HAL Id: hal-00589252
https://hal.archives-ouvertes.fr/hal-00589252
Submitted on 28 Apr 2011
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
Optimized Parameters of Optimized Schwarz Waveform Relaxation Methods for The Heat Equation
Minh-Binh Tran
To cite this version:
Minh-Binh Tran. Optimized Parameters of Optimized Schwarz Waveform Relaxation Methods for
The Heat Equation. 2011. �hal-00589252�
Methods for The Heat Equation
INTERNAL REPORT
Minh-Binh TRAN
Laboratoire Analyse G´ eom´ etrie et Applications Institut Galil´ ee, Universit´ e Paris 13, France
Email: binh@math.univ-paris13.fr
April 28, 2011
Contents
1 Introduction 3
2 Optimized Schwarz Waveform Relaxation Methods For The
One Dimensional Heat Equation 4
2.1 Optimized Schwarz Waveform Relaxation Methods For The One Dimensional Heat Equation With Robin Transmission
Condition . . . . 4
2.1.1 Proof of the Theorems in the Overlapping Case . . . . 11
2.1.2 Proof of the Theorems in the Non-Overlapping Case . 20 2.2 Optimized Schwarz Waveform Relaxation Methods For One Dimensional Heat Equation With First Order Transmission Condition . . . 23
2.2.1 Proof of the Theorems in the Overlapping Case . . . . 28
2.2.2 Proof of the Theorems in the Nonoverlapping Case . . 40
2.3 Numerical Results . . . 43
2.3.1 Test 1 . . . 43
2.3.2 Test 2 . . . 45
2.3.3 Test 3 . . . 47
2.3.4 Test 4 . . . 49
2.3.5 Test 5 . . . 52
2.3.6 Test 6 . . . 57
2.3.7 Test 7 . . . 60
2.3.8 Test 8 . . . 63
2.4 Optimization of The Convergence Factor: A Theoretical At- tempt . . . 65
2.4.1 The results . . . 65
2.4.2 Proofs of the results . . . 70
1
Condition . . . 99
3.1.1 Proof of the Theorems in the Overlapping Case . . . . 107
3.1.2 Proof of the Theorems in the Nonoverlapping Case . . 121
3.2 Optimized Schwarz Waveform Relaxation Methods For The Two Dimensional Heat Equation With Ventcell Transmission Condition . . . 128
3.2.1 Proof of the Theorems in the Overlapping Case . . . . 136
3.2.2 Proof of the Theorems in the Nonoverlapping Case . . 151
3.3 Numerical Results . . . 163
3.3.1 Test 1 . . . 163
3.3.2 Test 2 . . . 165
3.3.3 Test 3 . . . 176
3.3.4 Test 4 . . . 181
3.3.5 Test 5 . . . 184
3.3.6 Test 6 . . . 187
3.3.7 Test 7 . . . 190
3.3.8 Test 8 . . . 192
4 Acknowledgements. 194
2
Chapter 1 Introduction
The Schwarz domain decomposition methods is a procedure to parallelize and solve partial differential equations numerically, where each iteration involves the solutions of the original equations on smaller subdomains. It was original proposed by H. A. Schwarz [7] in 1870 as a technique to prove the existence of a solution to the Laplace equation on a domain which is a combination of a rectangle and a circle. The idea was then used by P. L. Lions [4], [5], [6] as parallel algorithms in solving partial differential equations. Since then, many kind of domain decomposition methods have been developped, to im- prove the performance of the classical domain decomposition method. One of the main streams in this direction is to replace the Dirichlet transmission condition by Robin and Ventcell transmission conditions and then calculate the convergence rates. Using different transmissions condition gives different convergence rates and we need to optimize to get the best transmission con- ditions, the methods are then called the optimized Schwarz methods. In [1]
and [2], D. Bennequin, M. Gander and L. Halpern show that the problem of optimizing the convergence rates is in fact a new class of best approximation problems and suggest a new method to solve this class of problems. The au- thors consider the model problem of optimizing the convergence factors for advection-diffusion equations. In this report, we use their methods to check the results announced in [3] and then extend the results to optimized Robin and Ventcell transmission conditions for 2 dimensional heat equations.
3
Optimized Schwarz Waveform Relaxation Methods For The One Dimensional Heat
Equation
2.1 Optimized Schwarz Waveform Relaxation Methods For The One Dimensional Heat Equation With Robin Transmission Con- dition
In this section, we consider the optimized Schwarz waveform relaxation method. The algorithm is
(∂
t− ν∂
xx)u
k1= f in Ω
1× (0, T ), u
k1(x, 0) = u
0(x) in Ω
1,
(∂
x+
2νp)u
k1(L, .) = (∂
x+
2νp)u
k2−1(L, .) in (0, T ),
(2.1.1)
(∂
t− ν∂
xx)u
k2= f in Ω
2× (0, T ), u
k2(x, 0) = u
0(x) in Ω
2,
(∂
x−
2νp)u
k2(0, .) = (∂
x−
2νp)u
k1−1(0, .) in (0, T ).
4
Let h
Land h
0be given in oH
34(0, T ). Let (e
1, e
2) be the solution in H
3,32(Ω
1× (0, T )) × H
3,32(Ω
2× (0, T )) of the problem
(∂
t− ν∂
xx)e
1= 0 in Ω
1× (0, T ), e
1(x, 0) = 0 in Ω
1,
(∂
x+
2νp)e
1(L, .) = h
Lin (0, T ),
(2.1.2)
(∂
t− ν∂
xx)e
2= 0 in Ω
2× (0, T ), e
2(x, 0) = 0 in Ω
2,
(∂
x−
2νp)e
2(0, .) = h
0in (0, T ).
We have
Fe
1(x, ω) = 2ν
√ 4iων + p Fh
Lexp(
r iω
ν (x − L)), (2.1.3) and
Fe
2(x, ω) = − 2ν
√ 4iων + p Fh
0exp( − r iω
ν x). (2.1.4) We have also
F(g
D(h
L, h
0)) = F((∂
xe
2+ p
2ν e
2)(L, .), (∂
xe
1− p
2ν e
1)(0, .)). (2.1.5) We have
∂
xFe
2(L, ω) + p
2ν Fe
2(L, ω) = − 2ν 2 √
iων + p ( − r iω
ν ) exp( − r iω
ν L)Fh
0− 2ν 2 √
iων + p p
2ν exp( − r iω
ν L)Fh
0= 2 √
iων − p 2 √
iων + p exp( − r iω
ν L)Fh
0, (2.1.6) and
∂
xFe
1(0, ω) − p
2ν Fe
1(0, ω) = 2ν 2 √
iων + p r iω
ν exp( − r iω
ν L)Fh
L− 2ν 2 √
iων + p p
2ν exp( − r iω
ν L)Fh
0= 2 √
iων − p 2 √
iων + p exp( − r iω
ν L)Fh
L. (2.1.7)
5
Therefore
F(g
2D(h
L, h
0)) = ( 2 √
iων − p 2 √
iων + p )
2exp( − 2L r iω
ν )F(h
L, h
0). (2.1.9) Consequently
| F(g
2D(h
L, h
0)) | = | ( 2 √
iων − p 2 √
iων + p )
2exp( − 2L r iω
ν ) || F(h
L, h
0) | (2.1.10)
= exp( − L q
2|ω|ν
)
(√
2|ω|ν−p)2+2|ω|ν (√
2|ω|ν+p)2+2|ω|ν
| F(h
L, h
0) | . Thus for k ∈ N ,
| Fg
2kD(h
L, h
0) | (ω) = exp( − kL r 2 | ω |
ν )( ( p
2 | ω | ν − p)
2+ 2 | ω | ν ( p
2 | ω | ν + p)
2+ 2 | ω | ν )
k| F(h
L, h
0) | (ω).
Therefore
|| g
2kDh
L||
H34(R)=
Z
+∞−∞
(1 + | ω |
2)
34| Fg
2kDh
L(ω) |
2dω
=
Z
+∞−∞
(1 + | ω |
2)
34exp( − 2kL r 2 | ω |
ν ) ×
× ( ( p
2 | ω | ν − p)
2+ 2 | ω | ν ( p
2 | ω | ν + p)
2+ 2 | ω | ν )
2k| Fh
L(ω) |
2dω.
Using the Lebesgue dominated convergence theorem with the notice that (
(√
2|ω|ν−p)2+2|ω|ν (√
2|ω|ν+p)2+2|ω|ν
)
2k< 1, we can see that {|| g
2kDh
L||
H34(R)} converges to 0 when k tends to ∞ . Similarly, {|| g
2kDh
0||
H34(R)} converges to 0 when k tends to ∞ .
For k ∈ N ,
| Fg
2kD(h
L, h
0) | (ω) = exp( − kL r 2 | ω |
ν )( ( p
2 | ω | ν − p)
2+ 2 | ω | ν ( p
2 | ω | ν + p)
2+ 2 | ω | ν )
k| F(h
L, h
0) | (ω).
6
We define the convergence factor by ρ(ω; p, L) = exp( − L
r 2 ω
ν ) (
νp− p
2
ων)
2+ 2
ων(
νp+ p
2
ων)
2+ 2
ων. (2.1.11) Put
ων= ¯ ω,
νp= ¯ p and ¯ ρ(¯ ω; ¯ p, L) = ρ(ω; p, L), we need to consider the following min-max problem
min
p¯∈Rmax
¯
ω∈[¯ωmin,¯ωmax]
ρ(¯ ¯ ω; ¯ p, L), (2.1.12) where ¯ ω
min=
2T νπ, ¯ ω
max=
2∆tνπ. Since
exp( − L √
2¯ ω) ( | p ¯ | − √
2¯ ω)
2+ 2¯ ω ( | p ¯ | + √
2¯ ω)
2+ 2¯ ω ≤ exp( − L √
2¯ ω) ( −| p ¯ | − √
2¯ ω)
2+ 2¯ ω ( −| p ¯ | + √
2¯ ω)
2+ 2¯ ω , so the minimum is attained for ¯ p ≥ 0, we only need to consider problem (2.1.12) in the case ¯ p ≥ 0, or the following problem
min
p¯≥0max
¯
ω∈[¯ωmin,¯ωmax]
ρ(¯ ¯ ω; ¯ p, L). (2.1.13) We have the following theorems for the overlapping case (L > 0)
Theorem 2.1.1. We suppose that L is small and ω ¯
maxis large.
a) For L(¯ ω
max)
34small (which means L ∽ C (¯ ω
max)
−34−γ, γ > 0), problem (2.1.13) has a unique solution
¯
p
∗∽ 2 √
2(¯ ω
minω ¯
max)
14then
|| ρ(¯ ¯ ω, p ¯
∗, L) ||
∞∽ 1 − 2 √
2( ω ¯
min¯ ω
max)
14,
where the asymptotic expansion is based on the scale of (ω
max)
−1.
b) For L(¯ ω
max)
34large (which means L ∽ C(¯ ω
max)
−34+γ, γ > 0) and L <
ν q
10√ 2−12
¯
ωmin
, the problem (2.1.13) has a unique solution
¯
p
∗∽ (4¯ ω
min)
13(L)
−13, then
|| ρ(ω, ¯ p ¯
∗, L) ||
∞∽ 1 − 4( ω ¯
min2 )
16(L)
13, where the asymptotic expansion is based on the scale of L.
7
¯
p
∗∽ 2 √ π
(2T C
2ν
2)
14∆x
−14, then
|| ρ(¯ ¯ ω, p ¯
∗, L) ||
∞∽ 1 − 2 √ 2C
1 4
2
2T
14∆x
41+ O(∆x
12).
For L = C
1∆x, ∆t = C
2∆x
2, ∆x ≤ min
π22T(C1ν−1)4 4C23 12,
(10π(C√21−ν−14)2T1)2 12, then problem (2.1.13) has a unique solution
¯
p
∗∽ 4π 2T C
1ν
13∆x
−13, then
|| ρ(¯ ¯ ω, p ¯
∗, L) ||
∞∽ 1 − 4 πC
12ν
−14T
16∆x
13+ O(∆x
23).
Remark 2.1.1.
8
Figure 2.1.1
Figure 2.1.2
9
the two boundaries to get p ¯
∗(figure 2.2).
For the non-overlapping case, we have the following result
Theorem 2.1.3. Problem (2.1.13) has one and only one solution which is given by ω ¯ = ¯ ω
minand ω ¯ = ¯ ω
max; p ¯ = 2
√π(2T ν2)14
∆t
−14. min
p¯max
¯
ω
ρ ¯ ∽ 1 − ( 32C
1T )
14∆t
14.
10
2.1.1 Proof of the Theorems in the Overlapping Case
Putting
h
L(¯ p) = max
¯
ω∈[¯ωmin,¯ωmax]
ρ(¯ ¯ ω, p, L) = ¯ || ρ(¯ ω; ¯ p, L) ||
∞,
we recall that (¯ p
∗, h
L(¯ p
∗)) is a strict local minimum of h
L(¯ p) if and only if there exists ǫ positive such that for all ¯ p in (¯ p
∗− ǫ, p ¯
∗+ ǫ), we have h
L(¯ p) >
h
L(¯ p
∗).
In order to prove the theorems, we need the following lemma as in [1]
Lemma 2.1.1. If (¯ p
∗, h
L(¯ p
∗)) is a strictly local minimum of h
L(¯ p), then it is the global minimum of h
L(¯ p) and p ¯
∗is the unique solution of (2.1.13).
Proof of Lemma 2.1.1
We denote D (z
0, δ) = { z ∈ C , |
z+zz−z00| < δ } , and D
δL= { p | h
L(p) ≤ δ } . We first prove that D
Lδis a convex set. Let ¯ p
1and ¯ p
2be to elements of D
Lδ, we have that
|| exp( − L √ i ω) ¯
¯ p1
ν
− 2 √ i¯ ω
¯
p
1+ 2 √
i ω ¯ ||
∞≤ δ.
Thus ∀ ω ¯ ∈ [¯ ω
min, ω ¯
max],
| exp( − L √
i¯ ω) p ¯
1− 2 √ i¯ ω
¯ p
1+ 2 √
i ω ¯ | ≤ δ.
Hence
exp( − L r ω ¯
2 ) | p ¯
1− 2 √ i ω ¯
¯
p
1+ 2 √
i ω ¯ | ≤ δ.
Therefore
| p ¯
1− 2 √ i¯ ω
¯ p
1+ 2 √
i¯ ω | ≤ δ exp(L r ω ¯
2 ).
This means ¯ p
1∈ D (2 √
i¯ ω, δ exp(L q
¯ ω 2
)).
Similarly, we have also ¯ p
2∈ D (2 √
i ω, δ ¯ exp(L q
¯ ω
2
)).
According to Lemma 2.1 in [1], D (z
0, δ) is the interior of the circle with center at
1+δ1−δ22z
0and radius
|1−2δδ2|| z
0| and the exterior otherwise.
11
If δ exp(L
ω¯2)) ≥ 1, using Lemma 2.1 in [1], we can see that for ¯ p
1, ¯ p
2≥ 0, θ ∈ [0, 1], we have θ p ¯
1+ (1 − θ) ¯ p
2∈ D (2 √
iω, δ exp(L q
¯ ω
2
)). Thus for θ ∈ [0, 1], we have θ p ¯
1+ (1 − θ) ¯ p
2∈ D
δL.
Therefore D
δLis convex.
Suppose that (¯ p
∗, h
L(¯ p
∗)) is a strictly local minimum of h
L(¯ p), we prove that it is a global minimum of h
L(¯ p). Suppose the contrary that there ex- ists (¯ p
∗∗, h
L(¯ p
∗∗)) such that h
L(¯ p
∗) ≥ h
L(¯ p
∗∗). Then there exists a convex neighborhood U of ¯ p
∗, such that ∀ s ∈ U , s 6 = ¯ p
∗and h
L(s) > h
L(¯ p
∗). Since
¯
p
∗∗∈ D
hLL(¯p∗∗)⊂ D
LhL(¯p∗),we have that ∀ θ ∈ [0, 1], θ p ¯
∗+ (1 − θ)¯ p
∗∗∈ D
LhL(¯p∗). For θ small enough, we have that θ p ¯
∗∗+ (1 − θ)¯ p
∗∈ U . This is a contradic- tion.
Thus ¯ p
∗is the unique solution of (2.1.13).
12
Proof of theorem 2.1.1
Case 1: For L(¯ ω
max)
34large and L < q
10√ 2−12
¯ ωmin
.
Firstly, we will prove that || ρ(¯ ¯ ω, p, L) ¯ ||
∞= max { ρ(¯ ¯ ω
min, p, L), ρ(¯ ¯ ω
2, p, L) ¯ } when ¯ p is closed enough to ¯ p
∗= (4¯ ω
min)
13L
−13.
We have that
¯
p
ω¯(¯ ω, p, L) = ¯ −
√ 2
2 exp( − L √
2¯ ω) 16L ω ¯
2− 16¯ p ω ¯ + L¯ p
4+ 4¯ p
3(4¯ ω + 2 √
2¯ ω p ¯ + ¯ p
2)
2. We consider the function:
f(¯ ω) = 16L ω ¯
2− 16¯ p¯ ω + L¯ p
4+ 4¯ p
3.
This is a quadratic equation in ¯ ω. We will prove that ∆
′= 64¯ p
2− 16L(L p ¯
4+ 4¯ p
3) = 16¯ p
2(4 − 4L p ¯ − L
2p ¯
2) > 0. Since ¯ p is closed to ¯ p
∗, we only need to prove that 4 − 4L¯ p
∗− L
2p ¯
2∗> 0, or L¯ p
∗< 2 √
2 − 2.
Since L <
q
10√ 2−12¯
ωmin
, we have that
L¯ p
∗= L(4¯ ω
min)
13(L)
−13= (4¯ ω
min)
13L
23< (4¯ ω
min)
13( 10 √ 2 − 14
¯ ω
min)
13= 2 √ 2 − 2.
Thus ∆
′> 0.
Therefore the equation f = 0 has the following solutions:
¯
ω
1= 2¯ p − p ¯ p
4 − 4L¯ p − L
2p ¯
24L = 2¯ p − p ¯ p
4 − 4L p ¯ − L
2p ¯
24L ,
¯
ω
2= 2¯ p + ¯ p p
4 − 4L¯ p − L
2p ¯
24L = 2¯ p + ¯ p p
4 − 4L¯ p − L
2p ¯
24L .
We will prove that in this case ¯ ω
max> ω ¯
2(¯ p). In order to do that, we only need to prove that ¯ ω
max> ω ¯
2(¯ p
∗).
Since L(¯ ω
max)
34is large, then L > (
2¯ω¯ωmin3max
)
14, we have L
4> ω ¯
min2¯ ω
max3. Thus
¯
ω
max3> ω ¯
min2 L
−4.
13
This means that in order to prove ¯ ω
max> ω ¯
2(¯ p
∗), we only have to prove that
16L¯ ω
2max− 16¯ p
∗ω ¯
max+ L¯ p
4∗+ 4¯ p
3∗> 0.
This inequality is equivalent to
16L¯ ω
2max− 16(4¯ ω
min)
13L
−13ω ¯
max+ L(4¯ ω
min)
43(L)
−43+ 16¯ ω
minL
−1> 0, or
¯
ω
max2L
23)
3− 2( ω ¯
min2 )
13ω ¯
maxL
23+ ( ω ¯
min2 )
43L
23+ ¯ ω
min=
= ¯ ω
maxL
23[¯ ω
maxL
43− 2( ω ¯
min2 )
13] + ( ω ¯
min2 )
43+ ¯ ω
min> 0.
This is true.
Hence ¯ ω
max> ω ¯
2(¯ p
∗) and ¯ ω
max> ω ¯
2(¯ p) for ¯ p closed enough to ¯ p
∗. Therefore || ρ(¯ ¯ ω, p, L) ¯ ||
∞= max { ρ(¯ ¯ ω
min, p, L), ¯ ρ(¯ ¯ ω
2, p, L) ¯ } .
Next, we will prove that ¯ p
∗is an asymptotic solution to the equation
¯
ρ(¯ ω
min, p, L) = ¯ ¯ ρ(¯ ω
2, p, L). Let ¯ p is a number closed to ¯ p
∗, and suppose that
¯
p has the form ¯ p ∽ C(
Lν)
−γ, we have that
¯
ρ(¯ ω
min, p, L) = exp( ¯ − L √
2¯ ω
min) ( √
2¯ ω
min− CL
−γ)
2+ 2¯ ω
min( √
2¯ ω
min+ CL
−γ)
2+ 2¯ ω
min= exp( − L √
2¯ ω
min) 4¯ ω
min− 2 √
2¯ ω
minCL
−γ+ C
2L
−2γ4¯ ω
min+ 2 √
2¯ ω
minCL
−γ+ C
2L
−2γ= exp( − L ν
√ 2ω
minν)
4ωminν
C2
(
Lν)
2γ−
2√2¯CωminL
γ+ 1
4¯ωmin
C2
L
2γ+
2√2¯CωminL
γ+ 1
∽ 1 − L √
2¯ ω
min− 4 √ 2¯ ω
minC L
γ+ 16¯ ω
minC
2L
2γ. We also have that
¯
ρ(¯ ω
2, p, L) = exp( ¯ − L √
2¯ ω
2) ( √
2¯ ω
2− CL
−γ)
2+ 2¯ ω
2( √
2¯ ω
2+ CL
−γ)
2+ 2¯ ω
2.
14
We have
exp( − L √
2¯ ω
2) = exp( − s
2L
22¯ p + ¯ p p
4 − 4L¯ p − L
2p ¯
24L )
= exp( − s
2¯ pL + ¯ pL p
4 − 4L¯ p − L
2p ¯
22 )
∽ 1 − p
2¯ pL ∽ 1 − √
2CL
1−2γ, and
( √
2¯ ω
2− p) ¯
2+ 2¯ ω
2ν ( √
2¯ ω
2+ ¯ p)
2+ 2¯ ω
2= 1 −
√2¯p¯ω2+
4¯p¯ω221 +
√2¯p¯ω2+
4¯p¯ω22∽ 1 − 2¯ p
√ 2¯ ω
2. Moreover, we have that
¯
√ p 2¯ ω
2= p ¯
q
2¯p+¯p√
4−4L¯p−L2p¯2 2L
=
s 2¯ pL 2 + p
4 − 4¯ pL − p ¯
2L
2∽ r C
2 L
1−2γ. Therefore
( √
2¯ ω
2− p) ¯
2+ 2¯ ω
2( √
2¯ ω
2+ ¯ p)
2+ 2¯ ω
2∽ 1 − √
2CL
1−γ2+ CL
1−γ. Thus
¯
ρ(¯ ω
2, p, L) ¯ ∽ (1 − √
2CL
1−γ2+ 2CL
1−γ)(1 − √
2CL
1−γ2+ CL
1−γ∽ 1 − 2 √
2CL
1−2γ+ 5CL
1−γ.
Equilibrate ¯ ρ(¯ ω
2, p, L) and ¯ ¯ ρ(¯ ω
min, p, L), we get ¯ ¯ p
∗.
Finally, we prove that this ¯ p
∗is a stricly local minimum of || ρ(¯ ω, p, L) ¯ ||
∞. We have that
∂
∂
p¯¯
ρ(¯ ω
2, p ¯
∗, L) = 4 √ 2 √
¯
ω
2exp ( − L √ 2 √
¯ ω
2) (¯ p
2∗+ 2 √
2 √
¯
ω
2ν p ¯
∗+ 4¯ ω
2ν)
2(¯ p
2∗− 4¯ ω
2) ω ¯
2¯ p < 0, and
∂
∂
p¯¯
ρ(¯ ω
min, p ¯
∗, L) = 4 √ 2 √
¯
ω
minexp ( − L √ 2 √
¯ ω
min) (¯ p
2∗+ 2 √
2 √
¯
ω
minp ¯
∗+ 4¯ ω
min)
2(¯ p
2∗− 4¯ ω
min) > 0.
15
cording to Lemma 2.1.1 it is also the global minimum. And
¯
p
∗∽ (4¯ ω
min)
13L
−13,
|| ρ(¯ ¯ ω, p ¯
∗, L) ||
∞∽ 1 − 4( ω ¯
min2 )
16L
13. Case 2: L¯ ω
3
max4
is small
In this case, we can see that ¯ ω
2> ω ¯
max. Thus
|| ρ(¯ ¯ ω, p, L) ¯ ||
∞= max { ρ(¯ ¯ ω
min, p, L), ¯ ρ(¯ ¯ ω
max, p, L) ¯ } .
As in the previous case, we will prove that ¯ p
∗is a solution of the equation
¯
ρ(¯ ω
min, p, L) = ¯ ¯ ρ(¯ ω
max, p, L). ¯
Let ¯ p be a number closed enough to ¯ p
∗. We have that
¯
ρ(¯ ω
min, p, L) = exp( ¯ − L √
2¯ ω
min) ( √
2¯ ω
min− p) ¯
2+ 2¯ ω
min( √
2¯ ω
min+ ¯ p)
2+ 2¯ ω
minν
∽ (1 − L √
2¯ ω
min+ L
22¯ ω
min)(1 − 4
√ 2¯ ω
min¯
p + 16¯ ω
min¯ p
2).
Since L¯ ω
3
max4
is small, then L < (
2¯ω¯ωmin3max
)
14, we have that L √
2¯ ω
min< ( ω ¯
min2¯ ω
max3)
14√
2¯ ω
min= 2
14( ω ¯
min¯ ω
max)
34,
and r
¯ ω
min¯ p
2∽
r ω ¯
min4 √
¯
ω
minω ¯
max= 1 2 ( ω ¯
min¯ ω
max)
14. Therefore
¯
ρ(¯ ω
min, p, L) ¯ ∽ 1 − 4 √ 2¯ ω
min¯
p .
16
We can suppose that 4¯ ω
min< ω ¯
max, then 2(¯ ω
minω ¯
max)
14< √
2¯ ω
max. Thus
¯ p < √
2¯ ω
max. Hence
¯
ρ(¯ ω
max, p, L) = exp( ¯ − L √
2¯ ω
max) ( √
2¯ ω
max− p) ¯
2+ 2¯ ω
max( √
2¯ ω
max+ ¯ p)
2+ 2¯ ω
max= exp( − L √
2¯ ω
max) (1 −
√2¯ωp¯max)
2+ 1 (1 +
√2¯ωp¯max
)
2+ 1
= exp( − L √
2¯ ω
max) 1 −
√2¯ωp¯max+
4¯ωp¯2max
1 +
√2¯ωp¯max
+
4¯ωp¯2max
∽ (1 − L √
2¯ ω
max+ L
22¯ ω
max)(1 − 2¯ p
√ 2¯ ω
max+ p ¯
2¯ ω
max).
Since L ∽ C ω ¯
−3 4+γ max
, then L √
2¯ ω
max< C ( ω ¯
min2¯ ω
max3)
41√
2¯ ω
maxω ¯
maxγ= C (¯ ω
min)
142
14ω ¯
−1 4+γ max
. We have also
¯ p
2¯
ω
max∽ 4 √
¯
ω
minω ¯
max¯ ω
max= 4
r ω ¯
min¯ ω
max. Therefore
¯
ρ(¯ ω
max, p, L) ¯ ∽ 1 − 2¯ p
√ 2¯ ω
max.
Equilibrate the two asymptotic expansion ¯ ρ(¯ ω
max, p, L) and ¯ ¯ ρ(¯ ω
2, p, L) we ¯ have the equation
2¯ p
√ 2¯ ω
max∽ 4 √ 2¯ ω
min¯
p .
Thus ¯ p ∽ 2(¯ ω
minω ¯
max)
14or ¯ p
∗is an asymptotic solution of the equation
¯
ρ(¯ ω
max, p, L) = ¯ ¯ ρ(¯ ω
2, p, L). Using the same argument as in the previous ¯ section we have that this ¯ p
∗is a global minimum of || ρ(¯ ¯ ω, p, L) ¯ ||
∞. And
|| ρ(¯ ¯ ω
min, p ¯
∗, L) ||
∞∽ 1 − 2¯ p
∗√ 2¯ ω
max∽ 1 − 4(¯ ω
minω ¯
max)
14√ 2¯ ω
max∽ 1 − 2 √
2( ω ¯
min¯ ω
max)
14.
17
¯
p is closed to ¯ p
∗. We have that
∂
ω¯p(¯ ¯ ω; ¯ p, L) = −
√ 2
2 exp( − L √
2¯ ω) 16L¯ ω
2− 16¯ p¯ ω + L p ¯
4+ 4¯ p
3(4¯ ω + 2 √
2¯ ω p ¯ + ¯ p
2)
2.
The function 16L¯ ω
2− 16¯ p ω ¯ + L¯ p
4+ 4¯ p
3, is a quadratic function of ¯ ω and it has ∆
′(¯ p) = 16¯ p
2(4 − 4L¯ p − L
2p ¯
2). We will prove that ∆
′(¯ p) > 0. Since ¯ p is closed to ¯ p
∗, we only need to prove that ∆
′(¯ p
∗) > 0. We have that
∆x < (17 − 12 √
2)2T C
2π
2(C
1ν
−12)
4 13. This implies
∆x < ( √
2 − 1)
43(2T C
2)
13( √
πC
1ν
−12)
43. Therefore
2 √
πνC
1ν
−1(2T C
2)
14∆x
34< 2 √ 2 − 2, or
L¯ p
∗< 2 √ 2 − 2.
Hence ∆
′(¯ p
∗) > 0.
Therefore the equation f = 0 has the following solutions:
¯
ω
1= 2¯ p − p ¯ p
4 − 4L¯ p − L
2p ¯
24L = 2¯ p − p ¯ p
4 − 4L p ¯ − L
2p ¯
24L ,
¯
ω
2= 2¯ p + ¯ p p
4 − 4L¯ p − L
2p ¯
24L = 2¯ p + ¯ p p
4 − 4L¯ p − L
2p ¯
24L .
We prove that for ¯ p closed to ¯ p
∗, we also have
2Lp¯>
∆tνπ, which implies
¯
ω
2>
∆tνπ. In fact, we only need to prove that
2Lp¯∗>
∆tνπ. Since
∆x < C
23π
2(C
1ν
−12)
42T ,
18
we have
2√ πν
(2T C2)14
∆x
−142C
1∆x > π C
2∆x , or
2Lp¯>
∆tνπ.
Therefore
|| ρ(¯ ¯ ω, p, L) ¯ ||
∞= max { ρ( ¯ π
2T ν ; ¯ p, L), ρ( ¯ π
∆tν ; ¯ p, L) } .
Using the same argument as in Theorem 2.1.1, we have that problem (2.1.13) has a unique solution
¯
p
∗∽ 2 √ π
(2T C
2)
14∆x
−14, then
|| ρ(¯ ¯ ω, p ¯
∗, L) ||
∞= 1 − 2 √ 2C
1 4
2
(2T )
14∆x
41+ O(∆x
12).
Case 2: L = C
1∆x, ∆t = C
2∆x
2, ∆x ≤ min
π22T(C1ν−1)4 4C23 12,
(10π(C√21−ν−14)2T1)2 12. Firstly, we prove that || ρ( ¯
2T νπ, p, L) ¯ ||
∞= max { ρ( ¯
2T νπ, p, L), ¯ ρ(¯ ¯ ω
2, p, L) ¯ } when ¯ p is closed to ¯ p
∗.
We have that
∂
ω¯p(¯ ¯ ω, p, L) = ¯ −
√ 2
2 exp( − L √
2¯ ω) 16L¯ ω
2− 16¯ p¯ ω + L p ¯
4+ 4¯ p
3(4¯ ω + 2 √
2¯ ω p ¯ + ¯ p
2)
2.
Similar as in the previous case, we prove that ∆
′(¯ p
∗) = 16¯ p
2∗(4 − 4L¯ p
∗− L
2p ¯
2∗) > 0.
We have that
∆x ≤ (10 √
2 − 14)2T π(C
1ν
−1)
2 12. This leads to
4π(C
1ν
−1)
22T ∆x
2< 8(5 √
2 − 7).
Thus 4π(C
1ν
−1)
22T ∆x
2< 8(5 √
2 − 7) = 8( √
2 − 1)
3.
19
We prove that in this case, ¯ ω
2<
∆tνπ. We have
∆x ≤ π
22T (C
1ν
−1)
44C
23 12. Hence
2π
T (C
1ν
−1)
4∆x
2≤ π
3C
23. Thus
(
2T C4πν21)
13∆x
−13C
1∆x ≤ π
C
2∆x
−2. Therefore
Lp¯≤
∆tνπ, which means ¯ ω
2<
∆tνπ. Therefore
|| ρ(¯ ¯ ω, p, L) ¯ ||
∞= max { ρ( ¯ π
2T ν ; ¯ p, L), ρ(¯ ¯ ω
2; ¯ p, L) } .
Equilibrate ¯ ρ(
2T νπ, p, L) and ¯ ¯ ρ(¯ ω
2, p, L), we get ¯ ¯ p
∗and using the same argument of the previous case, we can conclude that problem (2.1.13) has a unique solution
¯
p
∗∽ 2π T C
1 13∆x
−13, then
|| ρ(¯ ¯ ω, p ¯
∗, L) ||
∞= 1 − 4 πC
124T
16∆x
31+ O(∆x
23).
2.1.2 Proof of the Theorems in the Non-Overlapping Case
Proof of Theorem 2.1.3
Since ω ∈ [
2Tπ,
∆tπ], then x =
√2ωνpbelongs to [ q
2πν p22T
, q
2πν p2∆t
].
Thus
ρ(ω) = f(x) = 2x
2− 2x + 1
2x
2+ 2x + 1 . (2.1.14)
20
We have
f
′(x) = 4(2x
2− 1)
(2x
2+ 2x + 1)
2. (2.1.15) We have the following cases
Case 1: q
2νπ
p22T
≥
√12. We have that
max
ωρ = f (
r 2πν
p
2∆t ), (2.1.16)
Since q
2νπ
p22T
≥
√12, we have that q
2νπ
p2∆t
≥
√12q
2T
∆t
>
√1 2. Thus
min
pmax
ω
ρ = f ( 1
√ 2 r 2T
∆t ), (2.1.17)
when ω =
∆tπand p = 2 p
νπ2T
. Case 2: q
2νπ
p2∆t
≤
√12. We have that
max
ωρ = f ( r πν
p
2T ), (2.1.18)
Since q
2νπ
p2∆t
≤
√12, we have that q
νπp2T
≤
√12q
∆t
2T
<
√12. Thus
min
pmax
ω
ρ = f ( 1
√ 2 r ∆t
2T ) = f ( 1
√ 2 r 2T
∆t ), (2.1.19) when ω =
2Tπand p = 2 p
νπ∆t
. Case 3: q
2νπ p2∆t
>
√12
> q
2νπ p22T
.
We can see that if x > y and 2xy > 1, f(x) > f (y); and if x > y and 2xy < 1, f (x) < f (y); and if x > y and 2xy = 1, f(x) = f (y).
max
ωρ = max { f(
r 2πν p
22T ), f (
r 2πν
p
2∆t ) } . (2.1.20) Case 3.1: q
2νπ p2∆t
. q
2νπ
p22T
≥
12, or q
2νπ p2∆t
. q
2νπ
p2∆t
≥
12q
2T
∆t
, or q
2νπ p2∆t
≥
√1
2
(
2T∆t)
14>
√1 2.
max
ωρ = f (
r 2πν
p
2∆t ). (2.1.21)
21
when ω =
∆tand p = 2
(∆t2T)41
. Case 3.2: q
2νπ p2∆t
. q
νπ
p2T
≤
12, or q
νπ p2T
. q
νπ
p2T
≤
12q
∆t
2T
, or q
νπ p2T
≤
√1
2
(
∆t2T)
14<
√12.
max
ωρ = f ( r πν
p
2T ). (2.1.23)
Thus min
pmax
ω
ρ = f ( 1
√ 2 ( ∆t
2T )
14) = f ( 1
√ 2 ( 2T
∆t )
14) < f ( 1
√ 2 r 2T
∆t ), (2.1.24) when ω =
2Tπand p = 2
√νπ(∆t2T)14
.
22
2.2 Optimized Schwarz Waveform Relaxation Methods For One Dimensional Heat Equa- tion With First Order Transmission Con- dition
In this chapter, we consider the following algorithm
(∂
t− ν∂
xx)u
k1= f in Ω
1× (0, T ),
u
k1(x, 0) = u
0(x) in Ω
1,
(∂
x+
2νp+ 2q∂
t)u
k1(L, .) = (∂
x+
2νp+ 2q∂
t)u
k2−1(L, .) in (0, T ),
(2.2.1)
(∂
t− ν∂
xx)u
k2= f in Ω
2× (0, T ),
u
k2(x, 0) = u
0(x) in Ω
2,
(∂
x−
2νp− 2q∂
t)u
k2(0, .) = (∂
x−
2νp− 2q∂
t)u
k1−1(0, .) in (0, T ).
Similar as in the previous chapter, we consider the following problem
∂
te
1− ν∂
xxe
1= 0 in Ω
1× (0, T ),
e
1(x, 0) = u
0(x) in Ω
1,
(∂
x+
2νp+ 2q∂
t)e
k1(L, .) = (∂
x+
2νp+ 2q∂
t)e
2(L, .) in (0, T ),
(2.2.2)
∂
te
2− ν∂
xxe
2= 0 in Ω
2× (0, T ),
e
2(x, 0) = u
0(x) in Ω
2,
(∂
x−
2νp− 2q∂
t)e
2(0, .) = (∂
x−
2νp− 2q∂
t)e
1(0, .) in (0, T ).
From (2.2.2), we have that
iωFe
1− ν∂
xxFe
1= 0.
Therefore
Fe
1= C
1exp(
r iω
ν x) + C
2exp( − r iω
ν x),
23
Fe
1= C
1exp( iω ν x).
From (2.2.2), we have that
∂
xFe
1(L, ω) + p
2ν e
1(L, ω) + 2qiωFe
1(L, ω) = Fh
L(ω).
Thus
(C
1r iω ν + C
1p
2ν + C
12qωi) exp(
r iω
ν L) = Fh
L(ω).
Hence
C
1√ 4νωi + p + 4qωνi
2ν = Fh
Lexp( − r iω
ν L).
Thus
C
1= 2ν
√ 4νωi + p + 4qωνi Fh
Lexp( − r iω
ν L).
Therefore
Fe
1= 2ν
√ 4νωi + p + 4qωνi Fh
Lexp exp(
r iω
ν (x − L)).
From (2.2.2), we have that
iωFe
2− ν∂
xxFe
2= 0.
Therefore
Fe
2= D
1exp(
r iω
ν x) + D
2exp( − r iω
ν x), where Re( q
iω ν
) ≥ 0.
Since x ∈ (0, ∞ ) and Fe
2(x, .) ∈ L
2( R ), we have D
1= 0. Thus Fe
2= D
2exp( −
r iω ν x).
24
From (2.2.2), we have that
∂
xFe
2(0, ω) − p
2ν Fe
2(0, ω) − 2q
ν iωFe
2(0, ω) = h
0(ω).
Thus
(D
2r iω ν − D
2p
2ν − D
22qωi) = Fh
0(ω).
Hence
D
2√ 4νωi − p − 4qωi
2 = Fh
0.
Thus
D
2= 2ν
√ 4νωi − p − 4qωνi Fh
0. Therefore
Fe
2= 2ν
√ 4νωi − p − 4qωνi Fh
0exp(
r
− iω ν x).
Similar as in the previous chapter, we can define the convergence factor as
ρ(ω, p, q, L) = | 2 √
iων − p − 4qωνi 2 √
iων + p + 4qωνi exp( − √ iων L
ν ) |
2.
Put ¯ ω =
ων, ¯ p =
pνand ¯ ρ(¯ ω, p, q, L) = ¯ ρ(ω, p, q, L), we need to solve the problem
min
¯p,q∈R
max
¯
ω∈[¯ωmin,¯ωmax]
ρ(¯ ¯ ω, p, q, L). ¯ (2.2.3) Similar as in the previous chapter, we only need to solve the following problem
min
¯p,q≥0
max
¯
ω∈[¯ωmin,¯ωmax]
ρ(¯ ¯ ω, p, q, L). ¯ (2.2.4) We have the following theorems for the Overlapping Case
Theorem 2.2.1. If we fix ω ¯
minand ω ¯
max, then for L small satisfying that L
85ω ¯
maxis small and L¯ ω
maxis not small, the problem (2.2.4) has a unique solution
¯ p
∗∽ √
2¯ ω
3 8
min
ω ¯
1
max8
,
25
Theorem 2.2.2.
Case 1: For ∆x small enough L = C
1∆x, ∆t = C
2∆x, ω ¯ ∈ [
2T νπ,
Cπ2ν∆x
−1].
There exists a unique pair (¯ p
∗, q
∗) such that min
p,q¯ ≥0max
ω¯∈[ π2T ν,∆tνπ ]
ρ(¯ ¯ ω, p, q, L) = ¯ max
ω¯∈[2T νπ ,∆tνπ ]ρ(¯ ¯ ω, p ¯
∗, q
∗, L). Then
¯ p
∗∽ √
2( π
4T
3C
2ν
4)
18∆x
−18, q
∗∽ √
2( π
4ν
4T C
23)
−18∆x
83, min
¯p,q≥0
max
¯
ω∈[2T νπ ,∆tνπ ]
ρ(¯ ¯ ω, p, q, L) ¯ ∽ 1 − 4( C
22T )
18∆x
18+ O(∆x
14).
Case 2: For ∆x small enough L = C
1∆x, ∆t = C
2∆x
2, ω ¯ ∈ [
2T νπ,
C2πν4∆x
−2].
There exists a unique pair (¯ p
∗, q
∗) such that min
p,q¯ ≥0max
ω∈[2T νπ ,∆tνπ ]ρ(¯ ¯ ω, p, q, L) = ¯ max
ω¯∈[ π2T ν,∆tνπ ]
ρ(¯ ¯ ω, p ¯
∗, q
∗, L). Then
¯
p
∗∽ (4 π
2ν
24T
2(C
1ν
−1)
−1)
15∆x
−15, q
∗∽ (16(C
1ν
−1)
32T
πν )
15∆x
53, min
¯p,q≥0
max
¯
ω∈[2T νπ ,∆tνπ ]
ρ(¯ ¯ ω, p, q, L) ¯ ∽ 1 − 2
2110( πν
2T )
101(C
1ν
−1)
15∆x
15+ O(∆x
25).
Remark 2.2.1.
26
Figure 2.2.1.
Figure 2.2.2.
27
maximum point ω ¯
2, ω ¯
4of ρ) on the graph. In the first case ¯ ω ¯
4> ω ¯
max, we equilibrate the two boundaries and the maximal point ω ¯
2to get (¯ p
∗, q
∗).
We have the following theorem for the nonoverlapping case Theorem 2.2.3. The equation (2.2.4) has a unique solution
¯ p
∗= √
2¯ ω
3 8
min
ω ¯
1
max8
= √
2( πν
−12T )
38(πν
−1)
18∆t
−18, and
q
∗= √
2(ω
minν)
−18(ω
maxν)
−38= √ 2( πν
2T )
−18(πν)
−38∆t
38. Then, we have that
¯
min
p,q≥0
max
¯
ω∈[¯ωmin,¯ωmax]
ρ(¯ ω, p, q) = ¯ max
¯
ω∈[¯ωmin,¯ωmax]
ρ(¯ ω, p ¯
∗, q
∗) ∽ 1 − 4¯ ω
1 8
min
ω ¯
−1
max8
∽ 1 − 4( 1
2T )
18∆t
18.
2.2.1 Proof of the Theorems in the Overlapping Case
In this section, we will consider the problem of optimizing (¯ p, q) in the overlapping case.
Putting
h
L(¯ p, q) = max
¯
ω∈[¯ωmin,¯ωmax]
ρ(¯ ¯ ω, p, q, L) = ¯ || ρ(¯ ¯ ω, p, q, L) ¯ ||
∞,
we call that (¯ p
∗, q
∗, h
L(¯ p
∗, q
∗)) is a strictly local minimum of h
L(¯ p, q) if and only if there exists ǫ
1, ǫ
2positive such that for all (¯ p, q) in (¯ p
∗− ǫ
1, p ¯
∗+ ǫ
1) × (q
∗− ǫ
2, q
∗+ ǫ
2), we have h
L(¯ p, q) < h
L(¯ p
∗, q
∗).
In order to prove the theorems, we need the following lemma:
Lemma 2.2.1. If (¯ p
∗, q
∗, h
L(¯ p
∗, q
∗)) is a strictly local minimum of h
L(¯ p, q), then it is the global minimum of h
L(¯ p, q) and (¯ p
∗, q
∗) is the unique solution of (2.2.4).
28
Proof of Lemma 2.2.1
We denote D (z
0, δ) = { z ∈ C , |
zz+z−z00| < δ } , and D
δL= { (¯ p, q) | h
L(¯ p, q) ≤ δ } . We first prove that D
Lδis a convex set. Let (¯ p
1, q
1) and (¯ p
2, q
2) be to elements of D
Lδ, we have that
|| 2 √
i¯ ω − p ¯ − q ωi ¯ 2 √
i ω ¯ + ¯ p + q ωi ¯ exp( − √ i ω ¯ L
ν ) ||
∞≤ √ δ.
Thus ∀ ω ¯ ∈ [¯ ω
min, ω ¯
max],
| 2 √
i ω ¯ − p ¯ − q ωi ¯ 2 √
i¯ ω + ¯ p + q ωi ¯ exp( − r ω ¯
2 L) | ≤ √ δ.
Hence
| 2 √
i ω ¯ − p ¯ − q ωi ¯ 2 √
i¯ ω + ¯ p + q ωi ¯ | exp( − r ω ¯
2 L) ≤ √ δ.
Therefore
| 2 √
i ω ¯ − p ¯ − q ωi ¯ 2 √
i ω ¯ + ¯ p + q ωi ¯ | ≤ √
δ exp(L r ω ¯
2 ).
This means ¯ p
1+ q
1ωi ¯ ∈ D (2 √ i¯ ω, √
δ exp(L q
¯ ω 2
)).
Similarly, we have also ¯ p
2+ q
2ωi ¯ ∈ D (2 √ i ω, ¯ √
δ exp(L q
¯ ω 2
)).
If √
δ exp(L q
¯ ω
2
) < 1, using Lemma 2.1 in [1], we can see that D (2 √ i ω, ¯
√ δ exp(L q
¯ ω
2
)) is convex. Thus for θ ∈ [0, 1], we have θ(¯ p
1, q
2)+(1 − θ)(¯ p
2, q
2)
∈ D
δL. If √
δ exp(L q
¯ w
2
) ≥ 1, using Lemma 2.1 in [1], we can see that for
¯
p
1, ¯ p
2, q
1, q
2≥ 0, θ ∈ [0, 1], we have θ(¯ p
1+ q
1ωi) + (1 − θ)(¯ p
2+ q
2ωi) ¯
∈ D (2 √ i ω, ¯ √
δ exp(L q
¯ ω
2
)). Thus for θ ∈ [0, 1], we have θ(¯ p
1, q
1) + (1 − θ)(¯ p
2, q
2) ∈ D
δL.
Therefore D
δLis convex.
Suppose that (¯ p
∗, q
∗, h
L(¯ p
∗, q
∗)) is a strictly local minimum of h
L(¯ p, q ), we prove that it is a global minimum of h
L(¯ p, q). Suppose the contrary that there exists (¯ p
∗∗, q
∗∗, h
L(¯ p
∗∗)) such that h
L(¯ p
∗, q
∗) ≥ h
L(¯ p
∗∗, q
∗∗). Then there exists a convex neighborhood U of (¯ p
∗, q
∗), such that ∀ s ∈ U, s 6 = (¯ p
∗, q
∗) and h
L(s) > h
L(¯ p
∗, q
∗). Since (¯ p
∗∗, q
∗∗) ∈ D
hLL(¯p∗∗,q∗∗)
⊂ D
LhL(¯p∗,q∗)