparse
where [A
] has a sparsity pattern (either predened or found iteratively).
Hence each column of [A
] solves an uncoupled, small minimization problem:

A
k
= arg min
{E}
| e
k
E[A] | E R
1N
sparse
Simplication: choose number m of nonzero entries in each row of [A
] and
nd SPAI of [
A]:
i
= arg min
{
E}R
1,m
_
| E[
A
i
] |
2
2 trace(E[
A
i
]) + 1
_
(1i N)
where [
A] is the sparse matrix made of the m largest entries of [A].

Preconditioning
Further reading
Benzi, M.
Preconditioning techniques for large linear systems: a survey.
J. Comp. Phys., 182:418477 (2002).
Biros, G., Ying, L., Zorin, D.

A fast solver for the Stokes equations with distribuuted forces in complex
geometries.
J. Comp. Phys., 193:317348 (2003).
Bonnet, M.
Boundary Integral Equation Method for Solids and Fluids.
Wiley (1999).
orm, S., Grasedyck, L., Hackbusch, W.

Introduction to hierarchical matrices with applications.
Engng. Anal. with Bound. Elem., 27:405422 (2003).
Carpentieri, B., Duff, I. S., Giraud, L., Sylvand, G.

Combining fast mulrtipole techniques and an approximate inverse
precondtioner for large electromagnetism calculations.
SIAM J. Sci. Comput., 27:7747923 (2005).
Preconditioning
Chaillat, S., Bonnet, M., Semblat, J. F.

A multi-level fast multipole BEM for 3-D elastodynamics in the frequency
domain.
Comp. Meth. Appl. Mech. Engng. (2008, in press).
Cheng, H., Crutchfield, W. Y., Gimbutas, Z., Greengard, L. F.,

Ethridge, J. F., Huang, J., Rokhlin, V., Yarvin, N.
A wideband fast multipole method for the Helmholtz equation in three
dimensions.
J. Comp. Phys., 216:300325 (2006).
Chew, W. C. et al.
Fast integral equation solvers in computational electromagnetics of complex
structures.
Coifman, R., Rokhlin, V., Wandzura, S.

The Fast Multipole Method for the Wave Equation : A Pedestrian Prescription.
IEEE Antennas and Propagation Magazine, 35:712 (1993).
Preconditioning
Colton, D., Kress, R.

Integral Equation Method in Scattering Theory.
John Wiley and sons (1983).
Darve, E.
The fast multipole method: I. Error analysis and asymptotic complexity.
SIAM J. Numer. Anal., 38:98128 (2000).
Darve, E.
The fast multipole method: numerical implementation.
J. Comp. Phys., 160:195240 (2000).
Darve, E., Hav
e, P.
Ecient fast multipole method for low-frequency scattering.
J. Comp. Phys., 197:341363 (2004).
Dominguez, J.
Boundary elements in dynamics.
Comp. Mech. Publ., Southampton (1993).
Preconditioning
Epton, M.A., Dembart, B.

Multipole translation theory for the three-dimensional Laplace and Helmholtz
equations.
SIAM J. Sci. Comp., 16:865897 (1995).
Ergin, A. A., Shanker, B., Michielssen, E.

Fast evaluation of three-dimensional transient wave elds using diagonal
translation operators.
J. Comp. Phys., 146:157180 (1998).
Fischer, M., Gaul, L.

Application of the fast multipole bem for structural-acoustic simulations.
Journal of Computational Acoustics, 13:8798 (2005).
Fochesato, C., Dias, F.

A fast method for nonlinear three-dimensional free-surface waves.
Proc. Roy. Soc. A, 462:27152735 (2006).
Frayss
e, V., Giraud, L., Gratton, S., Langou, J.

A Set of GMRES Routines for Real and Complex Arithmetics on High
Performance Computers.
Tech. rep., CERFACS (2003).
Preconditioning
Fu, Y., Klimkowski, K. J., Rodin, G. J., Berger, E., Browne,

J. C., Singer, J. R., van der Greijn, R. A., Vemaganti, K. S.
A fast solution method for the three-dimensional many-particle problems of
linear elasticity.
Int. J. Num. Meth. in Eng., 42:12151229 (1998).
Fujiwara, H.
The fast multipole method for the integral equations of seismic scattering
problems.
Geophys. J. Int., 133:773782 (1998).
Fujiwara, H.
The fast multipole method for solving integral equations of three-dimensional
topography and basin problems.
Geophys. J. Int., 140:198210 (2000).
Greenbaum, A.
Iterative methods for solving linear systems.
SIAM, Philadelphia, USA (1997).
Preconditioning
Greengard, L., Huang, J., Rokhlin, V., Wandzura, S.

Accelerating fast multipole methods for the Helmholtz equation at low
frequencies.
IEEE Comp. Sci. Eng., 5:3238 (1998).
Greengard, L., Rokhlin, V.

A fast algorithm for particle simulations.
J. Comp. Phys., 73:325348 (1987).
Greengard, L., Rokhlin, V.

A new version of the fast multipole method for the Laplace equation in three
dimensions.
Acta Numerica, 6:229270 (1997).
Gumerov, N. A., Duraiswami, R.

Fast multipole methods for the Helmholtz equation in three dimensions.
Elsevier (2005).
Hackbusch, W., Nowak, Z. P.

On the fast matrix multiplication in the boundary element method by panel
clustering.
Numerische Mathematik, 54:463491 (1989).
Preconditioning
Hackbusch, W., Wittum, G. (eds.).

Boundary elements: implementation and analysis of advanced algorithms.
Wieweg, Braunschweig (1996).
Jiang, L. J., Chew, W. C.

A mixed-form fast multipole algorithm.
IEEE Trans. Antennas Propagat., 53:41454156 (2005).
Kellogg, O.D.
Fundations of potential theory (1929).
Kolk, K., Weber, W., Kuhn, G.

Investigation of 3D crack propagation problems via fast BEM formulations.
Comp. Mech., 37:3240 (2005).
Kupradze, V.D.
Dynamical problems in elasticity.
North-Holland, p. 259 (1963).
Kurz, S., Rain, ; O., Rjasanow, S.

The adaptive cross approximation technique for the 3D boundary element
method.
IEEE Transactions on Magnetics, 38:421424 (2002).
Preconditioning
Lage, C., Schwab, C.

Wavelet Galerkin algorithms for boundary integral equations.
SIAM J. Sci. Comput., 20:21952222 (1999).
Liu, Y., Nishimura, N.

The fast multipole boundary element method for potential problems: a tutorial.
Liu, Y., Nishimura, N., Otani, Y., Takahashi, T., Chen, X. L.,
Munakata, H.
A fast boundary element method for the analysis of ber-reinforced composites
based on a rigid-inclusion model.
ASME J. Appl. Mech., 72:115128 (2005).
Margonari, M., Bonnet, M.

Fast multipole method applied to the coupling of elastostatic BEM with FEM.
Computers and Structures, 83:700717 (2005).
Martinsson, P. G., Rokhlin, V.

A fast direct solver for boundarry integral equations in two dimensions.
J. Comp. Phys., 205:123 (2005).
Preconditioning
ed
elec, J. C.
Acoustic and electromagnetic equations.
Applied mathematical sciences (vol. 144). Springer Verlag (2001).
Nemitz, N., Bonnet, M.

Topological sensitivity and FMM-accelerated BEM applied to 3D acoustic
inverse scattering.
Engng. Anal. with Bound. Elem. (2008).
Nishimura, N.
Fast multipole accelerated boundary integral equation methods.
Appl. Mech. Rev., 55:299324 (2002).
Rokhlin, V.
Rapid solution of integral equations of classical potential theory.
J. Comp. Phys., 60:187207 (1985).
Rokhlin, V.
Rapid solution of integral equations of scattering theory.
J. Comp. Phys., 86:414439 (1990).
Preconditioning
Rokhlin, V.
Diagonal forms of translation operators for the Helmholtz equation in three
dimensions.
Appl. Comp. Harmonic Anal., 1:8293 (1993).
Rokhlin, V.
Sparse diagonal forms for translation operators for the Helmholtz equation in
two dimensions.
Appl. Comp. Harmonic Anal., 5:3667 (1998).
Saad, Y.
Iterative methods for sparse linear systems.
SIAM (2003).
Saad, Y., Schultz, M.H.

GMRES: a generalized minimal residual algorithm for solving nonsymmetric
linear systems.
SIAM J. Sci. Stat. Comput., 7:856869 (1986).
Preconditioning
Sakuma, T., Yasuda, Y..

Fast multipole boundary element method for large-scale steady-state sound
eld analysis. Part I: setup and validation.
Acta Acustica united with Acustica, 88:513525 (2002).
Schmindlin, G., Lage, C., Schwab, C.

Rapid solution of rst kind boundary integral equations in R
3
.
Song, J., Lu, C.C., Chew, W.C.

Multilevel fast multipole algorithm for electromagnetic scattering by large
complex objects.
IEEE Trans. Antennas Propag., 42:14881493 (1997).
Sylvand, G.
La methode multipole rapide en electromagnetisme : performances,
parallelisation, applications.
Ph.D. thesis, ENPC (2002).
Preconditioning
Takahashi, T., Kawai, A., Ebisuzaki, T.

Accelerating boundary-integral equation method using a special-purpose
computer.
Takahashi, T., Nishimura, N., Kobayashi, S.

A fast BIEM for three-dimensional elastodynamics in time domain.
Tausch, J.
Sparse BEM for potential theory and Stokes ow using variable order wavelets.
Comp. Mech., 32:3123185 (2003).
Wang, P. B., Yao, Z. H..

Fast multipole DBEM analysis of fatigue crack growth.
Comp. Mech., 38:223233 (2006).
Yarvin, N., Rokhlin, V.

Generalized Gaussian quadratures and singular value decompositions of integral
operators.
SIAM J. Sci. Comput., 20:699718 (1997).
Preconditioning
Ying, L., Biros, G., Zorin, D.

A kernel-independent adaptive fast multipole in two and three dimensions.
J. Comp. Phys., 196:591626 (2004).
Yoshida, K.I.
Applications of fast multipole method to boundary integral equation method.
Ph.D. thesis, University of Kyoto (2001).
Yoshida, S., Nishimura, N., Kobayashi, S.

Application of fast multipole Galerkin boundary integral equation method to
elastostatic crack problems in 3D.

Methode D Elements Frontieres

Загружено:

Сведения о документе

Исходное описание:

Оригинальное название

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Methode D Elements Frontieres

Загружено:

Авторское право:

Доступные форматы

Methode delements de fronti`ere

Ecole doctorale MODES Methodes numeriques avancees, 28 mars 2013

Well-known, and simple, physical setting

Allows to introduce important concepts of integral equation formulations

Said concepts will generalize to other settings (elasticity, electromagnetics...)

Coulomb interaction force:

Continuous charge distribution: dq = dV :

Poisson equation (Gauss theorem with E = V):

where () is the solid angle of the (closed) surface V from origin :

As already seen, 1[, V] satises the Poisson equation:

1[, V] is continuous everywhere, and in particular across the boundary V.

o[, S] is harmonic outside of S:

o[, S] is continuous everywhere, and in particular across S.

, (ii) opposite charge

Performing the limit h 0, one nds:

Marc Bonnet (POems, ENSTA) Methode delements de fronti`ere 11 / 183

Properties of electrostatic double-layer potentials

T[q, S] is harmonic outside of S:

T[q, S] is discontinuous across S, with

Electrostatics (u: electrostatic potential);

Potential uid ow (u: velocity potential);

Thermal equilibrium (u: temperature);

Torsion (u: warping function dened on 2-D section of shaft)

The single-layer potential o solves the Laplace equation in R

The volume potential 1 solves the Poisson equation, where

Continuity across (with [[f ]] := f

The following potential solve the Laplace equation in R

Boundary traces of double-layer potential:

) = (c(x) 1)(x) +T[, ](x),

Fundamental solution (again):

Greens function: fundamental solution with homogeneous BCs on :

Fundamental solution boundary unknowns u (on S

Greens function explicit integral representation formula (boundary

Seek solutions of Laplace equation using potentials:

Seek solutions of Poisson equation using potentials:

Constitutive equation (C: fourth-order tensor of elastic moduli):

For isotropic elasticity (: shear modulus, : Poisson ratio):

Governing eld equation for unknown displacement eld u():

Partially unknown contribution of (BC + unknown trace)

Known contribution of (if

Limit to the boundary approach

Direct approach using exclusion neighbourhood of z

Indirect regularization approach

Integral equation, regularized form:

Both forms require u C

Numerical implementation based on (well-documented) singular element

Boundary-only formulations in the absence of body forces and initial strains.

Treatments (double / multiple reciprocity methods) sometimes allow to

Volume potential (prescribed body forces):

(displacement eld created in R

Volume potential (prescribed initial strains):

(displacement eld created in R

Single-layer elastic potential:

Double-layer elastic potential:

( = 3 4, = 4(1 )(1 2))

Elastic ellipsoidal inhomogeneity (Eshelby problem), using

Half-space with free surface (Mindlin, 1936), Boussinesq as special case:

Two perfectly-bonded half spaces (Rongved, 1955) closed form

Fundamental solution (full space):

Dynamic (Helmholtz) and static (Laplace) fundamental solutions have same

Singular integral equation:

Singular integrals: invoke methods for handling static singular kernels: