18.336j/6.335j: fast methods for partial differential and...
TRANSCRIPT
18.336J/6.335J: Fast methods for partial differential and integral equations
Cambridge, September 21, 2017
MITMathematics Department
Instructor: Carlos Pérez Arancibia
Lecture 5
Single and double layer potentials
Single-layer potential:
Double-layer potential:
where
Sk : C(�) ! C2(Rd \ �)
Dk : C(�) ! C2(Rd \ �)
(x 2 Rd \ �)
(x 2 Rd \ �)
(� 2 C2)
(� 2 C2)
(Sk ) (x) :=
Z
�Ek(x� y) (y) dsy
(Dk ) (x) :=
Z
�
@Ek
@ny(x� y) (y) dsy
Ek(x� y) =
8>>>>>>><
>>>>>>>:
e
ik|x�y|
4⇡|x� y| if k � 0, d = 3,
� 1
2⇡log |x� y| if k = 0, d = 2,
i
4
H(1)0 (k|x� y|) if k > 0, d = 2.
where
Boundary integral operators
‣Notation:
andu|±(x) = lim✏>0,✏!0
u(x± ✏n(x))@u
@n
���±(x) = lim
✏>0,✏!0ru(x± ✏n(x)) · n(x)
where x 2 �
Dk[ ]��± =±
2+Dk[ ], Sk[ ]
��± = Sk[ ],
@Dk[ ]
@n
���±= Nk[ ],
@Sk[ ]
@n
���±=⌥
2+Kk[ ]
‣ Properties. For a densify function we have: 2 C(�)
(Sk )(x) :=
Z
�Ek(x� y) (y) dsy (Dk )(x) :=
Z
�
@Ek
@ny(x� y) (y) dsy
(Kk )(x) :=
Z
�
@Ek
@nx
(x� y) (y) dsy
(Nk )(x) :=
Z
�
@2Ek
@nx
@ny
(x� y) (y) dsy
Maue’s Integration by parts formula
The hypersingular operator has to be interpreted as a Hadamardfinite part integral that can be expressed in terms of a Cauchyprincipal value integral:
(Nk )(x) = k2Z
�Ek(x� y)n(x) · n(y) (y) ds
y
+p.v.
Z
�rx
sEk(x� y) ·ry
s (y) dsy
Calderón Identities
Assume � is of class C2. Then the single-layer (Sk), double-layer
(Dk), adjoint double-layer (Kk) and hypersingular (Nk) operators
satisfy:
DkSk = SkKk, NkDk = KkNk,
D2k � SkNk =
I
4
, K2k �NkSk =
I
4
.
Calderón Identities
Proof. Let ', 2 C1(�) and consider the function ue = �Dk'+
Sk . Evaluating ue|+ and @ue/@n|+ on � we obtain
ue|+ = �✓I
2
+Dk
◆'+Sk and
@ue
@n
��+= �Nk'+
✓�I
2
+Kk
◆ .
(1)
Since �ue + k2ue = 0 in Rd \ � and ue satisfies the radiation
condition, we have, by Green’s formula, that
ue|+ =
✓I
2
+Dk
◆ue|+ � Sk
@ue
@n
��+
and (2)
@ue
@n
��+= Nkue|+ �
✓�I
2
+Kk
◆@ue
@n
��+. (3)
Replacing (1) in (2) and (3), all four Calderon identities follow.
Weakly singular operators
Definition. A kernel K is said to be weakly singular if K is defined and
continuous for all x and y in �, x 6= y, and there exist constants ↵ 2 (0, 2] andM > 0 such that for all x,y 2 �, x 6= y, we have
|K(x,y)| M |x� y|↵�d+1
Theorem. Let � 2 C2, then there exists L > 0 such that
|n(y) · (x� y)| L|x� y|2 and
|n(x)� n(y)| L|x� y| 8x,y 2 �.
Consider the integral operator A : C(�) ! C(�), � 2 C2, defined
by
(A )(x) =
Z
�K(x,y) (y) dsy, x 2 �.
Nate that the single-layer (Sk), double-layer (Dk) and adjoint double-
layer (Kk) operators are weakly singular for a surface � 2 C2.
Compact operators
Theorem. The operators Sk : C(�) ! C(�), Dk : C(�) ! C(�), Kk : C(�) !C(�) and Nk �N0 : C(�) ! C(�) are compact.
Theorem. The space of complex-valued continuous functions C(�), � 2 C2,
equipped with the norm k�k = max
x2� |�(x)|, is a Banach space (i.e, a complete
normed vector space).
Definition. A linear operator A : X ! Y (X and Y are Banach spaces) is
compact if for each bounded sequence {�n} ⇢ X the sequence {A�n} contains
a convergent subsequence.
Note. The eigenvalues � of A (A compact) can only accumulate at 0.
Theorem. An integral operator A with continuous or weakly singular kernel is
a compact operator on C(�).
Spectral properties�(Sk) �(Dk)
�(Kk) �(Nk)
Interior problems
Interior Neumann. Find u : ⌦ ! C, u 2 C2(⌦) \ C1
(⌦), � = @⌦ 2 C2, such
that:
(IN)
8<
:
�u+ k2u = 0 in ⌦,
@u
@n= g on �.
Interior Dirichlet. Find u : ⌦ ! C, u 2 C2(⌦) \ C1
(⌦), � = @⌦ 2 C2, such
that:
(ID)
(�u+ k2u = 0 in ⌦,
u = f on �.
Theorem. The interior Dirichlet problem has at most one solution
for all k > 0 such that k 6=p�D where �D is a Laplace Dirichet
eigenvalue, while the interior Neumann problem has at most one
solution for all k > 0 such that k 6=p�N , where �N is a Laplace
Neumann eigenvalue.
Exterior problems
Exterior Dirichlet. Find u : Rd \ ⌦ ! C, u 2 C2(Rd \ ⌦) \ C1
(Rd \ ⌦),
� = @⌦ 2 C2, such that:
(ED)
8>>>><
>>>>:
�u+ k2u = 0 in Rd \ ⌦,
u = f on �,
lim
|x|!1|x|(d�1)/2
⇢@u
@|x| � iku
�= 0
Exterior Neumann. Find u : Rd \ ⌦ ! C, u 2 C2(Rd \ ⌦) \ C1
(Rd \ ⌦),
� = @⌦ 2 C2, such that:
(EN)
8>>>>><
>>>>>:
�u+ k2u = 0 in Rd \ ⌦,@u
@n= g on �,
lim
|x|!1|x|(d�1)/2
⇢@u
@|x| � iku
�= 0
Theorem. The exterior Dirichlet and Neumann problems have at
most one solution.
Green formula:
Interior Dirichlet problem
• Direct Integral equation formulations
second-kind integral equation for the unknown function
first-kind integral equation for the unknown function
Integral equation formulations: Interior problems
Interior Neumann problem:
=@u
@n
����
= u���
Sk[ ] =f
2
+Dk[f ] on �
2
+Dk[ ] = Sk[g] on �
(Dku) (x)�✓Sk
@u
@n
◆(x) =
⇢�u(x), x 2 ⌦,0, x 2 Rd \ ⌦.
Interior Dirichlet problem
• Direct Integral equation formulations
“first”-kind integral equation for the unknown function
second-kind integral equation for the unknown function
Integral equation formulations: Interior problems
Interior Neumann problem:
=@u
@n
����
= u���
Taking normal derivative of the Green’s representation formula we
obtain:
(Nku)(x)�✓Kk
@u
@n
◆(x) = �1
2
@u(x)
@n, x 2 �
� 2
+Kk[ ] = Nk[f ] on �
Nk[ ] = �g
2
+Kk[g] on �
Green formula:
Exterior Dirichlet problem:
• Direct integral equation formulations
second-kind integral equation for the unknown function
first-kind integral equation for the unknown function
Integral equation formulations: Exterior problems
Exterior Neumann problem:
= u��+
=@u
@n
���+
Sk[ ] = �f
2
+Dk[f ] on �
� 2
+Dk[ ] = Sk[g] on �
(Dku) (x)�✓Sk
@u
@n
◆(x) =
⇢0, x 2 ⌦,
u(x), x 2 Rd \ ⌦.
Exterior Dirichlet problem:
• Direct integral equation formulations
“first”-kind integral equation for the unknown function
second-kind integral equation for the unknown function
Integral equation formulations: Exterior problems
Exterior Neumann problem:
= u��+
=@u
@n
���+
Taking normal derivative of the Green’s representation formula we
obtain:
(Nku)(x)�✓Kk
@u
@n
◆(x) =
1
2
@u(x)
@n, x 2 �
2
+Kk[ ] = Nk[f ] on �
Nk[ ] =g
2
+Kk[g] on �
We look for a solution of the form:
and imposing the corresponding boundary condition we obtain:
• Indirect integral equation formulation
Integral equation formulations: Interior problems
Interior Dirichlet problem
Interior Neumann problem:second-kind integral equation for the unknown function
first-kind integral equation for the unknown functionSk[ ] = f on �
2
+Kk[ ] = g on �
u(x) = (Sk ) (x), x 2 Rd \ �
We look for a solution of the form:
and imposing the corresponding boundary condition we obtain:
• Indirect integral equation formulation
Integral equation formulations: Interior problems
Interior Dirichlet problem
Interior Neumann problem:“first”-kind integral equation for the unknown function
second-kind integral equation for the unknown function
u(x) = (Dk ) (x), x 2 Rd \ �
Nk[ ] = g on �
� 2
+Dk[ ] = f on �
We look for a solution of the form:
and imposing the corresponding boundary condition we obtain:
• Indirect integral equation formulation
Integral equation formulations: Exterior problems
Exterior Dirichlet problem
Exterior Neumann problem:second-kind integral equation for the unknown function
first-kind integral equation for the unknown functionSk[ ] = f on �
� 2
+Kk[ ] = g on �
u(x) = (Sk ) (x), x 2 Rd \ �
We look for a solution of the form:
and imposing the corresponding boundary condition we obtain:
• Indirect integral equation formulation
Integral equation formulations: Exterior problems
Exterior Dirichlet problem
Exterior Neumann problem:“first”-kind integral equation for the unknown function
second-kind integral equation for the unknown function
u(x) = (Dk ) (x), x 2 Rd \ �
Nk[ ] = g on �
2
+Dk[ ] = f on �
Calderón Preconditioners
Assume � is of class C2. Then the single-layer (Sk), double-layer
(Dk), adjoint double-layer (Kk) and hypersingular (Nk) operators
satisfy:
DkSk = SkKk, NkDk = KkNk,
D2k � SkNk =
I
4
, K2k �NkSk =
I
4
.
Consider the first kind integral equation Sk[ ] = f on �
Problem: Sk has eigenvalues clustered at 0.
second-kind integral equation
Solution: Precondition the integral equation from the left by Nk
and use the fact that NkSk = �I/4 +K2k :
NkSN [ ] =
✓�I
4
+K2k
◆[ ] = Nk[f ] on �
Calderón Preconditioners
�(Sk)
... GMRES convergence ...
krnk ✓1� �2
min
(1/2(A+AT))
�max
(ATA)
◆n/2
kr0
k
�
✓�I
4+K2
k
◆
Calderón Preconditioners
Assume � is of class C2. Then the single-layer (Sk), double-layer
(Dk), adjoint double-layer (Kk) and hypersingular (Nk) operators
satisfy:
DkSk = SkKk, NkDk = KkNk,
D2k � SkNk =
I
4
, K2k �NkSk =
I
4
.
Consider the “first” kind integral equation
second-kind integral equation
Nk[ ] = g on �
Problem: Nk has eigenvalues that tend to infinity.
Solution: Precondition the integral equation from the left by Sk
and use the fact that SkNk = �I/4 +D2k :
SNNk[ ] =
✓�I
4
+D2k
◆[ ] = Sk[g] on �
Calderón Preconditioners
... GMRES convergence ...
krnk ✓1� �2
min
(1/2(A+AT))
�max
(ATA)
◆n/2
kr0
k
�(Nk) �
✓�I
4+D2
k
◆
Uniqueness of solutionsThe first and second kind integral equations
Sk[ ] = f and
✓�I
2
+Kk
◆[ ] = g
have at most one solution for all k > 0 except when �k2 is an
eigenvalue of the interior Dirichlet problem for the Laplacian, i.e.,
when there exist uD 6= 0 and �D > 0 such that ��uD = �DuD in
⌦ and uD = 0 on � = @⌦.
The “first” and second kind integral equations
Nk[ ] = g and
✓I
2
+Dk
◆[ ] = f
have at most one solution for all k > 0 except when �k2 is an
eigenvalue of the interior Neumann problem for the Laplacian, i.e.,
when there exist uN 6= 0 and �N � 0 such that ��uN = �NuN in
⌦ and
@uN
@n= 0 on � = @⌦.
Discretization methods: Collocation Method
Say we want to solve a second-kind integral equation:
(x) +
Z
�K(x,y) (y) dsy = f(x) on �.
1. Approximate using certain basis functions {vn}Nn=1 (vn :
� ! C):
(x) ⇡ N (x) =
NX
n=1
cnvn(x).
2. Replace the approximation in the integral equation:
NX
n=1
cn
✓vn(x) +
Z
�K(x,y)vn(y) dsy
◆= f(x), x 2 �.
3. Evaluate the expressions on both sides of the identity above
at points {xm}Nm=1 on �:
NX
n=1
cn
✓vn(xm) +
Z
�K(xm,y)vn(y) dsy
◆= f(xm), m = 1, . . . , N.
4. Solve the linear system Ac = f where c = [c1, . . . , cN ]
T 2 CN,
f = [f(x1), . . . , f(xN )]
T 2 CNand
Am,n = vn(xm) +
Z
�K(xm, y)vn(y) dsy, n,m = 1, . . . , N.
Discretization methods: Collocation Method
Say we want to solve a second-kind integral equation:
(x) +
Z
�K(x,y) (y) dsy = f(x) on �.
1. Approximate using certain basis functions {vn}Nn=1 (vn :
� ! C):
(x) ⇡ N (x) =
NX
n=1
cnvn(x).
2. Replace the approximation in the integral equation:
NX
n=1
cn
✓vn(x) +
Z
�K(x,y)vn(y)
◆= f(x), x 2 �.
3. Evaluate the expressions on both sides of the identity above
at points {xm}Nm=1 on �:
NX
n=1
cn
✓vn(xm) +
Z
�K(xm,y)vn(y) dsy
◆= f(xm), m = 1, . . . , N.
4. Solve the linear system Ac = f where c = [c1, . . . , cN ]
T 2 CN,
f = [f(x1), . . . , f(xN )]
T 2 CNand
Am,n = vn(xm) +
Z
�K(xm, y)vn(y) dsy, n,m = 1, . . . , N.
Discretization methods
(y) ⇡NX
j=1
jvj(y)
• Boundary element method (Galerkin):i) Replace the curve by its discretization and expand the unknown function in terms of basis functions:
ii) Replace the approximate density in the integral operator
iii) Test against the same basis functions
� �h
y 2 �h
Z
�K(x,y) (y) dsy ⇡
NX
j=1
j
Z
�h
K(x,y)vj(y) dsy
NX
j=1
j
Z
�h
vi(x)
Z
�h
K(xi,y)vj(y) dsy dsx
=
Z
�h
f(x)vi(x) dsx
i = 1, . . . , Nfor
Discretization methods
• Nyström method:i) Parametrize and express the integral equation as
ii) Use a quadrature rule to approximate the integral operator:
iii) Evaluate the integral equation at the quadrature points:
�
i = 1, . . . , Nfor
(t) +
Z T
0K(t, ⌧) (⌧) d⌧ = g(t), t 2 [0, T )
Z T
0K(t, ⌧) (⌧) d⌧ ⇡
NX
j=1
K(t, tj) (tj)wj , t 2 [0, T )
i +NX
j=1
K(ti, tj) (tj)wj = g(ti)
In parametric form a second-kind integral equation is given by:
K(t, ⌧) = K1(t, ⌧) ln
✓4 sin2
t� ⌧
2
◆+K2(t, ⌧)where (K1 and K2 analytic)
Use the trapezoidal rule:
(t)�Z 2⇡
0K(t, ⌧) (⌧) d⌧ = g(t), 0 t 2⇡ ( is 2⇡-periodic)
Discretized integral equation:
It can be shown that for an analytic right-hand side it holds that:
tj =j⇡
n
Z 2⇡
0ln
✓4 sin2
t� ⌧
2
◆f(⌧) d⌧ ⇡
2n�1X
j=0
R(n)j (t)f(tj),
Z 2⇡
0f(⌧) d⌧ ⇡ ⇡
n
2n�1X
j=0
f(tj),
(n)(ti)�2n�1X
j=0
n
R(n)j (ti)K1(ti, tj) +
⇡
nK2(ti, tj)
o
(n)(tj) = g(ti), i = 0, . . . , 2n� 1
| (n)(t)� (t)| C e�n�, 0 t 2⇡, � > 0.
Martensen-Kussmaul Nyström method
tj =j⇡
n
Z 2⇡
0ln
✓4 sin2
t� ⌧
2
◆f(⌧) d⌧ ⇡
2n�1X
j=0
R(n)j (t)f(tj),
Nyström method
t 2 [0, 2⇡]
R(n)j (t) = �2⇡
n
n�1X
m=1
1
mcos(m(t� tj))�
⇡
n2cos(n(t� tj)), j = 0, . . . , 2n� 1
where
For the exterior Neumann problem, for example, we have:
K(t, ⌧) =ik
2{x0
2(⌧)[x1(⌧)� x1(t)]� x01(⌧)[x2(⌧)� x2(t)]}
H(1)1 (k|x(t)� x(⌧)|)|x(t)� x(⌧)|
K1(t, ⌧) =k
2⇡{x0
2(⌧)[x1(t)� x1(⌧)]� x01(⌧)[x2(t)� x2(⌧)]}
J1(k|x(t)� x(⌧)|)|x(t)� x(⌧)|
K2(t, ⌧) = K(t, ⌧)�K1(t, ⌧) ln
✓4 sin2
t� ⌧
2
◆
� = {(x1(t), x2(t)) 2 R2 : 0 t 2⇡}where
In detail