Optimization-based domain decomposition methods for...
Transcript of Optimization-based domain decomposition methods for...
Optimization-based domain decomposition methods for multidisciplinary
simulation
by
Jeehyun Lee
A dissertation submitted to the graduate faculty
in partial fulfillment of the requirements for the degree of
DOCTOR OF PHILOSOPHY
Major: Applied Mathematics
Major Professor: Max D. Gunzburger
Iowa State University
Ames, Iowa
2001
Copyright c© Jeehyun Lee, 2001. All rights reserved.
ii
Graduate CollegeIowa State University
This is to certify that the Doctoral dissertation of
Jeehyun Lee
has met the dissertation requirements of Iowa State University
Committee Member
Committee Member
Committee Member
Committee Member
Committee Member
Major Professor
For the Major Program
For the Graduate College
iii
TABLE OF CONTENTS
ACKNOWLEDGMENTS . . . . . . . . . . . . . . . . . . . . . . . . . . . iv
ABSTRACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
1 INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2 FLUID-STRUCTURE INTERACTION PROBLEMS . . . . . . . . 5
2.1 The model problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3 A weak formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.4 The existence of a weak solution . . . . . . . . . . . . . . . . . . . . . . . 9
2.5 Finite element approximations . . . . . . . . . . . . . . . . . . . . . . . . 14
2.6 Error estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3 OPTIMIZATION-BASED DOMAIN DECOMPOSITION . . . . . 20
3.1 The model problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.2 The existence of an optimal solution . . . . . . . . . . . . . . . . . . . . 21
3.3 Convergence with vanishing penalty parameter . . . . . . . . . . . . . . . 27
3.4 Optimality system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.5 Sensitivity derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.6 Gradient method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
BIBLIOGRAPHY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
iv
ACKNOWLEDGMENTS
I first would like to thank Dr. Max D. Gunzburger for the expert guidance, almost
interminable patience that he has provided throughout this research and the writing of
this thesis. I am also grateful to my committee members for their efforts and contribu-
tions to this work : Dr. Qiang Du, Dr. L. Steven Hou, Dr. Justin Peters, Dr. Janet
S. Peterson and Dr. Monlay Tidriri. Special thanks to Dr. Xioaming Wang who was
generous to share his knowledge and to answer any questions about partial differential
equations. Not a day of my life passes by without thoughts and fond memories of my
family. A big thank to my parents, parents-in-law and my sister, Jeena for their encour-
agement and for believing in me. I want to express my deepest thanks to my husband,
Wook for his love, friendship and unfailing support throughout my career in graduate
school and to my lovely boy, Daniel who is the joy of my life.
v
ABSTRACT
We consider a domain decomposition method for a fluid-structure interaction prob-
lem. The fluid-structure interaction problem involves two mathematical models, each
posed on a different domain, so that domain decomposition occurs naturally. Our ap-
proach to a domain decomposition method is based on a strategy in which unknown data
at the interface is determined through an optimization process. We prove that the solu-
tion of the optimization problem exists. And we show that the Lagrange multiplier rule
may be used to transform the constrained optimization problem into an unconstrained
one and that rule is applied to derive an optimality system from which optimal solutions
may be obtained. We then study a gradient method for solving optimization problem.
1
1 INTRODUCTION
Multidisciplinary simulation problems arise in a variety of settings in which more
than one media, or more than one mathematical model, or more than one dominant
effect are present. The direct solution of such problems presents a formidable challenge,
since they usually involve large, coupled system of partial differential equations. For
this reason, methods which uncouple the different disciplines are of interest. Here, we
discuss uncoupling procedures which are based on using an optimization strategy.
A main virtue of our approach is that it allows for the user to use existing codes for
each discipline as black boxes and only requires that the user write a simple code that
effects the coupling between the disciplines. One reason we are able to do this is that
our methodology allows for complete flexibility with regards to the boundary conditions
imposed on each discipline. Another virtue of the optimizaion-based decoupling is that
it allows for the use of efficient iterative strategies, e.g., quickly converging iterative pro-
cesses by which solutions of the coupled, multidisciplinary problem are determined. Our
methodology has other important virtues as well as allowing for the use of mismatched
grids and different discretization methods for each discipline.
Here, for the sake of concreteness, we will describe the optimization-based domain
decomposition method for a fluid-structure interaction problem. The subjects of fluid-
structure interaction problem have been extensively studied in the past and continue
to be the focus of much attention today. There are number of different types of math-
ematical models for fluid-structure interactions. We classify these models into three
categories.
2
Elementary fluid. The fluid motion is governed by potential equations, e.g., Laplace
equations or wave equations. In [34], a coupled system of a potential equation and a
wave equation is considered. Elementary fluids coupled with rigid cavity or moving wall
has been studied in [19] and with an elastic solid in [4].
Inviscid fluid. A few mathematical papers have appeared for fluid-structure interac-
tions modeled using inviscid fluid models, e.g., the Euler equations. Interactions between
linearized inviscid fluids and elastic solids have been analyzed in [2],[35]. An algorithm
for an inviscid nonlinear fluid coupled with rigid walls was given in [3].
Viscous fluid. There is an extensive literature on linearized viscous fluids coupled
with solids. Solids modeled using plate equations or shell equations are treated in
[15],[16],[17],[32]. The Stokes equations coupled with a beam equation has been ana-
lyzed in [21]. In [11], [31], interactions between linearized viscous fluid and elastic solids
are studied. See [7] interaction with rigid walls.
Also, there is a vast literature on fluid-structure interactions for which the fluid is
modeled using nonlinear viscous fluid models, e.g., the Navier-Stokes equations. Rigid
body motions of solids in a nonlinear viscous fluid have been studied in [8],[12], [22], [23].
In [18], a coupled problem of Navier-Stokes equations and a plate equation is studied.
Some works have treated interactions between nonlinear viscous fluids and elastic solids.
See, e.g., [8],[13],[14],[36],[37].
MODEL PROBLEM
In this paper, we consider elastic body motions in a fluid flow. Let Ωf and Ωs denote
the regions occupied by the fluid and solid, respectively. Let Γ0 denote the interface
between the fluid and solid and let Γf and Γs denote the boundaries of the fluid and
solid regions (other than the interface Γ0). In the fluid region, we apply the Stokes
System.
3
ρfvt +∇p− µf∆v = ρf f in Ωf
∇ · v = 0 in Ωf
v = 0 on Γf
v|t=0 = v0 in Ωf
(1.1)
Here, ρf and µf denote the (constant) fluid density and viscosity, v the fluid velocity,
p the fluid pressure, and v0 the initial velocity.
In the solid, we apply the equation of the linear elasticity.
ρsutt − µs∆u− (λs + µs)∇(∇ · u) = ρsb in Ωs
u = 0 on Γs
u|t=0 = u0 in Ωs
ut|t=0 = u1 in Ωs
(1.2)
Here, µf and λf are the Lame constants and ρs the constant density of solid, b
denotes a given loading force per unit mass, u the displacement of the solid, and u0 and
u1 are given initial data.
Along the fixed interface Γ0 between the fluid and solid, the velocity of the fluid and
solid are equal, as are the stress vector in the fluid and solid. Thus, we have
ut = v on Γ0 (1.3)
and
µs∇u · n + (λs + µs)(∇ · u)n = pn− µf∇v · n on Γ0 (1.4)
Solving (1.1)-(1.4) is a formidable challenge. Fluid-structure interaction problems
involve two different mathematical models, each posed on a different domain, so that
domain decomposition occurs naturally. Our approach to domain decomposition is based
4
on a strategy in which unknown data at the interface is determined through an opti-
mization process. We consider the following interface conditions.
v = g on Γ0 (1.5)
and
µs∇u · n + (λs + µs)(∇ · u)n = h on Γ0 (1.6)
Then we may solve (1.1) and (1.5) for v and p and solve (1.2) and (1.6) for u.
For an arbitrary choice of g and h, (1.1) and (1.2) are satisfied. However, (1.3) and
(1.4) will not be satisfied. On the other hand, we know that g and h exist such that
solutions of (1.1),(1.5) and (1.2),(1.6) are solutions of (1.1)-(1.4). We merely have to
choose g = v|Γ0 = ut|Γ0 and h = µs∇u · n + (λs + µs)(∇ · u)n = pn− µf∇v · n on Γ0,
where (v, p, u) is a solution of (1.1)-(1.4).
The optimization-based domain decomposition algorithm finds such g and h by min-
imizing the functional
J (v, p,g,u,h) =1
2
∫ T
0
∫Γ0
(ut − g)2 dΓ dt
+1
2
∫ T
0
∫Γ0
(pn− µf∇v · n− h)2dΓ dt
In the remainder of this chapter, we discuss the derivation of the model (1.1)-(1.4).
This thesis is organized as follows. In chapter 2, a coupled system of fluid-structure
interaction problem is studied. The existence of a weak solution is proved and finite ele-
ment approximations are discussed. In chapter 3, we introduce an optimization problem
to uncouple the system. We prove that the optimization problem has a solution and the
solution converges to the solution of the coupled system. The Lagrange multiplier rule
is used to derive an optimality system from which optimal solutions may be determined.
Finally, we define a gradient method for the solution of the optimality system and show
the convergence for the gradient method.
5
2 FLUID-STRUCTURE INTERACTION PROBLEMS
2.1 The model problems
We consider a coupled system of Stokes and elasticity problem that was introduced
in the previous chapter.
ρfvt +∇p− µf∆v = ρf f in Ωf
∇ · v = 0 in Ωf
v = 0 on Γf
ρsutt − µs∆u− (λs + µs)∇(∇ · u) = ρsb in Ωs
u = 0 on Γs
ut = v on Γ0
µs∇u · n + (λs + µs)(∇ · u)n = pn− µf∇v · n on Γ0
v|t=0 = v0 in Ωf
u|t=0 = u0 in Ωs
ut|t=0 = u1 in Ωs
(2.1.1)
2.2 Notation
Throughout this paper, C will denote a positive constant whose meaning and value
changes with context. Hs(D), s ∈ IR, denotes the standard sobolev space of order s
6
with respect to the set D, equipped with the standard norm ‖ · ‖s,D. Corresponding
sobolev spaces of vector-valued functions will be denoted by Hs(D). Dual spaces will
be denoted by (.)∗. We then define the subspaces
H1f (Ωf ) = v ∈ H1(Ωf ) : v = 0 on Γf
J = v ∈ D(Ωf ) : ∇ · v = 0 on Ωf
where D(Ωf ) be the space of C∞ functions with compact support contained in Ωf .
V = the closure of J in H1f (Ωf )
H = the closure of J in L2(Ωf )
and
H1s(Ωs) = u ∈ H1(Ωs) : u = 0 on Γs
We define, for (pq) ∈ L1(D) and (u · v) ∈ L1(D),
(p, q)D =∫
Dpq dD and (u,v)D =
∫D
u · v dD
respectively.
We define the bilinear forms
af (u,v) =∫Ωf
µf∇u : ∇v dΩ ∀u,v ∈ H1(Ωf )
as(u,v) =∫Ωs
µs∇u : ∇v + (λs + µs)(∇ · u)(∇ · v) dΩ ∀u,v ∈ H1(Ωs)
a(u,v) =∫Ωs
µs∇u : ∇v dΩ ∀u,v ∈ H1(Ωs)
and
b(v, q) = −∫Ωf
q divv dΩ ∀v ∈ H1(Ωf ), ∀q ∈ L2(Ωf )
It is well known that the forms af (·, ·), as(·, ·), a(·, ·) and b(·, ·) are continuous, i.e.
there exist positive constants ka and kb such that
|af (u,v)| ≤ ka‖u‖1,Ωf‖v‖1,Ωf
∀u,v ∈ H1(Ωf )
7
|as(u,v)| ≤ ka‖u‖1,Ωs‖v‖1,Ωs ∀u,v ∈ H1(Ωs)
|a(u,v)| ≤ ka‖u‖1,Ωs‖v‖1,Ωs ∀u,v ∈ H1(Ωs)
and
|b(v, q)| ≤ kb‖v‖1,Ωf‖q‖0,Ωf
∀v ∈ H1(Ωf ), ∀q ∈ L2(Ωf )
Also, af (·, ·), as(·, ·) and a(·, ·) satisfy the coercivity property and b(·, ·) satisfies the
inf-sup condition, which means there exist positive constants Ka and Kb such that
|af (u,u)| ≥ Ka‖u‖21,Ωf
∀u ∈ H1(Ωf )
|as(u,u)| ≥ Ka‖u‖21,Ωs
∀u ∈ H1(Ωf )
|a(u,u)| ≥ Ka‖u‖21,Ωs
∀u ∈ H1(Ωf )
and
inf0 6=q∈L2(Ωf )
sup0 6=v∈H1
0(Ωf )
b(v, q)
‖v‖1,Ωf‖q‖0,Ωf
≥ Kb
2.3 A weak formulation
We define a space of trial functions and test functions as
U = (v,u) : v ∈ L2(0, T ; H1f (Ωf )), u ∈ L2(0, T ; H1
s(Ωs)),
ut ∈ L2(0, T ; L2(Ωs)), such that v = ut on Γ0
A weak formulation corresponding to (2.1.1) is given by
For f ,v0,b,u0 and u1 given,
f ∈ L2(0, T ; H1f (Ωf )∗) (2.3.2)
v0 ∈ L2(Ωf ) (2.3.3)
b ∈ L2(0, T ; L2(Ωs)) (2.3.4)
u0 ∈ H1s(Ωs) (2.3.5)
u1 ∈ L2(Ωs) (2.3.6)
8
to find (v,u) ∈ U satisfying
ρf (vt,w)Ωf+ b(w, p) + af (v,w) + ρs(utt, θt)Ωs
+ as(u, θt) = ρf (f ,w)Ωf+ ρs(b, θt)Ωs ∀(w, θ) ∈ U (2.3.7)
b(v, q) = 0 ∀q ∈ L2(Ωf ) (2.3.8)
v|t=0 = v0 (2.3.9)
u|t=0 = u0 (2.3.10)
ut|t=0 = u1 (2.3.11)
We recall the definition of the spaces J, V and H, introduced in the previous section
and which will be the basic spaces for W .
J = v ∈ D(Ωf ) : ∇ · v = 0 on Ωf
V = the closure of J in H1f (Ωf )
H = the closure of J in L2(Ωf )
Now, we define a space of trial functions and test functions as
W = (v,u) : v ∈ L2(0, T ;V ), u ∈ L2(0, T ; H1s(Ωs)),
ut ∈ L2(0, T ; L2(Ωs)), such that v = ut on Γ0
Then a weak formulation (2.3.7)-(2.3.11) is equivalent to the following:
To find (v,u) ∈ W satisfying
ρf (vt,w)Ωf+ af (v,w) + ρs(utt, θt)Ωs + as(u, θt)
= ρf (f ,w)Ωf+ ρs(b, θt)Ωs ∀(w, θ) ∈ W (2.3.12)
v|t=0 = v0 (2.3.13)
u|t=0 = u0 (2.3.14)
ut|t=0 = u1 (2.3.15)
9
2.4 The existence of a weak solution
To show the existence of weak solutions for the coupled system, we introduce an
auxiliary problem involving an ’artificial viscosity’ term. The auxiliary problem is defined
as follows.
ρfvt +∇p− µf∆v = ρf f in Ωf
∇ · v = 0 in Ωf
v = 0 on Γf
ρsutt − εµs∆ut − µs∆u− (λs + µs)∇(∇ · u) = ρsb in Ωs
u = 0 on Γs
ut = v on Γ0
εµs∇ut · n + µs∇u · n + (λs + µs)(∇ · u)n = pn− µf∇v · n on Γ0
v|t=0 = v0 in Ωf
u|t=0 = u0 in Ωs
ut|t=0 = u1 in Ωs
(2.4.16)
We define a space of trial functions and test functions for the auxiliary problem as
U = (v,u) : v ∈ L2(0, T ; H1f (Ωf )), u ∈ L2(0, T ; H1
s(Ωs)),
ut ∈ L2(0, T ; H1s(Ωs)), such that v = ut on Γ0
A weak formulation corresponding to (2.4.16) is given by
For f ,v0,b,u0 and u1 given,
f ∈ L2(0, T ; H1f (Ωf )∗) (2.4.17)
v0 ∈ L2(Ωf ) (2.4.18)
10
b ∈ L2(0, T ; L2(Ωs)) (2.4.19)
u0 ∈ H1s(Ωs) (2.4.20)
u1 ∈ L2(Ωs) (2.4.21)
to find (v,u) ∈ U satisfying
ρf (vt,w)Ωf+ b(w, p) + af (v,w) + ρs(utt, θt)Ωs + εa(ut, θt)
+ as(u, θt) = ρf (f ,w)Ωf+ ρs(b, θt)Ωs ∀(w, θ) ∈ U (2.4.22)
b(v, q) = 0 ∀q ∈ L2(Ωf ) (2.4.23)
v|t=0 = v0 (2.4.24)
u|t=0 = u0 (2.4.25)
ut|t=0 = u1 (2.4.26)
Now, we define a space of trial functions and test functions as
W = (v,u) : v ∈ L2(0, T ;V ), u ∈ L2(0, T ; H1s(Ωs)),
ut ∈ L2(0, T ; H1s(Ωs)), such that v = ut on Γ0
Then a weak formulation (2.4.22)-(2.4.26) is equivalent to the following:
To find (v,u) ∈ W satisfying
ρf (vt,w)Ωf+ af (v,w) + ρs(utt, θt)Ωs + εa(ut, θt)
+ as(u, θt) = ρf (f ,w)Ωf+ ρs(b, θt)Ωs ∀(w, θ) ∈ W (2.4.27)
v|t=0 = v0 (2.4.28)
u|t=0 = u0 (2.4.29)
ut|t=0 = u1 (2.4.30)
We show the existence of solutions of the auxiliary problem in the next theorem.
11
Theorem 2.4.1 For given f ,v0,b,u0 and u1 which satisfy (2.4.17)-(2.4.21), there ex-
ists a unique solution (v,u) ∈ W which satisfy (2.4.27)-(2.4.30). Moreover v ∈ L∞(0, T ;H), u ∈
L∞(0, T ; H1s(Ωs)) and ut ∈ L∞(0, T ; L2(Ωs))
proof: We use the Galerkin method. Let (φn)n∈N be a basis of V and (θn)n∈N be a
basis of H1s(Ωs) such that φn = θn on Γ0. We define discrete spaces Mn = spanφm, 1 ≤
m ≤ n and Nn = spanθm, 1 ≤ m ≤ n. Also define a discrete space of trial and test
functions by
Vn = (v,u) ∈ Vn,f × Vn,s, ut = v on Γ0
with
Vn,f = v =n∑
i=1
ai(t)φi, ai(t) ∈ H1(0, T )
Vn,s = u =n∑
i=1
bi(t)θi, bi(t) ∈ H1(0, T )
The discrete problem is : Find (vn,un) ∈ Vn such that vn(0) ∈ Mn, u(0) ∈ Nn and
unt(0) ∈ Nn with vn(0) → v0 in L2(Ωf ), un(0) → u0 in H1(Ωs) and unt(0) → u1 in
L2(Ωs) and
ρf (vnt,wn)Ωf+ af (vn,wn) + ρs(untt, znt)Ωs + εa(unt, znt)
+as(un, znt) = ρf (f ,wn)Ωf+ ρs(b, znt)Ωs ∀(wn, zn) ∈ Vn
(2.4.31)
Since this is a linear system of ordinary differential equations, there is a unique solution.
We will obtain a priori estimates independent of n for the functions vn,un and then
pass to the limit. Set wn = vn and zn = un in (2.4.31). We get
ρf (vnt,vn)Ωf+ af (vn,vn) + ρs(untt,unt)Ωs + εa(unt,unt)
+as(un,unt) = ρf (f ,vn)Ωf+ ρs(b,unt)Ωs
(2.4.32)
and this gives
ρf
2
d
dt‖vn‖2
0,Ωf+Ka‖vn‖2
1,Ωf+ρs
2
d
dt‖unt‖2
0,Ωs+ εKa‖unt‖2
1,Ωs
+Ka
2
d
dt‖un‖2
1,Ωs≤ ρf‖f‖−1,Ωf
‖vn‖1,Ωf+ ρs‖b‖0,Ωs‖unt‖0,Ωs
12
Integrating this from 0 to t, we obtain
ρf
2(‖vn‖2
0,Ωf− ‖vn(0)‖2
0,Ωf) +Ka
∫ t
0‖vn‖2
1,Ωf+ρs
2(‖unt‖2
0,Ωs− ‖unt(0)‖2
0,Ωs)
+ εKa
∫ t
0‖unt‖2
1,Ωs+Ka
2(‖un‖2
1,Ωs− ‖un(0)‖2
1,Ωs)
≤ρ2
f
2Ka
∫ t
0‖f‖2
−1,Ωf+Ka
2
∫ t
0‖vn‖2
1,Ωf+ Tρs
∫ t
0‖b‖2
0,Ωs+ρs
2T
∫ t
0‖unt‖2
0,Ωs
Gronwall’s inequality may be used to conclude
suptρf‖vn‖2
0,Ωf+Ka
∫ T
0‖vn‖2
1,Ωf+ sup
tρs‖unt‖2
0,Ωs+ εKa
∫ T
0‖unt‖2
1,Ωs
+ suptKa‖un‖2
1,Ωs≤ C
∫ T
0‖f‖2
−1,Ωf+ ρsC
∫ T
0‖b‖2
0,Ωs
+ ρf‖v0‖20,Ωf
+ ρs‖u1‖20,Ωs
+Ka‖u0‖21,Ωs
Hence vn is uniformly bounded in L∞(0, T ;H) and L2(0, T ;V ); un is uni-
formly bounded in L∞(0, T ; H1s(Ωs)); unt is uniformly bounded in L∞(0, T ; L2(Ωs))
and L2(0, T ; H1s(Ωs)); Thus, there exist weakly convergent subsequences and by passing
to the limit, a solution of (2.4.27)-(2.4.30) exists.
Uniqueness. Let (v1,u1) and (v2,u2) be two solutions of (2.4.27)-(2.4.30). Then
energy estimates may be used to get
suptρf‖v1 − v2‖2
0,Ωf+Ka
∫ T
0‖v1 − v2‖2
1,Ωf+ sup
tρs‖u1t − u2t‖2
0,Ωs
+ εKa
∫ T
0‖u1t − u2t‖2
1,Ωs+ sup
tKa‖u1 − u2‖2
1,Ωs≤ 0
and hence v1 = v2 and u1 = u2.
Now we show the existence of the coupled system by taking the limit of solutions of
the auxiliary problem as ε→ 0
Theorem 2.4.2 For given f ,v0,b,u0 and u1 which satisfy (2.3.2)-(2.3.6), there exists a
unique solution (v,u) ∈ W which satisfy (2.3.12)-(2.3.15). Moreover v ∈ L∞(0, T ;H), u ∈
L∞(0, T ; H1s(Ωs)) and ut ∈ L∞(0, T ; L2(Ωs))
13
proof: ∀ε, there exists vε,uε which are solutions of (2.4.27)-(2.4.30). Using a priori
estimate obtained in the proof of Theorem 2.4.1,
suptρf‖vε‖2
0,Ωf+Ka
∫ T
0‖vε‖2
1,Ωf+ sup
tρs‖(uε)t‖2
0,Ωs
+ εKa
∫ t
0‖(uε)t‖2
1,Ωs+ sup
tKa‖uε‖2
1,Ωs≤ C(v0,u0,u1, f ,b)
From which we deduce
vε is uniformly bounded in L2(0, T ;V ) ∩ L∞(0, T ;H)
uε is uniformly bounded in L∞(0, T ; H1s(Ωs))
(uε)t is uniformly bounded in L∞(0, T ; L2(Ωs))
√ε(uε)t is uniformly bounded in L2(0, T ; H1
s(Ωs))
Therefore, we can extract a subsequence such that
vε v in L2(0, T ;V ) ∩ L∞(0, T ;H)
uε u in L∞(0, T ; H1s(Ωs))
(uε)t ut in L∞(0, T ; L2(Ωs))
Since√ε(uε)t is uniformly bounded in L2(0, T ; H1
s(Ωs)) then√ε(∆uε)t is uniformly
bounded in L2(0, T ; H1s(Ωs)
∗) and
limε→0
εas((uε)t,w) = 0 ∀w ∈ L2(0, T ; H1s(Ωs))
From (2.4.27),
ρf ((vε)t,w)Ωf+ ρs((uε)tt, θt)Ωs
= −af (vε,w)− εa((uε)t, θt)− as(uε, θt) + ρf (f ,w)Ωf+ ρs(b, θt)Ωs
→ −af (v,w)− as(u, θt) + ρf (f ,w)Ωf+ ρs(b, θt)Ωs ∀(w, θ) ∈ W
Moreover, since vε v in L∞(0, T ;H) and (uε)t ut in L∞(0, T ; L2(Ωs)) we have
ρf (vt,w)Ωf+ ρs(utt, θt)Ωs
= −af (v,w)− as(u, θt) + ρf (f ,w)Ωf+ ρs(b, θt)Ωs ∀(w, θ) ∈ W
14
Further,
vε|t=0 v|t=0 in H
uε|t=0 u|t=0 in H1s(Ωs)
uεt|t=0 ut|t=0 in L2(Ωs)
Thus the existence theorem is proved. The proof of uniqueness is exactly the same
as before.
2.5 Finite element approximations
Let h denote a discretization parameter tending to zero and, for each h, let Xhf ,
Sh and Xhs be finite dimensional spaces such that Xh
f ⊂ H1f (Ωf ), Sh ⊂ L2(Ωf ) and
Xhs ⊂ H1
s (Ωs). And
Uh = (vh,uh) : vh ∈ L2(0, T ; Xhf ), uh ∈ H1(0, T ; Xh
s ),
such that vh = uht on Γ0
We assume that the finite element spaces satisfy the standard approximation prop-
erties, i.e.,
infv∈Xh
f
‖v − vh‖1,Ωf≤ ‖v‖m+1,Ωf
∀v ∈ Hm+1(Ωf )
infp∈Sh
‖p− ph‖0,Ωf≤ ‖p‖m,Ωf
∀p ∈ Hm(Ωf )
and
infu∈Xh
s
‖u− uh‖1,Ωs ≤ ‖u‖m+1,Ωs ∀u ∈ Hm+1(Ωs)
We also assume the inf-sup condition (or LBB) condition,
inf0 6=qh∈Sh
sup0 6=vh∈Xh
f
b(vh, qh)
‖vh‖1,Ωf‖qh‖0,Ωf
≥ Kb
where Kb is a positive constant independent of h.
15
Finite element approximations of solutions of the coupled system (2.3.7)-(2.3.11) are
defined as follows: Seek (vh,uh, ph) ∈ Uh × L2(0, T ;Sh) such that
ρf (vht ,w
h)Ωf+ b(wh, ph) + af (vh,wh) + ρs(u
htt, θ
ht )Ωs + as(u
h, θht )
= ρf (f ,wh)Ωf+ ρs(b, θ
ht )Ωs ∀(wh, θh) ∈ Uh (2.5.33)
b(vh, qh) = 0 ∀qh ∈ L2(0, T ;Sh) (2.5.34)
vh|t=0 = v0h (2.5.35)
uh|t=0 = u0h (2.5.36)
uht |t=0 = u1h (2.5.37)
First, we show the convergence of finite element approximations. To show the con-
vergence, we consider finite element approximations of the auxiliary problem (2.4.22)-
(2.4.26): Seek (vεh,uεh, pεh) ∈ Uh × L2(0, T ;Sh) such that
ρf (vεht ,w
h)Ωf+ b(wh, pεh) + af (vεh,wh) + ρs(u
εhtt , θ
ht )Ωs
+ εa(uεht , θ
ht ) + as(u
εh, θht )
= ρf (f ,wh)Ωf+ ρs(b, θ
ht )Ωs ∀(wh, θh) ∈ Uh (2.5.38)
b(vεh, qh) = 0 ∀qh ∈ L2(0, T ;Sh) (2.5.39)
vεh|t=0 = v0h (2.5.40)
uεh|t=0 = u0h (2.5.41)
uεht |t=0 = u1h (2.5.42)
Lemma 2.5.1 For each ε ≥ 0, let (vε,uε, pε) denote solutions of auxiliary problem
(2.4.22)-(2.4.26) and (vεh,uεh, pεh) denote finite element approximations of solutions of
auxiliary problem. Then (vεh,uεh, pεh) → (vε,uε, pε) in L2(0, T ; H1(Ωf ))×H1(0, T ; H1(Ωs))×
H−1(0, T ;L2(Ωf )) as h→ 0.
proof: Set wh = vεh and θh = uεh in (2.5.38) and combine with (2.5.39), we obtain
suptρf‖vεh‖2
0,Ωf+Ka
∫ T
0‖vεh‖2
1,Ωf+ sup
tρs‖uεh
t ‖20,Ωs
16
+ εKa
∫ t
0‖uεh
t ‖21,Ωs
+ suptKa‖uεh‖2
1,Ωs
≤ C∫ T
0‖f‖2
−1,Ωf+ ρsC
∫ T
0‖b‖2
0,Ωs
+ ρf‖v0‖20,Ωf
+ ρs‖u1‖20,Ωs
+Ka‖u0‖21,Ωs
The same argument as in the proof of the Theorem 2.4.1 yields (vεh,uεh) → (vε,uε)
in L2(0, T ; H1(Ωf ))×H1(0, T ; H1(Ωs)) as h→ 0. If we take pεh satisfying (2.5.38) and pε
satisfying (2.4.22) then pεh → pε in H−1(0, T ;L2(Ωf )) as h→ 0. Thus lemma is proved.
It was proved in Theorem 2.4.2 that (vε,uε, pε) → (v,u, p) as ε → 0 and it can be
shown (vεh,uεh, pεh) → (vh,uh, ph) as ε → 0 in the same manner. Combining this with
Lemma 2.5.1 gives the following theorem.
Theorem 2.5.2 Let (v,u, p) denote solutions of the coupled system (2.3.7)-(2.3.11)
and (vh,uh, ph) denote solutions of (2.5.33)-(2.5.37). Then (vh,uh, ph) → (v,u, p) in
L2(0, T ; H1(Ωf ))× (L2(0, T ; H1(Ωs) ∩H1(0, T ; L2(Ωs))×H−1(0, T ;L2(Ωf )) as h→ 0.
2.6 Error estimates
For the purpose of the proof of next theorem, we introduce some spaces and a pro-
jection. Let Ω denote Ωf ∩ Ωs and Xh denote a finite dimensional space such that
Xh ∈ H1(Ω). We define a continous space G and a discrete space Gh by
G = z ∈ H1(Ω) : zf ≡ z|Ωf∈ V, zs ≡ z|Ωs ∈ H1(Ωs), and zf |Γ0 = zs|Γ0
Gh = z ∈ Xh : ψhf ≡ ψh|Ωf
∈ Xhf , ψ
hs ≡ ψh|Ωs ∈ Xh
s ,
b(ψhf , q
h) = 0 ∀qh ∈ Sh, and ψhf |Γ0 = ψh
s |Γ0
We define P h : G −→ Gh to be the projection with respect to the L2(Ω) inner
product, i.e. P hz = z if
(z, ψh)Ω = (z, ψh)Ω ∀ψ ∈ Gh, ∀z ∈ G
17
Then
‖z− z‖1,Ω ≤ Chm‖z‖m+1,Ω (2.6.43)
‖z− z‖0,Ω ≤ Chm+1‖z‖m+1,Ω (2.6.44)
Now we prove the following estimate for the error between the solutions of the semi-
discrete and continuous problems.
Let
ζ =
v on Ωf
ut on Ωs
And define v = P hζ|Ωf, ut = P hζ|Ωs ,
Theorem 2.6.1
ρf‖v − vh‖20,Ωf
+Ka
∫ t
0‖v − vh‖2
1,Ωfdt
+ρs‖ut − uht ‖2
0,Ωs+Ka‖u− uh‖2
1,Ωs
≤ Ch2m+2‖v‖2m+1,Ωf
+ Ch2m∫ t
0‖v‖2
m+1,Ωfdt
+Ch2m‖u‖2m+1,Ωs
+ Ch2m+2‖ut‖2m+1,Ωs
+Ch2m∫ t
0‖u‖2
m+1,Ωsdt+ Ch2m
∫ t
0‖p‖2
m,Ωfdt (2.6.45)
proof: By subtracting (2.5.33)-(2.5.34) from the corresponding equations of (2.3.7)-
(2.3.8) we obtain the “orthogonality conditions”.
ρf (vt − vht ,w
h)Ωf+ b(wh, p− ph) + af (v − vh,wh)
+ ρs(utt − uhtt, θ
ht )Ωs + as(u− uh, θh
t ) = 0 ∀(wh, θh) ∈ Uh (2.6.46)
b(v − vh, qh) = 0 ∀qh ∈ L2(0, T ;Sh) (2.6.47)
Using (2.6.46)-(2.6.47) we deduce that
ρf (vt − vht ,v − vh)Ωf
+ af (v − vh,v − vh)
18
+ρs(utt − uhtt,ut − uh
t )Ωs + as(u− uh,ut − uht )
= ρf (vt − vht ,v − v)Ωf
+ af (v − vh,v − v) + ρs(utt − uhtt,ut − ut)Ωs
+as(u− uh,ut − ut)− b(v − vh, p− ph)
+ρf (vt − vht , v − vh)Ωf
+ af (v − vh, v − vh) + ρs(utt − uhtt, ut − uh
t )Ωs
+as(u− uh, ut − uht ) + b(v − vh, p− ph)
= ρf (vt − vht ,v − v)Ωf
+ af (v − vh,v − v) + ρs(utt − uhtt,ut − ut)Ωs
+as(u− uh,ut − ut)− b(v − vh, p− ph) (2.6.48)
Then using the facts that (ψhf ,v − v) = 0 ∀ψh
f ∈ Xhf and that
b(v, qh) = 0 = b(vh, qh) ∀qh ∈ Sh (2.6.49)
We obtain (by also noting that (2.6.49) implies vt ∈ Xhf and vh
t ∈ Xhf )
ρf (vt − vht ,v − v)Ωf
= ρf (vt,v − v)Ωf= ρf (vt − vt,v − v)Ωf
(2.6.50)
Similarly,
ρs(utt − uhtt,ut − ut)Ωs = ρs(utt − utt,ut − ut)Ωs (2.6.51)
Combining (2.6.48)-(2.6.51) we deduce that for all qh ∈ L2(0, T ;Sh)
ρf
2
d
dt‖v − vh‖2
0,Ωf+Ka‖v − vh‖2
1,Ωf
+ρs
2
d
dt‖ut − uh
t ‖20,Ωs
+Ka
2
d
dt‖u− uh‖2
1,Ωs
≤ ρf
2
d
dt‖v − v‖2
0,Ωf+ ka‖v − vh‖1,Ωf
‖v − v‖1,Ωf
+ρs
2
d
dt‖ut − ut‖2
0,Ωs+ ka‖u− uh‖1,Ωs‖ut − ut‖1,Ωs
+kb‖v − v‖1,Ωf‖p− qh‖0,Ωf
+ kb‖v − vh‖1,Ωf‖p− qh‖0,Ωf
Therefore,
ρf‖v − vh‖20,Ωf
+Ka
∫ t
0‖v − vh‖2
1,Ωfdt+ ρs‖ut − uh
t ‖20,Ωs
+Ka‖u− uh‖21,Ωs
≤ C(‖v − v‖20,Ωf
+∫ t
0‖v − v‖2
1,Ωfdt+ ‖u(t1)− u(t1)‖2
1,Ωs
+‖ut − ut‖20,Ωs
+∫ t
0‖ut − ut‖2
1,Ωsdt+
∫ t
0‖p− qh‖2
0,Ωfdt) (2.6.52)
19
for all qh ∈ L2(0, T ;Sh), where ‖u(t1)− u(t1)‖ = maxt∈[0,T ] ‖u(t)− u(t)‖. Hence (2.6.45)
follows from (2.6.52), (2.6.43) and (2.6.44).
20
3 OPTIMIZATION-BASED DOMAIN DECOMPOSITION
In this chapter, we present an optimization-based domain decomposition method to
uncouple the computations.
3.1 The model problems
Let v, p,g,u and h satisfy the coupled system (2.1.1). Instead of constraints (2.1.1),
we consider the uncoupled system
ρfvt +∇p− µf∆v = ρf f in Ωf
∇ · v = 0 in Ωf
v = 0 on Γf
v = g on Γ0
ρsutt − µs∆u− (λs + µs)∇(∇ · u) = ρsb in Ωs
u = 0 on Γs
µs∇u · n + (λs + µs)(∇ · u)n = h on Γ0
v|t=0 = v0 in Ωf
u|t=0 = u0 in Ωs
ut|t=0 = u1 in Ωs
(3.1.1)
In this paper, we refer to g and h as controls. Our goal is to find g and h such
21
that the solutions of (3.1.1) coincide with solutions of (2.1.1). The optimization-based
domain decomposition algorithm finds such g and h by minimizing the functional
J (v, p,g,u,h) =1
2
∫ T
0
∫Γ0
(ut − g)2 dΓ dt
+1
2
∫ T
0
∫Γ0
(pn− µf∇v · n− h)2dΓ dt
In order to regularize the optimization problem, we instead minimize the penalized
functional
J δ(v, p,g,u,h) =1
2
∫ T
0
∫Γ0
(ut − g)2 dΓ dt
+1
2
∫ T
0
∫Γ0
(pn− µf∇v · n− h)2dΓ dt+δ
2
∫ T
0
∫Γ0
g2 dΓ dt
+δ
2
∫ T
0
∫Γ0
(∇Γg)2 dΓ dt+δ
2
∫ T
0
∫Γ0
g2t dΓ dt+
δ
2
∫ T
0
∫Γ0
h2 dΓ dt
where the penalty parameter δ is a positive constant that can be chosen to change the
relative importance of penalty terms in Jδ and ∇Γ denotes tangential gradient. Thus
the optimization problem we propose to solve is given by
Problem 1
Find (v,p,g,u,h) such that the functional Jδ is minimized subject to (3.1.1).
3.2 The existence of an optimal solution
We recall the definition of the spaces
H1f (Ωf ) = v ∈ H1(Ωf ) : v = 0 on Γf
J = v ∈ D(Ωf ) : ∇ · v = 0 on Ωf
V = the closure of J in H1f (Ωf )
H = the closure of J in L2(Ωf )
and
H1s(Ωs) = u ∈ H1(Ωs) : u = 0 on Γs
22
For functions defined on Γ0, we will use the subspaces
Y = g ∈ L2(0, T ; H1(Γ0)) ∩H1(0, T ; L2(Γ0)) : g = 0 on ∂Γ0
Z = L2(0, T ; L2(Γ0))
with norms
‖g‖2Y =
∫ T
0
∫Γ0
g2 dΓ dt+∫ T
0
∫Γ0
(∇Γg)2 dΓ dt+∫ T
0
∫Γ0
g2t dΓ dt
‖h‖2Z =
∫ T
0
∫Γ0
h2 dΓ dt
A weak formulation corresponding to (3.1.1) is given by : Seek v ∈ L2(0, T ; H1f (Ωf )), p ∈
L2(0, T ;L2(Ωf )), g ∈ Y, u ∈ L2(0, T ; H1s(Ωs)) ∩H1(0, T ; L2(Ωs)) and h ∈ Z satisfying
ρf (vt,w)Ωf+ b(w, p) + af (v,w) + (w, pn− µf∇v · n)Γ0
= ρf (f ,w)Ωf∀ w ∈ H1
f (Ωf ) (3.2.2)
b(v, q) = 0 ∀ q ∈ L2(Ωf ) (3.2.3)
(v, s)Γ0 − (g, s)Γ0 = 0 ∀ s ∈ H−1/2(Γ0) (3.2.4)
ρs(utt, θ)Ωs + as(u, θ) = ρs(b, θ)Ωs + (h, θ)Γ0
∀ θ ∈ H1s(Ωs) (3.2.5)
v|t=0 = v0 in Ωf (3.2.6)
u|t=0 = u0 in Ωs (3.2.7)
ut|t=0 = u1 in Ωs (3.2.8)
Next, we give a precise definition of an optimal solution. Let the admissibility set is
defined by
Uad = (v, p,g,u,h) ∈ L2(0, T ; H1f (Ωf ))× L2(0, T ;L2(Ωf ))×
Y × L2(0, T ; H1s(Ωs)) ∩H1(0, T ; L2(Ωs))× Z :
J δ(v, p,g,u,h) <∞ and (3.2.2)-(3.2.8) is satisfied (3.2.9)
23
Then (v, p, g, u, h) is called an optimal solution if there exists α ≥ 0 such that
J δ(v, p, g, u, h) ≤ J δ(v, p,g,u,h)
for all (v, p,g,u,h) ∈ Uad satisfying
‖g − g‖Y + ‖h− h‖Z ≤ α
To show the existence of optimal solutions, we introduce an auxiliary problem again.
ρfvt +∇p− µf∆v = ρf f in Ωf
∇ · v = 0 in Ωf
v = 0 on Γf
v = g on Γ0
ρsutt − εµs∆ut − µs∆u− (λs + µs)∇(∇ · u) = ρsb in Ωs
u = 0 on Γs
εµs∇ut · n + µs∇u · n + (λs + µs)(∇ · u)n = h on Γ0
v|t=0 = v0 in Ωf
u|t=0 = u0 in Ωs
ut|t=0 = u1 in Ωs
(3.2.10)
Problem 2
Find (v,p,g,u,h) such that the functional Jδ is minimized subject to (3.2.10).
A weak formulation corresponding to (3.2.10) is given by: Seek v ∈ L2(0, T ; H1f (Ωf )), p ∈
L2(0, T ;L2(Ωf )), g ∈ Y, u ∈ H1(0, T ; H1s(Ωs)) and h ∈ Z satisfying
ρf (vt,w)Ωf+ b(w, p) + af (v,w) + (w, pn− µf∇v · n)Γ0
= ρf (f ,w)Ωf∀ w ∈ H1
f (Ωf ) (3.2.11)
24
b(v, q) = 0 ∀ q ∈ L2(Ωf ) (3.2.12)
(v, s)Γ0 − (g, s)Γ0 = 0 ∀ s ∈ H−1/2(Γ0) (3.2.13)
ρs(utt, θ)Ωs + εa(ut, θ) + as(u, θ)
= ρs(b, θ)Ωs + (h, θ)Γ0 ∀ θ ∈ H1s(Ωs) (3.2.14)
v|t=0 = v0 in Ωf (3.2.15)
u|t=0 = u0 in Ωs (3.2.16)
ut|t=0 = u1 in Ωs (3.2.17)
And admissibility set be defined by
U εad = (v, p,g,u,h) ∈ L2(0, T ; H1
f (Ωf ))× L2(0, T ;L2(Ωf ))×
Y ×H1(0, T ; H1s(Ωs))× Z : J δ(v, p,g,u,h) <∞
and (3.2.11)-(3.2.17) is satisfied (3.2.18)
Using the properties of the bilinear forms we can obtain an a priori bounds for
solutions of the weak formulation (3.2.11)-(3.2.17). Let (v, p,g,u,h) satisfy (3.2.11)-
(3.2.17). From (3.2.11) and (3.2.14), we obtain
ρf (vt,w)Ωf+ b(w, p) + af (v,w)
+ρs(utt, θ)Ωs + εa(ut, θ) + as(u, θ)
= ρf (f ,w)Ωf+ ρs(b, θ)Ωs − (w, pn− µf∇v · n)Γ0 + (h, θ)Γ0
∀ w ∈ H1f (Ωf ), ∀ θ ∈ H1
s(Ωs)
or
ρf (vt,w)Ωf+ b(w, p) + af (v,w)
+ρs(utt, θ)Ωs + εa(ut, θ) + as(u, θ)
= ρf (f ,w)Ωf+ ρs(b, θ)Ωs − (w, pn− µf∇v · n)Γ0
+(w,h)Γ0 + (h, θ)Γ0 − (w,h)Γ0
25
∀ w ∈ H1f (Ωf ), ∀ θ ∈ H1
s(Ωs) (3.2.19)
Taking w = v and θ = ut in (3.2.19). Then because of (3.2.12)-(3.2.13) we have
ρf (vt,v)Ωf+ af (v,v) + ρs(utt,ut)Ωs + εa(ut,ut) + as(u,ut)
= ρf (f ,v)Ωf+ ρs(b,ut)Ωs − (g, pn− µf∇v · n)Γ0 + (g,h)Γ0
+(h,ut)Γ0 − (g,h)Γ0
This may be reduced to
ρf
2
d
dt‖v‖2
0,Ωf+Ka‖v‖2
1,Ωf+ρs
2
d
dt‖ut‖2
0,Ωs+ εKa‖ut‖2
1,Ωs+Ka
2
d
dt‖u‖2
1,Ωs
= −(g, pn− µf∇v · n− h)Γ0 + (h,ut − g)Γ0 + ρf (f ,v)Ωf+ ρs(b,ut)Ωs
≤ ‖g‖0,Γ0‖pn− µf∇v · n− h‖0,Γ0 + ‖h‖0,Γ0‖ut − g‖0,Γ0
+ρf‖f‖−1,Ωf‖v‖1,Ωf
+ ρs‖b‖0,Ωs‖ut‖0,Ωs
Integrating from 0 to t yields
ρf supt‖v‖2
0,Ωf+Ka
∫ T
0‖v‖2
1,Ωfdt+ ρs sup
t‖ut‖2
0,Ωs
+εKa
∫ T
0‖ut‖2
1,Ωsdt+Ka sup
t‖u‖2
1,Ωs
≤∫ T
0‖g‖2
0,Γ0dt+
∫ T
0‖pn− µf∇v · n− h‖2
0,Γ0dt+
∫ T
0‖h‖2
0,Γ0dt
+∫ T
0‖ut − g‖2
0,Γ0dt+ ρfC
∫ T
0‖f‖2
−1,Ωfdt+ ρsC
∫ T
0‖b‖2
0,Ωsdt
+ρf‖v0‖20,Ωf
+ ρs‖u1‖20,Ωs
+Ka‖u0‖21,Ωs
(3.2.20)
We take test functions w with ∇ ·w = 0 in equation (3.2.11), we get
ρf (vt,w)Ωf= −af (v,w)− (w, pn− µf∇v · n)Γ0 + ρf (f ,w)Ωf
∀w ∈ V
and this gives
|ρf (vt,w)Ωf| ≤ ka‖v‖1,Ωf
‖w‖1,Ωf
+ ‖w‖1,Ωf‖pn− µf∇v · n‖0,Γ0 + ρf‖f‖−1,Ωf
‖w‖1,Ωf
26
Therefore,
ρf
∫ T
0‖vt‖2
V ∗dt ≤ ka
∫ T
0‖v‖2
1,Ωfdt+
∫ T
0‖pn− µf∇v · n− h‖2
0,Γ0dt
+∫ T
0‖h‖2
0,Γ0dt+ ρf
∫ T
0‖f‖2
−1,Ωfdt (3.2.21)
From (3.2.11),
b(w, p) = −ρf (vt,w)Ωf− af (v,w)
− (w, pn− µf∇v · n)Γ0 + ρf (f ,w)Ωf∀w ∈ H1
f (Ωf )
Inf-sup condition may be used to get,
Kb
∫ T
0‖p‖2
0,Ωfdt ≤ ρf
∫ T
0‖vt‖2
V ∗dt+ ka
∫ T
0‖v‖2
1,Ωfdt
+∫ T
0‖pn− µf∇v · n‖2
0,Γ0dt+ ρf
∫ T
0‖f‖2
−1,Ωfdt (3.2.22)
Theorem 3.2.1 There exists an optimal solution (v, p, g, u, h) ∈ U εad for Problem 2.
proof: It is clear that U εad is not empty. Let (vn, pn,gn,un,hn) be a minimizing sequence
in Uad. i.e.
limn→∞
J δ(vn, pn,gn,un,hn) = inf
(v,p,g,u,h)∈Uad
J δ(v, p,g,u,h)
Thus from (3.2.18), we have that ‖gn‖ and ‖hn‖ are uniformly bounded in Y × Z.
Then since by (3.2.20) and (3.2.22) (vn, pn,un) ∈ L2(0, T ; H1f (Ωf ))×L2(0, T ;L2(Ωf ))×
H1(0, T ; H1s(Ωs)) is uniformly bounded. Thus, there exist subsequences, denoted by
(vn, pn,gn,un,hn) for simplicity, such that
vn v in L2(0, T ; H1f (Ωf ))
pn p in L2(0, T ; L2(Ωf ))
gn g in Y
un u in H1(0, T ; H1s(Ωs))
hn h in Z
27
for some (v, p, g, u, h) ∈ U εad.
Now, by passing to the limit, we have that (v, p, g, u, h) ∈ U εad satisfies (3.2.11)-
(3.2.17). Then by the fact that functional Jδ(·, ·, ·, ·, ·) is lower semi-continuous, we
conclude that
J δ(v, p, g, u, h) = inf(v,p,g,u,h)∈U ε
ad
J δ(v, p,g,u,h)
i.e. (v, p, g, u, h) is an optimal solution.
Now we take the limit as ε → 0 of optimal solutions of Problem 2 to get optimal
solutions of Problem 1.
Theorem 3.2.2 There exists an optimal solution (v, p, g, u, h) ∈ Uad for Problem 1.
The proof is the same as in the Theorem 2.2.
3.3 Convergence with vanishing penalty parameter
In the next theorem we show that optimal solutions of Problem 2 converges to the
weak solution of (2.4.16) as δ → 0.
Theorem 3.3.1 For each δ, let (vδ, pδ,gδ,uδ,hδ) denote an optimal solution of Problem
2 and (v, p, u) denote a solution of (2.4.22)-(2.4.26). Then ‖v − vδ‖1,Ωf→ 0, ‖p −
pδ‖0,Ωf→ 0, ‖u− uδ‖1,Ωs → 0, ‖ut − uδ
t‖1,Ωs → 0 as δ → 0.
proof: Let g = v|Γ0 and h = εµs∇ut · n + µs∇u · n + (λs + µs)(∇ · u)n on Γ0. For any
solution (v, p, u) of (2.4.16), we have
J δ(vδ, pδ,gδ,uδ,hδ) ≤ J δ(v, p, g, u, h)
i.e.
1
2
∫ T
0
∫Γ0
(uδt − g)2 dΓ dt+
1
2
∫ T
0
∫Γ0
(pδn− µf∇vδ · n− hδ)2dΓ dt
+δ
2
∫ T
0
∫Γ0
‖gδ‖2Y dΓ dt+
δ
2
∫ T
0
∫Γ0
‖hδ‖2ZdΓ dt
≤ J (v, p, g, u, h) +δ
2
∫ T
0
∫Γ0
‖g‖2Y dΓ dt+
δ
2
∫ T
0
∫Γ0
‖h‖2ZdΓ dt ∀δ
28
Then ‖gδ‖Y and ‖hδ‖Z are uniformly bounded, ‖uδt − g‖0,Γ0 → 0 and ‖pδn −
µf∇vδ · n − hδ‖0,Γ0 → 0 as δ → 0. We then obtain ‖vδ‖1,Ωf, ‖pδ‖0,Ωf
, ‖uδ‖1,Ωs and
‖uδt‖1,Ωs are uniformly bounded by (3.2.20) and (3.2.22). Hence, as δ → 0, there
exists a subsequence of (vδ, pδ,gδ,uδ,hδ) that converges to some (v, p, g, u, h) ∈
L2(0, T ; H1f (Ωf )) × L2(0, T ;L2(Ωf )) × Y ×H1(0, T ; H1
s(Ωs)) × Z. The fact that ‖uδt −
g‖0,Γ0 → 0 yields v = ut on Γ0, also ‖pδn − µf∇vδ · n − hδ‖0,Γ0 → 0 as δ → 0 implies
that εµs∇ut · n + µs∇u · n + (λs + µs)(∇ · u)n = pn− µf∇v · n on Γ0.
By passing to the limit, we have that (v, p, u) satisfy (2.4.22)-(2.4.26). By the unique-
ness of solutions of (2.4.22)-(2.4.26), (v, p, u) = (v, p, u) and hence theorem is proved.
The following theorem is achieved due to Theorem 2.4.2, Theorem 3.2.2 and Theorem
3.3.1.
Theorem 3.3.2 For each δ, let (vδ, pδ,gδ,uδ,hδ) denote an optimal solution of Problem
1 and (v, p, u) denote a solution of (2.3.7)-(2.3.11). Then ‖v − vδ‖1,Ωf→ 0, ‖p −
pδ‖0,Ωf→ 0, ‖u− uδ‖1,Ωs → 0, ‖ut − uδ
t‖1,Ωs → 0 as δ → 0.
3.4 Optimality system
We use the Lagrange multiplier rule to derive the first order necessary conditions
that optimal solutions must satisfy.
Let B1 = L2(0, T ;H1f (Ωf ))×L2(0, T ;L2(Ωf ))× Y ×H1(0, T ;H1
s (Ωs))×Z and B2 =
L2(0, T ;H1f (Ωf ))∗ × L2(0, T ;H1
f (Ωf ))× Y ×H1(0, T ;H1s (Ωs))
∗.
where, (·) denotes the dual space. Suppose the linear operator M : B1 → B2 denotes
the constraint operator, i.e., M : B1 → B2 is defined by M(v, p,g,u,h) = (f , z,d,b) if
and only if
∫ T
0[ρf (vt,w)Ωf
+ b(w, p) + af (v,w) + (w, pn− µf∇v · n)Γ0 ] dt
=∫ T
0ρf (f ,w)Ωf
dt ∀ w ∈ H1f (Ωf )
29
∫ T
0b(v, q) dt =
∫ T
0b(z, q) dt ∀ q ∈ L2(Ωf )∫ T
0[(v, s)Γ0 − (g, s)Γ0 ] dt =
∫ T
0(d, s)Γ0 dt ∀ s ∈ H−1/2(Γ0)
and∫ T
0[ρs(utt, θ)Ωs + εa(ut, θ) + as(u, θ)− (h, θ)Γ0 ] dt
=∫ T
0ρs(b, θ)Ωs dt ∀ θ ∈ H1
s (Ωs)
Note that the constraint equations (3.2.11)-(3.2.14) can simply be expressed as
M(v, p,g,u,h) = (f , 0, 0,b)
The operator M ′(v, p,g,u,h) ∈ L(B1, B2) may be defined as follows:
M ′(v, p,g,u,h) · (Λ, y,k, χ, l) = (f , z, d, b) if and only if∫ T
0[ρf (Λt,w)Ωf
+ b(w, y) + af (Λ,w) + (w, yn− µf∇Λ · n)Γ0 ] dt
=∫ T
0ρf (f ,w)Ωf
dt ∀ w ∈ H1f (Ωf )∫ T
0b(Λ, q) dt =
∫ T
0b(z, q) dt ∀ q ∈ L2(Ωf )∫ T
0[(Λ, s)Γ0 − (k, s)Γ0 ] dt =
∫ T
0(d, s)Γ0 dt ∀ s ∈ H−1/2(Γ0)
and∫ T
0[ρs(χtt, θ)Ωs + εa(χt, θ) + as(χ, θ)− (l, θ)Γ0 ] dt
=∫ T
0ρs(b, θ)Ωs dt ∀ θ ∈ H1
s (Ωs)
We also have that the operator Jδ ∈ L(B1, B2) may be defined by
Jδ′(v, p,g,u,h) · (Λ,y,k, χ, l) = a for (Λ,y,k, χ, l) ∈ B1
if and only if∫ T
0(ut − g, χt − k)Γ0 dt
+∫ T
0(pn− µf∇v · n− h, yn− µf∇Λ · n− l)Γ0 dt
+δ∫ T
0(g,k)Γ0 dt+ δ
∫ T
0(∇Γg,∇Γk)Γ0 dt+ δ
∫ T
0(gt,kt)Γ0 dt
+δ∫ T
0(h, l)Γ0 dt = a
30
Suppose (v, p, g, u, h) ∈ B1 is an optimal solution of Problem1. Then there exists a
nonzero Lagrange multiplier (µ, φ, τ, η) ∈ B∗2 satisfying the Euler equation.
−Jδ′(v, p,g,u,h) · (Λ,y,k, χ, l)
+ < (µ, φ, τ, η),M ′(v, p,g, t,u,h) · (Λ,y,k, χ, l) >= 0
∀(Λ,y,k, χ, l) ∈ B1 (3.4.23)
where, < ·, · > denotes the duality pairing between B2 and the dual space B∗2 .
∫ T
0[ρf (Λt, µ)Ωf
+ b(µ, y) + af (Λ, µ) + (µ, yn− µf∇Λ · n)Γ0 + b(Λ, φ)
+(Λ, τ)Γ0 − (k, τ)Γ0 + ρs(χtt, η)Ωs + εa(χt, η) + as(χ, η)− (l, η)Γ0 ] dt
=∫ T
0[(ut − g, χt − k)Γ0
+(pn− µf∇v · n− h, yn− µf∇Λ · n− l)Γ0 ] dt
+δ∫ T
0(g,k)Γ0 dt+ δ
∫ T
0(∇Γg,∇Γk)Γ0 dt+ δ
∫ T
0(gt,kt)Γ0 dt
+δ∫ T
0(h, l)Γ0 dt ∀(Λ,y,k, χ, l) ∈ B1 (3.4.24)
We may rewrite (3.4.24) in the form
−ρf (µt,Λ)Ωf+ af (Λ, µ) + b(Λ, φ) (3.4.25)
+(Λ, τ)Γ0 − (µf∇Λ · n, µ)
= −(pn− µf∇v · n− h, µf∇Λ · n) ∀ Λ ∈ H1f (Ωf )
b(µ, y) = 0 ∀ y ∈ L2(Ωf ) (3.4.26)
(µ, yn)Γ0 = (pn− µf∇v · n− h, yn)Γ0 ∀ y ∈ L20(Ωf ) (3.4.27)
µ|t=T = 0 (3.4.28)
ρs(ηtt, χ)Ωs + εa(χt, η) + as(χ, η)
= −(utt − gt, χ)Γ0 ∀ χ ∈ H1s (Ωs) (3.4.29)
η|t=T = 0 (3.4.30)
ηt|t=T = 0 (3.4.31)
31
(k, τ)Γ0 = (ut − g,k)Γ0 − δ(g,k)Γ0 (3.4.32)
−δ(∆Γg,∆Γk)Γ0 + δ(gtt,k)Γ0 ∀ k ∈ Y
(l, η)Γ0 = (pn− µf∇v · n− h, l)Γ0 − δ(h, l)Γ0 ∀ l ∈ Z (3.4.33)
Thus, solutions of the optimal problem are determined by solving the system (3.2.11)-
(3.2.17) and (3.4.25)-(3.4.33). This system of equations is called the optimality system.
We may replace (3.4.25) and (3.4.32) by
−ρf (µt,Λ)Ωf+ af (Λ, µ) + b(Λ, φ) (3.4.34)
+(Λ, φn− µf∇µ · n)Γ0 = 0 ∀ Λ ∈ H1f (Ωf )
(k, φn− µf∇µ · n)Γ0 = (ut − g,k)Γ0 − δ(g,k)Γ0 (3.4.35)
−δ(∆Γg,∆Γk)Γ0 + δ(gtt,k)Γ0 ∀ k ∈ Y
32
The optimality system is a weak formulation corresponding to
ρfvt +∇p− µf∆v = ρf f in Ωf
∇ · v = 0 in Ωf
v = 0 on Γf
v = g on Γ0
ρsutt − εµs∆ut − µs∆u− (λs + µs)∇(∇ · u) = ρsb in Ωs
u = 0 on Γs
εµs∇ut · n + µs∇u · n + (λs + µs)(∇ · u)n = h on Γ0
v|t=0 = v0 in Ωf
u|t=0 = u0 in Ωs
ut|t=0 = u1 in Ωs
−ρfµt +∇φ− µf∆µ = 0 in Ωf
∇ · µ = 0 in Ωf
µ = 0 on Γf
µ = pn− µf∇v · n− h on Γ0
ρsηtt − εµs∆ηt − µs∆η − (λs + µs)∇(∇ · η) = 0 in Ωs
η = 0 on Γs
εµs∇ηt · n + µs∇η · n + (λs + µs)(∇ · η)n = −(utt − gt) on Γ0
µ|t=T = 0 in Ωf
η|t=T = 0 in Ωs
ηt|t=T = 0 in Ωs
(1 + δ)g − δ∆Γg − δgtt = −φn + µf∇µ · n + ut on Γ0
(1 + δ)h = (pn− µf∇v · n− η) on Γ0
(3.4.36)
33
3.5 Sensitivity derivatives
The optimality system (3.2.11)-(3.2.17) and (3.4.25)-(3.4.33) may also be derived
using sensitivity derivatives instead of the Lagrange multiplier rule. The first derivatives
∂Jδ
∂g,∂Jδ
∂hof Jδ are defined through their actions on variations g and h as follows:
<∂Jδ
∂g, g >=
∫ T
0(ut − g,−g)Γ0dt (3.5.37)
+∫ T
0(pn− µf∇v · n− h, pn− µf∇v · n)Γ0dt
+δ∫ T
0(g, g)Γ0dt+ δ
∫ T
0(∇Γg,∇Γg)Γ0dt
+δ∫ T
0(gt, gt)Γ0dt (3.5.38)
where p, v are solutions of
ρf (vt,w)Ωf+ b(w, p) + af (v,w) (3.5.39)
+(w, pn− µf∇v · n)Γ0 = 0 ∀ w ∈ H1f (Ωf )
b(v, q) = 0 ∀ q ∈ H(Ωf ) (3.5.40)
(v, s)Γ0 − (g, s)Γ0 = 0 ∀ s ∈ H−1/2(Γ0) (3.5.41)
v|t=0 = 0 (3.5.42)
Setting Λ = v , yn = pn− µf∇v · n in (3.4.34),(3.4.27) and w = µ in (3.5.39) and
from (3.5.37)
<∂Jδ
∂g, g >=
∫ T
0(ut − g,−g)Γ0dt
+∫ T
0(g, φn− µf∇µ · n)Γ0dt+ δ
∫ T
0(g, g)Γ0dt (3.5.43)
−δ∫ T
0(∆Γg, g)Γ0dt− δ
∫ T
0(gtt, g)Γ0dt
<∂Jδ
∂h, h >=
∫ T
0(ut − g, ut)Γ0dt
+∫ T
0(pn− µf∇v · n− h,−h)Γ0dt+ δ
∫ T
0(h, h)Γ0dt (3.5.44)
34
where u is a solution of
ρs(utt, θ)Ωs + εa(ut, θ) + as(u, θ) = (h, θ)Γ0 ∀ θ ∈ H1s (Ωs) (3.5.45)
u|t=0 = 0 (3.5.46)
ut|t=0 = 0 (3.5.47)
Setting χ = u in (3.4.29) and θ = η in (3.5.45) and from (3.5.44)
<∂Jδ
∂h, h >=
∫ T
0(h, η)Γ0dt
+∫ T
0(pn− µf∇v · n− h,−h)Γ0dt+ δ
∫ T
0(h, h)Γ0dt (3.5.48)
Thus the first order necessary conditions ∂Jδ∂g
= 0 and ∂Jδ∂h
= 0 yield that
(1 + δ)g − δ∆Γg − δgtt = −φn + µf∇µ · n + ut on Γ0
(1 + δ)h = pn− µf∇v · n− η on Γ0
which are the same as in (3.4.36).
Note that equations (3.5.43) and (3.5.48) give an explicit formula for the gradient of
Jδ, i.e.,
∂Jδ
∂g=
∫ T
0[(1 + δ)g − δ∆Γg − δgtt + φn− µf∇µ · n− ut]dt (3.5.49)
∂Jδ
∂h=
∫ T
0[(1 + δ)h− pn + µf∇v · n + η]dt (3.5.50)
3.6 Gradient method
In this section, we study a gradient method to solve the optimization system (3.2.11)-
(3.2.17) and (3.4.25)-(3.4.33). The simple gradient method we consider is defined as
follows.
35
Given a starting guess g(0) and h(0)
g(n+1)
h(n+1)
=
g(n)
h(n)
− α
∂Jδ∂g
∂Jδ∂h
for n = 1, 2, · · · ,
where α is a step size. Combining with (3.5.49) and (3.5.50) yields, for n = 1, 2, · · · g(n+1)
h(n+1)
=
g(n)
h(n)
− α
∫ T0 [(1 + δ)g − δ∆Γg − δgtt + φn− µf∇µ · n− ut]dt∫ T0 [(1 + δ)h− pn + µf∇v · n + η]dt
Algorithm 3.6.1 1. Choose g(0) and h(0).
2. For n=0,1,2,· · ·,
(a) determine v(n), p(n) and u(n) from
ρf (v(n)t ,w)Ωf
+ b(w, p(n)) + af (v(n),w)
+(w, p(n)n− µf∇v(n) · n)Γ0 = ρf (f ,w)Ωf∀ w ∈ H1
f (Ωf )
b(v(n), q) = 0 ∀ q ∈ H2(Ωf )
(v(n), s)Γ0 − (g(n), s)Γ0 = 0 ∀ s ∈ H−1/2(Γ0)
v(n)|t=0 = v0 in Ωf
ρs(u(n)tt , θ)Ωs + εa(u
(n)t , θ) + as(u
(n), θ) = ρs(b, θ)Ωs + (h(n), θ)Γ0
∀ θ ∈ H1s (Ωs)
u(n)|t=0 = u0 in Ωs
u(n)t |t=0 = u1 in Ωs
36
(b) determine µ(n),φ(n) and η(n) from
−ρf (µ(n)t ,Λ)Ωf
+ af (Λ, µ(n)) + b(Λ, φ(n))
+(Λ, φnn− µf∇µ(n) · n)Γ0 = 0 ∀ Λ ∈ H1f (Ωf )
b(µ(n), q) = 0 ∀ q ∈ L2(Ωf )
(µ(n), s)Γ0 = (p(n)n− µf∇v(n) · n− h(n), s)Γ0 ∀ s ∈ H−1/2(Γ0)
µ(n)|t=T = 0
ρs(η(n)tt , χ)Ωs + εa(χt, η
(n)) + as(χ, η(n)) = −(u
(n)tt − g
(n)t , χ(n))Γ0
∀ χ ∈ H1s (Ωs)
η(n)|t=T = 0
η(n)t |t=T = 0
(c) determine g(n+1) and h(n+1) from
g(n+1) = g(n)
−α∫ T0 [(1 + δ)g(n) − δ∆Γg
(n) − δg(n)tt + φ(n)n− µf∇µ(n) · n− u
(n)t ]dt
h(n+1) = h(n) − α∫ T0 [(1 + δ)h(n) − p(n)n + µf∇v(n) · n + η(n)]dt
The following result is useful to determining sufficient conditions for the convergence
of the gradient method.
Theorem 3.6.1 Let X be a Hilbert space equipped with the inner product (·, ·)X and
norm ‖ · ‖X . Suppose M is a functional on x such that
1. M has a local minimum at x and is twice differentiable in an open ball B centered
at x;
2. | <M′′(u), (x, y) > | ≤M‖x‖X‖y‖Y , ∀u ∈ B, x ∈ X, y ∈ Y ;
3. | <M′′(u), (x, x) > | ≥ m‖x‖2X , ∀u ∈ B, x ∈ X,
37
where M and m are positive constants. Let R denote the Riesz map. Choose x(0)
sufficiently close to x and choose a sequence ρn such that 0 < ρ∗ ≤ ρn ≤ ρ∗ < 2m/M2.
Then the sequence x(n) defined by
x(n) = x(n−1) − ρnRM′(x(n−1)), for n = 1, 2, · · · ,
converges to x.
We examine the second derivatives of J to determine the constants M and m.
<∂2J∂g2
, (g,g) >=∫ T
0(g, g)Γ0dt
+∫ T
0(pn− µf∇v · n, pn− µf∇v · n)Γ0dt
+δ∫ T
0(g, g)Γ0dt+ δ
∫ T
0(∇Γg,∇Γg)Γ0dt+ δ
∫ T
0(gt, gt)Γ0dt
<∂2J∂g∂h
, (g,h) >=∫ T
0(−g,ut)Γ0dt
+∫ T
0(pn− µf∇v · n,−h)Γ0dt
<∂2J∂h∂g
, (h,g) >=∫ T
0(ut,−g)Γ0dt
+∫ T
0(−h, pn− µf∇v · n)Γ0dt
<∂2J∂h2
, (h,h) >=∫ T
0(ut, ut)Γ0dt+ (1 + δ)
∫ T
0(h, h)Γ0dt
where, v, p,u, v, p and u are solutions of
ρf (vt,w)Ωf+ b(w, p) + af (v,w)
+(w, pn− µf∇v · n)Γ0 = 0 ∀ w ∈ H1f (Ωf )
b(v, q) = 0 ∀ q ∈ L2(Ωf )
(v, s)Γ0 − (g, s)Γ0 = 0 ∀ s ∈ H−1/2(Γ0)
v|t=0 = 0
(3.6.51)
38
ρf (vt,w)Ωf+ b(w, p) + af (v,w)
+(w, pn− µf∇v · n)Γ0 = 0 ∀ w ∈ H1f (Ωf )
b(v, q) = 0 ∀ q ∈ L2(Ωf )
(v, s)Γ0 − (g, s)Γ0 = 0 ∀ s ∈ H−1/2(Γ0)
v|t=0 = 0
(3.6.52)
ρs(utt, θ)Ωs + εa(ut, θ) + as(u, θ) = (h, θ)Γ0 ∀ θ ∈ H1s (Ωs)
u|t=0 = 0
ut|t=0 = 0
(3.6.53)
ρs(utt, θ)Ωs + εa(ut, θ) + as(u, θ) = (h, θ)Γ0 ∀ θ ∈ H1s (Ωs)
u|t=0 = 0
ut|t=0 = 0
(3.6.54)
Then,
(g h)
∂2J∂g2
∂2J∂g∂h
∂2J∂h∂g
∂2J∂h2
g
h
= <
∂2J∂g2
, (g,g) > + <∂2J∂g∂h
, (g,h) > + <∂2J∂h∂g
, (h,g) > + <∂2J∂h2
, (h,h) >
=∫ T
0(g, g)Γ0dt+
∫ T
0(pn− µf∇v · n, pn− µf∇v · n)Γ0dt
+δ∫ T
0(g, g)Γ0dt+ δ
∫ T
0(∇Γg,∇Γg)Γ0dt+ δ
∫ T
0(gt, gt)Γ0dt
+∫ T
0(−g,ut)Γ0dt+
∫ T
0(pn− µf∇v · n,−h)Γ0dt
+∫ T
0(ut,−g)Γ0dt+
∫ T
0(−h, pn− µf∇v · n)Γ0dt
+∫ T
0(ut, ut)Γ0dt+ (1 + δ)
∫ T
0(h, h)Γ0dt
=∫ T
0(ut − g, ut − g)Γ0dt
39
+∫ T
0(pn− µf∇v · n− h, pn− µf∇v · n− h)Γ0dt
+δ∫ T
0(g, g)Γ0dt+ δ
∫ T
0(∇Γg,∇Γg)Γ0dt+ δ
∫ T
0(gt, gt)Γ0dt
+δ∫ T
0(h, h)Γ0dt
We show that ∣∣∣∣∣∣∣∣(g, h)
∂2J∂g2
∂2J∂g∂h
∂2J∂h∂g
∂2J∂h2
g
h
∣∣∣∣∣∣∣∣ ≤M‖x‖‖y‖
where x = (g, h)T , y = (g,h)T , ‖x‖ =√‖g‖2
Y + ‖h‖2Z and ‖y‖ =
√‖g‖2
Y + ‖h‖2Z
First, we obtain bounds for v and p. Let v, p satisfy (3.6.51), which is a weak
formulation of
ρfvt +∇p− µf∆v = 0 in Ωf
∇ · v = 0 in Ωf
v = g on Γ
v|t=0 = 0 in Ωf
(3.6.55)
where, Γ = Γf ∪ Γ0 and
g =
g on Γ0
0 on Γf
We introduce a space
Hs∆(Ωf ) = v ∈ Hs(Ωf ) : ∆v ∈ L2(Ωf )
with a norm
‖v‖s∆,Ωf= ‖v‖s,Ωf
+ ‖∆v‖0,Ωf
Then, for g ∈ L2(0, T ; H1(Γ)) ∩H1(0, T ; L2(Γ)), there is an extension R0g ∈
L2(0, T ; H3/2∆ (Ωf ))∩H1(0, T ; H
1/2∆ (Ωf )) such that∇·(R0g) = 0 in Ωf and R0g = g on Γ.
40
And
‖R0g‖3/2∆,Ωf+ ‖R0gt‖1/2∆,Ωf
≤ ‖g‖1,Γ + ‖gt‖0,Γ. (3.6.56)
Write v = v +R0g, then we have
ρf vt +∇p− µf∆v = −ρfR0gt + µf∆R0g in Ωf
∇ · v = 0 in Ωf
v = 0 on Γ
v|t=0 = −ρfR0g|t=0 in Ωf
(3.6.57)
Notice that −ρfR0gt + µf∆R0g ∈ L2(Ωf ) and we show that R0g|t=0 ∈ H1(Ωf ).
[X, Y ]θ denotes the interpolation space equipped with a norm
‖u‖[X,Y ]θ = ‖Λ1−θu‖Y
where Λ : Y → Y is an operator such that ‖u‖X = ‖Λu‖Y . And define a space W 1(0, T )
as
W 1(0, T ) = u ∈ L2(0, T ;X), ut ∈ L2(0, T ;Y )
Theorem 3.6.2 (Trace theorem) There exists a bounded operator γ0 : W 1(0, T ) →
[X, Y ]1/2 such that γ0u = u|t=0 and ‖γ0u‖[X,Y ]1/2≤ C‖u‖W 1(0,T )
proof: See Vivsik, Fursikov [40].
Lemma 3.6.3 If X = H3/2(Ωf ) and Y = H1/2(Ωf ) then [X, Y ]1/2 = H1(Ωf )
proof:
(Su1,u2)1/2,Ωf=
∫Ωf
(1 + |ξ|2)1/2Su1u2dξ
(u1,u2)3/2,Ωf=
∫Ωf
(1 + |ξ|2)3/2u1u2dξ
41
(Su1,u2)1/2,Ωf= (u1,u2)3/2,Ωf
implies S = 1 + |ξ|2 and hence Λ = (1 + |ξ|2)1/2 i.e.
‖Λu‖1/2,Ωf= ‖u‖3/2,Ωf
. By the definition of the interpolation space,
‖u‖2[X,Y ]1/2
= ‖Λ1/2u‖21/2,Ωf
= ‖u‖21,Ωf
Therefore, [X,Y ]1/2 = H1(Ωf )
If X = H3/2(Ωf ) and Y = H1/2(Ωf ) then we may rewrite the Trace Theorem as
follows :
There exists a bounded operator γ0 : W 1(0, T ) → H1(Ωf ) such that γ0u = u|t=0 and
‖γ0u‖21,Ωf
≤ C(∫ T0 ‖u‖2
3/2,Ωfdt+
∫ T0 ‖ut‖2
1/2,Ωfdt)
Since R0g ∈ L2(0, T ; H3/2∆ (Ωf )) ∩ H1(0, T ; H
1/2∆ (Ωf )), we have R0g|t=0 ∈ H1(Ωf ))
and
‖R0g|t=0‖21,Ω ≤ C(
∫ T
0‖R0g‖2
3/2,Ωfdt+
∫ T
0‖R0gt‖2
1/2,Ωfdt) (3.6.58)
Then (3.6.57) is an evolution Stokes equation and hence we obtain
v ∈ L2(0, T ; H2(Ωf ))
vt ∈ L2(0, T ; L2(Ωf ))
p ∈ L2(0, T ;H1(Ωf ))
Moreover,
∫ T
0‖vt‖2
0,Ωfdt ≤ ρ2
f
∫ T
0‖R0gt‖2
0,Ωfdt
µ2f
∫ T
0‖∆R0g‖2
0,Ωfdt+ ‖R0g|t=0‖2
1,Ω
≤ C(‖g‖21,Γ + ‖gt‖2
0,Γ)
by (3.6.56) and (3.6.58).
We then come back to (3.6.57) and we apply the regularity theorem of stationary
42
case. For almost every t in [0, T ],
∇p− µf∆v = −ρf vt − ρfR0gt + µf∆R0g in Ωf
∇ · v = 0 in Ωf
v = 0 on Γ
so that v(t) ∈ H2(Ω) and p(t) ∈ H1(Ω). Moreover,
∫ T
0‖v‖2
2,Ωfdt+
∫ T
0‖p‖2
1,Ωfdt
≤ C(ρ2f
∫ T
0‖vt‖2
0,Ωfdt+ ρ2
f
∫ T
0‖R0gt‖2
0,Ωfdt+ µ2
f
∫ T
0‖∆R0g‖2
0,Ωfdt)
≤ C(‖g‖21,Γ + ‖gt‖2
0,Γ)
Now we get bound for v.
∫ T
0‖v‖2
3/2∆,Ωfdt ≤
∫ T
0‖v‖2
3/2∆,Ωfdt+
∫ T
0‖R0g‖2
3/2∆,Ωfdt
≤ 2(‖v‖23/2,Ωf
dt+ ‖∆v‖20,Ωf
dt) +∫ T
0‖R0g‖2
3/2∆,Ωfdt
≤ C(‖g‖21,Γ + ‖gt‖2
0,Γ)
Then
∫ T
0‖pn‖2
0,Γ0dt ≤
∫ T
0‖p‖2
1,Ωfdt
≤ C(‖g‖21,Γ + ‖gt‖2
0,Γ) = C‖g‖2Y
and
∫ T
0‖∇v · n‖2
0,Γ0dt ≤
∫ T
0‖v‖2
3/2∆,Ωfdt
≤ C(‖g‖21,Γ + ‖gt‖2
0,Γ) = C‖g‖2Y
Now, we obtain a bound for u. Suppose u is a solution of (3.6.53). Then
ρs(utt, θ)Ωs + εa(ut, θ) + as(u, θ) = (h, θ)Γ0 ∀ θ ∈ H1s (Ωs) (3.6.59)
43
Setting θ = ut in (3.6.59) to get
suptρs‖ut‖2
0,Ωs+ εKa
∫ T
0‖ut‖2
1,Ωsdt+ sup
tKa‖u‖2
1,Ωs
≤ C∫ T
0‖h‖2
0,Γ0dt = C‖h‖2
Z
We determine the constants M using bounds we have obtained.∣∣∣∣∣∣∣∣(g, h)
∂2J∂g2
∂2J∂g∂h
∂2J∂h∂g
∂2J∂h2
g
h
∣∣∣∣∣∣∣∣
= |∫ T
0(ut − g, ut − g)Γ0dt
+∫ T
0(pn− µf∇v · n− h, pn− µf∇v · n− h)Γ0dt
+δ∫ T
0(g, g)Γ0dt+ δ
∫ T
0(∇Γg,∇Γg)Γ0dt+ δ
∫ T
0(gt, gt)Γ0dt
+δ∫ T
0(h, h)Γ0dt|
≤ 2(∫ T
0‖ut‖2
1,Ωsdt+
∫ T
0‖g‖2
0,Γ0dt)1/2(
∫ T
0‖ut‖2
1,Ωsdt+
∫ T
0‖g‖2
0,Γ0dt)1/2
+2(∫ T
0‖pn‖2
0,Γ0dt+ µ2
f
∫ T
0‖∇v · n‖2
0,Γ0dt+
∫ T
0‖h‖2
0,Γ0dt)1/2
(∫ T
0‖pn‖2
0,Γ0dt+ µ2
f
∫ T
0‖∇v · n‖2
0,Γ0dt+
∫ T
0‖h‖2
0,Γ0dt)1/2
+δ(∫ T
0‖g‖2
0,Γ0dt)1/2(
∫ T
0‖g‖2
0,Γ0dt)1/2
+δ(∫ T
0‖∇Γg‖2
0,Γ0dt)1/2(
∫ T
0‖∇Γg‖2
0,Γ0dt)1/2
+δ(∫ T
0‖gt‖2
0,Γ0dt)1/2(
∫ T
0‖gt‖2
0,Γ0dt)1/2
+δ(∫ T
0‖h‖2
0,Γ0dt)1/2(
∫ T
0‖h‖2
0,Γ0dt)1/2
≤ 2(C‖h‖2Z + ‖g‖2
Y )1/2(C‖h‖2Z + ‖g‖2
Y )1/2
+2(C‖g‖2Y + µ2
fC‖g‖2Y + ‖h‖2
Z)1/2(C‖g‖2Y + µ2
fC‖g‖2Y + ‖h‖2
Z)1/2
+δ‖g‖Y ‖g‖Y + δ‖h‖Y ‖h‖Y
≤ M‖x‖‖y‖
where, M = 2C + 2(1 + µ2f )C + δ.
44
Setting g = g and h = h to determine the constant m.
(g,h)
∂2J∂g2
∂2J∂g∂h
∂2J∂h∂g
∂2J∂h2
g
h
=
∫ T
0
∫Γ0
(ut − g)2 dΓ dt+∫ T
0
∫Γ0
(pn− µf∇v · n− h)2dΓ dt
+δ∫ T
0
∫Γ0
g2 dΓ dt+ δ∫ T
0
∫Γ0
(∇Γg)2 dΓ dt
+δ∫ T
0
∫Γ0
g2t dΓ dt+ δ
∫ T
0
∫Γ0
h2 dΓ dt
= J (v, p,g,u,h) + δ‖g‖2Y + δ‖h‖2
Z ≥ δ‖x‖2
where m = δ.
This proves the convergence of the Gradient method.
45
BIBLIOGRAPHY
[1] R. Adams, Sobolev Spaces, Academic, New York, 1975.
[2] A. Bermudez, R. Duran and R. Rodriguez, Finite element analysis of compressible
and incompressible fluid-solid systems. Math. Comp. 67(1998), no.221, pp.111-136.
[3] F.Blom, A monolithical fluid-structure interaction algorithm applied to the piston
problem. Comput. Methods Appl. Mech. Engrg.167(1998), no.3-4, pp.369-391
[4] J. Boujot, Mathematical formulation of fluid-structure interaction problems.
RAIRO Model. Math. Anal. Numer. 21(1987), no.2. pp.239-260.
[5] K. Chrysafinos, M. D. Gunzburger and L. S. Hou, Error estimates for semidiscrete
finite element approximations of linear and semilinear parabolic equations under
minimal regularity assumptions, to appear.
[6] P. Ciarlet, The finite element method for elliptic problems, North-Holland, Amster-
dam, 1978.
[7] C. Conca and M. Duran, A numerical study of a spectral problem in solid-fluid type
structures. Numer. Methods Partial Differential Equations 11(1995), no.4, pp.423-
444.
[8] C. Conca, J. Martin and J. Tucsnak, Motion of a rigid body in a viscous fluid. C.
R. Acad. Sci. Paris Ser. I Math. 328(1999), no.6, pp.473-478.
46
[9] R. Dautray and J.-L. Lions, Mathematical analysis and numerical methods for sci-
ence and technology, Volume 1 Physical origins and classical methods, Springer,
Berlin, 1990.
[10] R. Dautray and J.-L. Lions, Mathematical analysis and numerical methods for sci-
ence and technology, Volume 6 Evolution problems II, Springer, Berlin, 1988.
[11] S. Dasser, A Penalization method for the homogenization of a mixed fluid-structure
problem. C. R. Acad. Sci. Paris Ser. I Math. 320(1995), n0.6, pp.759-764.
[12] B. Desjardins and M. J. Esteban, On weak solutions for fluid-rigid structure inter-
action: compressible and incompressible models. Comm. Partial Differential Equa-
tions 25(2000), no.7-8, pp.1399-1413
[13] D. Errate, M. J. Esteban and Y. Maday, Couplage fluid-structure. Un modele
simplifie en dimension 1, C. R. Acad. Sci. Paris Ser. I Math. 318(1994), pp.275-281.
[14] C. Farhat, M. Lesoinne and P. LeTallec, Load and motion transfer algorithms for
fluid/structure interaction problems with non-matching discrete interfaces: momen-
tum and energy conservation, optimal discretization and application to aeroelastic-
ity. Comput. Methods Appl. Mech. Engrg. 157(1998), no.1-2, pp.95-114.
[15] F. Flori and P. Orenga, Fluid-structure interaction:analysis of a 3-D compressible
model. Ann. Inst. H. Poincare anl. Non Lineaire 17(2000), no.6, pp.753-777.
[16] F. Flori and P. Orenga, On a fluid-structure interaction problem. Trends in appli-
cations of mathematics to mechanics (Nice, 1998), pp.293-305
[17] F. Flori and P. Orenga, On a nonlinear fluid-structure interaction problem defined
on a domain depending on time. Nonlinear Anal. 38(1999), no.5, pp.549-569
47
[18] F. Flori and P. Orenga, Analysis of a nonlinear fluid-structure interaction problem
in velocity-displacement formulation. Nonlinear Anal. 35(1999) , no.5, pp.561-587.
[19] L. Gastaldi, Mixed finite element methods in fluid structure system. Numer. Math.
74(1996), no.2, pp.153-176.
[20] V. Girault and P. Raviart, Finite element methods for Navier-Stokes equations,
Springer, Berlin, 1986.
[21] C. Grandmont, Existence and uniqueness for a two-dimensional steady-state fluid-
structure interaction problem. C. R. Acad. Sci. Paris Ser. I Math. 326(1998), no.5,
pp.651-656.
[22] C. Grandmont and Y. Maday, Existence for an unsteady fluid-structure interaction
problem. M2AN Math. Model. Numer. Anal. 34(2000), no.3, pp.609-636.
[23] C. Grandmont and Y. Maday, Existence of solutions for a two-dimensional unsteady
fluid-structure interaction problem. C. R. Acad. Sci. Paris Ser. I Math. 326(1998),
no.4, pp.525-530 no.3, pp.609-636.
[24] M. D. Gunzburger, L. S. Hou and T. Svobodny, Analysis and finite element ap-
proximation of optimal control problems for the stationary Navier-Stokes equations
with Dirichlet control, Math. Model. Numer. Anal. 25(1991).
[25] M. D. Gunzburger, L. S. Hou and T. Svobodny, Optimal control problems for a
class of nonlinear equations with an application to the control of fluids; Optimal
control of viscous flows, ed. by S. Sritharan, SIAM, Philadelphia, 1998, pp.43-62.
[26] M. D. Gunzburger and J. Lee, A domain decomposition method for optimization
problems for partial differential equations, Comput. Math. Appl. 40(2000), pp.177-
192.
48
[27] M. D. Gunzburger and H. K. Lee, An optimization based domain decomposition
method for the Navier-Stokes equations, SIAM J. Numer. Anal. 37(2000), pp.1455-
1480.
[28] M. D. Gunzburger, J. S. Peterson and H. Kwon, An optimization based domain
decomposition method for partial differential equations , Comput. Math. Appl.
37(1999), pp.77-93.
[29] J. G. Heywood and R. Rannacher, Finite element approximations of the nonstation-
ary Navier-Stokes problem, I. Regularity solutions and second-order error estimates
for spatial discretization, SIAM J. Numer. Anal. 19(1982), pp.275-311.
[30] L. S. Hou, Error estimates form semidiscrete finite element approximations of the
Stokes equations under minimal regularity assumptions, to appear.
[31] G. Hsiao, R. Kleinman and G. Roach, Weak solutions of fluid-solid interaction
problems. Math. Nachr. 218(2000), pp.139-163
[32] P. LeTallec and S. Mani, Numerical analysis of a linearised fluid-structure interac-
tion problem, Numer. Math. 87(2000), no.2, pp.317-354
[33] J.-L. Lions and E. Magenes, Non-homogeneous boundary value problems and appli-
cations I, Springer-Verlag, Berlin, 1972.
[34] S. Micu and E. Zuazua, Asymptotics for the spectrum of a fluid/structure hybrid
system arising in the control of noise. SIAM J. Math. Anal. 29(1998), no.4, pp.967-
1001.
[35] R. Rodriguez and J. Solomin, The order of convergence of eigenfrequences in fi-
nite element approximations of fluid-structure interaction problems. Math. Comp.
65(1996), no.216, pp1463-1475.
49
[36] M.Rumpf, On equilibria in the interaction of fluids and elastic solids. Theory of the
Navier-Stokes equations, Ser. Adv. Math. Appl. Sci. 47(1998), pp.136-158.
[37] R. Schulkes, Interations of an elastic solid with a viscous fluid: eigenmode analysis.
J. Comput. Phys. 100(1992), no.2, pp.270-283.
[38] R. Temam, Navier-Sotkes equations, North-Holland, New York, 1977.
[39] V. Thomee, Galerkin finite element methods for parabolic equations, Springer,
Berlin, 1997.
[40] M. J. Vivsik and A.V.Fursikov, Mathematical problems of statistical hydromechan-
ics, Mathematics and its applications in physics and technology 41, 1986.