Control of Nonlinear Distributed Parameter Systems

edited by
Goong Chen, Texas A&M University, College Station, Texas
Irena Lasiecka, University of Virginia, Charlottesville, Virginia
Jianxin Zhou, Texas A&M University, College Station, Texas
 iii

Preface
This volume is an outgrowth of the conference “Advances in Control of
Nonlinear Distributed Parameter Systems”, held on October 22-23, 1999, at
Texas A&M University, College Station, Texas. The conference was jointly
sponsored by the National Science Foundation (NSF), The Institute of Math-
ematics and Its Applications (IMA) and Texas A&M University. Fifty-five
researchers attended and twenty-six talks were delivered during the two-day
event. Ten papers in this volume were written by those conference speakers.
To further broaden the scope and appeal of this volume, we have invited seven
additional papers from experts working in this field. Thus, a total of seventeen
papers have constituted the volume.

The mathematical theory of control is highly interdisciplinary—it is

a part of applied mathematics serving perhaps the most important link
between mathematics and technology: complex systems in aerospace, civil
and mechanical engineering must be controlled in order to achieve designated
mission or operational requirements. Many ultra-modern electronic and
optical devices are also designed for and dedicated to the purpose of acting
as control mechanisms and media, i.e., actuators and sensors. Most of those
devices are inherently nonlinear. The strong interest in mathematical con-
trol problems among mathematicians and engineers alike can be witnessed in
the large number of papers published in the various journals of IEEE and SIAM.

Even though steady progress has been made in the overall study of the
mathematics of control, and wider and wider applications to new problems
have been found, the leading edge of the field, as a mathematical subject, is
indisputably the area of control of distributed parameter systems (DPS). This
area concerns investigation of the control laws, stability and optimization of
systems and feedback syntheses for systems whose states are spatially and/or
temporally distributed and whose governing equations are partial differential
or functional (typically time delay) equations. Studies in the area also include
the associated questions of modelling, identification and estimation, analysis
and design, computation and visualization, etc., of DPS. Rapid progress has
occurred in this area since its inception during the 1960’s and its initial burst
of growth in the 1970’s.

After nearly three decades of research, though many interesting questions

remain open, control theory for linear DPS has attained a certain level of
maturity. The momentum of DPS research is now visibly moving toward
the study of control of nonlinear partial differential equations. Nonlinear
DPS (NDPS) are very much model-dependent. Since comprehensive, unified
theories are virtually nonexistent, research opportunities and challenges are
iv

extraordinarily numerous. Very substantial payoffs from the study of control

and optimization of wide-ranging, application-driven nonlinear DPS in various
areas of high technology may be expected to yield a substantial payoff through
operational economies and enhanced system performance. We hope the present
volume will stimulate active development of the mathematical theory in this
critically important area.

Two major influences are driving the recent sharp surge of interest in control
of nonlinear distributed parameter systems:

(A) Advances in “smart” materials, active actuators and sen-

sors, microelectromechanical systems (MEMS), etc.
Existing advanced, or “smart” materials largely consist of sophisticated
laminates incorporating specialized layers in an overall matrix form; they
are fabricated to achieve a variety of desirable properties. Actuators and/or
sensors consisting of piezoelectric/piezoceramic, opto-thermo-electric materials
or microprocessors can be bonded to external surfaces or embedded within the
layered structure itself. The response of such individual components is totally
nonlinear, resulting in an overall system of nonlinear partial differential equa-
tions as the operative mathematical model. For example, aircraft propellers
and helicopter rotor blades may be designed so that varying pitch is achieved by
torsional actuation within the blade itself rather than by a mechanically articu-
lated mechanism at the point where the blades are mounted. Another example
concerns active noise suppression in aircraft cabins. This is achieved by means
of actuator panels in the cabin walls, acting to achieve cancellation of high
amplitude noise signals propagated through the fuselage. Many more examples
of ultra-modern micromachined elastic structures in diverse applications may
be found in a large number of new technical journals. The complete list of
nonlinear distributed parameter systems finding applications in the area of ad-
vanced materials is much too long for us to cover in any representative way here.

(B) Advances in nonlinear PDEs and dynamical systems

The existing, now almost classical, theory of control of linear DPS is of
rather limited use in the nonlinear arena. New nonlinear methodology for
control, stabilization and optimization needs to be developed for such systems.
During the past thirty years, dramatic breakthroughs in theory and methods
for nonlinear PDEs have been made, including Lax’s entropy solution and
Glimm’s method for hyperbolic conservation laws, The Mountain Pass Lemma
of Ambrosetti and Rabinowitz, the method of viscosity solutions, Hopf
bifurcation phenomena in infinite dimensional spaces studied by Crandall and
Rabinowitz,. . . , enabling researchers to treat an increasing number of genuinely
 v

nonlinear PDEs with confidence. These equations, or systems of equations

often have unstable, multiple solutions, depending on the geometry of the
domain – a totally bewildering situation prior to recent developments. The
emergence of the new field of dynamical systems and chaos has likewise shifted
the focus of attention from the classical qualitative theory of ODEs and PDEs
to that of fractals, strange attractors, randomness, and their manipulations,
control and applications—these are some of the most intensively investigated
topics in the general scientific community at the present time.

Both exogenous and endogenous factors, i.e., (A) and (B), respectively,
above, are simultaneously at work, enriching and propelling the study of
control of nonlinear distributed parameter systems and cross-fertilizing other
intimately allied disciplines. These synergistic effects amply testify to the
timeliness of the publication of this volume.

The chapters in this volume cover interests in various aspects of NDPS.

For example, the paper by Seidman and Antman is related to Category (A)
above. The two papers by Ding and by Li and Zhou involve the application
of the Mountain Pass Lemma and are thus more associated with Category
(B). The paper by Chen, Huang, Juang and Ma studying chaotic phenomena
due to nonlinear boundary conditions has overlapping interests in both (A)
and (B). We hope the wide range of topics in these, and the other papers not
explicitly cited here, will provide a useful reference for the study of nonlinear
distributed parameter systems and stimulate further interest and research in
this important area.

We thank all the contributing authors for their work and and their patience
with repetitive revisions. We are grateful to Dr. Deborah Lockhart at NSF,
Professor Willard Miller, Jr. of IMA, and Professor Richard E. Ewing, Dean of
College of Science at Texas A&M University, for the financial support to the
conference. Finally, we thank Ms. Maria Allegra and Helen Paisner at Marcel
Dekker and Professor M. Zuhair Nashed of the University of Delaware for their
kind assistance in expediting the editorial and publication process.

Goong Chen and Jianxin Zhou

College Station, Texas

Irena Lasiecka
Charlottesville, Virginia
vi

Dedicated to

Professor David L. Russell

on the Occasion of his 60th Birthday

 Contents vii

Preface iii

1. Shape Sensitivity Analysis in Hyperbolic Problems with non

Smooth Domains
John Cagnol and J. Paul Zolesio 1

2. Unbounded Growth of Total Variations of Snapshots of the 1D

Linear Wave Equation due to the Chaotic Behavior of Iterates
of Composite Nonlinear Boundary Reflection Relations
Goong Chen, Tingwen. Huang, Jong Juang and Daowei Ma 15

3. Velocity method and Courant metric topologies in shape

analysis of partial differential equations
Michel Delfour and J. Pual Zolesio 45

4. Nonlinear Periodic Oscillations In Suspension Bridges

Zhonghai Ding 69

5. Canonical Dual Control for Nonconvex Distributed-Parameter

Systems: Theory and Method
David Y. Gao 85

6. Carleman estimate for a parabolic equation in a Sobolev space

of negative order and their applications
Oleg Imanuvilev and Masahiro Yamamoto 113

7. Bilinear control for global controllability of the semilinear

parabolic equations with superlinear terms
Alexander Khapalov 139

8. A Nonoverlapping Domain Decomposition for Optimal Bound-

ary Control of the Dynamic Maxwell System
John E. Lagnese 157

9. Boundary Stabilizibility of a Nonlinear Structural Acoustic

Model Including Thermoelastic Effects
Catherine Lebiedzik 177

10. On Modelling, Analysis and Simulation of Optimal Con-

trol Problems for Dynamic Networks of Euler-Bernoulli-and
Rayleigh-beams
Guenter Leugering and Wigand Rathman 199
viii Contents

11. Local Characterizations of Saddle Points and Their Morse

Indices
Yongxin Li and Jianxin Zhou 233

12. Static Buckling in a Supported Nonlinear Elastic Beam

David Russell and Luther White 253

13. Optimal control of a nonlinearly viscoelastic rod

Thomas Seidman and Stuart Antman 273

14. Mathematical Modeling and Analysis for Robotic Control

Sze-Kai Tsui 285

15. Optimal Control and Synthesis of Nonlinear Infinite Dimen-

sional Systems
Yuncheng You 299

16. Forced Oscillation of The Korteweg-De Vries-Burgers Equa-

tion and Its Stability
Bingyu Zhang 337
Shape Sensitivity Analysis in Hyperbolic Problems
with non Smooth Domains

John Cagnol1 , Université Léonard de Vinci, FST, DER-CS, 92916 Paris La

Défense Cedex, France, E-mail: [email protected]
Jean-Paul Zolésio, CNRS, Ecole des Mines de Paris, 06902 Sophia Antipolis
Cedex, France. E-mail: [email protected]

Abstract

The control with respect to the domain is inherently not linear due to
the non linear structure of the set of domains. In this paper we investigate
the weak shape differentiability of the solution to the generalized wave
equation when the domain has a Lipschitz continuous boundary. By the
means of the “hidden regularity”, a result for C 2 -boundary was obtained
recently, when the right hand side is in L2 . To extend that result to
Lipschitz continuous boundary, we first investigate the regularity of the
solution at the boundary. We need an exact estimate of the L2 -norm
of the normal derivative. Then, we build an increasing sequence of
smooth domains, and we establish the shape differentiability result as a
consequence of the situation for C 2 -boundary.

1 Introduction
The control with respect to the domain is inherently not linear due to the non
linear structure of the set of domains. In this paper we investigate the sensitivity
of the solution of an hyperbolic PDE with respect to the domain. This analysis
is carried out with the wave equation with an homogeneous Dirichlet boundary
condition. The novelty lies in the absence of regularity of the domain with
respect to which the analysis is done. In a sense we extend the result presented
in [6] to the case of Lipschitz-continuous domains.
Let N ≥ 2 be an integer and D be a bounded domain of RN . Throughout
this paper Ω will be an open domain, star-shaped, included in D whose
boundary Γ is assumed to be Lipschitz continuous. Moreover we will assume Ω
has a bounded perimeter. The family of such domains Ω shall be denoted O.

1
At the time this paper was presented, the first author was at the University of Virginia,
Charlottesville, VA. Research supported by the INRIA under grant 1/99017.

1
2 Cagnol and Zolésio

Let T be a non negative real and I = [0, T ] be the time interval. We

note Q =]0; T [×Ω the cylindrical evolution domain and Σ =]0, T [×Γ the lateral
boundary associated to any element Ω of the family O.

1.1 Shape Differentiability

Let E be the set of V ∈ C([0, S]; C 1 (D̄, RN )) with hV, n∂D i = 0 and free
divergence. For any V ∈ E we consider the flow mapping Ts (V ). At the point
x, V has the form as follows:
 
∂
(1) V (s)(x) = Ts ◦ Ts−1 (x)
∂s
For each s ∈ [0, S[, Ts is a one-to-one mapping from D onto D such that
i) T0 = I

ii) s 7→ Ts belongs to C 1 ([0; S[, C 1 (D̄; D̄)) with Ts (∂D) = ∂D

iii) s 7→ Ts−1 belongs to C([0; S[, C 1 (D̄; D̄))

We refer to [8] and [9] for further discussion on such mappings.
The family O is stable under the perturbations Ω 7→ Ωs (V ) = Ts (V )(Ω). We
denote by Qs the perturbed cylinder ]0; T [×Ωs (V ), Γs = ∂Ωs and Σs =]0, T [×Γs
the perturbed lateral boundary.
Let m ≥ 1 be an integer. Let f ∈ L1 (I, H m (D)) with its m-th time-
derivative in L1 (I, L2 (D)). Let ϕ ∈ H m+1 (D) and ψ ∈ H m (D). Let K be a
coercive and symmetric N × N -matrix whose coefficients belong to W 2,∞ (D).
To each element Ω ∈ O we associate the solution y = y(Ω) of the following
problem
 2
 ∂ y − div (K∇y) = f on Q
 t

y = 0 on Σ
(2)

 y(0) = ϕ on Ω

∂t y(0) = ψ on Ω

Throughout this paper we shall note P the operator ∂tt y − div (K∇).
A Galerking method proves

y ∈ H(I, Ω) = H 1 (I, L2 (Ω)) ∩ L2 (I, H01 (Ω))

For any V ∈ E and s ∈ [0; S] we set ys = y(Ωs ) ∈ L2 (Qs ). Following [5], [6],
[13] the mapping Ω 7→ y(Ω) is said to be shape differentiable in L2 (I, H m (D))

(3) ∃Y ∈ C 1 ([0; S], L2 (I, H m (D)))

(4) Y (s, ·, ·)|Qs = y(Ωs )

Sh. Sensitivity Analysis in Hyperbolic Pb. with non Smooth Domains 3

then ∂s Y (0, ·, ·)|Q which is the restriction to Q of the derivative with respect
to the perturbation parameter s at s = 0 is independent of the choice of Y
verifying (3) and (4). (cf. [13]).
Definition 1.1 (shape derivative). The shape derivative is that unique
element  
′ ∂ 
y (Ω; V ) = Y  ∈ L2 (Q)
∂s s=0 (t,x)∈Q

The weak shape differentiability can be defined analogously, replacing (3)

by the existence of Y in C 1 ([0; S], L2σ (I, H m (D))).

1.2 Known Results for C 2 -boundary

When the boundary is C 2 it was proven in [11] that (2) has a unique solution
in
Z m (I, Ω) = ∩m i
i=0 C (I, H
m−i
(Ω))
In [5], [6] the question of the shape differentiability is solved for various
conditions of regularity of the data, but the domain Ω needs to be C 2 . The
main result was
Theorem 1.1 (Cagnol-Zolsio, 1997). Let m be a positive integer and
let Ω be a domain with a C max{m,2} boundary.
i) If m ≥ 1 then the solution to (2) is shape differentiable at Ω, strongly in
L2 (I, H m−1 (D)).

ii) If m = 0 hen the solution to (2) is shape differentiable at Ω, weakly in

L1 (I, L2 (D)).
the shape derivative y ′ ∈ Z m (I, Ω) and is solution to
 2 ′

 ∂t y − div (K∇y ′ ) = 0 on Q
 ′ ∂y
y = − ∂n hV (0), ni on Σ
(5) ′

 y (0) = 0 on Ω

∂t y ′ (0) = 0 on Ω

1.3 Main Result

In this paper we extend the result of theorem 1.1 to the case of Lipschitz
continuous domains Ω. Problem (2) is well-posed and, as we said earlier, the
solution y lies in H(I, Ω). In [7] and [10] it is proven that the normal derivative
belongs to L2 (Σ). That leads to the well-posedness of (5). Hence looking for
the shape derivative in the case of Lipschitz continuous boundaries makes sense.
In this paper we shall prove the following result
Theorem 1.2. When m = 0, the solution to problem (2) is weakly shape
differentiable at Ω in L1 (I, L2 (D)). The shape derivative y ′ belongs to H(I, Ω)
and is solution to (5).
4 Cagnol and Zolésio

Remark 1.1. When m ≥ 1, the result can be improved to a weak

differentiability in L∞ (I, L2 (D)).

2 Mollification of the Domain

Given a Lipschitz continuous domain Ω, we build an increasing sequence
of smooth sub-domains converging to Ω with Haussdorf convergence of the
boundaries. See also [12].

2.1 Properties of Lipschitz Continuous Domains

Definition 2.1. An open set Ω ⊂ RN is said to have the cone property if
π
∃R > 0, ∃θ ∈]0, [, ∀x ∈ ∂Ω, ∃d, Cx (R, θ, d) ⊂ Ω
2
where Cx (R, θ, d) is the interior of a cone of revolution with the vertex at x,
height R cos(θ/2) and the axis pointing toward the versor d.
When Ω has a Lipschitz continuous boundary then Ω and RN r Ω have
the cone property (cf. [1], [2]). Let R(Ω) and θ(Ω) be the parameters arising
from the cone condition on Ω and R(RN r Ω) and θ(RN r Ω) be the parameters
arising from the cone condition on RN rΩ. We note R = min(R(Ω), R(RN rΩ))
and θ = min(θ(Ω), θ(RN r Ω)).
Remark 2.1. The reals R and θ do not depend on x.
Lemma 2.1. Let |X| denote the measure of X,
  
− + 1 M −  1  M +
∃M > 0, ∃M < 1, ∀κ ≥ , ∀x ∈ ∂Ω, N ≤ Ω ∩ B x, ≤ N
R κ κ  κ

Proof. Let x ∈ ∂Ω, the cone property yields the existence of a versor d such
that Cx ( κ1 , θ, d) ⊂ Ω. Since κ1 < R we get
   
1 1
Cx , θ, d ⊂ Ω ∩ B x,
κ κ

Let B(p) be the volume of the p-th dimensional ball of radius 1. We refer to [3,
pp. 208–210] for an expression of B(p) as a function of p. The volume of the
p-th dimensional ball of radius r is B(p)r p . Then, the volume of Cx ( κ1 , θ, d) is
1 1 1 N −1
N κ cos(θ/2)B(N − 1)( κ ) hence
    N −1
 
Ω ∩ B x, 1  ≥ 1 1 cos(θ/2)B(N − 1) 1
 κ  Nκ κ

therefore   
  −
Ω ∩ B x, 1  ≥ M
 κ  κN
Sh. Sensitivity Analysis in Hyperbolic Pb. with non Smooth Domains 5

with M − = N1 cos(θ/2)B(N − 1).

Considering the cone property for RN r Ω yields the existence of M > 0
such that (RN r Ω) ∩ B(x, κ1 )| ≥ κMN . Let M + = 1 − M , we obtain
   +
 
Ω ∩ B x, 1  ≤ M
 κ  κN

Let χ be the characteristic function of Ω and (ρκ ) be a mollifier. Let us

note ξκ = χ ∗ ρκ .
Proposition 2.1. There exists M − > 0 and M + < 1 such that
1
∀κ ≥ , ∀x ∈ ∂Ω, M − ≤ ξκ (x) ≤ M +
R
R R
Proof. One has ξκ (x) = R2 χ(t)ρκ (t − x) dt hence ξκ (x) = Ω ρκ (t − x) dt.
thus Z
ξκ (x) = ρκ (t − x) dt
1
Ω∩B(x, κ )

Using the symmetry property of ρκ we get

 
Ω ∩ B(x, 1 ) Z
κ
ξκ (x) =   ρκ (t) dt
B(x, 1 ) 1
B(0, κ )
κ

Lemma 2.1 applies and gives the result.

Lemma 2.2. supp ξκ = Ω + B(0, κ1 ) and supp (1 − ξκ ) = (RN r Ω) + B(0, κ1 )
Proof. The lemma is a consequence of supp (χ ∗ ρκ ) ⊂ supp χ + supp ρκ . We
use χ ≥ 0 and ρκ ≥ 0 to prove the first equality. The second equality can be
proven by the same techniques.
Proposition 2.2. Let κ ≥ R1 and x ∈ RN then

ξκ (x) > M + =⇒ x ∈ Ω

ξκ (x) < M − =⇒ x 6∈ Ω
Proof. From proposition 2.1 we have

ξκ−1 (]M + , +∞[) ∩ ∂Ω = ∅

therefore ξκ−1 (]M + , +∞[) ⊂ Ω or ξκ−1 (]M + , +∞[) ⊂ RN r Ω. Elements x of Ω

whose distance to the boundary is more than κ1 satisfy ξκ (x) = 1 thus

ξκ−1 (]M + , +∞[) ⊂ Ω

Analogous arguments show that ξκ−1 (] − ∞, M − [) ⊂ RN r Ω.

6 Cagnol and Zolésio

2.2 Definitions and Preliminary Results

Let Gκ ⊂ RN × R be the graph of ξκ . Since ξκ is C ∞ , the set Gκ is a C ∞
manifold. We note
π1 : RN × R : (x, y) 7→ x
π2 : RN × R : (x, y) 7→ y
The restriction of π2 to Gκ is injective. We note

−1 
Γ(κ, t) = π1 ◦ π2|Gκ (t) ⊂ RN

Lemma 2.3. Let κ be a positive integer and α and β be two reals such that
0 ≤ α < β ≤ 1. There exists t ∈]α, β[ such that Γ(κ, t) is C ∞ .
Proof. From the Sard’s theorem, the image of the critical points of π2κ has
measure 0 in R. Hence there exists t ∈]α, β[ such that (π2 |Gκ )−1 is not critical,
therefore (Γ(κ, t), t) is regular and Γ(κ, t) is C ∞ .

For a real t provided by lemma 2.3, let Ω(κ, t) = (π1 ◦ (π2 |Gκ )−1 )(]t, +∞[) ⊂
RN be the level set, then ∂Ω(κ, t) = Γ(κ, t).
Corollary 2.1. Under the hypothesis of lemma 2.3, there exists t ∈]α, β[
such that Ω(κ, t) is C ∞ .

2.3 Construction of a Sequence

The purpose of this section is to build an isotonic sequence of domains (Ωk )k≥0 ,
whose projective limit is Ω. Let α > M + .

Construction of the first term: Let κ0 be an integer larger that R1 . Let

β0 = 1, from lemma 2.3 there exists t ∈]α, β0 [ such that Ω(κ0 , t) is C ∞ . Let us
note Ω0 = Ω(κ0 , t) and β1 = t. The set Ω0 built that way satisfies
Ω0 ⊂ Ω
moreover the distance d0 = d(ξκ−1
0
(M + ), ξκ−1
0
(t)) > 0.

Construction of the next terms: Let κ1 ≥ max(κ0 + 1, d10 ). There exists

t ∈]α, β1 [ such that Ω(k1 , t) is C ∞ . Let Ω1 = Ω(κ1 , t) and β2 = t. We have
Ω1 ⊂ Ω
Since ξκ1 (x) = 1 for all x whose distance to the boundary of Ω is less than d0
we have ξκ1 (x) = 1 for all x ∈ Ω0 hence
Ω0 ⊂ Ω1
Let d1 = d(ξκ−1
0
(M + ), ξκ−1
1
(t)) > 0. Then we build Ω2 and so on so forth.

For each k ≥ 0, Γk = Γ(kκ , βκ ) which is also the boundary of Ωk .

Sh. Sensitivity Analysis in Hyperbolic Pb. with non Smooth Domains 7

2.4 Properties
Proposition 2.3. The sequence (Ωk )k≥0 has the subsequent properties
i) It is an increasing sequence of domains
ii) The limit ∪+∞ k
k=0 Ω is equal to Ω

Proof.
i) This is a consequence of the construction
ii) Since Ωk ⊂ Ω it is obvious that ∪+∞ k
k=0 Ω ⊂ Ω. Let x ∈ Ω, since Ω is
open there exists r > 0 such that B(x, r) ⊂ Ω. Let k be such that
κk ≥ max( 1r , k0 ) then ξk (x) = 1 hence x ∈ Ωk . It follows Ω ⊂ ∪+∞ k
k=0 Ω .

Proposition 2.4. Let K be a compact subset of Ω, there exists k ≥ k0 such

that K ⊂ Ωk .
Proof. Let r be the distance between K and Ω. Let k be such that
κk ≥ max( 1r , κ0 ) then for all x ∈ K we have ξκk (x) = 1 hence x ∈ Ωk .

2.5 Mollification of the Transported Domain

Transported domains Ωs = Ts (Ω) were considered in the introduction. They
are Lipschitz continuous so the construction which has been performed with Ω
can be repeated for those domains. That yields an isotonic sequence of domains
(Ωks )k≥0 which tends to Ωs . Property 2.4 holds when replacing Ω by Ωs . Once
s is given, all subsequent properties on Ω will hold for Ωs as well.
Remark 2.2. There is no reason to have Ts (Ωk ) = Ωks .
Let Qks = I × Ωks , Γks = ∂Ωks and Σks = I × Γks . We shall note ysk the solution
of the problem
 2
 ∂ y − div (K∇ysk ) = f on Qks
 kt

ys = 0 on Σks
(6)

 y k (0) = ϕ on Ωks
 s k
∂t ys (0) = ψ on Ωks

3 Continuity Result for the Wave Equation

In this section and the next one, we suppose m = 0, that is

f ∈ L1 (I, L2 (D)), ϕ ∈ H 1 (D), ψ ∈ L2 (D)

The aim of this section is to prove the solution to the wave equation in the
mollified domain tends to the solution of the wave equation in the Lipschitz
continuous domain. It is not a general continuity result (see [4]) since it
only works with the sequence of domain built in the previous section. In the
next section, that convergence will turn out to be enough to prove the shape
differentiability result that we are looking for.
8 Cagnol and Zolésio

3.1 Weak Convergence

For k ≥ 0 we note Qk = I × Ωk and Σk = I × Γk . Let us consider


 P y k = f on Qk
 k
y = 0 on Σk
(7)

 y k (0) = ϕ on Ωk

∂t y k (0) = ψ on Ωk

That problem has a unique solution in Z 1 (I, Ωk ).

The energy estimate gives the subsequent lemma (see [6, lemma 5]) Lemma
3.1. Let O be an open C 2 domain in D and µ ∈ Z 2 (I, O). We note
Z
a(µ) = kP µkL1 (I,L2 (O)) and b(µ) = (K∇µ(0).∇µ(0) + (∂t µ(0))2 )dx
O

then
p
(8) k∂t µkL∞ (I,L2 (O)) ≤ 2a(µ) + b(µ)

p
(9) kµkL∞ (I,H01 (O)) ≤ 2a(µ) + b(µ)

Proposition 3.1. Let O be an open C 2 domain in D and µ ∈ Z 1 (I, O)

with P µ ∈ L1 (I, L2 (Ω)). With the notations of lemma 3.1, identities (8) and
(9) hold.
Proof. This proposition is a consequence of lemma 3.1, the the density of
Z 2 (I, O) in Z 1 (I, O) and the continuity of the wave equation with respect to
the data.

The hypothesis of that proposition are satisfied for O = Ωk and µ = y k .

Let us note
Z
k k
a = kf kL1 (I,L2 (Ωk )) and b = (K∇ϕ.∇ϕ + ψ 2 )dx
Ωk

then √
k∂t y k kL∞ (I,L2 (Ωk )) ≤ 2ak + bk
√
ky k kL∞ (I,H01 (Ωk )) ≤ 2ak + bk
R
Let a∗ = kf kL1 (I,L2 (D)) and b∗ = D (K∇ϕ.∇ϕ + ψ 2 )dx then for all k we have
ak ≤ a∗ and bk ≤ b∗ . Moreover y k can be extended by 0 on RN r Ωk . hence
√
ky k kW 1,∞ (I,L2 (Ω))∩L∞ (I,H01 (Ω)) ≤ 2a∗ + b∗
Sh. Sensitivity Analysis in Hyperbolic Pb. with non Smooth Domains 9

that yields ky k kH(I,Ω) is bounded, hence there exists a converging subsequence

weakly in H 1 (Q) Let us note y ∗ such an2 element. We have
(10) y ∗ ∈ H(I, Ω)
Remark 3.1. As a corollary of (10) we have y ∗ = 0 on Σ.
Proposition 3.2. One has P y ∗ = f on Q
Proof. Let θ ∈ C0∞ (Q), since P y k = f on Qk we get
Z
∞
∀θ ∈ C0 (Q), (P y k )θ − f θ = 0
Qk

using proposition 2.4 we obtain the subsequent identity, when k is large enough
Z
∞
∀θ ∈ C0 (Q), (P y k )θ − f θ = 0
Q
R
lemma 3.4 yields ∀θ ∈ C0∞ (Q), ∗ ∗
Q (P y )θ − f θ = 0 therefore P y = f on Q.
Proposition 3.3. One has y k (0) = ϕ and ∂t y k (0) = ψ on Ω.
The proof of that lat proposition is analogous to the proof of proposition
3.2. Then we obtain


 P y ∗ = f on Q
 ∗
y = 0 on Σ
(11)
 ∗ (0) = ϕ on Ω
 y

∂t y ∗ (0) = ψ on Ω
Since that problem is well-posed we have y ∗ = y. The subsequent lemma
follows:
Proposition 3.4. y k ⇀ y weakly in H(I, Ω) as Qk → Q

3.2 Strong Convergence

We consider
Z D E
1
Ek (t) = K∇y k (t), ∇y k (t) + (∂t y k (t))2
2 Ωk

and E∞ (t) the corresponding energy when replacing Ωk by Ω and y k by y.

Lemma 3.2. Ek (0) → E∞ (0) when k → +∞.
Proof. One has Z
1
Ek (0) = hKϕ, ϕi + ψ 2
2 Ωk
since supp ϕ ⋐ Ω and supp ψ ⋐ Ω, proposition 2.4 gives
Z
1
Ek (0) = hKϕ, ϕi + ψ 2
2 Ω
when k is large enough.

2
At this point we do not know it is unique
10 Cagnol and Zolésio

Lemma 3.3. One has

Z T Z
kykH(I,Ω) = T E∞ (0) + (T − τ )f ∂t y
0 Ω
Z T Z
k
ky kH(I,Ωk ) = T Ek (0) + (T − τ )f ∂t y k
0 Ωk
Proof. We shall do the proof for the second identity, the proof of the first
one is analogous. The energy estimates gives
Z
∂t Ek (t) = f ∂t y
Ωk
Rτ R
this gives Ek (τ ) = Ek (0) + 0 Ωk P y k ∂t y k we get
Z T Z T Z τ Z
k
ky kH(I,Ωk ) = Ek (τ ) dτ = T Ek (0) + P y k ∂t y k
0 0 0 Ωk

Proposition 3.5. When k tends to +∞ we have

ky k kH(I,Ω) → kykH(I,Ω)

Proof. Lemma 3.2 and proposition 3.4 give

Z TZ Z T Z
T Ek (0) + (T − τ )f ∂t y → E∞ (0) + (T − τ )f ∂t y
0 Ωk 0 Ω

lemma 3.3 yields

ky k kH(I,Ωk ) → kykH(I,Ω)

since (Ωk ) is an increasing sequence of domains and y k is extend by 0 out of Ωk

we get
ky k kH(I,Ωk ) = ky k kH(I,Ω)
Corollary 3.1. We have

y k → y strongly in H(I, Ω) as Qk → Q

Remark 3.2. The same proof gives

ysk → ys strongly in H(I, Ωs ) as Qks → Qs

for all s ∈ [0, S].

Sh. Sensitivity Analysis in Hyperbolic Pb. with non Smooth Domains 11

4 Shape Differentiability
4.1 Absolute Continuity
Let θ ∈ L1 (I, L2 (D)), we note
Z
hk (s) = ysk θ dx dt
Qks
Z
h(s) = yθ dx dt
Qs

the shape differentiability for smooth domains gives

Z
′k
h (s) = ys′ θ dx dt
Qks

Let ȳ be the solution to the subsequent well-posed problem



 P (ȳ) = 0 on Q
 ∂y
ȳ = − ∂n hV (0), ni on Σ
(12)

 ȳ(0) = 0 on Ω

∂t ȳ(0) = 0 on Ω

at this point we do not know that ȳ is the shape derivative of the state function
y, and it is precisely what we are going to prove. Let us note
Z
h̄(s) = ȳs θ dx dt
Qs

The absolute continuity of hk gives

Z s
(13) ∀k ∈ N∗ , ∀s ∈ [0, S], hk (s) = hk (0) + h′k (σ) dσ
0

From proposition 3.4, the left hand side and the first term of the right hand
side converge to h(s) and h(0) Rrespectively. To prove the Rabsolute continuity of
s s
h it is sufficient to prove that 0 h′k (σ) dσ converges to 0 h̄(σ) dσ as k tends
to +∞. To achieve that goal let us introduce the following adjoint problem


 P (Λk ) = θ on Qks
 k s
Λs = 0 on Σks
(14)

 Λk (T ) = 0 on Ωks
 s k
∂t Λs (T ) = 0 on Ωks

From proposition 3.1 we get

Λks → Λs strongly in H(I, Ωs ) as Qks → Qs

12 Cagnol and Zolésio

Following [6] we have

Lemma 4.1. Let Ks = (DTs )−1 (K ◦ Ts )(∗ DTs−1 ) then
Z  2  2 ! D ED E
′k 1 ∂(ysk + Λks ) ∂(ysk − Λks ) k k k
h (s) = − K n
s s , n s V (s), n s dΓ dt
4 Σks ∂nks ∂nks

For the sake of shortness we suppose K = I. Following [6] we have

Lemma 4.2. Let µks ∈ L2 (I, H01 (Ωks )) ∩ H1 (I, L2 (Ωks )) such that P µks ∈
L2 (Qks ) then
Z  k 2 D E
∂µs k
V (s), n s =
Σks ∂nks
Z D E Z D E
−2 ∂t µks (0) ∇µks (0), V (s)(0) − (div (V (s) − 2ε(V ))∇µks , ∇µks
Ωks Qks
Z D E Z D E
2
−2 P µks ∇µks , V (s) + ∂t µks div (V − s) − 2∂t µks ∇µks , V (s)
Qks Qks

Let µks,α = ysk + αΛks with α ∈ {−1; 1}. It satisfies the hypothesis of lemma
4.2 since P ysk = f and P Λks = θ, Moreover ysk and Λks as well as their time
derivative and its gradient vanish on Ωs r Ωks , therefore the integrals on Ωks and
Qks of lemma 4.2 can be replaced by integrals on Ω s and Qs respectively. It
follows that h̄k (s) converges to
Z     !
∂(ys + Λs ) 2 ∂(ys − Λs ) 2
− hKs ns , ns i hV (s), ns i dΓ dt
Σs ∂ns ∂ns

it follows the
Proposition 4.1. One has

lim h̄k (s) = h′ (s)

k→+∞

From lemma 3.1, the real hk (s) is dominated by a constant independent of

s and k. Since aks and bks are bounded by a∗ and b∗ respectively.
Corollary 4.1. The function h is absolutely continuous.

4.2 Differentiability
Lemma 4.3. When s → 0 one has

ys ⇀ y in H 1 (I × D)

Proof. Proposition 3.1 works with O = Ωs and µ = ys . Since we extend ys

by zero out of Ωs we get
√
kys kW 1,∞ (I,L2 (D))∩L∞ (I,H 1 (D)) ≤ 2a∗ + b∗
Sh. Sensitivity Analysis in Hyperbolic Pb. with non Smooth Domains 13

Since a∗ and b∗ depend only on the hold-all D, we extract a subsequence

converging to an element y∗ . The last point to be proven is that y∗ is solution
of (2). Even though we do not have an isotonic sequence, property 2.4 holds
when replacing Ωk by Ωs and “k large enough” by “s small enough”. The only
problem is to prove y∗ vanishes on the lateral boundary Σ. Let dX denote
the distance to the set X. The convergence of the boundary gives dΩs → dΩ
strongly in L2 (I × D), hence

dΩs ys → dΩ y

in L2 (I × D) as s → 0. Since ys = 0 on D r Ω̄s and dΩs = 0 on Ωs , we get

dΩ y = 0 hence y = 0 almost everywhere in D r Ω. As Ω is Lipschitz continuous,
it has the Keldysh stability property, therefore y = 0 quasi everywhere in Dr Ω̄.
That yields y = 0 on ∂(D r Ω̄). We end up with y∗ = 0 on Σ.
Proposition 4.2. When s → 0 one has

ys → y strongly in H 1 (I × D)

Proof. The proof is based on the ideas of section 3.2. The energy to be
considered is Z
1
hK∇ys (t), ∇ys (t)i + (∂t ys (t))2
2 Ωs
Again, we do not have an isotonic sequence, however because Ωs is the image
of Ω by the flow mapping Ts , each compact of Ω is included in Ωs for s small
enough. We derive kys kH(I,Ωs ) → kykH(I,Ω) when s → 0. That leads to
kys kH(I,D) → kykH(I,D) .

Since a∗ and b∗ depend only on D, the domination of kys kH(I,D) is

straightforward. That proves

lim h̄(s) = h̄(0)

s→0

That gives the weak shape differentiability in L1 (I, L2 (D)) of the state function.
Remark 4.1. When m = 0, taking θ ∈ L∞ (I, L2 (Ω)) is required because
the weak shape differentiability for C 2 -boundary takes place in L1 (I, L2 (Ωk )).
When m ≥ 1, the test θ can be taken L1 (I, L2 (θ)), that leads to the shape
differentiability in L∞ (I, L2 (D)) of the state function.

References

[1] S. Agmon, Lectures on Elliptic Boundary Value Problems, Van Nostrand Mathe-
matical Studies, 1965.
14 Cagnol and Zolésio

[2] C. Baiocchi and A. Capelo, Variational and quasivariational inequalities, John

Wiley & Sons Inc., New York, 1984.
[3] M. Berger and B. Gostiaux, Differential geometry: manifolds, curves, and
surfaces, Springer-Verlag, New York, 1988.
[4] D. Bucur, Contrôle par rapport au domaine dans les EDP, PhD thesis, Ecole des
Mines de Paris, 1995.
[5] J. Cagnol and J.-P. Zolésio, Hidden shape derivative in the wave equation, in
Systems modelling and optimization (Detroit, MI, 1997), Chapman & Hall/CRC,
Boca Raton, FL, 1999, pp. 42–52.
[6] , Shape derivative in the wave equation with Dirichlet boundary conditions,
J. Differential Equations, 158 (1999), pp. 175–210.
[7] A. Chaïra, Equation des ondes et régularité sur un ouvert lipschitzien, Comptes
Rendus de l’Académie des Sciences, Paris, series I, Partial Differential Equations,
316 (1993), pp. 33–36.
[8] M. C. Delfour and J.-P. Zolésio, Structure of shape derivatives for non smooth
domains, Journal of Functional Analysis, 104 (1992), pp. 1–33.
[9] , Shape analysis via oriented distance functions, Journal of Functional
Analysis, 123 (1994), pp. 129–201.
[10] , Hidden boundary smoothness in hyperbolic tangential problems on non-
smooth domains, in Systems modelling and optimization (Detroit, MI, 1997),
Chapman & Hall/CRC, Boca Raton, FL, 1999, pp. 53–61.
[11] I. Lasiecka, J.-L. Lions, and R. Triggiani, Non homogeneous boundary value
problems for second order hyperbolic operators, Journal de Mathématiques pures
et Appliquées, 65 (1986), pp. 149–192.
[12] J. Nečas, Sur les domaines de type N, Czechoslovak Math, 12 (1962), pp. 274–287.
(Russian with a French summary).
[13] J.-P. Zolésio, Introduction to shape optimization and free boundary problems, in
Shape Optimization and Free Boundaries, M. C. Delfour, ed., vol. 380 of NATO
ASI, Series C: Mathematical and Physical Sciences, Kluwer Academic Publishers,
1992, pp. 397–457.
Unbounded Growth of Total Variations of Snapshots
of the 1D Linear Wave Equation due to the Chaotic
Behavior of Iterates of Composite Nonlinear
Boundary Reflection Relations

Goong Chen(1),(2) , Texas A&M University, College Station, Texas

Tingwen Huang(1) , Texas A&M University, College Station, Texas
Jonq Juang(3) , National Chiao Tung University, Hsinchu, Taiwan, ROC
Daowei Ma(4) , Wichita State University, Wichita, Kansas

Abstract

Consider a linear one-dimensional wave equation on an interval. If

the boundary conditions are also linear, then the total variation of the
gradient (wx (·, t), wt (·, t)) on the spatial interval remains bounded as
t → ∞, provided that the initial condition (w(·, 0), wt (·, 0)) has finite total
variation. However, if we let the left-end boundary condition pump energy
into the system linearly, while the right-end boundary condition be self-
regulating of the van der Pol type with a cubic nonlinearity, then chaotic
vibrations occur when the parameters enter a certain regime. In this paper,
we characterize the chaotic behavior of the gradient (wx (·, t), wt (·, t)) by
proving that its total variation grows unbounded (with generically given
initial conditions) as t → ∞, even though the initial condition has a finite
total variation. The proofs are obtained by the technique of interval
covering sequences based on Stefan cycles and homoclinic orbits of the
composite nonlinear boundary reflection map.

(1) E-mails: [email protected] and [email protected].

(2) Supported in part by Texas A&M University Interdisciplinary Research
Initiative IRI 99-22.
(3) Work completed while on sabbatical at Texas A&M University. Supported
in part by a grant from NSC of R.O.C. E-mail: [email protected].
(4) E-mail: [email protected].

15
16 Chen et al.

1 Introduction
In this paper, we study a special property of chaotic vibration of the
wave equation, that of unbounded growth of total variations of snapshots
(wx (·, t), wt (·, t)) on the spatial interval of the one-dimensional (1D) wave
equation as t → ∞.
Earlier, in a series of papers [3–6], we have studied chaotic vibration of the
1D wave equation

(1.1) wxx (x, t) − wtt (x, t) = 0, 0 < x < 1, t > 0,

subject to the following boundary conditions

(1.2)
left-end x = 0 : wt (0, t) = −ηwx (0, t), η > 0, η 6= 1, t > 0;
(1.3)
right-end x = 1 : wx (1, t) = αwt (1, t) − βwt3 (1, t), 0 < α ≤ 1, β > 0,

where the boundary condition (1.2) signifies energy injection or pumping into
the system, while (1.3) signifies a feedback with cubic nonlinearity of the van
der Pol type. Note that in (1.1) we have set the spatial domain to be the unit
interval I ≡ (0, 1) just for convenience. Two initial conditions

(1.4) w(x, 0) = w0 (x), wt (x, 0) = w1 (x), 0 < x < 1,

are also prescribed. Then it was established in [5] that for fixed α, β, the
composite reflection map Gη ◦ Fα,β : I¯ → I¯ is chaotic (cf. [5, (9)–(12)] or
(1.9)–(1.10) below for Gη and Fα,β ) and, therefore, for initial conditions (1.4)
of generic type, (wx (x, t), wt (x, t)) displays chaotic behavior. Here, we follow
Devaney’s definition of chaos [8]; see also [2].
To make this paper sufficiently self-contained, let us repeat the solution
procedure for (1.1)–(1.4) from [4] using the method of characteristics. Define
1 1
(1.5) u(x, t) = [wx (x, t) + wt (x, t)], v(x, t) = [wx (x, t) − wt (x, t)].
2 2
Then (u, v) satisfies the following initial-boundary value problem (IBVP), a
first-order diagonalized symmetric hyperbolic system
     
∂ u(x, t) 1 0 ∂ u(x, t)
(1.6) = , 0 < x < 1, t > 0,
∂t v(x, t) 0 −1 ∂x v(x, t)

with boundary conditions

(1.7) [u(0, t) − v(0, t)] = −η[u(0, t) + v(0, t)],

(1.8) u(1, t) + v(1, t) = α[u(1, t) − v(1, t)] + β[u(1, t) − v(1, t)]3 .
 Unbounded Growth of Total Variations 17

The algebraic equations (1.7) and (1.8) define the reflection relations
1+η
(1.9) v(0, t) = Gη (u(0, t)) ≡ u(0, t),
1−η
(1.10) u(1, t) = Fα,β (v(1, t)),

at, respectively, the left-end x = 0 and the right-end x = 1, where in (1.10),

Fα,β : R → R is a nonlinear mapping such that for each given v ∈ R, u ≡ Fα,β (v)
is the unique real solution of the cubic equation

(1.11) β(u − v)3 + (1 − α)(u − v) + 2v = 0.

The initial conditions are now

u(x, 0) = u0 (x) = 12 [w0′ (x) + w1 (x)],
(1.12) 0 < x < 1.
v(x, 0) = v0 (x) = 12 [w0′ (x) − w1 (x)],

From time to time, we also need that u0 and v0 satisfy the compatibility
conditions

(1.13) v0 (0) = Gη (u0 (0)), u0 (1) = Fα,β (v0 (1)).

Using the maps Fα,β and Gη , we can represent the solution (u, v) of (1.6)
explicitly as follows [5, (13), (14), p. 425]: for t = 2k + τ , k = 0, 1, 2, . . . ,
0 ≤ τ < 2 and 0 ≤ x ≤ 1,

(1.14)
 
 (Fα,β ◦ Gη )k (u0 (x + τ )), τ ≤ 1 − x, 

u(x, t) = −1 k+1
G ◦ (Gη ◦ Fα,β ) (v0 (2 − x − τ )), 1 − x < τ ≤ 2 − x, 


 η k+1 (u (τ + x − 2)),


 (Fα,β ◦ Gη ) 0 2 − x < τ < 2;
k
 (Gη ◦ Fα,β ) (v0 (x − τ )), τ ≤ x, 



v(x, t) = Gη ◦ (Fα,β ◦ Gη )k (u0 (τ − x)), x < τ ≤ 1 + x, 

 
(Gη ◦ Fα,β )k+1 (v0 (2 + x − τ )), 1 + x < τ < 2,

where (Gη ◦Fα,β )n = (Gη ◦Fα,β )◦(Gη ◦Fα,β )◦· · ·◦(Gη ◦Fα,β ), the n-times iterative
composition of Gη ◦ Fα,β . Since the solution representation (1.14) depend on
(Gη ◦ Fα,β )n , it constitutes a natural Poincaré section for the solution of (1.6).
We say that the solution of (1.6) is chaotic if the map Gη ◦ Fα,β : I¯ → I¯ (or
Gη ◦ Fα,β : I → I, where I is an invariant subset of Gη ◦ Fα,β contained in I) ¯
is chaotic. Since (wx , wt ) is topologically conjugate to (u, v) in the sense of [4,
§5], we also say that the gradient w of the system (1.1)–(1.4) is chaotic.
The orbit diagram of the map Gη ◦ Fα,β , where α and β are held fixed, say
α = 1/2, β = 1, and only η is varying, can be seen from [5, Fig. 3, p. 433] (for
0 < η < 1) and [5, Fig. 4, p. 434] (for η > 1), wherein period-doubling cascades
are manifest. For the purpose of making this paper sufficiently self-contained,
18 Chen et al.

we reproduce these two figures in Figs. 1 and 2, respectively. Furthermore, the

existence of homoclinic orbits has been established for the parameter range
(1.15)
     . 
1+α . 1+α 1+α 1+α
1− √ 1+ √ ≤ η < 1, 1<η≤ 1+ √ 1− √ .
3 3 3 3 3 3 3 3
See [5, Theorems 4.1 and 4.2, pp. 436–439].
Let us now observe a few snapshots of the solution (u, v) of (1.6)–(1.12).
Choose initial conditions
π 
(1.16) w(x, 0) = 0.2 sin x , wt (x, 0) = 0.2 sin(πx), x ∈ [0, 1],
2
and, thus
(1.17)
hπ π  i hπ π  i
u0 (x) = (0.1) · cosx + sin(πx) , v0 (x) = (0.1) · cos x − sin πx
2 2 2 2
are the initial conditions for (1.6)–(1.12). We display the snapshots of u(·, t)
and v(·, t) using (1.14) in Figs. 3∼8, with the following parameter values:
(1.18) α = 0.5, β = 1,
(1.19) Figs. 3 ∼ 4 : η = 0.525, t = 52,
(1.20) Figs. 5 ∼ 6 : η = 0.525, t = 102,
(1.21) Figs. 7 ∼ 8 : η = 1.52 t = 52.
For the value η = 0.525 used for Figs. 3–6, the map Gη ◦Fα,β has just completed
its period-doubling cascade, as can be seen from the orbit diagram [5, Fig. 3,
loc. cit.]. Regarding the profiles of u and v, we see that as t increases from
t = 52 in (1.19) to t = 102 in (1.20), there is a very noticeable increase of
“oscillatory ripples” from Figs. 3∼4 to Figs. 5∼6 (with the presence of some
“macroscopically coherent periodic structure”). Let us further look at Figs. 7
and 8, where for the parameter value η = 1.52 in (1.21), (1.15) implies the
existence of a homoclinic orbit (cf. (1.15)2 ); the profiles of u and v therein
look extremely oscillatory at time t = 52, resembling something akin to “white
noise”, along with the disappearance of any coherent patterns.
The highly oscillatory behavior of u and v as displayed in these figures
motivated us to pose the following question:
“[Q] Assume that the composite reflection map Gη ◦ Fα,β is chaotic. Does
there exist a large class of initial conditions (u0 , v0 ) for (1.6), (1.9), (1.10),
and (1.12) such that VI¯(u0 , v0 ) < ∞ but
(1.22) lim [VI¯((u(·, t)) + VI¯(v(·, t))] = ∞?”
t→∞

In (1.22), V[a,b] (f ) denotes the total variation of a given function f on interval

[a, b]; see the definition in [1, p. 165], for example.
 Unbounded Growth of Total Variations 19

Fig. 1. The orbit diagram of Gη ◦ Fα,β , with α = 0.5, β = 11, and η being the
varying parameter, 0 < η < 11.

Fig. 2. The orbit diagram of Gη ◦ Fα,β , with α = 0.5, β = 11, and η being the
varying parameter, η > 11.
20 Chen et al.

Fig. 3. The profile of u(x, t) at t = 52

52, with α = 0.5
0.5, β = 1, η = 0.525
0.525, for the
system (1.6), (1.7), (1.8) and (1.17). (Reprinted from [5, p. 436], courtesy of World
Scientific, Singapore.)

Fig. 4. The profile of v(x, t) at t = 52

52, with α = 0.5, β = 1, η = 0.525
0.525, for the
system (1.6) (1.7), (1.8) and (1.17). (Reprinted from [5, p. 346], courtesy of World
Scientific, Singapore.)
 Unbounded Growth of Total Variations 21

Fig. 5. The profile of u(x, t) at t = 102

102, with α = 0.5, β = 1, η = 0.525
0.525, for the
system (1.6) (1.7), (1.8) and (1.17). (Reprinted from [5, p. 437], courtesy of World
Scientific, Singapore.) The chaos here is due to the period-doubling cascade.

Fig. 6. The profile of v(x, t) at t = 102

102, with α = 0.5, β = 1, η = 0.525
0.525, for the
system (1.6) (1.7), (1.8) and (1.17). (Reprinted from [5, p. 437], courtesy of World
Scientific, Singapore.) As with Fig. 5, the chaos here is due to the period-doubling
cascade.
22 Chen et al.

Fig. 7. The profile of u(x, t) at t = 52 52, with α = 0.5, β = 1, η = 1.52

1.52, for the
system (1.6) (1.7), (1.8) and (1.17). (Reprinted from [5, p. 442], courtesy of World
Scientific, Singapore.) The chaos here is due to a homoclinic orbit of Gη ◦ Fα,β .

Fig. 8. The profile of v(x, t) at t = 52 52, with α = 0.5, β = 1, η = 1.52

1.52, for the
system (1.6) (1.7), (1.8) and (1.17). (Reprinted from [5, p. 442], courtesy of World
Scientific, Singapore.) As with Fig. 7, the chaos here is due to a homoclinic orbit of
Gη ◦ Fα,β .
 Unbounded Growth of Total Variations 23

In this paper, we give some informative answers to the question [Q] posed
above.
The rest of the paper is divided into three parts. In §2, we present a few
facts about linear vibration in order to show the contrasts between linearity and
nonlinearity. The main theorems are established in §3. In §4, miscellaneous
discussions are given. A useful proposition, which was used in §2, is given
separately in the Appendix near the end of the paper.

2 Bounds on the Total Variation of (u(·, t), v(·, t)) of the Linear
Wave Equation as t → ∞
Consider the wave equation (1.1), but with linear boundary conditions such
as

(2.1) left-end x = 0 : w(0, t) = 0, t > 0,

(2.2) right-end x = 1 : wx (1, t) = 0, t > 0,

in lieu of (1.2) and (1.3). Let the initial conditions satisfy

(2.3) w(x, 0) = w0 (x) ∈ H01 (0, 1), wt (x, 0) = w1 (x) ∈ L2 (0, 1),

where
H01 (0, 1) = {f : (0, 1) → R | f (0) = 0; f, f ′ ∈ L2 (0, 1)}
is a Sobolev space with norm
Z 1 1/2
kf kH01 (0,1) = (f 2 + f ′2 )dx .
0

Then for the system (1.1), (2.1)–(2.3), the energy

Z
1 1 2
(2.4) E(t) = [w (x, t) + wt2 (x, t)]dx
2 0 x
is conserved and, therefore, we have the estimate

(2.5) kw(·, t)kH01 (0,1) + kwt (·, t)kL2 (0,1) ≤ C(kw0 kH01 (0,1) + kw1 kL2 (0,1) )

for some constant C > 0 independent of (w0 , w1 ). This type of Sobolev

estimates is quite well known for the IBVP of the linear wave equation.
Not so well known are similar types of estimates in terms of total variations
for the linear wave equation. Let us convert (1.1), and (2.1)–(2.3) into a first-
order diagonalized symmetric hyperbolic system (1.6) through (1.5). Then (2.1)
and (2.2) lead to the reflection relations

(2.6)
v(0, t) = G(u(0, t)) ≡ u(0, t), u(1, t) = F (v(1, t)) ≡ −v(1, t), t > 0.
24 Chen et al.

Assume that the initial conditions u0 and v0 (cf. (1.12)) are continuous on I¯
and satisfy the compatibility conditions

(2.7) v0 (0) = G(u0 (0)), u0 (1) = F (v0 (1)).

Then from the representation formula (1.14) we easily obtain

(2.8) VI¯(u(·, t)) + VI¯(v(·, t)) = VI¯(u0 ) + VI¯(v0 ), ∀ t > 0,

i.e., the sum of the total variations of u and v at any t on I¯ is conserved.

If u0 and v0 are continuous on I¯ but the compatibility conditions in (2.7) are
violated, then discontinuities can propagate along characteristics x − t = −k,
x + t = 1 + k for k = 0, 1, 2, . . . so (2.8) needs to be modified to

(2.9) VI¯(u(·, t)) + VI¯(v(·, t)) ≤ VI¯(u0 ) + VI¯(v0 ) + C, ∀ t > 0,

for some constant C:

(2.10) C ≡ |u0 (0) − v0 (0)| + |u0 (1) + v0 (1)|.

Using (1.5) and (2.9), we further deduce that

(2.11) VI¯(wx (·, t)) + VI¯(wt (·, t)) ≤ 2[VI¯(w0′ ) + VI¯(w1 ) + C]

e

e > 0. From
for some constant C
Z x
w(x, t) = wx (ξ, t)dξ
0

and for any {xk ∈ [0, 1] | k = 0, 1, . . . , n} satisfying 0 = x0 < x1 < x2 < · · · <
xn = 1,
 
n−1
X X Z xk+1
n−1

|w(xk+1 , t) − w(xk , t)| =  wx (ξ, t)dξ 

k=0 k=0 xk
n−1
X Z
xk+1 Z xk+1 
1
≤ dξ + wx2 (ξ, t)dξ
2 xk xk
k=0
Z 1
1
(2.12) = + wx2 (x, t)dx,
2 0

we obtain Z 1
1
VI¯(w(·, t)) ≤ + wx2 (x, t)dx.
2 0
Therefore (2.11) can be furthered strengthened. We summarize the above in
the following.
 Unbounded Growth of Total Variations 25

Theorem 2.1. Consider the system (1.1), (2.1), (2.2) and (2.3), with
w0 ∈ C 1 ([0, 1]) and w1 ∈ C 0 ([0, 1]). Then
(2.13)
ee
VI¯(w(·, t)) + VI¯(wx (·, t)) + VI¯(wt (·, t)) ≤ 2[VI¯(w0′ ) + VI¯(w1 ) + E(0) + C],
ee
for some C > 0 depending only on C in (2.10).
¯ and w1 ∈ C 0 (I),
From the estimate (2.13) we see that if w0 ∈ C 1 (I) ¯ then
as t → ∞, the left-hand side (LHS) of (2.13) remains bounded, provided that
initially w0′ and w1 have bounded total variations. The LHS of (2.13) can grow
unbounded (when and) only when initially (at least one of) w0′ and w1 have
unbounded total variations. This is possible, as shown in the following example.
Example 2.1. Choose
Z x
1
w0 (x) = ξ(ξ − 1) sin dξ, w1 (x) = 0; 0 < x < 1.
0 ξ
Then (w0 , w1 ) ∈ H01 (0, 1) × L2 (0, 1). The compatibility conditions (2.7)1 and
(2.7)2 are satisfied. Therefore the solution (u, v) of (1.6), (2.6)–(1.12) is
continuous for any (x, t) ∈ [0, 1] × [0, ∞). Here
1
w0′ (x) = x(x − 1) sin , 0 < x < 1,
x
is easily verified to have
VI¯(w0′ ) = ∞.
We see that the LHS of (2.13) is ∞ for any t > 0. 
Next, let us consider, again, linear boundary conditions but somewhat
different from the ones in (2.1) and (2.2). The IBVP system is


 wxx (x, t) − wtt (x, t) = 0, x ∈ (0, 1), t > 0,

wt (0, t) + γw(0, t) = 0, t > 0,
(2.14)
 w (1, t) = 0, t > 0,
 x

w(x, 0) = w0 (x), wt (x, 0) = w1 (x), x ∈ (0, 1).
Note that the boundary condition (2.14)2 is integrable along the t-direction:
(2.15) w(0, t) = w(0, 0)e−γt , t ≥ 0.
Again, converting (2.14) into a first-order diagonalized symmetric hyperbolic
system using (1.5) and utilizing (2.15), we obtain the following snapshots at
t = 1, 2, . . . , inductively:
(2.16)


 u0 (x), v0 (x) are given (according to (1.12)); and w(0, 0) is also known,

u(x, k + 1) = −v(1 − x, k),

 v(x, k + 1) = u(1 − x, k) + γak e−γ e−γ(1−x) ,

ak+1 = w(0, k + 1) = e−γ w(0, k), w(0, 0) ≡ a0 .
26 Chen et al.

If γ < 0, then (2.15) implies that w(x, t) can grow unbounded in general and,
therefore, the total variations of w, wx , wt , u and v can not be expected to remain
bounded even if the initial condition w0 and w1 (or u0 , v0 ) have bounded total
variations at t = 0. This is a classical instability case. So let us only consider
the case γ > 0. (The case γ = 0 is already covered in Theorem 2.1.)
Theorem 2.2. For (2.14), let u and v be defined by (1.5). Then there exist
a constant C > 0, depending only on γ > 0 and the right side of (2.10), such
that

(2.17) VI¯(u(·, t)) + VI¯(v(·, t)) ≤ VI¯(u0 ) + VI¯(v0 ) + C.

Proof. We need only establish (2.17) for integral values of t. Let us use
(2.16) to verify that the following holds:

VI¯(u(·, 2k)) ≤ VI¯(u0 ) + γ|a0 |Ak , 


VI¯(u(·, 2k + 1)) ≤ VI¯(v0 ) + γ|a0 |Bk ,
(2.18)
VI¯(v(·, 2k)) ≤ VI¯(v0 ) + γ|a0 |Bk , 


VI¯(v(·, 2k + 1)) ≤ VI¯(u0 ) + γ|a0 |Ak+1 ,

for k = 1, 2, . . . , where

e−γ − e−(2k−1)γ e−2γ − e−2(k+1)γ

Ak = , Bk = .
1 − e−2γ 1 − e−2γ
We prove by induction. When k = 1, (2.16) gives

 u(x, 1) = −v0 (1 − x),
v(x, 1) = u0 (1 − x) + γa0 e−γ e−γ(1−x) ,

a1 = a0 e−γ ;

 u(x, 2) = −u0 (x) − γa0 e−γ e−γx ,
v(x, 2) = −v0 (x) + γa0 e−2γ e−γ(1−x) ,

a2 = a0 e−2γ ;

 u(x, 3) = v0 (1 − x) − γa0 e−2γ e−γx ,
v(x, 3) = −u0 (1 − x) − γa0 e−γ e−γ(1−x) + γa0 e−3γ e−γ(1−x) ,

a3 = a0 e−3γ .

Therefore, (2.18) is clear for k = 1. Suppose (2.18) is valid for k ≤ n. We now

prove (2.18) for k = n + 1. By (2.16),

u(x, 2(n + 1)) = −v(1 − x, 2n + 1);

(2.19) VI¯(u(·, 2(n + 1))) = VI¯(v(·, 2n + 1)) ≤ VI¯(u0 ) + γ|a0 |An+1 ;
v(x, 2(n + 1)) = u(1 − x, 2n + 1) + γa2n+1 e−γ e−γ(1−x)
= u(1 − x, 2n + 1) + γa0 e−(2n+2)γ e−γ(1−x) .
 Unbounded Growth of Total Variations 27

Thus

VI¯(v(·, 2(n + 1))) ≤ VI¯(u(·, 2n + 1)) + VI¯(γa0 e−2(n+1)γ e−γ(1−x) )

≤ VI¯(u(·, 2n + 1)) + γ|a0 |e−2(n+1)γ
≤ VI¯(v0 ) + γ|a0 |B2n + γ|a0 |e−2(n+1)γ
" #
e−2γ − e−2(n+1)γ −2(n+1)γ
= VI¯(v0 ) + γ|a0 | · +e
1 − e−2γ
e−2γ − e−2(n+2)γ
= VI¯(v0 ) + γ|a0 | ·
1 − e−2γ
(2.20) = VI¯(v0 ) + Bn+1 .

Therefore, (2.18)1 and (2.18)3 are verified by (2.19) and (2.20), respectively.
The proof of (2.18)2 and (2.18)4 can be done in the same way. Therefore we
have proved (2.17).
Corollary 2.1. For the IBVP (2.14), assuming that w0 ∈ C 1 ([0, 1]) and
w1 ∈ C 0 ([0, 1]). Then the estimate (2.13) holds for t > 0. 
If, instead of the linear boundary conditions pair (2.14)2 and (2.14)3 , we
consider

wx (0, t) − γw(0, t) = 0, γ > 0, t > 0,
(2.21)
wt (1, t) = 0, t > 0,

where one of them is of Robin type [7, §1.2 and 1.3], then the treatment
becomes much more challenging. After some extra efforts, we have succeeded
in establishing an estimate similar to (2.13), as shown below.
Theorem 2.3. Consider the system (1.1), (2.21), with initial conditions
w0 ∈ C 1 ([0, 1]) and w1 ∈ C 0 ([0, 1]) in (1.4). Then there exist two positive
constants C1 , C2 such that

(2.22) VI (w(·, t)) + VI (wx (·, t)) + VI (wt (·, t)) ≤ C1 [VI (w0′ ) + VI (w1 )] + C2

for all t > 0.

Proof. First, we will establish the inequality

(2.23) e1 [V (u0 ) + V (v0 )] + C

VI (u(·, t)) + VI (v(·, t)) ≤ C e2
I I

for some positive constants C e1 and C e2 , for all t > 0, from which (2.22) will
naturally follow. Here, as before, u, v, u0 and v0 are defined through (1.5) and
(1.12).
28 Chen et al.

From (2.21)1 , we have

Z t 
[u(0, t) + v(0, t)] − γ [u(0, τ ) − v(0, τ )]dτ + w(0, 0) = 0,
0
Z t  Z t 
v(0, t) + γ v(0, τ )dτ = − u(0, t) − γ u(0, τ )dτ + γw(0, 0),
0 0
 Z t   Z t 
−γt d γt γt d −γt
e e v(0, τ )dτ = −e e u(0, τ )dτ + γw(0, 0),
dt 0 dt 0

which, after further simplification and integrations by parts, leads to

Z t
(2.24) v(0, t) = −u(0, t) + 2γ e−γ(t−τ ) u(0, τ )dτ + γw(0, 0)e−γt , t > 0.
0

This is the reflection relation at the left end x = 0. At the right end x = 1, the
reflection relation is

(2.25) u(0, t) = v(0, t), t > 0.

From (2.24) and (2.25), it is clear (based on a representation similar to (1.14))

that (2.23) will hold if we can prove that there exist two positive constants Cb1
b
and C2 such that

(2.26) b1 · V[0,T ] (u(0, ·)) + C

V[0,T ] (v(0, ·)) ≤ C b1 , for all T > 0,

under the assumption of (2.24). (The reflection relation (2.25) is easy and
simple so it does not require any separate consideration.)
But (2.24) implies (2.26), following the application of a technical Proposi-
tion A in the Appendix near the end of the paper.
We leave out the details that (2.23) yields (2.22).
Note that if (2.21)2 is replaced by wx (1, t) = 0, then (2.25) correspondingly
will be changed to
u(0, t) = −v(0, t)
and the arguments from (2.24) through (2.26) also need to be adapted
accordingly. Nevertheless, such modifications are straightforward and the
estimate (2.22) remains valid.
As a summary of this section, we have successfully shown that for major
typical homogeneous linear boundary conditions of the wave equation, the total
variations of the snapshot of the gradient as well as the state of the wave
equation at any time t on I will remain uniformly bounded in time, if the
initial total variation on I is finite.
 Unbounded Growth of Total Variations 29

3 Unbounded Growth of the Total Variations of (u(·, t), v(·, t))

as t → ∞ When Chaotic Vibration Occurs
This is the main section of the paper where we will treat question [Q].
Let us first estimate the growth of total variations of (u(·, t), v(·, t)) due to
the post period-doubling of the reflection map Gη ◦ Fα,β .
We will utilize special properties of the Stefan Cycle as given in Robinson
[11, pp. 67–70]. Let fµ : J → J be continuous on a finite closed interval J ⊂ R.
Assume that fµ has completed a period-doubling cascade as the parameter µ
crosses a value µ0 . Therefore, we may now assume that fµ has a prime periodic
point with period n = m ·2k , where m is an odd integer greater than or equal to
3, as such an n is above the doubling cascade portion · · · 2j 2j−1 2j−2 · · ·2
in Sharkovskii’s ladder. From now on, to simplify notation, we just write fµ as
f.
k
Define g = f 2 . Then g has a periodic orbit
(3.1) O = {xj | j = 1, 2, . . . , m}
of prime odd period m such that
(3.2) g(xj ) = xj+1 , for j = 1, 2, . . . , m − 1,
satisfying either
(3.3) xm < xm−2 < · · · < x3 < x1 < x2 < x4 < · · · < xm−1
or
(3.4) xm−1 < xm−3 < · · · < x4 < x2 < x1 < x3 < · · · < xm−2 < xm .
Let us just treat the case (3.3) below because relation (3.4) is just a reflection
of (3.3) centered at x1 and all the arguments for (3.3) will also go through for
(3.4) after a straightforward modification. Now define ([11, p. 67]) m − 1 closed
intervals
(3.5) I1 = [x1 , x2 ], I2 = [x3 , x1 ], I3 = [x2 , x4 ], . . . , I2j−1 = [x2j−2 , x2j ],
I2j = [x2j+1 , x2j−1 ], . . . , for j = 2, . . . , (m − 1)/2.
For two closed intervals J1 and J2 , we say that J1 f -covers J2 , in notation
f
J1 −→ J2 , if J2 ⊆ f (J1 ). Then we have the following.
Proposition 3.1. ([11, p. 68]) Assume that J is a finite closed interval
and g : J → J is continuous with a prime periodic orbit of odd period m.
Then there exist m − 1 closed subintervals I1 , I2 , . . . , Im−1 of J defined through
(3.1)–(3.3) and (3.5) that overlap at most at endpoints such that
g g g g g g
(3.6) I1 −→ I2 −→ I3 −→ · · · −→ Im−1 −→ I1 −→ I1 ∪ I2 .

30 Chen et al.

Theorem 3.1. Assume that J is a closed interval and g : J → J is

continuous with a prime periodic orbit of odd period m. Then
(3.7) lim VJ (gn ) = ∞.
n→∞

Proof. We first show that

(3.8) lim VI1 (gkm ) = ∞; see I1 defined in (3.5)).
k→∞

Utilizing (3.6), we have I1 gm -covers I1 ∪ I2 . Therefore

VI1 (gm ) ≥ ℓ(I1 ) + ℓ(I2 )
(3.9) = (x2 − x1 ) + (x1 − x3 ) = x2 − x3 , by (3.5),

where ℓ(Ij ) denotes the length of the interval Ij . Also, since I1 gm -covers I1 ∪I2 ,
we can find two subintervals I1,1 and I1,2 of I1 , overlapping at most at endpoints,
such that
(3.10) gm (I1,1 ) = I1 , gm (I1,2 ) = I2 .

Next, from (3.6) and (3.10), we have

gm g g g g g
(3.11) I1,1 −→ I1 −→ I2 −→ I3 −→ · · · Im−1 −→ I1 −→ I1 ∪ I2 ,
gm g g g g g g
(3.12) I1,2 −→ I2 −→ I3 −→ I4 −→ · · · Im−1 −→ I1 −→ I1 −→ I1 ∪ I2 ,

and, therefore, I1,1 has two subintervals I1,11 and I1,12 such that
(3.13) g2m (I1,11 ) = I1 and g2m (I1,12 ) = I2 ,

with I1,11 and I1,12 overlapping at most at endpoints. Similarly, from (3.12),
we obtain two closed subintervals I1,21 and I1,22 of I1,2 , overlapping at most at
endpoints, such that

g2m (I1,21 ) = I1 , g2m (I1,22 ) = I2 .

We therefore obtain

VI1 (g2m ) ≥ VI1,11 (g2m ) + VI1,12 (g2m ) + VI1,21 (gm ) + VI1,22 (gm )
≥ ℓ(I1 ) + ℓ(I2 ) + ℓ(I1 ) + ℓ(I2 )
= 2(x2 − x3 ).
This process can be continued indefinitely. In general, from a subinterval
I1,a1 a2 ...ak where aj = 1 or 2 for j = 1, 2, . . . , k, we have
g km g g g g g
I1,a1 a2 ...ak−1 ,1 −→ I1 −→ I2 −→ I3 −→ · · · −→ Im−1 −→ I1 ∪ I2 ,
g km g g g g g g
I1,a1 a2 ...ak−1 ,2 −→ I2 −→ I3 −→ I4 −→ · · · −→ Im−1 −→ I1 −→ I1 ∪ I2 .
 Unbounded Growth of Total Variations 31

From either of the above two sequences we can find two subintervals I1,a1 ...ak 1
and I1,a1 ...ak 2 of I1,a1 ...ak , overlapping at most at endpoints, such that

g(k+1)m (I1,a1 ...ak 1 ) = I1 and g(k+1)m (I1,a1 ... ,ak 2 ) = I2 ,

and because the collection of subintervals {I1,a1 a2 ...ak ak+1 | aj = 1, 2; j =
1, 2, . . . , k + 1} has non-overlapping interior,
X
VI1 (g(k+1)m ) ≥ VI1 ,a1 a2 ...ak ak+1 (g(k+1)m) )
aj =1,2
j=1,... ,k+1

(3.14) ≥ (k + 1)(x2 − x3 ) → ∞ as k → ∞.
Therefore, we have established (3.8).
To show (3.7), it is sufficient to show
lim VI1 (gkm+j ) = ∞, for j = 1, 2, . . . , m − 1.
k→∞

We utilize the covering sequence

gj g g g g g g g g
I1 −→ Ij+1 −→ Ij+2 −→ · · · −→ Im−1 −→ I1 −→ I1 −→ · · · −→ I1 −→ I1 ∪ I2
| {z }
I1 appearing
j+1 times

(j) (j)
to deduce that I1 has two closed subintervals I1,1 and I1,2 , overlapping at most
at endpoints, such that
(j) (j)
gm+j (I1,1 ) = I1 , gm+j (I1,2 ) = I2 .
(j)
Inductively, if I1,a1 ... ,ak is constructed, satisfying either

(j) g km+j g g g g g g
I1,a1 ... ,ak−1 1 −−−−−−−→ I1 −→ I2 −→ I3 −→ · · · −→ Im−1 −→ I1 −→ I1 ∪ I2 ,
or
(j) g km+j g g g g g g g
I1,a1 ...ak−1 2 −−−−−−−→ I2 −→ I3 −→ I4 −→ · · · −→ Im−1 −→ I1 −→ I1 −→ I1 ∪I2 ,
(j)
depending, respectively, on ak = 1 or 2, then we can find I1,a1 ... ,ak ’s two
(j) (j)
subintervals I1,a1 ... ,ak 1 and I1,a1 ... ,ak 2 , overlapping at most at endpoints, such
that
(j) (j)
g(k+1)m+j (I1,a1 ...ak 1 ) = I1 , g(k+1)m+j (I1,a1 ...ak 2 ) = I2 .
Again, we have
X
VI1 (g(k+1)m+j ) ≥ VI1 ,a1 ...ak+1 (g(k+1)m+j) ≥ (k +1)(x2 −x3 ) → ∞ as k → ∞.
aj =1,2
j=1,... ,k+1

The proof of (3.7) is now complete.

32 Chen et al.

Corollary 3.1. Assume that J is a closed interval f : J → J is continuous

with a prime periodic orbit of period m · 2k , where m is odd. Then

(3.15) lim VJ (f n ) = ∞.
n→∞

Proof. Let O1 = {f ℓ (ξ) | ℓ = 0, 1, . . . , m · 2k − 1} be an orbit of f with prime

k k
period m · 2k . Then O2 = {f j·2 (ξ) | j = 1, 2, . . . , m} is an orbit of g ≡ f 2
with prime period m. Write O2 in the form of (3.1) such that (3.2), (3.3) and
(3.5) are satisfied. Therefore, we have O2 = {xj | j = 1, 2, . . . , m} and for some
integer j1 : 0 < j1 ≤ m,
k k k
x1 = f j1·2 (ξ), x2 = f (j1 +1)·2 (ξ), . . . , xm = f (j1 +m)·2 (ξ).

The main idea of the proof is to show that

k )+ℓ)
(3.16) lim VIe0 (f j·(m·2 =∞
j→∞

for any ℓ = 0, 1, 2, . . . , m · 2k − 1, for some subinterval Ie0 ⊆ J (where Ie0 depends

on given ℓ). Given any such ℓ ∈ {0, 1, 2, . . . , m · 2k − 1}, we can find a positive
integer ℓ̂ > 0 such that

ℓ + ℓ̂ = j1 · 2k (modm · 2k ).

Define
(
ℓ̂+2k [y1 , y2 ], if y1 < y2 ,
(3.17) y1 = f ℓ̂ (ξ), y2 = f (ξ), Ie0ℓ =
[y2 , y1 ], if y1 > y2 .

Then
k k k
f ℓ (y1 ) = f ℓ+ℓ̂ (ξ) = f j1 ·2 (ξ) = x1 , f ℓ (y2 ) = f ℓ+ℓ̂+2 (ξ) = f (j1 +1)·2 (ξ) = x2 ,

and we have the covering sequence

ℓ
f g g g g g g
Ie0ℓ −→ I1 −→ I2 −→ I3 −→ · · · −→ Im−1 −→ I1 −→ I1 ∪ I2 ,
k ℓ and Ieℓ , overlapping at
where g = f 2 . So Ie0ℓ contains two subintervals Ie0,1 0,2
most at endpoints, such that
k +ℓ k +ℓ
f m·2 (Ie0,1
ℓ
) = I1 , f m·2 (Ie0,2
ℓ
) = I2 .

In general, if Ie0,a
ℓ
1 ...ap
is constructed, for aj = 1, 2, j = 1, 2, . . . , p, satisfying the
following covering sequences
 Unbounded Growth of Total Variations 33

(i) if ap = 1, then

(3.18)
fpm·2k +ℓ g g g g g
Ie0,a
ℓ
1 ... ,ap−1 1
−−−−−−−→ I1 −→ I2 −→ · · · −→ Im−1 −→ I1 −→ I1 ∪ I2 ;

(ii) if ap = 2, then

(3.19)
f pm·2k +ℓ g g g g g g
Ie0,a
ℓ
1 ...ap−1 2
−−−−−−−→ I2 −→ I3 −→ · · · −→ Im−1 −→ I1 −→ I1 −→ I1 ∪ I2 .

From (3.18) and (3.19), we have two subintervals Ie0,a

ℓ
1 ...ap 1
and Ie0,a
ℓ
1 ...ap 2
of
Ie0,a ...ap , such that
1

k +ℓ k +ℓ
f (p+1)·m·2 (Ie0,a
ℓ
1 ...ap 1
) = I1 , f (p+1)·m·2 (Ie0,a
ℓ
1 ...ap 2
) = I2 .

The rest of the arguments follows in the same way as in the proof of Theorem 3.1.
Therefore (3.16) follows from each ℓ ∈ {0, 1, . . . , m·2k −1}. Hence (3.15) follows.
Theorem 3.2. Consider the IBVP (1.6)–(1.8), and (1.12). Assume
that the composite reflection map f = Gη ◦ Fα,β has a periodic orbit O =
{f ℓ (ξ) | ℓ = 0, 1, . . . , m · 2k − 1}, with prime period m · 2k , where m is odd.
Further, assume that the initial conditions u0 and v0 in (1.12) are continuous
and satisfy the compatibility conditions in (1.13) such that for some integer
j0 : 0 ≤ j0 ≤ m · 2k − 1,
k
(3.20) f j0 (ξ), f j0 +2 (ξ) ∈ Range z, z ≡ u0 or z ≡ v0 .

Then

(3.21) lim [VI¯(u(·, t)) + VI¯(v(·, t))] = ∞.

t→∞

k
Proof. Let us assume that {f j0 (ξ), f j0 +2 (ξ)} ⊆ Range u0 . Then we can
construct a subinterval Ie0ℓ as in (3.17) by letting ℓ = j0 therein. From the proof
of Cor. 3.1 and (1.14), we easily deduce that

lim [VI¯(u(·, n)) + VI¯(v(·, n))] = ∞.

n→∞

It is also easy to show that for any τ : 0 < τ < 1, by using the continuity of
the total variations with respect to τ , that

lim [VI¯(u(·, n + τ )) + VI¯(v(·, n + τ ))] = ∞.

n→∞

Therefore (3.21) follows.

34 Chen et al.

Remark 3.1. It seems natural to believe that Theorem 3.2 remains valid
even if condition (3.20) is weakened to the following:
“there exist integers j1 and j2 : 0 ≤ j1 < j2 ≤ m · 2k − 1, such that
(3.22) f j1 (ξ), f j2 (ξ) ∈ (Range u0 ) ∪ (Range v0 ).”
However, in order to establish (3.21) under condition (3.22), the arguments
used in the proof of Cor. 3.1 also need to be considerably strengthened in order
to take care of the laborious “bookkeeping” details of finer interval covering
sequences, which we do not yet have an elegant way to handle so far. 
Next, we study the growth of total variations of snapshots (u(·, t), v(·, t))
when the composite reflection map Gη ◦ Fα,β has homoclinic orbits. There are
two cases to be considered: (i) η > 1, and (ii) 0 < η < 1.
Write f = Gη ◦ Fα,β . Here we only consider the case that f has a bounded
invariant interval J such that f : J → J. For this to hold, we need [5,
Lemma 2.5] either
r r
1+η1+α 1+α 1 + η 1 + αη
(3.23) (i) M ≡ ≤ , if 0 < η < 1,
1−η 3 3β 2η βη
or
r r
1+η1+α 1+α 1+η α+η
(3.24) (ii) M ≡− ≤ , if η > 1,
1−η 3 3β 2 β
in addition to (1.15), with
(3.25) J = [−M, M ].
Two graphs of f are provided in Figs. 9 and 10, where η = 0.552, 1.812,
respectively.
Theorem 3.3. Assume that 0 < α ≤ 1, β > 0, η > 0 and η 6= 1. Assume
also that (1.15), (3.23)–(3.25) are also satisfied so that J = [−M, M ] is a
bounded invariant interval of Gη ◦ Fα,β . Then
lim VJ ((Gη ◦ Fα,β )n ) = ∞.
n→∞
Proof. We first consider the case η > 1. Define a sequence of points
{xi ∈ J | i = 0, 1, 2, . . . } as follows. Let
(3.26)
q 
x0 = vI = 1+α
β , the positive v-axis intercept of f [5, (32), p. 428], 




x1 = −1
min{f (x0 )}, 



x2 = f −1 (x1 ), 
.. 
. 



xn+1 = f −1 (x n ), 

.. 


.
 Unbounded Growth of Total Variations 35

Then for n = 0, 1, 2, . . . , xn ∈ J and xn ↓ 0 as n → ∞. Also, define subintervals

(3.27) I0 = [x1 , x0 ], I1 = [x2 , x1 ], . . . , In = [xn+1 , xn ], . . . .

Then because f (I0 ) = [0, x1 ] we have the following covering sequence

n
[
f f f f f f
(3.28) In −→ In−1 −→ In−2 −→ · · · −→ I1 −→ I0 −→ Ij
j=1
n
[
f f f f f f
(3.29) −→ Ik −→ Ik−1 −→ · · · −→ I1 −→ I0 −→ Iℓ , for k = 1, . . . , n.
ℓ=1

Therefore from (3.28), In has n subintervals In,j , j = 1, 2, . . . , n, overlapping

at most at endpoints, such that

(3.30) f n (In,j ) = Ij , j = 1, 2, . . . , n.

Fig. 9. A degenerate homoclinic orbit of the map f = Gη ◦ Fα,β , where

α = 0.5, β = 1 and η = 0.552
0.552. (Reprinted from [5, p. 426], courtesy of World Scientific,
Singapore.)

Using the second part of the covering sequence in (3.29)

n
[
f f f f f f f f f
Ik −→ Ik−1 −→ · · · −→ I0 −→ In−k −→ · · · −→ I2 −→ I1 −→ I0 −→ Iℓ ,
ℓ=1
36 Chen et al.

Fig. 10. A degenerate homoclinic orbit of the map f = Gη ◦ Fα,β , where

α = 0.5
0.5, β = 1 and η = 1.812
1.812. (Reprinted from [5 p. 426], courtesy of World Scientific,
Singapore.)

we also obtain n subintervals Ik,1 , . . . , Ik,n of Ik , overlapping at most at

endpoints, such that

(3.31) f n (Ik,j ) = Ij , j = 1, . . . , n; for k = 1, 2, . . . , n − 1.

From (3.30) and (3.31), we obtain

V[0,x0] (f n ) ≥ V[xn+1 ,x0 ] (f n )

X n n
X
≥ VIk,j (f n ) ≥ [(x0 − x1 ) + (x1 − x2 ) + · · · + (xn − xn+1 )]
k,j=1 k=1

(3.32) = n(x0 − xn+1 ) → ∞, as n → ∞.

Next, we consider the case 0 < η < 1. Let us modify (3.26) only slightly by
redefining (3.26)2 as

(3.33) x1 = max{f −1 (x0 )}, x1 < 0.

The rest of (3.26) remains unchanged. Now, define intervals

I0 = [x2 , x0 ], I1 = [x1 , x3 ], I2 = [x4 , x2 ], I3 = [x3 , x5 ], . . . ,

(3.34) I2n = [x2n+2 , x2n ], I2n+1 = [x2n+1 , x2n+3 ], . . . ,

using the fact that

x1 < x3 < x5 < · · · < x2n+1 < · · · < 0 < · · · < x2n < x2n−2 < · · · < x4 < x2 < x0 .
 Unbounded Growth of Total Variations 37

Then because f (I0 ) = [x1 , 0], we have the following covering sequence
n
[
f f f f f f
I2n+1 −→ I2n −→ I2n−1 −→ · · · −→ I1 −→ I0 −→ I2j+1 .
j=0

The rest of the proof can be done in the same way as in (3.28)–(3.32).
Theorem 3.4. Consider the IBVP (1.6)–(1.8), (1.12). Assume that η
satisfies either (3.23) or (3.24) so that the composite reflection map f =
Gη ◦ Fα,β has a bounded invariant interval J = [−M, M ] and a homoclinic
orbit in J. Further, assume that the initial conditions u0 and v0 in (1.12)
satisfy the compatibility conditions in (1.13) such that

(3.35)
Range z ⊇ In , z ≡ u0 or z ≡ v0 for some n ∈ {0, 1, 2, . . . , }, cf. (3.27) or (3.34).

Then
lim [VI¯(u(·, t)) + VI¯(v(·, t))] = ∞.
t→∞
Proof. Same as that of Theorem 3.2.
Remark 3.2. We believe that (3.35) can be weakened at least to

Range u0 ∪ Range v0 ⊇ In , for some n ∈ {0, 1, 2, . . . }. 

Remark 3.3. The proof of Theorem 3.3 is essentially similar to those of

Theorem 3.1 and Corollary 3.1, and is based on the following fact:

“Let J be a finite closed interval and f : J → J is continuous.

(3.36) If f has a homoclinic orbit in J, then lim VJ (f n ) = ∞”.
n→∞

Actually, (3.36) above stands alone as a theorem itself and can also be proved
by quoting the proofs of Theorem 3.1 and Corollary 3.1, provided that the
homoclinic orbit in (3.36) is nondegenerate, because by Theorem 1.16.5 in
DevaneyP [8, p. 124], the map f is then topologically conjugate to the shift map
k
σ on 2 and, therefore, f has some periodic orbits of prime period m · 2 ,
with m being odd and k ∈ {0, 1, 2, . . . }. Hence the proofs of Theorem 3.1 and
Corollary 3.1 apply.
When the homoclinic orbit in (3.36) is degenerate, then f is “more chaotic”
(than the case when the homoclinic orbit is nondegenerate) and has homoclinic
bifurcations. The renormalization procedure for the “model case” quadratic map
fµ (x) = µx(1−x) as µ → 4 (the degenerate homoclinic orbit case) as mentioned
in [8, §1.16] suggests that for µ = 4, f4 should be in the “post period doubling
era” and therefore, f4 has many period-m · 2k orbits, with m being odd. It is
quite obvious that our f in (3.36) ought to also have this kind of period-m · 2k
orbits (when the homoclinic orbit in (3.36) is degenerate) and, therefore, the
38 Chen et al.

proofs of Theorem 3.1 and Corollary 3.1 again apply. Nevertheless, we could
not locate a precise reference to this effect.
In passing, we may also note that the condition (3.35) is stated quite
differently from (3.20), in the sense that the end-points of intervals In in (3.35)
are not periodic points. (Or rather, the end-points of In have an “infinite
periodicity”.) 

4 Miscellaneous Remarks
In this paper, we have successfully shown that when chaotic vibration occurs
for the wave equation caused by the nonlinear boundary condition specified
here, the total variations of snapshots tend to infinity as t → ∞ for a large
class of initial data, even though the total variation of any such initial data
is finite at time t = 0. Our theorems in §3 have covered the case of “stable”
chaos on a bounded invariant interval. A different type of “unstable” chaos,
discussed in [5, §5], happens on an invariant Cantor set (rather than a bounded
invariant interval, because the map Gη ◦ Fα,β does not have one for that set of
α, β and η values). In that case, it is trivial to show that the total variations of
snapshots also tend to infinity as t → ∞ for a large class of initial data, even
though initially, the total variation of the state is finite.
One may ask a converse question to [Q]:
“[−Q] Assume that there exist initial conditions (u0 , v0 ) for (1.6), (1.9), (1.10)
and (1.12) and an invariant interval I¯ of Gη ◦ Fα,β such that

(4.1) VI¯(u0 , v0 ) < ∞, VI¯(u(·, t)) + VI¯(v(·, t)) → ∞ as t → ∞.

Is the map Gη ◦ Fα,β necessarily chaotic?”

The answer is negative, as the following counterexample has shown.
Example 4.1. Let α = 0.5, β = 1, either η ∈ (0, 0.433) or η ∈ (2.312, ∞).
Then as Figs. 1 and 2 ([5, Figs. 3 and 4, pp. 433–434]) have shown, the map
Gη ◦ Fα,β has a locally attracting period-2 orbit near 0, which is repelling.
Let g(x) = x2 . For x ∈ [0, 1],

(4.2) 
 1


 x2 , if x = , n = 1, 2, . . . ,

 n

 2n + 1

 0, if x = , n = 1, 2, . . . ,
2n(n + 1)
u0 (x) = 2(n + 1) 2n + 1 h 2n + 1 1 i

 x− , if x ∈ ,

 n n2


 2n 2n + 1 h 2n(n
1
+ 1) n
2n + 1 


 − x+ , if x ∈ , .
n+1 (n + 1)2 n + 1 2n(n + 1)
Then u0 (x) is continuous. Choose any v0 , continuous of bounded total
h variation
i
1
satisfying the compatibility condition (1.13). On each subinterval n+1 , n1 , the
 Unbounded Growth of Total Variations 39

total variation of u0 is 1/n2 + 1/(n + 1)2 . Therefore u0 has bounded total

variation on the interval [0,1]. Let the period-2 orbit of Gη ◦ Fα,β be {p, −p},
where p > 0. Then for each y ∈ [0, 1], y 6= n1 for n = 1, 2, . . . , we have

lim |(Gη ◦ Fα,β )n (u0 (y))| = p.

n→∞
h i
1 1
Therefore, on each subinterval n+1 , n , the total variation of (Gη ◦Fα,β )n tends
to p, and

(4.3) lim [VI¯(u(·, t)) + VI¯(v(·, t))] = ∞, but VI¯(u0 ) + VI¯(v0 ) < ∞.
t→∞


The above negative result seems to have weakened the connection between
chaotic vibration and unbounded growth of total variations of snapshots.
However, we may take note of the following recent result in [10]. Let
f : J → J be chaotic (according to Devaney [8, p. 50]), on the finite closed
interval J. Then f has sensitive dependence on initial data [2]. This sensitive
dependence on initial data is regarded as a major feature of chaotic maps; [10]
has proved the following:

“(i) Let f : J → J has sensitive dependence on initial data. Then

lim VJ ′ (f n ) = ∞ for every closed subinterval J ′ of J.
n→∞

(ii) Let f : J → J be continuous with finitely many extremal points,

satisfying lim VJ ′ (f n ) = ∞ for every closed subinterval J ′ of J. Then f
n→∞
has sensitive dependence on initial data.”

The theorems in [10] actually explains why the breakdown (4.3) happens: the
initial data in (4.2) has infinitely many extremal points, i.e., there are infinitely
many oscillations on a finite closed interval and, thus, it is “too oscillatory”.
The growth rate of VI¯(f n ) as estimated in (3.14) and (3.32) are linear with
respect to n. Sharper estimates may also be possible, at least for certain special
cases. In [9], examples of exponential growth have been found. Related issues
such as Remarks 3.1 and 3.2 and others are also being investigated in [9].

Appendix A Key Proposition

In this section, we prove the following.
Proposition A. Assume that u and v are related through
Z t
(A.1) v(t) = αu(t) + β e−γ(t−τ ) u(τ )dτ + f (t), t ≥ 0,
0
40 Chen et al.

where α, β ∈ R, γ > 0. Then

 
|β| |β|
(A.2) V[0,T ] (v) ≤ |α| + V[0,T ] (u) + |u(0+)| + V[0,T ] (f ),
γ γ
for all T > 0, where u(0+) = lim u(t).
t→0+
We first establish the following technical lemma.
Lemma B. Let γ > 0 and
Z t
g(t) = [Qu](t) ≡ e−γ(t−τ ) u(τ )dτ, ∀ t ≥ 0.
0

Then
1
(A.3) V[0,T ] (g) ≤ [|u(0+)| + V[0,T ] (u)], ∀ T > 0.
γ

Proof.
(1) We first assume that u is increasing and continuous. Then
Z t
′
g (t) = −γ e−γ(t−s) u(s)ds + u(t)
0
s=t Z t

= −e−γ(t−s) u(s) + e−γ(t−s) du(s) + u(t)
s=0 0
Z t
(A.4) = e−γt u(0) + e−γ(t−s) du(s).
0

Note that the integral in (A.4) is a Stieltjes integral [1, Chap 9]. Since g
is absolutely continuous, we see that
Z T Z T Z TZ t
′ −γt
V[0,T ] (g) = |g (t)|dt ≤ |u(0)| e dt + e−γ(t−s) du(s)dt
0 0 0 0
Z T Z T 
1
≤ |u(0)| + e−γt dt eγs du(s)
γ 0 s
Z
1 1 T −γs
= |u(0)| + (e − e−γT )eγs du(s)
γ γ 0
Z
1 1 T
≤ |u(0)| + du(s)
γ γ 0
1 1
= |u(0) + [u(T ) − u(0)]
γ γ
1
= [|u(0)| + V[0,T ] (u)].
γ
Therefore (A.3) is true when u is increasing and continuous.
 Unbounded Growth of Total Variations 41

(2) We now assume that u is increasing and left-continuous. Then

X
(A.5) u = u0 + rc H c ,
c∈J

where u0 is increasing and continuous, u0 (0) = u(0+), J is a (possibly

empty) countable set of nonnegative real numbers, rc > 0, and

0, 0 ≤ t ≤ c,
Hc (t) =
1, t > c,

is the Heaviside function. Then

X
V[0,T ] (u) = V[0,T ] (u0 ) + rc ,
c∈J,c<T

Z t  0, 0 ≤ t ≤ c,
hc (t) ≡ [QHc ](t) = e−γ(t−s) Hc (s)ds = 1
0  [1 − e−γ(t−c) ], c < t.
γ

Since hc (t) is increasing, V[0,t] (hc ) = hc (T ) and, thus,


= 0, if T ≤ c,
V[0,T ] (hc ) 1
≤ , for all T ≥ 0.
γ

From (A.5), g = Qu0 + Σrc QHc , so

X
V[0,T ] (g) ≤ V[0,T ] (Qu0 ) + rc V[0,T ] (QHc )
1 X rc
≤ [|u0 (0)| + V[0,T ] (u0 )] + (by part (1))
γ γ
c<T
1
= [|u(0+)| + V[0,T ] (u)].
γ

Therefore (A.2) is proved for increasing, left-continuous function u.

(3) Now we consider a left-continuous function u. Define

1
u1 (t) = [V[0,t] (u) + u(t) + u(0)],
2
1
u2 (t) = [V[0,t] (u) − u(t) + u(0)].
2
Then u1 and u2 are both left-continuous, increasing functions, and u =
u1 − u2 , V[0,T ] (u) = V[0,T ] (u1 ) + V[0,T ] (u2 ), u1 (0+) = u(0+), u2 (0+) = 0.
42 Chen et al.

By part (2) above,

V[0,T ] (g) ≤ V[0,T ] (Qu1 ) + V[0,T ] (Qu2 )

1 1
≤ [|u1 (0+)| + V[0,T ] (u1 )] + [|u2 (0+)| + V[0,T ] (u2 )]
γ γ
1 1
= |u(0+)| + V[0,T ] (u).
γ γ

(4) We can now consider the general case. Let


 lim u(s), for t > 0,
û(t) = s→t−
u(0), if t = 0.

Then û is left-continuous, Qû = Qu, V[0,T ] (û) ≤ V[0,T ] (u) and û(0+) =
u(0+). By part (3),

V[0,T ] (g) = V[0,T ] (Qu) = V[0,T ] (Qû)

1 1
≤ [|û(0+)| + V[0,T ] (û)] ≤ [|u(0+)| + V[0,T ] (u)].
γ γ

Therefore (A.3) is proved.

Proof. [Proof of Proposition A] From (A.1), we have, by (A.3),

1
V[0,T ] (v) ≤ |α|V[0,T ] (u) + |β| · [|u(0+)| + V[0,T ] (u)] + V[0,T ] (f ).
γ

Therefore, (A.2) holds.

References

[1] T.M. Apostal, Mathematical Analysis, Addison Wesley, Reading, MA, 1957.
[2] J. Banks, J. Brooks, G. Cairns, G. Davis and P. Stacey, On Devaney’s definition
of chaos, Amer. Math. Monthly 99 (1992), 332–334.
[3] G. Chen, S.B. Hsu and J. Zhou, Linear superposition of chaotic and orderly
vibrations on two serially connected strings with a van der Pol joint, Int. J.
Bifurcation and Chaos 6 (1996), 1509–1527.
[4] G. Chen, S.B. Hsu and J. Zhou, Chaotic vibrations of the one-dimensional
wave equation due to a self-excitation boundary condition, Part I: Controlled
hysteresis, Trans. Amer. Math. Soc. 350 (1998), 4265–4311.
[5] G. Chen, S.B. Hsu and J. Zhou, Chaotic vibration of the one-dimensional wave
equation due to a self-excitation boundary condition, Part II: Energy injection,
period doubling and homoclinic orbits, Int. J. Bifurcation and Chaos 8(3) (1998),
423–445.
 Unbounded Growth of Total Variations 43

[6] G. Chen, S.B. Hsu and J. Zhou, Chaotic vibrations of the one-dimensional wave
equation due to a self-excitation boundary condition, Part III: Natural hysteresis
memory effect, Int. J. Bifurcation and Chaos 8(3) (1998), 447–470.
[7] G. Chen and J. Zhou, Vibration and Damping in Distributed Systems, Vol I:
Analysis, Estimation, Attenuation and Design, CRC Press, Boca Raton, FL, 1993.
[8] R.L. Devaney, An Introduction to Chaotic Dynamical Systems, Addison-Wesley,
New York, 1989.
[9] T.W. Huang, Ph.D. Dissertation, Department of Mathematics, Texas A&M
University, in progress.
[10] T.W. Huang and G. Chen, Chaotic behavior of interval maps as characterized by
unbounded growth of total variations of their n-th iterates as n → ∞, preprint.
[11] C. Robinson, Dynamical Systems, Stability, Symbolic Dynamics and Chaos, CRC
Press, Boca Raton, FL, 1995.
44 Chen et al.
Velocity method and Courant metric topologies in
shape analysis of partial differential equations

Michel C. Delfour1 , Université de Montréal, C. P. 6128, Montréal Qc, Canada

H3C 3J7, E-mail: [email protected]
Jean-Paul Zolésio, CNRS, Centre de Mathématiques Appliquées, Ecole des
Mines, INRIA, 2004 route des Lucioles, BP 93, 06902 Sophia Antipolis Cedex,
France. E-mail: [email protected]
Abstract
The object of this paper is to give the equivalence between the
continuity of a shape function with respect to the generic complete Courant
metric topology introduced by A.M. Micheletti in 1972 and its continuity
along the flow of non-autonomous velocity fields. This result also clarifies
the connection between spaces of domains endowed with a Courant metric
and the Velocity Method used in the definition of semiderivatives of shape
functions.

1 Introduction
The Shape Analysis of partial differential equations deals with problems where
the underlying domain has a free boundary, changes with time, or is a control
or design variable. Spaces of shapes or geometries are usually nonlinear and
nonconvex spaces. The dependence of the solution of a pde with respect to
the domain as a variable is also nonlinear. Perhaps the first appearance of the
geometry as a time-dependent variable in Control Theory is due to Bardos and
Chen [1] who showed that stabilization of the wave equation could be achieved
by moving (time like) the boundary of the domain. Truchi and Zolésio [8]
extended this result to membrane equations by periodic motion of the boundary.
More recently D.L. Russell introduced the notion of formability [7]2 which can
be seen as an equivalent of controllability for shapes. There are many ways
to tackle this general class of problems. For instance the Velocity Method was

1
This paper was completed while the first author was on sabbatical leave at INRIA-
Rocquencourt (France) in the projects MACS and SOSSO. The research of the first author has
been supported by National Sciences and Engineering Research Council of Canada research
grant A–8730 and by a FCAR grant from the Ministère de l’Éducation du Québec.
2
The ”theory of formability” characterizes the ability to deform a body into a prescribed
shape through control of the microstructure of the material.

45
46 M.C. Delfour and Jean-Paul Zolésio

used by Truchi and Zolésio while the Courant metric was used by Micheletti [5]
to study the dependence of the first eigenvalue of an elliptic pde with respect
to the domain.
The admissible domains are often defined as the set of images of a fixed
subset of RN through a family of transformations of RN . The composition
of transformations induces a natural group structure on images and the full
power of function analytic methods is available to define a topology on the
group. The choice is obviously very much problem dependent. In 1972 A.M.
Micheletti [5] gave perhaps one of the first complete metric topologies on a
family of domains of class C k which are the images of a fixed bounded open
connected domain of class C k through a family of C k diffeomorphisms of RN .
Her analysis culminates with the construction of the Courant metric which
makes the associated quotient group of diffeomorphisms a complete metric
space. Her construction and results are quite general and even generic for
other families of transformations of RN including the ones considered by Murat
and Simon [6] later in 1976.
Another point of view is the so called Velocity Method (cf. [10, 9, 3, 4]),
where the transformations of RN which define the variations of the shape are
constructed from the flow of non-autonomous velocity fields. This has been
extensively used to define semiderivatives of a shape function with respect to
the shape of the underlying domain.
This paper deals with the long standing issue of the connection between the
Velocity Method and topologies generated by Courant metrics. Its main object
is to give the equivalence between the continuity of a shape function with respect
to the generic complete Courant metric topology and its continuity along the
flow of non-autonomous velocity fields. We specialize the equivalence for the
transformations associated with the spaces C0k+1 (RN , RN ), C k+1 (RN , RN ), and
C k,1 (RN , RN )3 .

2 Courant metric topology on C k domains

In this section we recall and quote the main constructions and results in [5].
def
Given an integer k ≥ 1, denote by C k (RN ) = C k (RN , RN ) the space of k-times
continuously differentiable transformations f of RN . When endowed with the
norm
def def
kf kC k = max k∂ α f kC , k∂ α f kC = sup |∂ α f |,
0≤|α|≤k x∈RN

For k ≥ 0, C0k+1 (RN , RN ) is the space of all (k + 1)-times continuously differentiable

3

mappings f : RN → RN for which f and all its partial derivatives up to order k + 1 vanish
at infinity. C k+1 (RN , RN ) is the space of mappings f : RN → RN for which f and ∂ α f are
bounded and uniformly continuous on RN for all α, 1 ≤ |α| ≤ k + 1. C k,1 (RN , RN ) is the
space of mappings f : RN → RN for which f and ∂ α f are bounded and uniformly Lipschitz
continuous on RN for all α, 1 ≤ |α| ≤ k.
 Velocity method and Courant metric topologies 47
def
C0k (RN ) = C0k (RN , RN ) is a Banach space. Associate with C0k (RN ) the space
def
n o
F0k = F : RN → RN : F − I ∈ C0k (RN ) and F −1 ∈ C k (RN ) .

Theorem 2.1. For k ≥ 1 the space F0k is a group with respect to the composition
◦ of transformations.
Lemma 2.1. Given f and g in C k (RN ), let ψ = f ◦ (I + g). Then for each
x ∈ RN

|ψ(x)| = |f (x + g(x))|
(1)
|ψ (x)| ≤ |f (1) (x + g(x))| [1 + |g(1) (x)|]
|ψ (i) (x)| ≤ |f (1) (x + g(x))| |g(i) (x)|
i
X
+ |f (j) (x + g(x))| aj (|g(1) (x)|, . . . , |g(i−1) (x)|)
j=2

for i = 2, . . . , k where aj is a polynomial.

Lemma 2.2. Given integers r ≥ 0 and s > 0 there exists a constant c(r, s) > 0
with the following property: if the sequence f1 , . . . , fn in C r (RN ) is such that
Xn
(2.1) kfi kC r < α, 0 < α < s,
i=1

then for the map F = (I + fn ) ◦ · · · ◦ (I + f1 )

(2.2) kF − IkC r ≤ α c(r, s).

Associate with a fixed open or closed subset4 Ω0 of RN the following family

of images of Ω0 by the elements of F0k

def
n o
X (Ω0 ) = F (Ω0 ) ⊂ RN : ∀F ∈ F0k

This induces a bijection

(2.3) [F ] 7→ F (Ω0 ) : F0k /G(Ω0 ) → X (Ω0 ).

between X (Ω0 ) and the quotient group of F0k by the subgroup

def
n o
(2.4) G(Ω0 ) = F ∈ F0k : F (Ω0 ) = Ω0

4
In her paper [5] A.M. Micheletti assumes that Ω0 is a bounded connected open domain
of class C k in order to make all the images F (Ω0 ) bounded connected open domains of class
C k . However the construction of the Courant metric only requires that Ω0 be closed or open.
48 M.C. Delfour and Jean-Paul Zolésio

of transformations which map Ω0 onto Ω0 .

The next step is to construct a complete metric space topology on F0k which
induces a complete metric space topology on the quotient group F0k /G(Ω0 ). The
space X (Ω0 ) is then identified with the topological quotient group F0k /G(Ω0 ).
This quotient metric is called the Courant metric by A.M. Micheletti. The
construction of the metric topology on F0k /G(Ω0 ) is not as straightforward as
it might appear at first sight. The obvious candidates for the metric do not
usually satisfy the triangle inequality and only yield a pseudo-metric on the
quotient group. In order to get around this difficulty Micheletti introduced the
following construction. Given F ∈ F0k , consider finite factorizations of F and
F −1 of the form

F = (I + fn ) ◦ · · · ◦ (I + f1 ) and F −1 = (I + gm ) ◦ · · · ◦ (I + g1 )

Define
n
X m
X
def
(2.5) d(I, F ) = inf kfi kC k + inf kgi kC k
(f1 ,...,fn ) (g1 ,...,gm )
i=1 i=1

where the infima are taken with respect to all finite factorizations of F and F −1
in F0k . Extend this definition to all pairs F and G in F0k

def
(2.6) d(F, G) = d(I, G ◦ F −1 )

By definition d is right-invariant since for all F , G and H in F0k

(2.7) d(F, G) = d(F ◦ H, G ◦ H)

and the three axioms which define a metric on F0k are verified.
Theorem 2.2. F0k is a complete metric group.
Corollary 2.1. The topology induced by the metric d on the topological group
F0k coincides with the topology which has as a basis of neighborhoods of the
identity in F0k the sets
n o
def
E(ε) = F ∈ F0k : kF − IkC k + kF −1 − IkC k < ε .

It is readily seen that we have the following properties.

Lemma 2.3. Given an open or a closed5 subset Ω0 of RN , the family
n o
def
G(Ω0 ) = F ∈ F0k : F (Ω0 ) = Ω0

is a closed subgroup of F0k .

5
As noted earlier it is not necessary that Ω0 be a bounded connected open C k domain.
 Velocity method and Courant metric topologies 49

By definition for each Ω ∈ X (Ω0 ) there exists F ∈ F0k such that Ω = F (Ω0 ).
Therefore the following map is well-defined and bijective

(2.8) Ω 7→ χ(Ω) = F ◦ G(Ω0 ) : X (Ω0 ) → F0k /G(Ω0 ).

Using χ we now introduce the following complete metric on X (Ω0 ).

Theorem 2.3. Given an open or a closed subset Ω0 of RN , the function
def
(2.9) δ(F ◦ G(Ω0 ), H ◦ G(Ω0 )) = inf d(F ◦ G, H ◦ G̃)
G,G̃∈G(Ω0 )

is a metric on F0k /G(Ω0 ). The topology induced by δ coincides with the quotient
topology of F0k /G(Ω0 ) and the space F0k /G(Ω0 ) is complete.
Finally it is natural to define on X (Ω0 ) the following metric

def
(2.10) ρ(Ω1 , Ω2 ) = δ(χ(Ω1 ), χ(Ω2 ))

induced by the bijection χ on X (Ω0 ). With that metric, X (Ω0 ) is a complete

metric space. This metric is called the Courant metric (metrica di Courant) in
[5].

3 The generic framework of Micheletti

The construction of a complete metric topology on the group and a Courant
metric on the quotient group naturally extends to other underlying spaces Θ
than the Banach space C0k (RN ). Given a Banach space Θ of transformations of
RN , define the space

def 
F(Θ) = F : RN → RN : F − I ∈ Θ and F −1 − I ∈ Θ .

Associate with F ∈ F(Θ) the distance

n
X m
X
def
(3.1) d(I, F ) = inf kfi kΘ + inf kgi kΘ ,
(f1 ,...,fn ) (g1 ,...,gm )
i=1 i=1

where the infima are taken over all finite factorizations in F(Θ) of the form

F = (I + fn ) ◦ · · · ◦ (I + f1 ) and F −1 = (I + gm ) ◦ · · · ◦ (I + g1 ), fi , gi ∈ Θ.

Extend this definition to all pairs F and G in F(Θ)

def
(3.2) d(F, G) = d(I, G ◦ F −1 )
50 M.C. Delfour and Jean-Paul Zolésio

Define for some fixed open or closed subset Ω0 of RN the subgroup

def
G(Ω0 ) = {F ∈ F(Θ) : F (Ω0 ) = Ω0 }

and the Courant metric on the equivalence classes of F(Θ)/G(Ω0 )

def
(3.3) ∀F, G ∈ F(Θ), δ([F ], [H)] = inf d(F ◦ G, H ◦ G̃)
G,G̃∈G(Ω0 )

def
(3.4) ∀F1 , F2 ∈ F(Θ), ρ(F1 (Ω0 ), F2 (Ω0 )) = δ([F1 ], [F2 ])

on the images F (Ω0 ) of Ω0 by elements F of F(Θ). Of course F(Θ)/G(Ω0 )

will be a well-defined complete metric space for the Courant metric only for
appropriate choices of spaces Θ.
As an illustration consider the spaces of transformations of RN introduced
by Murat and Simon [6] in 1976 in the construction of metric spaces of domains.
For integers k ≥ 0, they correspond to the following choices of the space Θ:

W k+1,c(RN , RN )
def
n o
= f ∈ W k+1,∞(RN , RN ) : ∀ 0 ≤ |α| ≤ k + 1, ∂ α f ∈ C(RN , RN )

and W k+1,∞(RN , RN ). The first space W k+1,c (RN , RN ) algebraically and

def
topologically coincides with the space C k+1 (RN ) = C k+1 (RN , RN ). The
corresponding space of transformations

def 
F k+1 (RN ) = F : RN → RN :

F − I ∈ C k+1 (RN ) and F −1 − I ∈ C k+1 (RN )

is a topological group for the metric d as in the case of F0k+1 in § 2. The

def
second space W k+1,∞ (RN , RN ) coincides with C k,1 (RN ) = C k,1 (RN , RN ). The
associated space of transformations

def 
F k,1 (RN ) = F : RN → RN :
(3.5)
F − I ∈ C k,1 (RN ) and F −1 − I ∈ C k,1 (RN )

is complete for the topology induced by the metric d, but is not a topological
group. In both cases the Courant metric defines a complete metric topology on
the corresponding quotient space. Also recall from [6] that F 1 (RN ) transports
locally Lipschitzian (graph) domains onto locally Lipschitzian (graph) domains
but that F 0,1 (RN ) does not.
 Velocity method and Courant metric topologies 51

3.1 Approach of Murat and Simon

The construction introduced by Murat and Simon [6] to obtain a complete
metric topology on the quotient spaces are different from the ones of Micheletti
[5] which were seemingly not known to them. They worked with the pseudo-
metric
def
dp (F2 , F1 ) = kF2 ◦ F1−1 − IkW k+1,∞ + kF1 ◦ F2−1 − IkW k+1,∞

rather than the metric defined by the infima over finite factorizations of F2 ◦F1−1
and F1 ◦ F2−1 . They recover a metric from the pseudo-metric by using an
auxiliary construction which depends on the third property of a pseudo-metric.
We briefly recall the definition and the result.
Definition 3.1. A pseudo-metric on a space E is a function δ : E × E → R+
with the following properties
(i) δ(F2 , F1 ) = 0 ⇐⇒ F2 = F1
(ii) δ(F2 , F1 ) = δ(F1 , F2 ) for all F1 and F2
(iii) δ(F1 , F3 ) ≤ δ(F1 , F2 ) + δ(F2 , F3 ) + δ(F1 , F2 ) δ(F2 , F3 ) P (δ(F1 , F2 ) +
δ(F2 , F3 )) for all Fi ’s, where P : R+ → R+ is a continuous increas-
ing function.
Proposition 3.1. Let δ be a pseudo-metric on E. For all α, 0 < α < 1, there
exists a constant ηα > 0 such that the function δ(α) : E × E → R+ defined as
def
δ(α) (F1 , F2 ) = inf{δ(F1 , F2 ), ηα }α
is a metric on E.

3.2 Approach of Micheletti

The pseudo-metric can be completely by-passed. By combining the construction
of [5] with the properties established in [6], we readily get the completeness of
the (Micheletti) metric for the group of transformations and of the Courant
metric for the quotient group. In both cases the results can also be obtained
directly by adapting with obvious technical changes in the case C k,1 ((RN ) the
sequence of lemmas and theorems of § 2. The first case is technically analogous
to F0k by choosing Θ = C k (RN ).
Theorem 3.1. Let k ≥ 1 be an integer.
(i) The topology induced by d on F k (RN ) makes it a complete metric
topological group. Moreover around the identity I for all 0 < ε < 1

n E(ε) ⊂ S(ε) ⊂ E(2c(k, 1)ε) o

def N ) : kF − Ik k + kF −1 − Ik k < ε .
E(ε) = F ∈ F k (R
n C o C
def
S(ε) = F ∈ F k (RN ) : d(I, F ) < ε .
52 M.C. Delfour and Jean-Paul Zolésio

and the topology coincides with the topology which has as a basis of
neighborhoods of the identity in F k (RN ) the sets E(ε) (the constant c(k, 1)
is as specified in Lemma 2.2).
(ii) Given an open or closed subset Ω0 of RN ,
def
(3.6) δ(F ◦ G, H ◦ G) = inf d(F ◦ G, H ◦ G̃)
G,G̃∈G(Ω0 )

is a metric on F k (RN )/G(Ω0 ). The topology induced by δ is complete and

coincides with the quotient topology of F k (RN )/G(Ω0 ).
In the other case Θ = W k+1,∞ (RN , RN ) is equal to C k,1 (RN ) which is a Banach
space when endowed with the norm
def
kf kC k,1 = kf kC k + ck (f )
(3.7) def |f (y) − f (x)| def |∂ α f (y) − ∂ α f (x)|
c(f ) = sup ck (f ) = max sup .
x6=y |y − x| |α|=k x6=y |x − y|

We get a similar result for F k,1 (RN ) except that it is not a topological group.
Theorem 3.2. Let k ≥ 1 be an integer.
(i) F k,1 (RN ) is a group under composition. The function d induces a
complete metric topology on F k,1 (RN ). Moreover around the identity I
for all 0 < ε < 1

n E(ε) ⊂ S(ε) ⊂ E(c(k) ε) o

def k,1 N ) : kF − Ik k,1 + kF −1 − Ik k,1 < ε .
E(ε) = F ∈ F (R n C o C
def
S(ε) = F ∈ F k,1 (RN ) : d(I, F ) < ε
for some constant c(k) independent of ε.
(ii) Given an open or closed subset Ω0 of RN ,
def
(3.8) δ(F ◦ G, H ◦ G) = inf d(F ◦ G, H ◦ G̃)
G,G̃∈G(Ω0 )

is a metric on F k,1 (RN )/G(Ω0 ). The topology induced by δ is complete

and coincides with the quotient topology of F k,1 (RN )/G(Ω0 ).

4 Unconstrained families of domains

In this section we study equivalences between the Velocity Method (cf. [9],
[10]) and methods using a family of transformations. In § 4.1 we give some
general conditions to construct a family of transformations of RN from a
non-autonomous velocity field. Conversely we show how to construct a non-
autonomous velocity field from a family of transformations of RN . In § 4.2 the
equivalences of § 4.1 are specialized to velocities in C0k+1 (RN ), C k+1 (RN ), and
C k,1 (RN ), k ≥ 0.
 Velocity method and Courant metric topologies 53

4.1 Equivalence between velocities and transformations

Let the real number τ > 0 and the map V : [0, τ ] × RN → RN be given. The
map V can be viewed as a family {V (t) : 0 ≤ t ≤ τ } of non-autonomous velocity
fields on RN defined by
def
(4.1) x 7→ V (t)(x) = V (t, x) : RN 7→ RN .
Assume that


∀x ∈ RN , V (·, x) ∈ C [0, τ ]; RN
(4.2) (V)
∃c > 0, ∀x, y ∈ RN , kV (·, y) − V (·, x)kC([0,τ ];RN ) ≤ c|y − x|
where V (·, x) is the function t 7→ V (t, x). Associate with V the solution x(t; V )
of the ordinary differential equation
dx

(4.3) (t) = V t, x(t) , t ∈ [0, τ ], x(0) = X ∈ RN ,
dt
and introduce the homeomorphism
def
(4.4) X 7→ Tt (V )(X) = xV (t; X) : RN → RN .
and the maps
def
(4.5) (t, X) 7→ TV (t, X) = Tt (V )(X) : [0, τ ] × RN → RN ,
def
(4.6) (t, x) 7→ TV−1 (t, x) = Tt−1 (V )(x) : [0, τ ] × RN → RN .

In the sequel we shall drop the V in TV (t, X), TV−1 (t, x) and Tt (V ) whenever
no confusion arises.
Theorem 4.1.
(i) Under assumption (V) the map T has the following properties:


∀X ∈ RN , T (·, X) ∈ C 1 [0, τ ]; RN and ∃c > 0,
(T1)
∀X, Y ∈ RN , kT (·, Y ) − T (·, X)kC 1 ([0,τ ];RN ) ≤ c|Y − X|,
(4.7) (T2) ∀t ∈ [0, τ ], X 7→ Tt (X) = T (t, X) : RN → RN is bijective,


∀x ∈ RN , T −1 (·, x) ∈ C [0, τ ]; RN and ∃c > 0,
(T3)
∀x, y ∈ RN , kT −1 (·, y) − T −1 (·, x)kC([0,τ ];RN ) ≤ c|y − x|.

(ii) Given a real number τ > 0 and a map T : [0, τ ] × RN → RN verifying

assumptions (T1), (T2) and (T3), then the map
def ∂T

(4.8) (t, x) 7→ V (t, x) = t, Tt−1 (x) : [0, τ ] × RN → RN
∂t
verifies assumption (V), where Tt−1 is the inverse of X 7→ Tt (X) =
T (t, X). If, in addition, T (0, ·) = I, then T (·, X) is the solution of (4.3).
54 M.C. Delfour and Jean-Paul Zolésio

(iii) Given a real number τ > 0 and a map T : [0, τ ] × RN → RN satisfying

assumptions (T1), (T2) and T (0, ·) = I, then there exists τ ′ > 0 such that
the conclusions of part (ii) hold on [0, τ ′ ].
This equivalence theorem says that we can either start from a family of velocity
fields {V (t)} on RN or a family of transformations {Tt } of RN provided that
the map V , V (t, x) = V (t)(x), verifies (V) or the map T , T (t, X) = Tt (X),
verifies (T1) to (T3).
Starting from V , the family of homeomorphisms {Tt (V )} generates the
family
(4.9) Ωt = Tt (V )(Ω) = {Tt (V )(X) : X ∈ Ω}.
of perturbations of the initial domain Ω. Interior (resp. boundary) points of Ω
are mapped onto interior (resp. boundary) points of Ωt . This is the basis of the
Velocity method.

4.2 Equivalence for special families of velocities

In this section we specialize Theorem 4.1 to velocities in C k,1 (RN ), C0k+1 (RN ),
and C k+1 (RN ), k ≥ 0. The following notation will be helpful
def dTt def
f (t) = Tt − I, f ′ (t) = , g(t) = Tt−1 − I,
dt
whenever Tt−1 exists and the identities
g(t) = −f (t) ◦ Tt−1 = −f (t) ◦ [I + g(t)]
dTt
V (t) = ◦ Tt−1 = f ′ (t) ◦ Tt−1 = f ′ (t) ◦ [I + g(t)].
dt
Recall also the notation c(f ) and ck (f ) in (3.7).
Theorem 4.2. Let k ≥ 0 be an integer.
(i) Given τ > 0 and a velocity field V such that
(4.10) V ∈ C([0, τ ]; C k (RN , RN )) and ck (V (t)) ≤ c
for some constant c > 0 independent of t, the map T given by (4.3)-(4.5)
satisfies conditions (T1), (T2), and
(4.11)
f ∈ C 1 ([0, τ ]; C k (RN , RN )) ∩ C([0, τ ]; C k,1 (RN , RN )), ck (f ′ (t)) ≤ c
for some constant c > 0 independent of t. Moreover condition (T3) is
satisfied and there exists τ ′ > 0 such that
(4.12) g ∈ C([0, τ ′ ]; C k (RN , RN )), ck (g(t)) ≤ ct
for some constant c independent of t.
 Velocity method and Courant metric topologies 55

(ii) Given τ > 0 and T : [0, τ ] × RN → RN satisfying conditions (4.11) and

T (0, ·) = I, there exists τ ′ > 0 such that the velocity field V (t) = f ′ (t)◦Tt−1
satisfies conditions (V) and (4.10) in [0, τ ′ ].

Proof. We prove the theorem for k = 0. The general case is obtained by

induction over k. (i) By assumption on V , the conditions (V) given by (4.2) are
satisfied and by Theorem 4.1 the corresponding family T satisfies conditions
(T1) to (T3).
(Conditions (4.11) on f ). For any x and s ≤ t
Z t
Tt (x) − Ts (x) = V (r) ◦ Tr (x) dr
Z t s
|Tt (x) − Ts (x)| ≤ c|Tr (x) − Ts (x)| + |V (r) ◦ Ts (x)| dr
s
Z t
|f (t)(x) − f (s)(x)| ≤ c|f (r)(x) − f (s)(x)| + kV (r)kC dr.
s

By assumption on V and Gronwall’s inequality

∀t, s ∈ [0, τ ], kf (t) − f (s)kC ≤ c |t − s|

for another constant c independent of t. Moreover

|(f (t) − f (s))(y) − (f (t) − f (s))(x)| = |(Tt − Ts )(y) − (Tt − Ts )(x)|

Z t
≤ |V (r) ◦ Tr (y) − V (r) ◦ Tr (x)| dr
s
Z t
≤ c|(Tr − Ts )(y) − (Tr − Ts )(x)| + c|Ts (y) − Ts (x)| dr
s
Z t
≤ c|(f (r) − f (s))(y) − (f (r) − f (s))(x)| + cc′ |y − x| dr
s

for some other constant c′ by the second condition (T1). Again by Gronwall’s
inequality’s there exists another constant c such that

|(f (t) − f (s))(y) − (f (t) − f (s))(x)| ≤ c|t − s| |y − x|

⇒ c(f (t) − f (s)) ≤ c |t − s|

(4.13) ⇒ f ∈ C([0, τ ]; C 0,1 (RN , RN )) and kf (t) − f (s)kC 0,1 ≤ c |t − s|.

Moreover f ′ (t) = V (t) ◦ Tt and

|f ′ (t)(x) − f ′ (s)(x)| ≤ |V (t)(Tt (x)) − V (s)(Tt (x))| + |V (s)(Tt (x)) − V (s)(Ts (x))|
≤ kV (t) − V (s)kC + c(V (s)) kTt − Ts kC
≤ kV (t) − V (s)kC + c kf (t) − f (s)kC .
56 M.C. Delfour and Jean-Paul Zolésio

Finally

|f ′ (t)(y) − f ′ (t)(x)| ≤ |V (t)(Tt (y)) − V (t)(Tt (x))|

≤ c(V (t)) |Tt (y) − Ts (x)| ≤ c c(Tt ) |y − x|

and c(f ′ (t)) ≤ c for some new constant c independent of t. Therefore

f ∈ C 1 ([0, τ ]; C(RN , RN )) and c(f ′ (t)) ≤ c.

(Conditions (4.12) on g). Since conditions (T1) and (T2) are satisfied there
exists τ ′ > 0 such that conditions (T3) are satisfied by Theorem 4.1 (iii).
Moreover from conditions (4.11)

|g(t)(y) − g(t)(x)| ≤ |f (t)(Tt−1 (y)) − f (t)(Tt−1 (x))|

≤ c(f (t))|Tt−1 (y) − Tt−1 (x)|
≤ c(f (t)) (|g(t)(y) − g(t)(x)| + |y − x|)
⇒ (1 − c(f (t))) |g(t)(y) − g(t)(x)| ≤ c(f (t)) |y − x| ≤ ct |y − x|.

Choose a new τ ′′ = min{τ ′ , 1/(2c)}. Then for 0 ≤ t ≤ τ ′′ , c(g(t)) ≤ 2ct. Now

g(t) − g(s) = −f (t) ◦ [I + g(t)] + f (s) ◦ [I + g(s)]

kg(t) − g(s)kC ≤kf (t) ◦ [I + g(t)] − f (t) ◦ [I + g(s)]kC
+ kf (t) ◦ [I + g(s)] − f (s) ◦ [I + g(s)]kC
≤c(f (t))kg(t)] − g(s)kC + kf (t) − f (s)kC
≤ct kg(t) − g(s)kC + kf (t) − f (s)kC .

For t in [0, τ ′′ ], ct ≤ 1/2, and

kg(t) − g(s)kC ≤ 2kf (t) − f (s)kC

⇒ g ∈ C([0, τ ′′ ]; C(RN , RN )) and c(g(t)) ≤ 2ct.

Thus conditions (4.11) on f are satisfied for k = 0. For k = 1 we start from the
equation
Z t
DTt − DTs = DV (r) ◦ Tr DTr dr
s

and use the fact that DTt−1 = [DTt ]−1 ◦ Tt−1 in connection with the identity

Dg(t) = −Df (t) ◦ Tt−1 DTt−1 = −(Df (t)[DTt ]−1 ) ◦ Tt−1 .

(ii) From conditions (4.11) on f the transformation T satisfies conditions (T1).

To check condition (T2) we consider two cases: k ≥ 1 and k = 0. For k ≥ 1
 Velocity method and Courant metric topologies 57

the function t 7→ Df (t) = DTt − I : [0, τ ] → C k−1 (RN , RN )N is continuous.

Hence t 7→ det DTt : [0, τ ] → R is continuous and det DT0 = 1. So there
exists τ ′ > 0 such that Tt is invertible for all t in [0, τ ′ ] and (T2) is satisfied
in [0, τ ′ ]. In the case k = 0 consider for any Y the map h(X) = Y − f (t)(X).
For any X1 and X2 , |h(X2 ) − h(X1 )| ≤ c(f (t)) |X2 − X1 |. But by assumption
f ∈ C([0, τ ; C 0,1 (RN )) and c(f (0)) = 0 since f (0) = 0. Hence there exists τ ′ > 0
such that c(f (t)) ≤ 1/2 for all t in [0, τ ′ ] and h is a contraction. So for all Y in
RN there exists a unique X such that

Y − [Tt (X) − X] = h(X) = X ⇐⇒ Tt (X) = Y,

Tt is bijective, and (T2) is satisfied in [0, τ ′ ]. By Theorem 4.1 (iii) from (T1)
and (T2), there exists another τ ′ > 0 for which conditions (T3) on g and (V) on
V (t) = f ′ (t) ◦ Tt−1 are also satisfied. Moreover we have seen in the proof of part
(i) that conditions (4.12) on g follow from (T2) and (4.11). Using conditions
(4.11) and (4.12)

|V (t)(y) − V (t)(x)| ≤|f ′ (t)(Tt−1 (y)) − f ′ (t)(Tt−1 (x))|

≤c(f ′ (t)) |Tt−1 (y) − Tt−1 (x)|
≤c(f ′ (t)) [1 + c(g(t))] |y − x| ≤ c′ |y − x|

and c(V (t)) ≤ c′ . Also

|V (t)(x) − V (s)(x)| =|f ′ (t)(Tt−1 (x)) − f ′ (s)(Ts−1 (x))|

≤|f ′ (t)(Tt−1 (x)) − f ′ (t)(Ts−1 (x))|
+ |f ′ (t)(Ts−1 (x)) − f ′ (s)(Ts−1 (x))|
≤c(f ′ (t)) |Tt−1 (x) − Ts−1 (x)| + kf ′ (t) − f ′ (s)kC
≤c kg(t) − g(s)kC + kf ′ (t) − f ′ (s)kC .

Therefore since both g and f ′ are continuous

V ∈ C([0, τ ′ ]; C(RN , RN )) and c(V (t)) ≤ c

for some constant c independent of t. This completes the proof for k = 0. As

in part (i) for k = 1 we use the identity

DV (t) = Df ′ (t) ◦ Tt−1 DTt−1 = (Df ′ (t)[DTt ]−1 ) ◦ Tt−1

and proceed in the sane way. The general case is obtained by induction over k.
We now turn to the case of velocities in C0k (RN ).
Theorem 4.3. Let k ≥ 1 be an integer.
58 M.C. Delfour and Jean-Paul Zolésio

(i) Given τ > 0 and a velocity field V such that

(4.14) V ∈ C([0, τ ]; C0k (RN , RN )),

the map T given by (4.3)–(4.5) satisfies conditions (T1), (T2), and

(4.15) f ∈ C 1 ([0, τ ]; C0k (RN , RN ))

Moreover conditions (T3) is satisfied and there exists τ ′ > 0 such that

(4.16) g ∈ C([0, τ ′ ]; C0k (RN , RN )).

(ii) Given τ > 0 and T : [0, τ ] × RN → RN satisfying conditions (4.15) and

T (0, ·) = I, there exists τ ′ > 0 such that the velocity field V (t) = f ′ (t)◦Tt−1
satisfies conditions (V) and (4.14) on [0, τ ′ ].

Proof. It will be convenient to use the notation C0k for the space C0k (RN ). As
in the proof of Theorem 4.2 we only prove the theorem for k = 1. The general
case is obtained by induction on k, the various identities on f , g, f ′ and V , and
the techniques of Theorem 2.1 and Lemmas 2.1 and 2.2.
(i) By the embedding C01 (RN , RN ) ⊂ C 1 (RN , RN ) ⊂ C 0,1 (RN , RN ), it
follows from (4.14) that V ∈ C([0, τ ]; C 0,1 (RN , RN )) and condition (4.10) of
Theorem 4.2 are satisfied. Therefore conditions (4.11) and (4.12) of Theorem 4.2
are also satisfied in some interval [0, τ ′ ].
(Conditions (4.15) on f ). It remains to show that f (t) and f ′ (t) belong to the
subspace C0 (RN , RN ) of C(RN , RN ) and prove the appropriate properties for
Df (t) and Df ′ (t). Recall from the proof of the previous theorems that there
exists c > 0 such that
Z t Z t
|f (t)(x)| ≤ c |V (r)(x)| dr ≤ c |(V (r) − V (0))(x)| dr + c t |V (0)(x)|.
0 0

By assumption on V (0), for ε > 0 there exists a compact set K such that

∀x ∈ ∁K, |V (0)(x)| ≤ ε/(2c)

and there exists δ, 0 < δ < 1, such that

∀0 ≤ t ≤ δ, kV (r) − V (0)kC ≤ ε/(2c)

⇒ ∀0 ≤ t ≤ δ, ∀x ∈ ∁K, |f (t)(x)| ≤ ε ⇒ f (t) ∈ C0 .

Proceeding in this fashion from the interval [0, δ] to the next interval [δ, 2δ]
using the inequality
Z t
|(f (t) − f (s))(x)| ≤ c |(V (r) − V (δ))(x)| dr + c |t − δ| |V (δ)(x)|,
s
 Velocity method and Courant metric topologies 59

the uniform continuity of V

∀t, s, |t − s| < δ, kV (t) − V (s)kC ≤ ε/(2c)

and the fact that V (δ) ∈ C0 , that is, there exists a compact set K(δ) such that

∀x ∈ ∁K(δ), |V (δ)(x)| ≤ ε/(2c),

we get f (t) ∈ C0 , δ ≤ t ≤ 2δ, and hence f ∈ C([0, τ ]; C0 ). For f ′ (t) we make use
of the identity f ′ (t) = V (t) ◦ Tt . Again by assumption for any ε > 0 there exists
a compact set K(t) such that |V (t)(x)| ≤ ε on ∁K(t). Thus by choosing the
compact Kt′ = Tt−1 (K(t)), |f ′ (t)(x)| ≤ ε on ∁Kt′ and f ′ ∈ C([0, τ ]; C0 ). In order
to complete the proof, it remains to establish the same properties for Df (t)
and Df ′ (t). The matrix Df (t) is solution of the equations
d
Df (t) = DV (t) ◦ Tt DTt , Df (0) = 0
dt
(4.17) ⇒ Df ′ (t) = DV (t) ◦ Tt Df (t) + DV (t) ◦ Tt .

From the proof of Theorem 2.1 in [5] for each t the elements of the matrix
def
A(t) = DV (t) ◦ Tt = DV (t) ◦ [I + f (t)]

belong to C0 since DV (t) and f (t) do. By assumption V ∈ C([0, τ ]; C0k ) and V
and all its derivatives ∂ α V are uniformly continuous in [0, τ ] × RN . Therefore
for each ε > 0 there exists δ > 0 such that

∀ |t − s| < δ, ∀ |y ′ − x′ | < δ, |DV (t)(y ′ ) − DV (s)(x′ )| < ε.

Pick 0 < δ′ < δ such that

∀ |t − s| < δ′ , kTt − Ts kC = kf (t) − f (s)kC < δ

⇒ ∀x, ∀ |t − s| < δ′ , |Tt (x) − Ts (x)| < δ
⇒ |DV (t)(Tt (x)) − DV (s)(Ts (x))| < ε
⇒ kA(t) − A(s)kC < ε ⇒ A ∈ C([0, τ ]; (C0 )N ).

For each x, Df (t)(x) is the unique solution of the linear matrix equation (4.17).
To show that Df (t) ∈ (C0 )N we first show that Df (t)(x) is uniformly continuous
for x in RN . For any x and y
Z t
|Df (t)(y) − Df (t)(x)| ≤ |V (r, Tr (y)) − V (r, Tr (x))| dr
0
Z t
≤ c|Tr (y) − Tr (x)| dr
0
Z t
≤c |f (r)(y) − f (r)(x)| + |y − x| dr.
0
60 M.C. Delfour and Jean-Paul Zolésio

But f ∈ C([0, τ ]; C0 ) is uniformly continuous in (t, x): for each ε > 0 there
exists δ, 0 < δ < ε/(2cτ ) such that

∀ |t − s| < δ, ∀ |y − x| < δ, |f (t)(y) − f (s)(x)| < ε/(2cτ ).

Substituting in the previous inequality for each ε > 0, there exists δ > 0 such
that

∀ t, ∀ |y − x| < δ, |Df (t)(y) − Df (t)(x)| < ε.

Hence Df (t) is uniformly continuous in RN . Furthermore from the equation

(4.17) we have the following inequality
Z t
|Df (t)(x)| ≤ |DV (r)(Tr (x)| |Df (r)(x)| + |DV (r)(Tr (x))| dr
(4.18) 0 Z t
≤ c (|Df (r)(x)| + 1) dr
0

since V ∈ C([0, τ ]; C01 ). By Gronwall’s inequality

|Df (t)(x)| ≤ ct

for some other constant c independent of t. Thence Df (t) ∈ C(RN , RN )N .

Finally to show that Df (t) vanishes at infinity we start from the integral form
of (4.17)
Z t
Df (t)(x) = DV (r)(Tr (x)) DTr (x) dr
Z t 0
|Df (t)(x)| ≤c |DV (r)(Tr (x)) − DV (r)(x)| + |DV (r)(x)| dr
0 Z t
′
≤c |f (r)(x)| + |DV (r)(x)| dr.
0

By the same technique as before for f (t), it follows that the elements of Df (t)
belong to C0 since both f (s) and DV (r) do. Finally for the continuity with
respect to t
Z t
Df (t) − Df (s) = A(r)Df (r) + A(r) dr
s
Z t
kDf (t) − Df (s)kC ≤ kA(r)kC kDf (r) − Df (s)kC
s
+ kA(r)kC (1 + kDf (r)kC ) dr.

Again, by Gronwall’s inequality, there exists another constant c such that

kDf (t) − Df (s)kC ≤ c |t − s|.

 Velocity method and Courant metric topologies 61

Therefore Df ∈ C([0, τ ]; (C0 )N ) and f ∈ C([0, τ ]; C01 ). For Df ′ we repeat the

proof for f ′ using the identity

Df ′ (t) = DV (t) ◦ Tt Df (t) + DV (t) ◦ Tt

to get

Df ′ ∈ C([0, τ ]; (C0 )N ) ⇒ f ′ ∈ C([0, τ ]; C01 ).

(Conditions (4.16) on g). From the remark at the beginning of part (i) of the
proof, the conclusion of Theorem 4.2 are true for g and it remains to check the
other conditions on g and Dg using the identities

g(t) = −f (t) ◦ [I + g(t)], Dg(t) = −Df (t) ◦ [I + g(t)] (I + Dg(t)).

By the proof of Theorem 2.1 in [5], g(t) ∈ C0 since Df (t) and g(t) do. Therefore
g(t) ∈ C01 . The continuity follows by the same argument as for f ′ and hence
g ∈ C([0, τ ]; C01 ).
(ii) By assumption from conditions (4.15) conditions (T1) are satisfied. For
(T2) observe that for k ≥ 1 the function t 7→ Df (t) = DTt − I : [0, τ ] →
C k−1 (RN , RN )N is continuous. Hence t 7→ det DTt : [0, τ ] → R is continuous
and det DT0 = 1. So there exists τ ′ > 0 such that Tt is invertible for all t in
[0, τ ′ ] and (T2) is satisfied in [0, τ ′ ]. Furthermore we have seen in the proof of
part (i) that conditions (T1), (T2) and (4.15) imply conditions (T3) and (4.16)
on g in [0, τ ′ ], τ ′ > 0. Therefore the velocity field

V (t) = f ′ (t) ◦ Tt−1 = f ′ (t) ◦ [I + g(t)]

satisfies the conditions (V) specified by (4.2). From the proof of Theorem 2.1 in
[5] V (t) ∈ C0k since f ′ (t) and g(t) belong to C0k . By assumption f ∈ C 1 ([0, τ ]; C0k ).
Hence f ′ and all its derivatives ∂ α f ′ , |α| ≤ k, are uniformly continuous on
[0, τ ] × RN , that is, given ε > 0, there exists δ > 0 such that

∀ |t − s| < δ, ∀ |y ′ − x′ | < δ, |∂ α f ′ (t)(y ′ ) − ∂ α f ′ (s)(x′ )| < ε.

Similarly g ∈ C([0, τ ′ ]; C0k ) and there exists 0 < δ′ ≤ δ such that

∀ |t − s| < δ′ , ∀ |y − x| < δ′ , |∂ α g(t)(y) − ∂ α g(s)(x)| < δ.

Therefore for |t − s| < δ′

kTt−1 − Ts−1 kC = kg(t) − g(s)kC < δ

and since δ′ < δ

∀x, |f ′ (t)(Tt−1 (x)) − f ′ (t)(Ts−1 (x))| < ε

⇒ kV (t) − V (s)kC < ε ⇒ V ∈ C([0, τ ′ ]; C0 ).
62 M.C. Delfour and Jean-Paul Zolésio

We then proceed to the first derivative of V

DV (t) = Df ′ (t) ◦ Tt−1 DTt−1 = Df ′ (t) ◦ [I + g(t)] [I + Dg(t)]

and by uniform continuity of the right-hand side V ∈ C([0, τ ′ ]; C01 ). By

induction on k we finally get V ∈ C([0, τ ′ ]; C0k ).
The proof of the last theorem is based on the fact that the vector functions
involved are uniformly continuous. The fact that they vanish at infinity is not an
essential element of the proof. Therefore the theorem is valid with C k (RN , RN )
in place of C0k (RN , RN ).
Theorem 4.4. Let k ≥ 1 be an integer.
(i) Given τ > 0 and a velocity field V such that

(4.19) V ∈ C([0, τ ]; C k (RN , RN )),

the map T given by (4.3)–(4.5) satisfies conditions (T1), (T2), and

(4.20) f ∈ C 1 ([0, τ ]; C k (RN , RN )).

Moreover conditions (T3) are satisfied and there exists τ ′ > 0 such that

(4.21) g ∈ C([0, τ ′ ]; C k (RN , RN )).

(ii) Given τ > 0 and T : [0, τ ] × RN → RN satisfying conditions (4.20) and

T (0, ·) = I, there exists τ ′ > 0 such that the velocity field V (t) = f ′ (t)◦Tt−1
satisfies conditions (V) and (4.19) on [0, τ ′ ].

5 Continuity of shape functions

In this section we give a characterization of the continuity of a shape function6

(5.1) Ω 7→ J(Ω) : A ⊂ P(RN ) → B.

defined on a family A in P(RN ) with values in a Banach space B with respect

to the Courant metric in terms of its continuity along the flows generated
by a family of velocity fields using the equivalence Theorems 4.2, 4.3 and
4.4. Checking the continuity along flows is usually easier and more natural.
We specifically consider the continuity of shape functions with respect to the
Courant metric associated with the quotient spaces of transformations F0k /G(Ω)
of § 2 and F k (RN )/G(Ω) and F k,1 (RN )/G(Ω) of § 3 corresponding to the families
of velocity fields C0k (RN ), C k (RN ), and C k,1 (RN ).

6
To be well-defined on the quotient spaces the shape function J must be independent of
the choice of the representative in the equivalence class.
 Velocity method and Courant metric topologies 63

5.1 Courant metrics and flows of velocities

We start with the space C0k (RN ) used in [5].
Theorem 5.1. Let k ≥ 1 be an integer, B a Banach space, and Ω a non-empty
open subset of RN . Consider a shape function J : NΩ ([I]) → B defined in a
neighborhood NΩ ([I]) of [I] in F0k /G(Ω). Then J is continuous at Ω for the
Courant metric if and only if

(5.2) lim J(Tt (Ω)) = J(Ω)

tց0

for all families of velocity fields {V (t) : 0 ≤ t ≤ τ } satisfying the condition

(5.3) V ∈ C([0, τ ]; C0k (RN , RN )).

Proof. It is sufficient to prove the theorem for a real valued function J. The
Banach space case is readily obtained by considering the new real valued
function j(T ) = |J(T (Ω)) − J(Ω)|.
(i) If J is δ-continuous at Ω, then for all ε > 0 there exists δ > 0 such that

∀T, [T ] ∈ NΩ ([I]), δ([T ], [I]) < δ, |J(T (Ω)) − J(Ω)| < ε.

Condition (5.3) on V coincides with condition (4.14) of Theorem 4.3 which

implies conditions (4.15) and (4.16):

f ∈ C 1 ([0, τ ]; C0k (RN , RN )) and g ∈ C([0, τ ]; C0k (RN , RN ))

⇒ kTt − IkC k (RN ) → 0 and kTt−1 − IkC k (RN ) → 0 as t → 0.

But by definition of δ

δ([Tt ], [I]) ≤ kTt−1 − IkC k + kTt − IkC k → 0 as t → 0.

and we get the convergence (5.2) of the function J(Tt (Ω)) to J(Ω) as t goes to
zero for all V satisfying (5.3).
(ii) Conversely it is sufficient to prove that for any sequence {[Tn ]} such that
δ([Tn ], [I]) goes to zero there exists a subsequence such that

J(Tnk (Ω)) → J(I(Ω)) = J(Ω) as k → ∞.

Indeed let

ℓ = lim inf J(Tn (Ω)) and L = lim sup J(Tn (Ω)).

n→∞ n→∞

By definition of the liminf, there is a subsequence, still indexed by n, such that

ℓ = lim inf n→∞ J(Tn (Ω)). But since there exists a subsequence {Tnk } of {Tn }
such that J(Tnk (Ω)) → J(Ω), then necessarily ℓ = J(Ω). The same reasoning
64 M.C. Delfour and Jean-Paul Zolésio

applies to the limsup and hence the whole sequence J(Tn (Ω)) converges to J(Ω)
and we have the continuity of J at Ω.
We now prove that we can construct a velocity V associated with a
subsequence of {Tn } verifying conditions (4.14) of Theorem 4.3 and hence
conditions (5.3). By Corollary 2.1 to Theorem 2.2 and the same technique as
in the proof of Theorem 2.3 and Theorem 2.2 in [5] associate with a sequence
{Tn } such that δ([Tn ], [I]) → 0 a subsequence, still denoted {Tn }, such that

kfn kC k + kgn kC k = kTn−1 − IkC k + kTn − IkC k ≤ 2−2(n+2) .

For n ≥ 1 set tn = 2−n and observe that tn − tn+1 = −2−(n+1) . Define the
following C 1 -interpolation in (0, 1/2]: for t in [tn+1 , tn ]
 
def tn+1 − t def
Tt (X) = Tn (X) + p (Tn+1 (X) − Tn (X)), T0 (X) = X
tn+1 − tn

where p ∈ P 3 [0, 1] is the polynomial of order 3 on [0, 1] such that p(0) = 1 and
p(1) = 0 and p(1) (0) = 0 = p(1) (1).
(Conditions on f .) By definition for all t, 0 ≤ t ≤ 1/2, f (t) = Tt − I ∈ C0k (RN ).
Moreover for 0 < t ≤ 1/2
∂T ∂T
Ttn (X) = Tn (X),
Ttn+1 (X) = Tn+1 (X), (tn , X) = 0 = (tn+1 , X)
 ∂t ∂t
∂T Tn+1 (X) − Tn (X) (1) tn+1 − t ∂T
(t, X) = p ⇒ f ′ (t) = (t, ·) ∈ C0k (RN )
∂t |tn − tn+1 | tn+1 − tn ∂t

and f (·)(X) = T (·, X) − I ∈ C 1 ((0, 1/2]; RN ). By definition f (0) = 0. For each

0 < t ≤ 1/2 there exists n ≥ N such that tn+1 ≤ t ≤ tn and
 
tn+1 − t
kf (t) − f (0)kC k = kf (t)kC k = kfn + p (fn+1 − fn )kC k
tn+1 − tn
≤ 2kfn kC k + kfn+1 kC k ≤ 2 2−2(n+2) + 2−2(n+3) ≤ 2−(n+1) ≤ t.

Define at t = 0 f ′ (t) = 0. By the same technique there exists a constant c > 0,

and for each 0 < t ≤ 1/2 there exists n ≥ N such that tn+1 ≤ t ≤ tn and

kf ′ (t) − f ′ (0)kC k = kf ′ (t)kC k

∂T kTn+1 − Tn kC k kfn+1 − fn kC k
=
∂t (t, ·) k ≤ c |tn+1 − tn | = c
C 2−(n+1)
−2(n+2)
≤ c22 /2−(n+1) ≤ c 2−1 2−(n+1) ≤ c 2−(n+1) ≤ c t
⇒ kf ′ (tkC k ≤ ct.

So for each X the functions t 7→ f (t)(X) and t 7→ Tt (X) belong to

C 1 ([0, 1/2]; RN ). By uniform C k -continuity of the Tn ’s and the continuity
 Velocity method and Courant metric topologies 65

with respect to t for each X, it follows that f ∈ C 1 ([0, 1/2]; C0k (RN )) and the
condition (4.15) of Theorem 4.3 is satisfied. Hence the corresponding velocity
V satisfies conditions (4.14). Finally V satisfies conditions (5.3) and by (5.2)
for all ε > 0 there exists δ > 0 such that

∀0 ≤ t ≤ δ, |J(Tt (Ω)) − J(Ω)| < ε.

In particular there exists N > 0 such that for all n ≥ N , tn ≤ δ and

∀n ≥ N, |J(Tn (Ω)) − J(Ω)| = |J(Ttn (Ω)) − J(Ω)| < ε

and this proves the δ-continuity for the subsequence {Tn }.

The case of the Courant metric for the space C k (RN ) is a corollary to
Theorem 5.1.
Theorem 5.2. Let k ≥ 1 be an integer, B a Banach space, and Ω a non-empty
open subset of RN . Consider a shape function J : NΩ ([I]) → B defined in a
neighborhood NΩ ([I]) of [I] in F k (RN )/G(Ω). Then J is continuous at Ω for
the Courant metric if and only if

(5.4) lim J(Tt (Ω)) = J(Ω)

tց0

for all families of velocity fields {V (t) : 0 ≤ t ≤ τ } satisfying the condition

(5.5) V ∈ C([0, τ ]; C k (RN , RN )).

The proof of the theorem for the Courant metric topology associated with
the space C k,1 (RN ) is similar to the proof of the first theorem with obvious
changes.
Theorem 5.3. Let k ≥ 0 be an integer, Ω a non-empty open subset of RN , and
B a Banach space. Consider a shape function J : NΩ ([I]) → B defined in a
neighborhood NΩ ([I]) of [I] in F k,1 (RN )/G(Ω). Then J is continuous at Ω for
the Courant metric if and only if

(5.6) lim J(Tt (Ω)) = J(Ω)

tց0

for all families {V (t) : 0 ≤ t ≤ τ } of velocity fields in C k,1 (RN , RN ) satisfying

the conditions

(5.7) V ∈ C([0, τ ]; C k (RN , RN )) and ck (V (t)) ≤ c

for some constant c independent of t.

Proof. As in the proof of Theorem 5.1, it is sufficient to prove the theorem for
a real valued function J.
(i) If J is δ-continuous at Ω, then for all ε > 0 there exists δ > 0 such that

∀T, [T ] ∈ NΩ ([I]), δ([T ], [I]) < δ, |J(T (Ω)) − J(Ω)| < ε.

66 M.C. Delfour and Jean-Paul Zolésio

Under condition (5.7) from Theorem 4.2

f = T· − I ∈ C([0, τ ]; C k,1 (RN )) and kTt − IkC k,1 → 0 as t → 0

g(t) = Tt−1 − I ∈ C k,1 (RN ) and kTt−1 − IkC k,1 ≤ ct → 0 as t → 0.

But by definition of δ

δ([Tt ], [I]) ≤ kTt−1 − IkC k,1 + kTt − IkC k,1 → 0 as t → 0.

and we get the convergence (5.6) of the function J(Tt (Ω)) to J(Ω) as t goes to
zero for all V satisfying (5.7).
(ii) Conversely, as in the proof of Theorem 5.1, it is sufficient to prove that
given any sequence {[Tn ]} such that δ([Tn ], [I]) → 0 there exists a subsequence
such that

J(Tnk (Ω)) → J(I(Ω)) = J(Ω) as k → ∞.

By the same technique as in the proof of Theorem 2.3 and Theorem 2.2 in [5]
associate with a sequence {Tn } such that δ([Tn ], [I]) → 0 a subsequence, still
denoted {Tn }, such that

kfn kC k,1 + kgn kC k,1 = kTn−1 − IkC k,1 + kTn − IkC k,1 ≤ 2−2(n+2) .

For n ≥ 1 set tn = 2−n and observe that tn − tn+1 = −2−(n+1) . Define the
following C 1 -interpolation in (0, 1/2]: for t in [tn+1 , tn ]
 
def tn+1 − t def
Tt (X) = Tn (X) + p (Tn+1 (X) − Tn (X)), T0 (X) = X
tn+1 − tn

where p ∈ P 3 [0, 1] is the polynomial of order 3 on [0, 1] such that p(0) = 1 and
p(1) = 0 and p(1) (0) = 0 = p(1) (1).
(Conditions on f .) By definition for all t, 0 ≤ t ≤ 1/2, f (t) = Tt −I ∈ C k,1 (RN ).
Moreover for 0 < t ≤ 1/2
∂T ∂T
Ttn (X) = Tn (X),
Ttn+1 (X) = Tn+1 (X), (tn , X) = 0 = (tn+1 , X)
 ∂t ∂t
∂T Tn+1 (X) − Tn (X) (1) tn+1 − t ∂T
(t, X) = p ⇒ f ′ (t) = (t, ·) ∈ C k,1 (RN )
∂t |tn − tn+1 | tn+1 − tn ∂t

and f (·)(X) = T (·, X) − I ∈ C 1 ((0, 1/2]; RN ). By definition f (0) = 0. For each

0 < t ≤ 1/2 there exists n ≥ N such that tn+1 ≤ t ≤ tn and
 
tn+1 − t
kf (t) − f (0)kC k,1 = kf (t)kC k,1 = kfn + p (fn+1 − fn )kC k,1
tn+1 − tn
≤ 2kfn kC k,1 + kfn+1 kC k,1 ≤ 2 2−2(n+2) + 2−2(n+3) ≤ 2−(n+1) ≤ t.
 Velocity method and Courant metric topologies 67

Define at t = 0 f ′ (t) = 0. By the same technique there exists a constant c > 0,

and for each 0 < t ≤ 1/2 there exists n ≥ N such that tn+1 ≤ t ≤ tn and

kf ′ (t) − f ′ (0)kC k,1 = kf ′ (t)kC k,1

∂T kTn+1 − Tn kC k,1 kfn+1 − fn kC k,1
= (t, ·) ≤c =c
∂t k,1 |tn+1 − tn | 2−(n+1)
C
≤ c 2 2−2(n+2) /2−(n+1) ≤ c 2−1 2−(n+1) ≤ c 2−(n+1) ≤ c t
⇒ kf ′ (tkC k,1 ≤ ct and ck (f ′ (t)) ≤ ct.

and for each X the functions t 7→ f (t)(X) and t 7→ Tt (X) belong to

C 1 ([0, 1/2]; RN ). By uniform C k -continuity of the Tn ’s and the continuity with
respect to t for each X, it follows that f ∈ C 1 ([0, 1/2]; C k (RN )). Moreover it
can be shown that

ck (f ′ (t)) ≤ ct ⇒ ∀t, s ∈ [0, τ ], ck (f (t) − f (s)) ≤ c′ |t − s|

for some c′ > 0. The result is straightforward for k = 0 and then the general
case follows by induction on k. As a result f ∈ C([0, 1/2]; C k,1 (RN )) and the
condition (4.11) of Theorem 4.2 is satisfied. Hence the corresponding velocity
V satisfies conditions (4.10). Finally the velocity field V satisfies conditions
(5.7) and by (5.6) for all ε > 0 there exists δ > 0 such that

∀t ≤ δ, |J(Tt (Ω)) − J(Ω)| < ε.

In particular there exists N > 0 such that for all n ≥ N , tn ≤ δ and

∀n ≥ N, |J(Tn (Ω)) − J(Ω)| = |J(Ttn (Ω)) − J(Ω)| < ε

and we have the δ-continuity for the subsequence {Tn }.

Remark 5.1. The conclusions of Theorems 5.1, 5.2 and 5.3 are generic.
They also have their counterpart in the constrained case. For instance a
generalization of Theorem 5.1 in the constrained case has been announced by
[2] for an open subset D of RN of class C 2 . The difficulty lies in the second
part of the theorem which requires a special construction to make sure that the
family of transformations {Tt : 0 ≤ t ≤ τ } constructed from the sequence {Tn }
are homeomorphisms of D.

References

[1] C. Bardos and G. Chen Control and stabilization for the wave equation. III.
Domain with moving boundary, SIAM J. Control Optim. 19 (1981), no. 1, 123–138.
68 M.C. Delfour and Jean-Paul Zolésio

[2] S. Boisgerault and N. Gomez, personal communication.

[3] M.C. Delfour and J.-P. Zolésio, Velocity method and Lagrangian Formulation for
the computation of the shape Hessian, SIAM J. Control Optim. 29 (1991), no. 6,
1414–1442.
[4] M.C. Delfour and J.-P. Zolésio, Structure of shape derivatives for nonsmooth
domains, J. Funct. Anal. 104 (1992), no. 1, 1–33.
[5] A.M. Micheletti, Metrica per famiglie di domini limitati e proprietà generiche
degli autovalori, Ann. Scuola Norm. Sup. Pisa Ser. III, 26 (1972), 683–694.
[6] F. Murat and J. Simon, Sur le contrôle par un domaine géométrique, Rapport
76015, Université Pierre et Marie Curie, Paris, 1976.
[7] D.L. Russell, Introduction to Formation Theory of linear elastic materials, Lecture
notes, Ecole CEA-EDF-INRIA, Pole Universitaire Leonard de Vinci, Paris, 1997.
[8] J.-P.- Zolésio and C.Truchi, Shape stabilization of wave equation, in “Boundary
control and boundary variations” (Nice, 1986), 372–398, Lecture Notes in Comput.
Sci., 100, Springer, Berlin-New York, 1988.
[9] J. Sokolowski and J.-P. Zolésio, Introduction to shape optimization. Shape sensitiv-
ity analysis, Springer Series in Computational Mathematics, 16. Springer-Verlag,
Berlin, 1992.
[10] J.-P. Zolésio, Identification de domaines par déformation, thèse de doctorat d’état,
Université de Nice, France, 1979.
Nonlinear Periodic Oscillations In Suspension
Bridges

Zhonghai Ding1 , University of Nevada, Las Vegas, Nevada

Abstract
In this paper, we investigate nonlinear periodic oscillations in a
suspension bridge system which is described by the nonlinearly coupled
wave and beam equations. By applying the Mountain Pass Theorem
to a dual variational formulation of the problem, it is proved that the
suspension bridge system has at least two periodic oscillation solutions.

1 Introduction
The suspension bridge is a common type of civil engineering structures. It
is well known that a suspension bridge may display certain oscillations under
external aerodynamic forces. Under the action of a strong wind, in particular, a
narrow and very flexible suspension bridge can undergo dangerous oscillations.
The collapse of the Tacoma Narrows suspension bridge caused by a wind blowing
at a speed of 42 miles per hour in the State of Washington on November 7, 1940,
is one of the most striking examples. The Federal Works Agency Report [3] on
the collapse has created a widespread demand for a comprehensive investigation
of dynamic oscillation problems in suspension bridges in order to understand
the causes of such destructive oscillations, and to develop design techniques to
prevent their recurrence in future. A systematic study of the mathematical
theory of suspension bridges appears to be initiated by Bleich, McCullough,
Rosecrans and Vincent [5] in 1950. Since then, the extensive studies on
dynamics of suspension bridges were carried out by many researchers (see, for
example, [1], [15], [17]-[20] and the references therein), and more recently by
Lazer and McKenna [12], [13]. Based upon the observation of the fundamental
nonlinearity in suspension bridges that the stays connecting the supporting
cables and the roadbed resist expansion, but do not resist compression, new
models describing oscillations in suspension bridges have been developed by
Lazer and McKenna in [12], [13]. The new models are described by systems

1
This research is supported in part by NSF Grant DMS 96-22910.

69
70 Ding

of coupled nonlinear partial differential equations. Multiple large amplitude

periodic oscillations have been found theoretically and numerically in the single
one-dimensional Lazer-McKenna suspension bridge equation (see [6], [7], [9],
[10], [12]-[14] and references therein). However, there is very little discussion
on multiple large amplitude periodic oscillations in suspension bridge systems
of coupled nonlinear partial differential equations in the existing literature.
The objective of this paper is to study nonlinear large amplitude periodic
oscillations in a suspension bridge model described by two coupled nonlinear
partial differential equations.
Consider a simplified suspension bridge configuration: the roadbed of length
L is modeled by a horizontal vibrating beam with both ends being simply
supported; the supporting cable is modeled by a horizontal vibrating string
with both ends being fixed; and the vertical stays connecting the roadbed to
the supporting cable are modeled by one-sided springs which resist expansion
but do not resist compression. Let u(x, t) and w(x, t) denote the downward
deflections of the cable and the roadbed, respectively. The following suspension
bridge model has been proposed by Lazer-McKenna [12]:

(1.1)


 mc utt − Quxx − K(w − u)+ = mc g + f1 (x, t), 0 < x < L, t > 0,

mb wtt + EIwxxxx + K(w − u)+ = mb g + f2 (x, t), 0 < x < L, t > 0,

 u(0, t) = u(L, t) = 0,

w(0, t) = w(L, t) = 0, wxx (0, t) = wxx (L, t) = 0,
where (w − u)+ = max{w − u, 0}; mc and mb are the mass densities of the cable
and the roadbed, respectively; Q is the coefficient of cable tensile strength; EI
is the roadbed flexural rigidity; K is the Hooke’s constant of the stays; f1 and
f2 represent the external aerodynamic forces. The total energy E(t), including
all kinetic and potential energies, is given by

Z L
1
E(t) = [mc u2t + Qu2x ] + [mb wt2 + EIwxx
2
] + K((w − u)+ )2 dx.
2 0

In the absence of nonconservative external forces, i.e., f1 (x, t) = 0 and

f2 (x, t) = 0, system (1.1) is conservative, i.e., E ′ (t) = 0. Except a recent paper
[2] by Ahmed and Harbi who investigated the asymptotic stability of (1.1) with
nonlinear damping, the model (1.1) has not yet received in-depth study in the
literature. In this paper, we study nonlinear periodic oscillations in (1.1). We
are interested in periodic oscillation being symmetric about x = L/2,

u(x, t + T ) = u(x, t), w(x, t + T ) = w(x, t), 0 ≤ x ≤ L, t > 0

(1.2)
u(x, t) = u(L − x, t), w(x, t) = w(L − x, t), 0 ≤ x ≤ L/2, t > 0,
 Nonlinear Oscillations In Suspension Bridges 71

where T is the period of periodic oscillations. By rescaling and translating x

and t, system (1.1) with (1.2) can be written in an equivalent form:

(1.3)


 mc utt − Quxx − K(w − u)+ = mc g + f1 (x, t), −π/2 < x < π/2, t > 0,

 mb wtt + EIwxxxx + K(w − u)+ = mb g + f2 (x, t), −π/2 < x < π/2, t > 0,



u(−π/2, t) = u(π/2, t) = 0, t > 0,

 w(−π/2, t) = w(π/2, t) = 0, wxx (−π/2, t) = wxx (π/2, t) = 0, t > 0,



 u(−x, t) = u(x, t), w(−x, t) = w(x, t), 0 ≤ x ≤ π/2, t > 0,

u(x, t + π) = u(x, t), w(x, t + π) = w(x, t), −π/2 ≤ x ≤ π/2, t > 0.

By applying the Mountain Pass Theorem to a dual variational formulation of

the problem, we prove that, under some periodic external forces, the above
problem has at least two periodic solutions. One is a near-equilibrium periodic
oscillation, and the other is a nonlinear periodic oscillation.
The organization of this paper is as follows. In Section 2, we study the
near-equilibrium periodic oscillation. In Section 3, we formulate an equivalent
dual problem. In Section 4, we prove that the dual problem has at least two
periodic solutions by applying the Mountain Pass Theorem.

2 Near-equilibrium periodic oscillation

To investigate the suspension bridge system (1.3), we assume throughout
this paper that

(2.1) Q ≤ mc , EI ≤ mb .

These assumptions hold naturally for suspension bridges in civil engineering

applications.
The equation for equilibrium oscillation (we , ue ) of (1.3) is



 −Quxx − K(w − u)+ = mc g, −π/2 < x < π/2,


 EIwxxxx + K(w − u)+ = mb g, −π/2 < x < π/2,
(2.2) u(−π/2) = u(π/2) = 0,



 w(−π/2) = w(π/2) = 0, wxx (−π/2) = wxx (π/2) = 0,

u(−x) = u(x), w(−x) = w(x), 0 ≤ x ≤ π/2.

The following three propositions follow from a direct and tedious calculation.
Let
 
(mb + mc )g  π 2 2 EI(mb + mc )g
h(x) = −x − .
2Q 2 Q2
72 Ding

4Q2
Proposition 2.1. If K > , and if
EI
π π
mc ω1 ω2 ω1 tanh ω2 − ω2 tanh ω1
< 2 · 2 2,
mb ω1 + ω22 ω1 tanh ω1 π − ω2 tanh ω2 π
2 2
where
" r # " r #
1 4Q 2K 1 4Q 2K
ω12 = K + K2 − > 0, ω22 = K − K2 − > 0,
2Q EI 2Q EI

then the equilibrium solution of (2.2) is given by


 EI


 ue (x) =
 A1 ω12 cosh ω1 x + A2 ω22 cosh ω2 x + h(x),
Q

 mg

 we (x) = A1 cosh ω1 x + A2 cosh ω2 x + b + h(x),
K

where A1 and A2 are determined from the boundary conditions in (2.2).

Furthermore, w(x) − u(x) > 0 for |x| < π/2.
4Q2
Proposition 2.2. If K = , and if
EI
mc 1 sinh ω0 π − ω0 π
< ·
mb 2 sinh ω0 π + ω0 π
s
K
where ω0 = , then the equilibrium solution of (2.2) is given by
2Q
  

 EI
 ue (x) = 2B1 + 2B2 Q ω0 cosh ω0 x + 2B2 x sinh ω0 x + h(x),




 we (x) = B1 cosh ω0 x + B2 x sinh ω0 x + mb g + h(x),

K

where B1 and B2 are determined from the boundary conditions in (2.2).

Furthermore, w(x) − u(x) > 0 for |x| < π/2.
4Q2
Proposition 2.3. If 0 < K < , and if
EI

mc Qπ 2
≤ ,
mb 8EI
 Nonlinear Oscillations In Suspension Bridges 73

then the equilibrium solution of (2.2) is given by

  

 ue (x) = C1 (ω32 − ω42 ) − 2C2 ω3 ω4 sinh ω3 x sin ω4 x



  
+ 2C1 ω3 ω4 + C2 (ω32 − ω42 ) cosh ω3 x cos ω4 x + h(x),




 we (x) = C1 sinh ω3 x sin ω4 x + C2 cosh ω3 x cos ω4 x + mb g + h(x),

K

where
r s ! r s !
1 K KEI 1 K KEI
ω32 = 1+ , ω42 = 1− ,
2 EI 4Q2 2 EI 4Q2

C1 and C2 are determined from the boundary conditions in (2.2). Furthermore,

w(x) − u(x) > 0 for |x| < π/2.
For the purpose of our investigation, let

(2.3) f1 (x, t) = 0, f2 (x, t) = ε sin 2t cos x,

where ε is a small parameter to be specified later. Let

λ10 µ10
λ10 = Q − 4mc < 0, µ10 = EI − 4mb < 0, σ10 = .
λ10 + µ10

The near-equilibrium solution of system (1.3) with (2.3) can be obtained also
from a direct and careful calculation.
Proposition 2.4. Assume K > −σ10 and (ue , we ) being the equilibrium
solution of (2.1) given in Propositions 2.1-2.3. If |ε| < ε0 , where ε0 > 0 is
a constant determined by K, mc , mb , Q and EI, then the suspension bridge
system (1.3) with the external forces (2.3) admits a near-equilibrium solution
(u0 , w0 ) given by



 εK

 u0 (x, t) = ue (x) + cos x sin 2t,
 (K + σ10 )(λ10 + µ10 )
(2.4)

 ε(K + λ10 )


 w0 (x, t) = we (x) + cos x sin 2t.
(K + σ10 )(λ10 + µ10 )

Furthermore, w0 (x, t) − u0 (x, t) > 0 for |x| < π/2 and t > 0.
74 Ding

3 A duality formulation
The objective of the rest of this paper is to show that, in addition to the
near-equilibrium solution given in (2.4), the suspension bridge system (1.3) with
(2.3) has at least one nonlinear periodic solution. To prove the existence of such
a solution, we derive first an equivalent duality formulation of system (1.3) with
(2.3).
Define the wave operator L1 by


 L1 u = mc utt − Quxx ,
u(−π/2, t) = u(π/2, t) = 0,

u(x, t) = u(−x, t), u(x, t + π) = u(x, t).
Define the beam operator L2 by



 L2 w = mb wtt + EIwxxxx ,

w(−π/2, t) = w(π/2, t) = 0,
 w (−π/2, t) = wxx (π/2, t) = 0,
 xx

w(x, t) = w(−x, t), w(x, t + π) = w(x, t).
Denote by {λmn } the eigenvalues of L1 and by {µmn } the eigenvalues of L2 .
Then it follows from a direct calculation that
λmn = Q(2n + 1)2 − 4mc m2 , m, n = 0, 1, 2, · · · ,
(3.1)
µmn = EI(2n + 1)4 − 4mb m2 , m, n = 0, 1, 2, · · · .
The eigenfunctions of L1 corresponding to eigenvalue λmn are the same as that
of L2 corresponding to eigenvalue µmn , which are given by

ϕ0n (x, t) = cos(2n + 1)x, n ≥ 0,

ϕmn (x, t) = cos(2n + 1)x cos 2mt, m ≥ 1, n ≥ 0,
ψmn (x, t) = cos(2n + 1)x sin 2mt, m ≥ 1, n ≥ 0.
Assume throughout the rest of this paper that the material parameters mc , mb ,
Q and EI are chosen such that
(3.2)
 r r

 Q EI
 both and are rational numbers;
mc mb
 λ = Q(2n + 1)2 − 4mc m2 6= 0, µmn = EI(2n + 1)4 − 4mb m2 6= 0,
 mn

λmn + µmn 6= 0, for m ≥ 1, n ≥ 1.
Let Ω = [−π/2, π/2] × [−π/2, π/2], and H be a Hilbert space defined by

H = {u ∈ L2 (Ω) | u(−x, t) = u(x, t)}.

 Nonlinear Oscillations In Suspension Bridges 75

It is easy to check that the set of eigenfunctions {ϕmn , ψmn } is an orthogonal

basis of H. Under the assumption (3.2), L1 , L2 and L1 + L2 have p compact
inverses from H to H. Noting that the assumption of both
p Q/mc and
EI/mb being rational is necessary due to the well-known fact that certain
number theoretical difficulties may be encountered.
Under the above notations, system (1.3) with (2.3) can be written as


L1 u − K(w − u)+ = mc g,
(3.3)
L2 w + K(w − u)+ = mb g + ε sin 2t cos x.

Let ū = u − u0 and w̄ = w − w0 where (u0 , w0 ) is the near-equilibrium solution

given by (2.4), then (3.3) becomes


L1 ū − K [((w̄ − ū) + (w0 − u0 ))+ − (w0 − u0 )] = 0,
(3.4)
L2 w̄ + K [((w̄ − ū) + (w0 − u0 ))+ − (w0 − u0 )] = 0.

Since system (1.3) with (2.3) is equivalent to system (3.4), the problem of finding
a nonlinear periodic solution in system (1.3) with (2.3) becomes the problem of
finding a nontrivial periodic solution of system (3.4).
From (3.4), one has

L1 ū + L2 w̄ = 0.

By applying L−1 −1
1 L2 to both sides of this equation, we have

L−1 −1
2 ū + L1 w̄ = 0.

Let w̃ = L−1 −1
1 w̄ and ũ = L2 ū, then w̃ + ũ = 0, ū = L2 ũ and w̄ = L1 w̃.
Substituting them into the second equation of (3.4), we obtain
 
L2 L1 w̃ + K ((L1 + L2 )w̃ + (w0 − u0 ))+ − (w0 − u0 ) = 0.

Let ŵ = (L1 + L2 )w̃ and g0 = w0 − u0 , then the above equation can be written
as
 
(3.5) L2 L1 (L1 + L2 )−1 ŵ + K (ŵ + g0 )+ − g0 = 0.

Let Aŵ = L2 L1 (L1 + L2 )−1 ŵ. The eigenvalues of A are given by

λmn µmn
σmn = ,
λmn + µmn
where the corresponding eigenfunctions are {ϕmn , ψmn }. Under assumption
(2.1), it is easy to check that
76 Ding

σ20 < σ10 < 0 < σ00 .

Assume throughout this paper that

the only eigenvalue of A in the interval (σ20 , σ00 ) is σ10 .

Let β be a given constant satisfying
(3.6) −σ10 < K < β < −σ20 .
Let

Aβ ŵ = Aŵ + β ŵ,
Fβ (ŵ) = β ŵ − K [(ŵ + g0 )+ − g0 ] .
Under the assumption (3.6), Aβ has a compact inverse from H to H. Equation
(3.5) can be written as

(3.7) −Aβ ŵ + Fβ (ŵ) = 0.

By (3.6), Fβ : ℜ −→ ℜ is a monotone increasing function. Thus Fβ has a
monotone increasing inverse given by


 1

 ŵ, if ŵ > −(β − K)g0
 β−K
Fβ−1 (ŵ) =

 1

 (ŵ − Kg0 ), if ŵ ≤ −(β − K)g0
β

1 1
= [ŵ + (β − K)g0 ]+ − [ŵ + (β − K)g0 ]− − g0 ,
β−K β
where u− = max{−u, 0}. Let v = Aβ ŵ, then equation (3.7) can be written as
(3.8) −A−1 −1
β v + Fβ (v) = 0.

Note that v = 0 is a trivial solution of (3.8).

Proposition 3.1. Let condition (3.6) be satisfied. If equation (3.8) has a
nontrivial solution v in H, then the suspension bridge system (3.4) admits a
nontrivial periodic solution (ū, w̄) such that (ū, w̄) ∈ H 2 (Ω) × H 3 (Ω).
The proof of this proposition is based on analyzing
ū = −L2 (L1 + L2 )−1 A−1
β v, w̄ = L1 (L1 + L2 )−1 A−1
β v,

and bootstrapping regularities.

 Nonlinear Oscillations In Suspension Bridges 77

4 Nonlinear periodic oscillation

In this section, we prove that (3.8) admits a nontrivial solution in H by
applying the Mountain Pass Theorem. Define a functional I(v) : H −→ ℜ by
Z  
1 −1
I(v) = − Aβ v · v + F(v) dxdt,
Ω 2

where

F(v)
1
2 1
2 β−K 2
= [v + (β − K)g0 ]+ + [v + (β − K)g0 ]− − g0 v − g0 .
2(β − K) 2β 2

Lemma 4.1. I(v) is continuous Frechet differentiable with

Z h i
I ′ (v)ϕ = −A−1
β v + Fβ
−1
(v) ϕdxdt,
Ω

for any v, ϕ ∈ H. Consequently, the solutions of (3.8) in H correspond to

critical points of I(v) in H.
The Mountain Pass Theorem due to Ambrosetti and Rabinowitz [4] has been
used to prove the existence of critical points of functionals satisfying a condition
called the Palais-Smale (PS) condition, which occurs repeatedly in the critical
point theory. We say that a functional J satisfies the (PS) condition if any
sequence {vn } ⊂ H for which J(vn ) is bounded and J ′ (vn ) → 0 possesses a
convergent subsequence.
Mountain Pass Theorem. Let E be a real Banach space. J ∈ C 1 (E, ℜ)
satisfies the (PS) condition. Suppose J(0) = 0 and
(a). there are constants ρ, α > 0 such that J|∂Bρ ≥ α,
(b). there is an e ∈ E\B ρ such that J(e) ≤ 0.
Then J possesses a critical value c ≥ α. Moreover c can be characterized as

c = inf max J(u),

g∈Γ u∈g([0,1])

where Γ = {g ∈ C([0, 1], E) | g(0) = 0, g(1) = e}.

In the next several lemmas, we show that I(v) satisfies all conditions in the
Mountain Pass Theorem.
K − σ10
Lemma 4.2. If (3.6) is satisfied, β < , and |ε| < ε0 , where ε0 is
2
defined in Proposition 2.4, then v = 0 is a strict local minimum of I(v) in H.
78 Ding

Proof. It is easy to verify that I(0) = 0 and I ′ (0) = 0. We follow an

approach used in [8] to prove v = 0 is a strict local minimum of I(v). For s > 0
and ϕ ∈ H with kϕkL2 = 1, we have

I ′ (sϕ)ϕ
Z h
1
= − sA−1
β ϕ·ϕ+ (sϕ + (β − K)g0 )+ ϕ
Ω β − K
1 i
− (sϕ + (β − K)g0 )− ϕ − g0 ϕ dxdt
β
Z Z
s
= −s A−1β ϕ · ϕdxdt + ϕ2 dxdt
Ω β − K Ω
 Z
1 1
+ − (sϕ + (β − K)g0 )− ϕdxdt,
β −K β Ω

where we have used the relation u = u+ − u− in the last equality. Let

  
 −1 1
H1 = span ϕmn , ψmn  |β + σmn | > ,
2β
which is finite dimensional. Let H = H1 ⊕ H2 where H2 is the orthogonal
complement of H1 in H. In fact,
  
 −1 1

H2 = the closure of span ϕmn , ψmn  |β + σmn | ≤ .
2β
For any ψ ∈ H2 , we have
Z  Z
 
 A−1 ψ · ψdxdt ≤ 1 ψ 2 dxdt.
 β  2β
Ω Ω

Thus, by letting ϕ = ϕ1 + ϕ2 where ϕ1 ∈ H1 and ϕ2 ∈ H2 , we have

I ′ (sϕ)ϕ
Z h i Z
s  
= −s A−1
β ϕ1 · ϕ1 + A−1
β ϕ2 · ϕ2 dxdt + ϕ21 + ϕ22 dxdt
Ω β−K Ω
 Z
1 1
+ − (s(ϕ1 + ϕ2 ) + (β − K)g0 )− (ϕ1 + ϕ2 )dxdt
β−K β Ω
 Z  Z
1 1 2 1 1
≥ s − ϕ1 dxdt + s − ϕ22 dxdt
β−K |β + σ20 | Ω β−K 2β Ω
 Z
1 1
− − (s(ϕ1 + ϕ2 ) + (β − K)g0 )− (|ϕ1 | + |ϕ2 |)dxdt.
β−K β Ω
 Nonlinear Oscillations In Suspension Bridges 79

Under the assumption (3.6) and |ε| < ε0 , we know from Proposition 2.4 that
g0 (x, t) = w0 (x, t) − u0 (x, t) > 0 for |x| < π/2, and gx (−π/2, t) > 0 and
gx (π/2, t) < 0. Since kϕ1 kL2 ≤ kϕkL2 = 1, there is a small s0 > 0, which is
dependent of β, K and g0 only, such that for any 0 < s < s0 ,

1 π π
s|ϕ1 (x, t)| ≤ (β − K)g0 (x, t), − ≤ x ≤ .
2 2 2

Let Ωs = {(x, t) ∈ Ω | sϕ2 ≤ − 12 (β − K)g0 }. Thus

I ′ (sϕ)ϕ
 Z  Z
1 1 2 1 1
≥ s − ϕ1 dxdt + s − ϕ22 dxdt
β−K |β + σ20 | Ω β − K 2β Ω
 Z
1 1 1
− − (sϕ2 + (β − K)g0 )− (|ϕ1 | + |ϕ2 |)dxdt
β−K β Ω 2
 Z  Z
1 1 1 1
≥ s − ϕ21 dxdt + s − ϕ22 dxdt
β−K |β + σ20 | Ω β − K 2β Ω
 Z
1 1
−s − ϕ22 dxdt
β−K β
  Z Ωs
1 1 1
− − (tϕ2 + (β − K)g0 )− |ϕ1 |dxdt
β−K β Ωs 2
Z
1
≥ C1 s − C2 (sϕ2 + (β − K)g0 )− |ϕ1 |dxdt,
Ωs 2

where

 
1 1 1 1 1
C1 = min − , > 0, C2 = − > 0.
β−K |β + σ20 | 2β β−K β
1
For any δ T> 0, define Ωδ = {(x, t) ∈ Ω | 2 (β − K)g0 ≤ δ}. Hence for any
(x, t) ∈ Ωt (Ω \ Ωδ ), −sϕ2 ≤ −δ. Thus,
Z Z \
δ2
1= (ϕ21 + ϕ22 )dxdt ≥ T(Ω\Ω ) ϕ22 dxdt ≥ mes(Ω s (Ω \ Ωδ )).
Ω Ωs δ
s2

Then

\ s2
mes(Ωs (Ω \ Ωδ )) ≤ .
δ2
80 Ding

Since kϕ1 kL2 ≤ 1, it is easy to check that kϕ1 kL∞ ≤ C3 where C3 is a

positive constant independent of ϕ1 . From Propositions 2.1-2.3, we know that
g0 (x, t) = w0 (x, t) − u0 (x, t) = 0 when x = ±π/2, g0 (x, t) > 0 for |x| < π/2,
and g0 (x, t) is continuous on Ω. Then for any δ1 > 0, one can choose δ > 0
sufficiently small such that kϕ1 kL∞ (Ωs T Ωδ ) ≤ δ1 . Thus
I ′ (sϕ)ϕ

Z Z !
1
≥ C1 s − C2 T + T (sϕ2 + (β − K)g0 )− |ϕ1 |dxdt
Ωs Ω
δ Ωs (Ω\Ω )
δ
2
Z Z !
1
≥ C1 s − C2 δ1 T +C2 C3 T (sϕ2 + (β − K)g0 )− dxdt
Ωs Ωδ Ωs (Ω\Ωδ ) 2
q \
≥ C1 s − C2 δ1 skϕ2 kL2 (Ωs T Ωδ ) mes(Ωs Ωδ )
q \
−C2 C3 skϕ2 kL2 (Ωs T(Ω\Ωδ ) mes(Ωs (Ω \ Ωδ )

C 2 C3 s 2
≥ C1 s − C2 δ1 sπ − .
δ
By choosing sufficiently small δ > 0, one has C1 − C2 δ1 π > C1 /2. Then fix δ,
and choose s1 > 0 small enough such that
C 1 C 2 C3 s C1
− ≥ , for 0 ≤ s ≤ s1 .
2 δ 4
Thus for 0 < s ≤ min{s0 , s1 }, we have
Z s Z s
C1 C1 2
I(sϕ) = I ′ (τ ϕ)ϕdτ ≥ τ dτ = s .
0 0 4 8
Thus v = 0 is a strictly local minimum of I(v) in H.
Lemma 4.3. If −σ10 < K < −σ20 , the following equation
(4.1) Av + Kv + = 0
admits only the trivial solution v = 0 in H.
Proof. Under the assumption −σ10 < K < −σ20 and by applying an abstract
symmetry theorem due to Lazer and McKenna [11], all solutions of (4.1) can be
expressed as v(x, t) = g(t) cos x. It is straightforward to check that if v ∈ H is a
solution of (4.1), then u = −L2 (L1 + L2 )−1 v ∈ H and w = L1 (L1 + L2 )−1 v ∈ H
is a solution to the following system

L1 u − K(w − u)+ = 0,
L2 w + K(w − u)+ = 0.
 Nonlinear Oscillations In Suspension Bridges 81

Since v = g(t) cos x, one has u = g1 (t) cos x and w = g2 (t) cos x. Thus (g1 , g2 )
satisfies

 mc g1′′ + Qg1 − K(g2 − g1 )+ = 0,
(4.2) m g′′ + EIg2 + K(g2 − g1 )+ = 0,
 c 2
g1 (t + π) = g1 (t), g2 (t + π) = g2 (t).

A simple phase plane analysis of (4.2) shows that it admits only the trivial
solution g1 (t) = g2 (t) = 0.
Lemma 4.4. If (3.6) is satisfied, then I(v) satisfies the (PS) conditions.
Proof. Assuming {vn } ⊂ H such that I(vn ) is bounded and I ′ (vn ) −→ 0
strongly in L2 (Ω), we need to show that {vn } has a convergent subsequence.
We claim that {vn } is bounded in L2 (Ω). Assume the contrary, kvn kL2 −→
vn
∞ as n −→ ∞. Let wn = . Since I ′ (vn ) −→ 0, i.e.,
kvn kL2

−A−1 −1
β vn + Fβ (vn ) −→ 0,

we have
 +  −
1 β−K 1 β−K 1
−A−1
β wn + wn + g0 − wn + g0 − g0 −→ 0.
β−K kvn kL2 β kvn kL2 kvn kL2
Since kwn kL2 = 1, there exists a subsequence of {wn }, denoted by itself, such
that wn −→ w weakly in L2 (Ω). Since A−1 β is compact from H to H, we have

lim A−1 −1
β wn = Aβ w in L2 (Ω).
n→∞

Then
 +  −
1 β−K 1 β−K
wn + g0 − wn + g0 −→ A−1
β w.
β−K kvn kL2 β kvn kL2
By applying Proposition B.1 in [16], we have
 
β−K
lim wn + = (β − K)[A−1 + −1 −
β w] − β[Aβ w] strongly in L2 (Ω).
n→∞ kvn kL2

Since the weak limit of {wn } is unique, we have lim wn = w strongly in L2 (Ω)
n→∞
and

w = (β − K)[A−1 + −1 −
β w] − β[Aβ w] .

Let u = A−1
β w, then u statisfies

Au + Ku+ = 0.
82 Ding

By Lemma 4.3, we have u = 0. Hence w = 0. Thus lim wn = 0 strongly in

n→∞
L2 (Ω), which contradicts to kwn kL2 = 1.
Since {vn } is bounded, there exists a subsequence of {vn }, denoted by itself,
such that lim vn = v weakly in L2 (Ω). By repeating the same argument as
n→∞
above, we have lim vn = v strongly in L2 (Ω).
n→∞

Lemma 4.5. Assume (3.6) is satisfied, and |ε| < ε0 , where ε0 is defined in
Proposition 2.4. If β satisfies further
2 1 1
(4.3) − + + < 0,
β + σ10 β − K β

then lim I(sϕ10 ) = −∞.

s→∞
Proof. It is easy to check that
Z Z Z
+ 2 − 2 1 π2
(ϕ10 ) dxdt = (ϕ10 ) dxdt = (ϕ10 )2 dxdt = .
Ω Ω 2 Ω 8
Thus for any η > 0 and s being sufficiently large, we have

I(sϕ10 )
 "
Z  + #2
s2 1 1 β − K
= − ϕ2 + ϕ10 + g0
2 Ω  β + σ10 10 β − K s
" 
 − #2  Z Z
1 β−K β−K
ϕ10 + g0 dxdt − s ϕ21 g0 dxdt − g02 dxdt
β s  Ω 2 Ω

Z  i
s2 1 2 1 h +
2 i 1h

− 2
≤ − ϕ + ϕ10 + η + ϕ10 + η dxdt
2 Ω β + σ10 10 β − K β
Z Z
β−K
−s ϕ21 g0 dxdt − g02 dxdt
Ω 2 Ω
   
s2 π 2 2 1 1 1 1
= − + + 8η +
16 β + σ10 β − K β β−K β
Z Z
β−K
−s ϕ21 g0 dxdt − g02 dxdt.
Ω 2 Ω

Since η > 0 is arbitrary and (4.3) holds, we have

lim I(sϕ10 ) = −∞.

s→∞
 Nonlinear Oscillations In Suspension Bridges 83

By applying the Mountain Pass Theorem, we obtain from Lemmas 4.1-4.5

the following theorem.
K − σ10
Theorem 4.1. If (3.6) and (4.3) are satisfied, β < , and |ε| < ε0 ,
2
where ε0 is defined in Proposition 2.4, then I(v) admits at least one nontrivial
critical point v ∈ H.
To apply the Mountain Pass Theorem to I(v), we only need that there
exists an e ∈ H and e 6= 0 such that I(e) ≤ 0. Thus Lemma 4.5 provides more
than what we need. The assumption (4.3) is not necessary needed if one can
find some s0 > 0 such that I(s0 ϕ10 ) ≤ 0.
Since β is introduced to formulate the dual problem, one can derive easily
from (3.6) and (4.3) that the only condition on K needed is

−σ10 < K < min{−σ20 , −2σ10 }.

Hence, as a consequence of Theorem 4.1 and Lemma 3.1, we obtain the following
main result of this paper.
Theorem 4.2. If −σ10 < K < min{−σ20 , −2σ10 }, and |ε| < ε0 , where ε0
is defined in Proposition 2.4, then the suspension bridge system (1.3) with (2.3)
admits, in addition to an explicit near-equilibrium periodic oscillation
T given in
Proposition 2.4, at least one nonlinear periodic oscillation in H 2 (Ω) H 3 (Ω).

5 Conclusion
In this paper, we have studied nonlinear periodic oscillations in a suspension
bridge system governed by the nonlinearly coupled wave and beam equations.
It is shown that the suspension bridge system has at least two periodic
oscillations: one is an explicit near-equilibrium oscillation, and the other is
a nonlinear periodic oscillation. More theoretical and numerical results on
multiple nonlinear periodic oscillations in suspension bridges are being reported
in several other papers.

References

[1] A. M. Abdel-Ghaffer, Suspension bridge vibration: continuum formulation, J.

Engrg. Mech., 108(1982), pp. 1215-1232.
[2] N. U. Ahmed and H. Harbi, Mathematical analysis of dynamic models of
suspension bridges, SIAM J. Appl. Math., 58(1998), pp. 853-874.
[3] O. H. Amann, T. Von Karman and G. B. Woodruff, The failure of the Tacoma
Narrows Bridge, Federal Works Agency, Washington D. C., 1941.
[4] A. Ambrosetti and P. H. Rabinowitz, Dual variational methods in critical point
theory and applications, J. Funct. Anal., 14(1973), pp. 349-381.
84 Ding

[5] F. Bleich, C. B. McCullough, R. Rosecrans and G. S. Vincent, The mathematical

theory of suspension bridges, Bureau of Public Roads, U. S. Department of
Commence, Washington D. C., 1950.
[6] Q. H. Choi, T. Jung and P. J. McKenna, The study of a nonlinear suspension
bridge equation by a variational reduction method, Appl. Analysis, 50(1993), pp.
73-92.
[7] Y. S. Choi, K. C. Jen and P. J. McKenna, The structures of the solution set for
periodic oscillations in a suspension bridge model, IMA J. Appl. Math., 47(1991),
pp. 283-306.
[8] Y. S. Choi, P. J. McKenna and M Romano, A mountain pass method for the
numerical solution of semilinear wave equations, Numer. Math., 64(1993), pp.
487-509.
[9] J. Glover, A. C. Lazer and P. J. McKenna, Existence and stability of large scale
nonlinear oscillations in suspension bridges, Z. Angew. Math. Phys., 40(1989),
pp. 172-200.
[10] L. D. Humphreys, Numerical mountain pass solutions of a suspension bridge
equation, Nonlinear Analysis, Vol. 28(1997), pp. 1811-1826.
[11] A. C. Lazer and P. J. McKenna, A symmetry theorem and applications to
nonlinear partial differential equations, J. Diff. Equations, 72(1988), pp. 95-106.
[12] , Large-amplitude periodic oscillations in suspension bridges: Some new
connections with nonlinear analysis, SIAM Review, 32(1990), pp. 537-578.
[13] , Large scale oscillation behavior in loaded asymmetric systems, Ann. Inst.
H. Poincare Anal. Non Lineaire, 4(1987), pp. 243-274.
[14] P. J. McKenna and W. Walter, Nonlinear oscillations in a suspension bridge,
Arch. Rational Mech. Anal., 98(1987), pp. 167-177.
[15] B. Pittel and V. Jakubovic, A mathematical analysis of the stability of suspension
bridges based on the example of the Tacoma bridge, Vestnik Leningrad. Univ., 24
(1969), pp. 80-91.
[16] P. H. Rabinowitz, Minimax methods in critical point theory with applications to
differential equations, Regional Conference Series in Mathematics, No. 65, AMS,
1986.
[17] R. H. Scanlan, The action of flexible bridges under wind. Part I: Flutter theory,
J. Sound Vibration, 60(1978), pp. 187-199.
[18] , The action of flexible bridges under wind. Part II: Buffeting theory, J.
Sound Vibration, 60(1978), pp. 201-211.
[19] A. Selberg, Oscillation and aerodynamic stability of suspension bridges, Acta
Polytech. Scand., 13(1961), pp. 308-377.
[20] E. G. Wiles, Report of aerodynamic studies on proposed San Pedro-Terminal
Island suspension bridge, California, Research, Bureau of Public Roads, U. S.
Department of Commerce, Washington, D. C., 1960.
Canonical Dual Control for Nonconvex
Distributed-Parameter Systems: Theory and
Method

David Y. Gao, Virginia Polytechnic Institute and State University, Blacks-

burg, VA 24061. E-mail: [email protected]

Abstract

This paper presents a potentially powerful canonical dual transfor-

mation method and associated duality theory for solving fully nonlinear
distributed-parameter control problems. The extended Lagrange duality
and the interesting triality theory proposed recently in finite deformation
theory are generalized into nonconvex dynamical systems. A bifurcation
criterion is proposed, which leads to an effective dual feedback control
against the chaotic vibration in Duffing system.

1 Problems and Motivations

We shall study a duality approach for solving the following very general abstract
distributed parameter problem ((P) for short),

(1.1) (P) : ρu,tt + A(u, µ) = 0 ∀u ∈ Uk ,

where the feasible space Uk is a convex, non-empty subset of a reflexive Banach

space U over an open space-time domain Ωt = Ω × (0, tc ) ⊂ Rn × R+ , in which,
certain essential boundary-initial conditions are prescribed. We assume that
for a given distributed parameter control field µ(x, t) over Ωt , the mapping
A(u, µ) is a potential operator from Uk into its dual space U ∗ , i.e., there exists
a Gâteaux differentiable potential functional Pµ (u) = P (u; µ), such that the
directional derivative of P at ū ∈ Uk in the direction δu can be written as

δPµ (ū; δu) = hDPµ (ū), δui ∀δu ∈ Uk ,

where the operator DPµ (ū) = A(ū, µ) is the Gâteaux derivative of Pµ at the
point ū; the bilinear form h·, ·i : U × U ∗ → R places U and U ∗ in duality. By
nonlinear operator theory we know that the mapping A : Uk → U ∗ is monotone
if and only if P is convex on Uk .
85
86 Gao

The problem (P) is said to be exactly controllable if for certain given initial
data (u0 (x), v0 (x)) in Uk and the final state (ūc (x), v̄c (x)) there exists suitable
control function µ(x, t) such that the solution u(x, t) of the problem (P) satisfies

(1.2) u(x, tc ) = ūc (x), u,t (x, tc ) = v̄c (x) ∀x ∈ Ω.

Dually, the problem (P) is said to be observable if for certain given input
control µ(x, t), there exists an output function h(u) such that the initial state
(uo (x), vo (x)) can be uniquely determined from the output z = h(u(x, t)) over
any interval 0 < t < tc .
The abstract form of problem (P) covers a great variety of situations. Very
often, the total potential Pµ (u) can be written as

Pµ (u) = Φµ (u, Λ(u)) = Wµ (Λ(u)) − Fµ (u),

where Λ is a Gâteaux differentiable operator from U into another Banach space

E; the functional Wµ (ξ) is the so-called stored (or internal) potential; while the
functional Fµ (u) represents the external potential of the system.
In convex Hamilton systems, the total potential Pµ (u) is convex and
its Gâteaux derivative A(u; µ) = DPµ (u) is usually an elliptic operator in
conservative problems. In linear field theory of mathematical physics, Λ is
usually a gradient-like operator, say Λ = grad, and Wµ (ξ) is a quadratic
functional, for example,
Z
1
Pµ (u) = a(x)|∇u|2 dΩ − Fµ (u),
Ω 2
where a(x) > 0 ∀x ∈ Ω. In this case, the governing equation (1.1) reads

(1.3) ρu,tt = ∇ · (a(x)∇u) + DFµ (u) ∀(x, t) ∈ Ωt .

It is a linear wave equation if Fµ (u) is a linear functional, say Fµ (µ) =

hu , u∗ (µ)i, where u∗ (µ) is a given function of the input control field µ(x, t). If
Fµ (u) is nonlinear, then the governing equation (1.3) is semi-linear. In boundary
control problems, the distributed-parameter µ also appears in the feasible set
Uk . In applications of engineering mechanics, the state variable u could be also
a vector-valued function and Λ is a tensor type operator. For example, in the
shear control of extended beam structures, the actuators are filaments attached
to the upper and lower beam surfaces (y = ±h). The external signals effect a
change of the properties of these filaments in such way that they produce shear
forces µ± (x, t). Thus, µ± (x, t) is, in effect, the applied distributed-control, and
the composite beam/actuator system is then an instance of an active, or “smart”
structure. Since the repeated operation of these actuator devices results large
shear deformations, the traditional Timoshenko beam model can not be used to
the study of these phenomena because it assumes that the shear deformation is
 Canonical Dual Control for Nonconvex Systems 87

a function of x and t alone and does not vary in the lateral beam direction. In
order to study the control problems of smart structures, several extended beams
models have been proposed recently by Gao and Russell (1994, 1996), where the
state variable space U = C 1 (Ωt ; R2 ) is a displacement space over the space time
domain Ωt = (0, ℓ) × (−h, h) × (0, tc ). The element u = {χ(x, y, t), w(x, t)} ∈ U
is a continuous, differentiable vector in R2 with domain Ωt , where χ(x, y, t)
measures the shear deformation of the beam at the point (x, y), while w(x, t) is
the deflection of the beam. In the case that the elastic beam subjected to the
transverse load f (x, t) undergone infinitesmall deformation, the total potential
is a quadratic functional
Z
1
Pµ (χ, w) = [χ2 + β(χ,y + w,x )2 ] dΩ
2 Ω ,x
Z ℓ
− (µ+ (x, t)χ(x, h, t) + µ− (x, t)χ(x, −h, t) + f (x, t)w) dx.
0

If the beam is clamped at x = 0 and simply supported at x = ℓ, and subjected

to a compressive load at x = ℓ, the kinematical admissible space Uk ⊂ U can
be definded as
  
  χ(x, −y, t) = −χ(x, y, t), w(0, t) = w(ℓ, t) = 0, 

    

χ  χ(0, y, t) = χ,x (ℓ, y, t) = 0 ∀y ∈ [−h, h],
Uk = ∈ U  .

 w  (χ, w) = (χ0 , w0 ), (χ,t , w,t ) = (χ̇0 , ẇ0 ) 

  
∀(x, y) ∈ Ω, t = 0

In this case, the abstract governing equation (1) is a linear coupled partial
differential system

ρχ χ,tt = χ,xx + βχ,yy ,

β
(1.4) ρw w,tt = βw,xx + 2h [χ,x (x, h, t) − χ,x (x, −h, t)] + f (x, t),

χ,y (x, ±h, t) + w,x (x, t) = ±µ± (x, t).

Since the total potential of this system is strictly convex, for the given input
control function µ± (x, t), this system possesses a unique stable solution.
Due to the efforts of more than thirty years research by many well-
known mathematicians and scientists, the mathematical theory for distributed-
parameter control systems have been well-established for convex Hamilton
systems governed by partial differential equations (cf. e.g., Russell, 1973, 1978,
1986, 1996; Chen et al, 1991; Komornik, 1994; Lasiecka and Triggiani, 1999)
with substantial applications in mechanics and structures (see, for examples,
Lagnese and Lions, 1988; Lasiecka, 1998a; Lasiecka and Triggiani, 1987, 1999;
Zuazua, 1996). In linear systems, there exists a very elegant duality relationship
between the controllability and observability (see Dolecki and Russell, 1977).
88 Gao

If the system reversible, the well-known Russell principle states that the
stabilizability implies its exact controllability. The celebrated review articles
by Russell (1978) and Lions (1988) still serve the excellent introductions to
the mathematical aspects of controllability, stabilization and perturbations for
distributed-parameter systems.
Duality is a fundamental concept that underlies almost all natural phenom-
ena. In classical optimization and calculus of variation, duality methods possess
beautiful theoretical properties, potentially powerful alternative performances
and pleasing relationships to many other fields. The associated theory and
extremality principles have been well studied for convex static and Hamilton
systems (cf. e.g., Toland, 1978, 1979; Auchmuty, 1983, 1989, 1997; Strang, 1986;
Rockafellar and Wets, 1997). There is a growing interest in studying and ap-
plications of convex duality theory in optimal control (cf., e.g., Mossino (1975),
Chan and Ho (1979), Chan (1985), Chan and Yung (1987), Barron (1990),
Tanimoto (1992), Lee and Yung (1997), Bergounioux et al (1999), Arada and
Raymond (1999) and many others). The interesting one-to-one analogy be-
tween the optimal control and engineering structural mechanics was discovered
by Zhong et al (1993, 1999). Recently, the so-called primal-dual interior-point
(PDIP) method has been considered as a revolution in linear constrained opti-
mization problems (cf. e.g., Gay et al, 1998; Wright, 1998). It was shown by
Helton et al (1998) that the fundamental H ∞ optimization problem of control
can be naturally treated with the PDIP methods.
However, the beautiful duality relationship in convex systems is broken in
nonconvex problems. In many applications of engineering and sciences, the
total potential of system is usually nonconvex, and even nonsmooth. The
exact controllability and stability for nonconvex/nonsmooth systems are much
more difficult. For example, in the well-known von Kármán thin plate model,
the state variable u is a vector-valued function u = {χ(x, t), w(x, t)} over
Ωt ⊂ R2 × R, where χ = {χα } (α = 1, 2) is an in-plane displacement vector,
while w(x, t) stands for the deflection of the plate at (x, t) ∈ Ωt . The total
potential is a nonlinear functional
Z
1 1
(1.5) P (χ, w) = a(w, w) + b(ξ(χ, w), ξ(χ, w)) − f w dΩ,
2 2 Ω

where a(w, w) and b(ξ, ξ) are two bilinear forms, defined respectively by
Z
a(w, w) = K [(1 − ν)(∇∇w)(∇∇w) + ν∆w∆w] dΩ,
Z Ω
b(ξ, ξ) = hξαβ Cαβγθ ξγθ dΩ,
Ω
and ξ is a Cauchy-Green type strain tensor, defined by
1
ξαβ = (χα,β + χβ,α + w,α w,β ), α, β = 1, 2.
2
 Canonical Dual Control for Nonconvex Systems 89

The governing equations for dynamical von Kármán plate are coupled nonlinear
partial differential system
ρw w,tt = h∇ · (σ · ∇w) − K0 ∆∆w + f,
(1.6)
ρχ χ,tt = ∇ · σ, σ = Cξ(χ, w).
This coupled nonlinear partial differential system is a typical example in finite
deformation mechanics. The mathematical control theory for large deformation
plates and shells has emerged as the most challenging and active research field
in recent years. In a series of papers by Lasiecka and her colleagues (see,
for examples, Horn and Lasiecka, 1994, 1995; Favini et al, 1996; Lasiecka,
1998, 1999), many important contributions and open questions have been
addressed for stabilizability of the so-called full von Kármán system with
nonlinear boundary feedback (see Lasiecka, 1998). A detailed documentation
on mathematical control theory of coupled nonlinear PDE’s has been given in
a lecture note by Lasiecka (1999). Since the von Kármán model is valid only
for plates subjected to the moderately large deflections, only the second-order
nonlinear term w,α w,β is considered, and the governing equation is linearly
dependent on the in-plane deformation χ. In many engineering applications,
the acceleration term ρχ,tt can usually be ignored. Thus, the second equation
in (1.6) reads ∇ · σ(χ, w) = 0. If the plate is subjected to compressive load on
the boundary, the plate will be in the post-buckling state when the compressive
load reaches its critical point. In this case, the total potential is nonconvex
(i.e. the so-called double-well energy) (see Gao, 1995). In one-dimensional
problems, the in-plane equilibrium condition σ,x = 0 leads to a constant stress
σ = −λ everywhere in the domain Ω = (0, ℓ) ⊂ R. In this case, the nonlinear
von Kármán model (1.6) in R2 reduces to a linear equation in one-dimensional
“beam” problem, i.e.,
ρw w,tt = hλw,xx − K0 w,xxxx + f.
The main reason behind this von Kármán “paradox” is that the seconder order
nonlinear term w,α w,β is considered for in-plane strain ξ, but it is ignored in
the thickness direction. It may be appropriate for thin plates, but for one-
dimensional beam models, this is wrong! It is shown in Gao (1996) that the
strain in the thickness direction of the beam is proportional to the second-order
term w,x2 , and cannot be ignored when the beam is subjected to moderately large

rotations. Thus, an extended large deformation beam model was proposed as

1 2
(1.7) ρw w,tt + K0 w,xxxx − k0 (λ − w,x )w,xx − f = 0 in Ωt = (0, ℓ) × (0, tc ),
2
where k0 > 0 is a positive material constant. The total potential energy
associated with this nonlinear beam theory is a nonlinear functional
Z   Z
1 2 1 2 2
(1.8) P (w) = K0 w,xx + k0 ( w,x − λ) dx − f w dx.
I 2 2 I
90 Gao

In static problem, if the beam is clamped at x = 0, simply supported at

x = ℓ, the kinematically admissible space Uk can be written as

Uk = {w ∈ C 2 (0, ℓ)| w(0) = w,x (0) = 0, w(ℓ) = w,xx (ℓ) = 0}.

It is clear that for the given Euler pre-buckling load λc > 0,

Z Z
2 2
K0 w,xx dx ≥ λc w,x dx, ∀w ∈ Uk .
I I

Thus, on Uk ,
Z   Z
1 2 1 2
P (w) ≥ λc w,x + k0 ( w,x − λ)2 dx − f w dx
I 2 2 I
= Pµ (w) + λℓλc /k0 − ℓλ2c /(2k02 ),

where Pµ (w) is a nonlinear functional

Z Z
1 1 2
(1.9) Pµ (w) = k0 ( w,x − µ)2 dx − f w dx,
I 2 2 I

and µ = λ − λc /k0 ∈ R is
R a1 constant. Clearly, when the parameter µ > 0, the
1 2
stored energy Wµ (ε) = I 2 k0 ( 2 ε − µ)2 dx is the well-known van der Waals
double-well function (see Figure 1a) of the linear strain ε = w,x , the beam is
in a post-buckled (bifurcation) state. In this case, the total potential Pµ is
nonconvex. It has three critical points: two local minimizers, corresponding to
two possible stable buckled states, and one local maximizer, corresponding to
an unstable buckled state. The global minimizer of Pµ depends on the lateral
load f (see Figure 1b).

(a) Graph of Wµ (ε) (b) Graphs of Pµ (u) (f > 0 solid, f < 0 dashed)

Fig. 1. Double-well energy and nonconvex potential

 Canonical Dual Control for Nonconvex Systems 91

If the beam is subjected to a periodic dynamical load f (x, t), the two local
minimizers of Pµ become extremely unstable, and the beam is in dynamical
post-buckling state. In this case, the governing equation (1.7) is replaced by
3 2
(1.10) ρw w,tt = k0 ( w,x − µ)w,xx + f (x, t) in Ωt = (0, ℓ) × (0, tc ),
2
If the deflection w(x, t) can be separated into w(x, t) = u(t)v(x), this post-
buckling dynamical beam model is equivalent to the well-known Duffing
equation:
1
(1.11) u,tt = au(µo − u2 ) + µ(t).
2
where a > 0 and µo ∈ R are constants. This equation is extremely sensitive to
the initial data. It is known that for certain give parameter µo and the driving
input µ(t), this equation may produce the so-called chaotic solutions.
The problem of controlling chaotic systems is of significant practical
importance and has attracted considerable attention during the last years.
Mathematically speaking, the total potential of the chaotic system is usually
nonconvex or even nonsmooth. Very small perturbations of the system’s initial
conditions and parameters may lead the system to different operating points
with significantly different performance characteristics. This is the one of main
reasons why the traditional perturbation analysis, the direct approaches and
many standard control techniques cannot successfully be applied to chaotic
systems. Based upon these observations and in order to handle the nonlinear
problem, a school of new techniques has been developed (see, e.g., Fowler, 1989;
Ott et al, 1990; Chen and Dong, 1992, 1993; Ogorzalek, 1993; Antoniou et al,
1996; Ghezzi and Piccardi, 1997; Mertzios and Koumboulis, 1996; Koumboulis
and Mertzios, 2000). In the shear control of large deformation extended beam
model, the equation (1.4) can be replaced by (see Gao, 2000a)

(1.12)
χ,xx + βχ,yy = 0,
 2 
β
ρw w,tt = 3α2 w,x
2 + β − λα w
,xx + 2h [χ,x (x, h, t) − χ,x (x, −h, t)] + f (x, t),

χ,y (x, ±h, t) + w,x (x, t) = ±µ± (x, t),

where α > 0 is a given material constant and λ > 0 represents the axial load.
The total potential associated with this model is a nonconvex functional
Z
1 1 2
Pµ (χ, w) = [(χ2,x + αw,x − λ)2 + β(χ,y + w,x )2 ] dΩ
2 Ω 2
Z ℓ
− (µ+ (x, t)χ(x, h, t) + µ− (x, t)χ(x, −h, t) + f (x, t)w) dx.
0
92 Gao

In order to control the chaotic vibration of this nonconvex dynamical beam

system, an efficient canonical dual feedback control method has been proposed
recently by the author (Gao, 2000e).
The duality theory in fully nonlinear variational problems was originally
studied by Gao and Strang (1989) for large deformation nonsmooth mechanics.
In order to recover the broken symmetry in fully nonlinear systems (see Defini-
tion 2), a so-called complementary gap function was introduced. It was realized
recently in post-buckling analysis of nonlinear beam theory (Gao, 1996) that
this function recovered the duality gap between the nonconvex primal problems
and the Fenchel-Rockafellar dual problems. A self-contained comprehensive pre-
sentation of the mathematical theory for general nonconvex systems was given
recently by Gao (1999). A so-called canonical dual transformation method and
associated triality theory have been proposed for solving nonconvex/nonsmooth
variational-boundary value problems. Compared with the traditional analytic
methods and direct approaches, the main advantages of this canonical dual
transformation method are
(1) converting nonconvex/nonsmooth constrained variational problems into
smooth unconstrained dual problems;
(2) transforming certain fully nonlinear partial differential equations into
algebraic systems;
(3) providing powerful and efficient primal-dual alternative approaches.
The aim of this article is to generalize the author’s previous results on
nonconvex variational problems into distributed-parameter control systems.
The rest of this paper is divided into four main sections. The next section
set up the notation used in the paper. A general framework in fully nonlinear
systems are discussed. Section 3 presents an extended Lagrangian critical point
theorem and associated triality theory in general nonconvex dynamical systems.
The critical points in fully nonlinear systems are classified. Section 4 is devoted
mainly to the construction of dual action in fully nonlinear systems. The
nice tri-duality proposed in static boundary value problems is generalized into
control problems. Section 5 discusses the application in Duffing system. A
bifurcation criterion is proposed which can be used for feedback controlling
against chaotic vibrations.

2 Framework for Canonical Systems and Classification

Let U and U ∗ be two locally convex topological real linear spaces, placed in
separating duality by a bilinear form h·, ·i : U × U ∗ → R. Let P : Us → R be
a given functional, well-defined on a convex domain Us ⊂ U such that for any
given u ∈ Us , P (u) is Gâteaux differentiable. Thus, the Gâteaux derivative DP
of P at u ∈ Us is a mapping from Us into U ∗ . Let Us∗ ⊂ U ∗ be the range of the
mapping DP : Us → U ∗ . If the relation u∗ = DP (u) is reversible on Us , then,
for any given u∗ ∈ Us∗ , the classical Legendre conjugate functional P ∗ : Us∗ → R
 Canonical Dual Control for Nonconvex Systems 93

of P (u) is defined by

P ∗ (u∗ ) = hu(u∗ ), u∗ i − P (u(u∗ )).

The conjugate pair (u, u∗ ) is called the Legendre duality pair on Us ×Us∗ ⊂ U ×U ∗
if and only if the equivelant relations

(2.1) u∗ = DP (u) ⇔ u = DP ∗ (u∗ ) ⇔ P (u) + P ∗ (u∗ ) = hu, u∗ i.

hold on Us × Us∗ .
The following notations and definitions, used in Gao (1999), will be of
convenience in nonconvex control problems.
Definition 2.1. The set of functionals P : U → R which are either convex
or concave is denoted by Γ(U). In particular, let Γ̌(U) denote the subset of
functionals P ∈ Γ(U) which are convex and Γ̂(U) the subset of P ∈ Γ(U) which
are concave.
The canonical functional space ΓG (Us ) is a subset of functionals P ∈ Γ(Us )
which are Gâteaux differentiable on Us ⊂ U, such that the relation u∗ = DP (u)
is reversible for any given u ∈ Us . ♦
Clearly, if P ∈ ΓG (Us ) and Us∗ is the range of the mapping DP : Us → U ∗ ,
then the Legendre duality relations (2.1) hold on Us × Us∗ .
Let (E, E ∗ ) be an another pair of locally convex topological real linear spaces
paired in separating duality by the second bilinear form h· ; ·i : E ×E ∗ → R. The
so-called geometrical operator Λ : U → E is a continuous, Gâteaux differentiable
operator such that for any given u ∈ Ua ⊂ U, there exists an element ξ ∈ Ea ⊂ E
satisfying the geometrical equation

ξ = Λ(u).

The directional derivative of ξ at ū in the direction u ∈ U is then definded by

ξ(ū + θu) − ξ(ū)

(2.2) δξ(ū; u) := lim = Λt (ū)u,
θ→0+ θ

where Λt (ū) = DΛ(ū) : U → E denotes the Gâteaux derivative of the operator

Λ at ū. For a given ξ ∗ ∈ E ∗ , GΛ (u) = hΛ(u) ; ξ ∗ i is a real-valued functional of
u on U. Its Gâteaux derivative at ū ∈ U in the direction u ∈ U reads

δGΛ (ū; u) = hΛt (ū)u ; ξ ∗ i = hu , Λ∗t (ū)ξ ∗ i,

where Λ∗t (ū) : E ∗ → U ∗ is the adjoint operator of Λt associated with the two
bilinear forms.
Let V and V ∗ be the velocity and momentum spaces, respectively, placed in
duality by the third bilinear form h∗ , ∗i : V ×V ∗ → R. For Newtonian systems,
94 Gao

the kinetic energy K : V → R and its Legendre conjugate K ∗ : V ∗ → R are

quadratic forms
Z Z
1 2 ∗ 1 −1 2
K(v) = ρv dΩ, K (p) = ρ p dΩ.
Ω 2 Ω 2

Thus the canonical physical relations between V and V ∗ are linear:

p = DK(v) = ρv ⇔ v = DK ∗ (p) = ρ−1 p.

Let Va ⊂ V be a subspace defined by

(2.3) Va = {v ∈ V| v(x, 0) = v0 ∀x ∈ Ω}.

Finally, we let M be an admissible control space over Ωt . For any given

µ ∈ M, we assume that there exists a Gâteaux differentiable functional
Φµ : Ua × Ea ⊂ U × E → R, such that the total potential P (u; µ) of the
system can be written as

(2.4) Pµ (u) = P (u; µ) = Φµ (u, Λ(u)),

and the total action of the system

Z tc
(2.5) Πµ (u) = [K(u,t ) − Φµ (u, Λ(u))] dt
0

is well-defined on the feasible space Uk given by

(2.6) Uk = {u ∈ Ua | Λ(u) ∈ Ea , u,t ∈ Va }.

The following classification for distributed parameter control systems was

originally introduced in nonlinear variational/boundary value problems by Gao
(1998, 1999).
Definition 2.2. Suppose that for the problem (P) given in (1), the
associated total potential Pµ (u) is well-defined on its domain Us ⊂ U. If
the geometrical operator Λ : U → E can be chosen in such a way that
Pµ (u) = Φµ (u, Λ(u)), Φµ ∈ ΓG (Ua ) × ΓG (Ea ) and Us = {u ∈ Ua | Λ(u) ∈ Ea }.
Then
(1) the transformation {P ; Us } → {Φµ ; Ua × Ea } is called the canonical
transformation, and Φµ : Ua × Ea → R is called the canonical functional
associated with Λ;
(2) the problem (P) is called geometrically nonlinear (or linear) if Λ : U → E
is nonlinear (or linear); it is called physically nonlinear (resp. linear) if the
duality mapping DΦµ : Ua × Ea → Ua∗ × Ea∗ is nonlinear (resp. linear); it is
called fully nonlinear if it is both geometrically and physically nonlinear. ♦
 Canonical Dual Control for Nonconvex Systems 95

v∈ V  hv , pi - V∗ ∋ p
d
6
−d
dt
? dt

u∈ U  hu , u∗ i - U ∗ ∋ u∗
6Λ∗ = (Λ − Λ )∗
Λt + Λc = Λ
?
t c

ξ∈ E  hξ ; ξ ∗ i - E ∗ ∋ ξ∗

Fig. 2. Framework in fully nonlinear Newtonian systems

The canonical transformation plays a fundamental role in duality theory of

nonconvex systems. Clearly, if Φµ ∈ ΓG (Ua ) × ΓG (Ea ) is a canonical functional,
the Gâteaux derivative DΦµ : Ua × Ea → Ua∗ × Ea∗ ⊂ U ∗ × E ∗ is a monotone
mapping, i.e., the duality relations

(2.7) u∗ = Du Φµ (u, ξ), ξ ∗ = Dξ Φµ (u, ξ)

are reversible between the paired spaces (Ua , Ua∗ ) and (Ea , Ea∗ ), where Du Φµ and
Dξ Φµ denote the partial Gâteaux derivatives of Φµ with respect to u and ξ,
respectively. Thus, on Uk , the directional derivative of Pµ at ū in the direction
u ∈ Uk can be written as

δPµ (ū; u) = hu , Du Φµ (ū, Λ(ū))i + hΛt (ū)u ; Dξ Φµ (ū, Λ(ū))i

= hu , ū∗ i + hu ; Λ∗t (ū)ξ̄ ∗ i ∀u ∈ Uk .

In terms of canonical variables, the governing equation (1) for the fully nonlinear
problems can be written in the tri-canonical forms, namely,
(1) geometrical equations: v = u,t , ξ = Λ(u),
(2.8) (2) physical relations: p = ρv, (u∗ , ξ ∗ ) = DΦµ (u, ξ),
(3) balance equation: p,t + u∗ + Λ∗t (u)ξ ∗ = 0.
The framework for the fully nonlinear system is shown in Figure 2. Extensive
illustrations of the canonical transformation and the tri-canonical forms in
mathematical physics and variational analysis were given in the monograph
by Gao (1999).
In geometrically linear systems, where Λ : U → E is linear, we have Λ = Λt .
For dynamical problems, if the total potential Pµ is convex, the total action
96 Gao

associated with the problem (P) is a d.c. functional, i.e., the difference of
convex functionals:
Z tc
Πµ (u) = [K(u,t ) − Pµ (u)] dt.
0

It was shown by Gao (1999) that the critical point of Πµ either minimizes or
maximizes Πµ over the kinetically admissible space. The classical Hamiltonian
associated with this d.c. functional Πµ is a convex functional on the phase
space U × V ∗ , i.e.
(2.9) H(u, p) = K ∗ (p) + Pµ (u),

The classical canonical forms for convex Hamilton systems are well-known
d d
u = Dp H(u, p), − p = Du H(u, p).
dt dt
Furthermore, if Φµ (u, ξ) = 12 hξ ; Cξi − hu , µi is a quadratic quadratic
functional, where C : E → E ∗ is a linear operator, then the governing equations
for linear system can be written as

ρū,tt + Λ∗ CΛu = µ.
For conservative systems, the operator Λ∗ CΛ is usually symmetric.
In geometrically nonlinear systems, Λ 6= Λt , and the total potential Pµ (u)
is usually a nonconvex functional. In this case, we have the following operator
decomposition
(2.10) Λ(u) = Λt (u)u + Λc (u),
where Λc : U → E is called the complementary operator of the Gâteaux
derivative operator Λt . By this decomposition, we have

(2.11) hΛ(ū) ; ξ̄ ∗ i = hū , Λ∗t (ū)ξ̄ ∗ i − G(ū, ξ¯∗ )

where G : U × E ∗ → R is so-called complementary gap functional, defined by
(2.12) G(u, ξ ∗ ) = h−Λc (u) ; ξ ∗ i : U × E ∗ → R.

This functional was first introduced by Gao and Strang (1989) in finite
deformation theory, which plays a key role in nonconvex variational problems.
As a typical example in nonconvex dynamical systems, let us consider the
post-buckling dynamical beam model (1.10) discussed in section 1. For a given
feasible space Uk , we consider the following nonconvex variational problem over
the domain Ωt = (0, ℓ) × (0, tc )
Z  
1 2 1 1 2 2
(2.13) Πµ (u) = ρu,t − a( u,x − µ) + uf dx dt → sta ∀u ∈ Uk ,
Ωt 2 2 2
 Canonical Dual Control for Nonconvex Systems 97

where a, µ are given positive constants. This nonconvex problem also appears
very often in phase transitions and hysteresis.
First, we let Λ = d/ dx be a linear operator, and Pµ (u) = Wµ (Λu) − Fµ (u)
with
Z ℓ Z ℓ
1 1 2 2
Wµ (ε) = a( ε − µ) dx, F (u) = uf dx,
0 2 2 0

Thus, Wµ (ε) is the so-called van der Waals’ double-well function of the linear
“strain” ε = u,x . Since Wµ (ε) is not a canonical functional, the constitutive
equation ε∗ = DWµ (ε) is not one-to-one. Thus, the Legendre conjugate of
Wµ (ε) does not have a simple algebraic expression. The Fenchel conjugate
Wµ∗ (ε∗ ) of the double-well energy Wµ (ε), defined by

Wµ∗ (ε∗ ) = sup{hε ; ε∗ i − Wµ (ε)},

ε

is always a convex, lower semi-continuous functional. However, the well-known

Fenchel-Young inequality

Wµ (u,x ) ≥ hu,x ; ε∗ i − Wµ∗ (ε∗ )

leads to a so-called duality gap between the primal problem and the Fenchel-
Rockafellar dual problem (see Gao, 1999). This nonzero duality gap indicates
that the well-established Fenchel-Rockafellar duality theory can only be used
for solving convex variational problems.
¿From the theory of continuum mechanics we know that in finite deforma-
tion problems, ε = u,x is not a strain measure (it does not satisfy the axiom of
material frame-indifference (cf. e.g., Gao, 1999). In order to recover this duality
gap, we need to choose a suitable geometrical operator Λ, say, Λ(u) = 12 u2,x − µ,
so that the nonconvex problem (2.13) can be put in our framework. In contin-
uum mechanics, this quadratic measure ξ = Λ(u) is a Cauchy-Green type strain.
Thus, in terms of u and ξ, Φµ (u, ξ) = Wµ (ξ) − Fµ (u) = 12 hξ ; aξi − hu , f i
is a canonical functional. The Legendre conjugate of the quadratic functional
Wµ (ξ) = 12 hξ ; aξi is simply defined by W ∗ (ξ ∗ ) = 21 ha−1 ξ ∗ ; ξ ∗ i. The operator
decomposition (2.10) for this quadratic operator reads

1
Λ(u) = Λt (u)u + Λc (u), Λt (u)u = u,x u,x , Λc (u) = − u2,x − µ.
2
The complementary gap functional associated with this quadratic operator is a
quadratic functional of u
Z ℓ
∗ ∗ 1 2 ∗
G(u, ξ ) = h−Λc (u) ; ξ i = u ξ dx.
0 2 ,x
98 Gao

For homogeneous boundary conditions, we have

Z ℓ Z ℓ
∗ ∗
hΛt (u)u ; ξ i = u,x u,x ξ dx = − u(u,x ξ ∗ ),x dx = hu , Λ∗t (u)ξ ∗ i,
0 0

which leads to the adjoint operator Λ∗t of Λt . Thus, the tri-canonical equations
for this nonconvex problem can be listed as the following.
1
v = u,t , ξ = au2,x − µ,
2

p = ρv, ξ ∗ = DWµ (ξ) = aξ, u∗ = DFµ (u) = f

p,t = −Λ∗t (u)ξ ∗ + u∗ = (u,x ξ ∗ ),x + f.

Since the geometrical operator Λ is nonlinear, and the canonical constitutive

equations are linear, the nonconvex problem (2.13) is a geometrically nonlinear
system.

3 Extended Lagrangian and Triality Theory

The triality theory in nonconvex problems was originally proposed by the author
(Gao, 1996, 1997, 1999, 2000) in static finite deformation theory and global
optimization. In this section, we will generalize this interesting result into
fully nonlinear dynamical systems. We assume that for a given fully nonlinear
system, there exists a Gâteaux differentiable operator Λ : Ua → Ea such that
the total potential of the system can be written as

(3.1) Pµ (u) = Wµ (Λ(u)) − Fµ (u),

where Wµ ∈ Γ̌G (Ea ) is a convex canonical functional, while Fµ : Ua → R is a

linear functional. Thus, the primal problem (P) can be reformulated as the
following.
Problem 1 (Primal Distributed-Parameter Control Problem).
For any given primal feasible space Uk = {u ∈ Ua | u,t ∈ Va , Λ(u) ∈ Ea }
and the final state (ūc (x), v̄c (x)), find the control field µ(x, t) ∈ M such that
the solution ū(x, t) of the variational problem
Z tc
(3.2) (P) : Πµ (u) = [K(u,t ) − Wµ (Λ(u)) + Fµ (u)] dt → sta ∀u ∈ Uk
0

satisfying the controllability condition

(ū(x, tc ), ū,t (x, tc )) = (ūc (x), v̄c (x)) ∀x ∈ Ω.

 Canonical Dual Control for Nonconvex Systems 99

It is easy to check that the critical point condition DΠµ (ū) = 0 leads to the
the canonical governing equation

(3.3) ρū,tt = DFµ (ū) − Λ∗t (ū)DWµ (Λ(ū)).

By the Legendre-Fenchel transformation, the conjugate of Wµ (ξ) is defined

by

Wµ∗ (ξ ∗ ) = sup{hξ ; ξ ∗ i − Wµ (ξ)}.

ξ∈E

Since Wµ : Ea → R is a convex canonical functional, Wµ∗ (ξ ∗ ) is well-defined on

the range Ea∗ of the duality mapping DWµ∗ : Ea → E ∗ , the Legendre duality
relation

ξ ∗ = DWµ (ξ) ⇔ ξ = DWµ∗ (ξ ∗ ) ⇔ Wµ (ξ) + Wµ∗ (ξ) = hξ ; ξ ∗ i

holds on Ea × Ea∗ . Moreover, we have Wµ∗∗ (ξ) = Wµ (ξ) for all ξ ∈ Ea . Let
Z = U × V ∗ × E ∗ be the so-called extended canonical phase space.
Definition 3.1. Suppose that for a given problem (P), there exists a
Gâteaux differentiable operator Λ : U → E and canonical functionals Wµ ∈
Γ(E), Fµ ∈ Γ(U) such that Pµ (u) = Wµ (Λ(u)) − Fµ (u). Then
(1) the functional Hµ : Z → R defined by

(3.4) Hµ (u, p, ξ ∗ ) = K ∗ (p) − Wµ∗ (ξ ∗ ) + Fµ (u) ∈ Γ(U) × Γ(V ∗ ) × Γ(E ∗ )

is called the extended canonical Hamiltonian density associated with Πµ ;

(2) the functional Lµ : Z → R definded by

(3.5) Lµ (u, p, ξ ∗ ) = hu,t , pi − hΛ(u) ; ξ ∗ i − Hµ (u, p, ξ ∗ )

is called the extended Lagrangian density of (P) associated with Λ;

(3) the functional Ξµ : Z → R definded by
Z tc
∗
(3.6) Ξµ (u, p, ξ ) = Lµ (u, p, ξ ∗ ) dt
0

is called the extended Lagrangian form of (P). It is called the canonical

Lagrangian form if Ξµ ∈ Γ(U) × Γ(V ∗ ) × Γ(E ∗ ). ♦
∗
A point (ū, p̄, ξ̄ ) ∈ Z is said to be a critical point of Ξµ if Ξµ is Gâteaux-
differentiable at (ū, p̄, ξ¯∗ ) and DΞµ (ū, p̄, ξ¯∗ ) = 0. It is easy to find out that the
criticality condition DΞµ (ū, p̄, ξ¯∗ ) = 0 leads to the following canonical Lagrange
equations

Λ(ū) = Dξ ∗ Wµ∗ (ξ̄ ∗ ), ū,t = DK ∗ (p̄),
(3.7) DΞµ (ū, p̄, ξ̄ ∗ ) = 0 ⇒
p̄,t = DFµ (ū) − Λ∗t (ū)ξ̄ ∗ .
100 Gao

Since Wµ and Fµ are canonical functionals, we know that, by the Legendre

duality theory, any critical point of Ξµ solves the variational problem (P).
Since Fµ (u) : Ua → R is a linear functional, by the Riesz representation
theory, there exists an element ū∗ (µ) ∈ U ∗ such that Fµ (u) = hu , ū∗ (µ)i.
Thus, the extended Lagrangian associated with (P) can be written as

(3.8)
Z tc
Ξµ (u, p, ξ ∗ ) = [hu,t , pi − hΛ(u) ; ξ ∗ i − K ∗ (p) + W ∗ (ξ ∗ ) + hu , ū∗ (µ)i] dt.
0

Note that Ξµ : Va∗ × Ea∗ → R is a saddle functional for any given u ∈ Ua , we

have always the equality

(3.9) inf sup Ξµ (u, p, ξ ∗ ) = sup ∗inf ∗ Ξµ (u, p, ξ ∗ ) ∀u ∈ Ua .

ξ ∗ ∈Ea∗ p∈V ∗ p∈Va∗ ξ ∈Ea
a

However, for any given (p, ξ ∗ ) ∈ Va∗ × Ea∗ , the convexity of Ξµ (·, p, ξ ∗ ) → R
depends on the operator Λ. Let Lc ⊂ Za = Ua × Va∗ × Ea∗ be a critical point set
of Ξµ , i.e.

Lc = {(ū, p̄, ξ̄ ∗ ) ∈ Za | δΞ(ū, p̄, ξ¯∗ ; u, p, ξ ∗ ) = 0 ∀(u, p, ξ ∗ ) ∈ Za }.

For any given critical point (ū, p̄, ξ̄ ∗ ) ∈ Lc , we let Zr = Ur × Vr∗ × Er∗ ⊂ Za
be its neighborhood such that on Zr , (ū, p̄, ξ̄ ∗ ) is the only critical point of Ξµ .
The following extremum results are of fundamental importance in the stability
analysis of nonlinear dynamical systems.
Theorem 3.1 (Triality Theorem). Suppose that (ū, p̄, ξ¯∗ ) ∈ Lc , and Zr
is a neighborhood of (ū, p̄, ξ¯∗ ).
If hΛ(u) ; ξ¯∗ i is concave on Ur , then on Zr ,

(3.10) Ξµ (ū, p̄, ξ̄ ∗ ) = min max min Ξµ (u, p, ξ ∗ ) = max min min Ξµ (u, p, ξ ∗ ).
∗ u p ξ ∗ p u ξ

However, if hΛ(u) ; ξ̄ ∗ i is convex on Ur , then on Zr we have either

Ξµ (ū, p̄, ξ¯∗ ) = min max min Ξµ (u, p, ξ ∗ ) = min max min Ξµ (u, p, ξ ∗ )
∗ u p ξ ∗ p u ξ
∗
(3.11) = min max Ξµ (u, p, ξ ) = min max Ξµ (u, p, ξ ∗ ).
∗ ξ ,u p ∗ p,ξ u

or

Ξµ (ū, p̄, ξ̄ ∗ ) = max min max Ξµ (u, p, ξ ∗ ) = max min max Ξµ (u, p, ξ ∗ )
∗ u ξ p ∗ p ξ u
∗
(3.12) = min max Ξµ (u, p, ξ ) = max min Ξµ (u, p, ξ ∗ ).
∗ ξ u,p ∗ u,p ξ
 Canonical Dual Control for Nonconvex Systems 101

Proof. Since Wµ∗ ∈ Γ̌(Ea∗ ), K ∗ ∈ Γ̌(Va∗ ), if hΛ(u) ; ξ̄ ∗ i is concave on Ur ,

then for the given ξ̄ ∗ , Ξµ ∈ Γ̌(Ur ) × Γ̂(Va∗ ) is a saddle functional. Thus the
equality (3.10) follows from the saddle-Lagrangian duality theorem (cf. e.g.,
Gao, 1999). However, if hΛ(u) ; ξ¯∗ i is convex on Ur , then for any given ξ ∗ ∈ Er∗ ,
the extended Lagrangian Ξµ ∈ Γ̂(Ur ) × Γ̂(Va∗ ) is a super-critical functional (see
Gao, 1999). By the super-Lagrangian duality theorem proved in Gao (1999), we
have either (3.11) or (3.12).

4 Dual Action and Tri-Duality Theory

The goal of this section is to develop a dual approach for solving the distributed
parameter control problem (P). For any given u ∈ Uk , the extended Lagrangian
density Ξµ (u, p, ξ ∗ ) is a saddle functional on V ∗ × E ∗ , and we have

(4.1) Πµ (u) = sup ∗inf ∗ Ξµ (u, p, ξ ∗ ) ∀u ∈ Uk .

p∈V ∗ ξ ∈E

On the other hand, the dual action Πdµ : Va∗ × Ea∗ → R can be defined by

Πdµ (p, ξ ∗ ) = sta{Ξµ (u, p, ξ ∗ )| ∀u ∈ Ua }

Z tc
(4.2) = FµΛ (p, ξ ∗ ) − [K ∗ (p) − Wµ∗ (ξ ∗ )] dt, ∀(p, ξ ∗ ) ∈ Va∗ × Ea∗ .
0

where FµΛ (p, ξ ∗ ) is the so-called Λ-dual functional of Fµ (u) defined by

Z tc
(4.3) FµΛ (p, ξ ∗ ) = sta [hu,t , pi − hΛ(u) ; ξ ∗ i + Fµ (u)] dt.
u∈Ua
0

Since Fµ (u) = hu , ū∗ (µ)i is a linear functional, for any given (p, ξ ∗ ) ∈ Va∗ × Ea∗
and the applied control µ ∈ M, the solution ū of this stationary problem (4.3)
satisfies the balance equation

(4.4) p,t + Λ∗t (ū)ξ ∗ = ū∗ (µ) in Ωt .

For geometrically linear systems, where Λ is a linear operator, we have

(4.5) FµΛ (p, ξ ∗ ) = up|t=tc ∗ ∗ ∗

t=0 , s.t. Λ ξ + p,t = ū (µ).

In this case,
Z tc
(4.6) Πdµ (p, ξ ∗ ) = up|t=t
t=0
c
+ [Wµ∗ (ξ ∗ ) − K ∗ (p)] dt
0

is the classical complementary action in linear engineering dynamical systems

(see Tabarrok and Rimrott, 1994) defined on the dual feasible space

Ts = {(p, ξ ∗ ) ∈ Va × Ea∗ | p,t + Λ∗ ξ ∗ = ū∗ (µ)}.

102 Gao

In fully nonlinear systems, we let Ts ⊂ Va∗ × Ea∗ be a subspace such that

for any given (p, ξ ∗ ) ∈ Ts , the critical point ū can be determined by (4.4) as
ū = ū(p, ξ ∗ ) and the dual action Πdµ is well defined by (4.2). Thus, by the
operator decomposition Λ = Λt + Λc , we have
Z tc
∗ t=tc
(4.7) Λ
Fµ (p, ξ ) = up|t=0 + Gd (p, ξ ∗ ) dt, s.t. Λ∗t (ū)ξ ∗ + p,t = u∗ (µ),
0

where Gd (p, ξ ∗ )
= h−Λc (ū) ; ξ ∗ i is the so-called pure complementary gap
functional. Then, the problem dual to the primal control problem (P) can
be proposed as the following.
Problem 2 (Dual Distributed-Parameter Control Problem). For
a given dual feasible space Ts and the final state (uc (x), vc (x)), find the control
field µ(x, t) ∈ M such that the dual solution (p̄(x, t), ξ¯∗ (x, t)) of the dual
variational problem
(4.8) (P d ) : Πdµ (p, ξ ∗ ) → sta ∀(p, ξ ∗ ) ∈ Ts
and the associated state ū(x, t) satisfying the controllability condition
(4.9) (ū(x, tc ), ρ−1 p̄(x, tc )) = (uc (x), vc (x)) ∀x ∈ Ω.
The following lemma plays a key role in duality theory for nonlinear
dynamical systems.
Lemma 4.1. Let Ξµ (u, p, ξ ∗ ) be a given extended Lagrangian associated
with (P) and Πdµ (p, ξ ∗ ) the dual action defined by (4.2). Suppose that Zr =
Ur × Vr∗ × Er∗ is an open subset of Za and (ū, p̄, ξ̄ ∗ ) ∈ Zr is a critical point of
Ξµ on Zr , Πµ is Gâteaux differentiable at ū, and Πdµ is Gâteaux differentiable
at (p̄, ξ¯∗ ). Then DΠµ (ū) = 0, DΠdµ (p̄, ξ¯∗ ) = 0, and

(4.10) Πµ (ū) = Ξµ (ū, p̄, ξ̄ ∗ ) = Πdµ (p̄, ξ¯∗ ).

The proof of this lemma can be found in Gao (1998) in parametrical
variational analysis. Lemma 4 shows that the critical points of the extended
Lagrangian are also the critical points for both the primal and dual variational
problems.
Theorem 4.1 (Tri-Duality Theorem). Suppose that (ū, p̄, ξ¯∗ ) ∈ Lc is
a critical point of Ξµ and Zr = Ur × Vr∗ × Er∗ is a neighborhood of (ū, p̄, ξ¯∗ ) such
that Vr∗ × Er∗ ⊂ Ts . If hΛ(u) ; ξ̄ ∗ i is concave on Ur , then
(4.11) Πµ (ū) = min Πµ (u) iff Πdµ (p̄, ξ̄ ∗ ) = max∗ min
∗ ∗
Πdµ (p, ξ ∗ ).
u∈Ur p∈Vr ξ ∈Er

However, if hΛ(u) ; ξ¯∗ i is convex on Ur , then

(4.12) Πµ (ū) = min Πµ (u) iff Πdµ (p̄, ξ¯∗ ) = min Πdµ (p, ξ ∗ );
u∈Ur (p,ξ ∗ )∈Ts

(4.13) Πµ (ū) = max Πµ (u) iff Πdµ (p̄, ξ¯∗ ) = max∗ min
∗ ∗
Πdµ (p, ξ ∗ ).
u∈Ur p∈Vr ξ ∈Er
 Canonical Dual Control for Nonconvex Systems 103

Proof. This theorem can be proved by combining Lemma 1 and the triality
theorem.

5 Feedback Control Against Chaotic Duffing System

As a typical example, let us consider the very simple nonconvex dynamical
problem over the time domain I = (0, tc )
Z
′ 1 1
(5.1) Πµ (u) = [ρu 2 − a( u2 − µo )2 + µu] dt → sta ∀u ∈ Uk .
I 2 2

The kinematically admissible space Uk for the initial-value problem of this one-
dimensional dynamical system is given simply as

Uk = {u ∈ L4 (0, tc )| u′ ∈ L2 (0, tc ), u(0) = u0 , u′ (0) = v0 }.

The criticality condition for Πµ leads to the well-known Duffing equation

1
(5.2) ρu′′ = au(µo − u2 ) + µ(t), ∀t ∈ I, u ∈ Uk .
2
In terms of the nonlinear canonical measure ξ = Λ(u) = 12 u2 , the energy density
Wµ (ξ) and its conjugate Wµ∗ (ς) are convex functions:

1 1 2
Wµ (ξ) = a(ξ − µo )2 , Wµ∗ (ς) = ς + µo ς.
2 2a
The extended Lagrangian for this nonconvex system is
Z   Z
′ 1 2 1 2 1 2
(5.3) Ξµ (u, p, ς) = pu − ς( u − µo ) − p + ς dt + µu dt.
I 2 2ρ 2a I

The criticality condition Du Ξµ (ū, p, ς) = 0 leads to the equilibrium equation

p′ + ūς = µ ∀t ∈ I.

Clearly, the critical point ū = (µ − p′ )/ς is well-defined for any nonzero ς. Thus,
the dual feasible space can be defined as
  
 p(0) = ρv0 , −µo a ≤ ς(t) < +∞,
1
Ts = (p, ς) ∈ C (I)   .
ς(t) 6= 0 ∀t ∈ I, ς(0) = a( 21 u20 − µo )

Substituting ū = (µ − p′ )/ς into Ξdµ , the dual action is obtained as

Πdµ (p, ς) = sta Ξµ (u, p, ς)

u∈Ua
Z
1 2 (p′ − µ)2 1
(5.4) = p(tc )u(tc ) − ρv0 u0 + [ ς + µo ς + − p2 ] dt,
I 2a 2ς 2ρ
104 Gao

which is well defined on Ts . The criticality condition for Πdµ leads to the dual
Duffing system in the time domain I ⊂ R
 ′
1 ′ 1
(5.5) (p − µ) + p = 0,
ς ρ

1 1
(5.6) ς 2 ( ς + µo ) = (µ − p′ )2 .
a 2
This system consists of the so-called differential-algebraic equations (DAE’s),
which arise naturally in many applications (cf., e.g., Brenan et al, 1996;
Beardmore and Song, 1998). Although the numerical solution of these types of
systems has been the subject of intense research activity in the past few years,
the solvability of each problem depends mainly on the so-called index of the
system. Clearly, the algebraic equation (5.6) has zero solution ς = 0 if and only
if σ = (µ − p′ ) = 0. Otherwise, for any nonzero σ(t) = µ(t) − p′ (t), the algebraic
equation (5.6) has at most three real roots ςi (t) (i = 1, 2, 3), each of them leads
to the state solution ui (t) = (µ(t) − p′i (t))/ςi (t).
Theorem 5.1 (Stability and Bifurcation Criteria). For a given
parameter µo > 0, initial data (u0 , v0 ) and the input control µ(t), if at a certain
time period Is ⊂ I = (0, tc ),
 2/3
3 µ(t) − p′ (t)
(5.7) µc (t) = > µ o , t ∈ Is
2 a

then the Duffing system possesses only one solution set (ū(t), p̄(t), ς¯(t)) satisfy-
ing ς¯(t) > 0 ∀t ∈ Is , and over the period Is ,

(5.8) Πµ (ū) = min Πµ (u) iff Πdµ (p̄, ς¯) = min Πdµ (p, ς),

(5.9) Πµ (ū) = max Πµ (u) iff Πdµ (p̄, ς¯) = max min Πdµ (p, ς).
p ς

However, if at a certain time period Ib ⊂ I = (0, tc ) we have µc (t) < µo ,

then, the system possesses three sets of different solutions (ūi , p̄i (t), ς¯i (t)), i =
1, 2, 3. In the case that the three solutions ςi (t) are in the following ordering

(5.10) −aµo ≤ ς¯3 (t) ≤ ς¯2 (t) ≤ 0 ≤ ς¯1 (t) ∀t ∈ Ib ,

then over the period Ib , the solution set (ū1 (t), p̄1 (t), ς¯1 (t)) satisfies either (5.8)
or (5.9); while the solution sets (ūi (t), p̄i (t), ς¯i (t)) (i = 2, 3) satisfy

(5.11) Πµ (ūi ) = min Πµ (u) = max min Πdµ (p, ς) = Πdµ (p̄i , ς¯i ) i = 2, 3.
u p ς
 Canonical Dual Control for Nonconvex Systems 105

This theorem can be proved by combining the multi-solution theorem given

by Gao (1999, Theorem 3.4.4) and the triality theorem.
Remark. By Theorem 3.4.4 proved by the author (Gao, 1999), for any
given continuous function σ(t), if ς¯i (t) (i = 1, 2, 3) are the three solutions of the
dual Euler-Lagrange equation (5.6) in the order of (5.10), then the associated
ū1 (t) is a global minimizer of the total potential
Z Z
1 1 2 2
Pµ (u) = a( u − µo ) dt − σ(t)u dt;
I 2 2 I

while ū2 (t) is a local minimizer of Pµ and ū3 (t) is a local maximizer of Pµ .
In algebraic geometry, the dual Euler-Lagrange equation (5.6) is the so-
called singular algebraic curve in (ς, σ)-space, i.e. ς = 0 is on the curve (see
Silverman & Tate, 1992, p. 99). With a change of variables, the singular cubic
curve (5.6) can be given by the well-known Weierstrass equation

y 2 = x3 + αx2 + βx + γ,

where α, β, γ ∈ R are constants. If we let Cns be a set consisting of non-

singular points on the curve, then Cns is an Abelian group. This fact in algebraic
geometry is very important in understanding the stability of the nonconvex
dynamical systems. Actually, from Figure 3 we can see clearly that for a
given input control, if µc (t) < µo , the cubic algebraic equation (5.6) possesses
three different real solutions for ς(t). The two negative solutions ς¯(t) are the
sources that lead to the chaotic motion of the system. Thus, the inequality
(5.7) provides a bifurcation (or chaotic) criterion for the Duffing system. Fig.
3 also shows that if the continuous function σ(t) = µ(t) − p′ (t) is one-signed on
certain time interval Ib ⊂ I = (0, tc ), each root ς¯(t) of (5.6) is also one-signed
on Ii .

σ µo < µc
0.4

µo = µc
0.2
µo > µc

-1 -0.8 -0.6 -0.4 -0.2 0.2 0.4 ς

-0.2

-0.4

Fig. 3. Singular algebraic curve for the dual Duffing equation (5.6)

Theoretically speaking, for the same initial conditions, the Duffing equation
(5.2) and its dual system (5.5-5.6) should have the same solution set. Numer-
ically, the primal and dual Duffing problems give complementary bounding
106 Gao

approaches to the real solution. For the given data a = 1, µo = 1.5, u0 = 2, v0 =

1.4 and µ = 0, Figures 4 and 5 show the numerical primal (solid line) and
dual (dashed line) solutions. From the dual trajectories in the dual phase space
ς-p-p,t (Fig. 5(c-d)) we can see that at the point ς3 (t) = −aµo , if the function
σ(t) = µ(t) − p,t (t) changes its sign, the state u(t) crosses the origin goes to
another potential well in the phase space Z = U × V ∗ , and the bifurcation is
then occurred. Thus, based on the canonical dual transformation method and
theorems developed in this paper, the dual feedback control against the chaotic
vibration of the Duffing system can be suggested as the following.
1. Periodic vibration on the whole phase plane.
Choosing the controller µ(t) such that the function σ(t) = µ(t) − p′ (t)
changes its sign at the point ς¯3 (t) = −aµo .
2. Unilateral vibrations on half phase planes (either u(t) > 0 or u(t) < 0).
There are two methods: (1) choosing the controller µ(t) such that the
function σ(t) = µ(t) − p′ (t) does not change its sign at the point ς¯3 (t) = −aµo ;
(2) choosing µ(t) such that either

1/2
(5.12) µ(t) > p′ (t) + a(2µo /3)3 ∀t ∈ I,

or

1/2
(5.13) µ(t) < p′ (t) − a(2µo /3)3 ∀t ∈ I.

Detailed study on the exact controllability and stability for the Duffing
system will be given in other papers (cf. e.g., Gao, 2000d).

6 Concluding Remarks
The concept of duality is one of the most successful ideas in modern mathe-
matics and science. The inner beauty of duality theory owes much to the fact
that the nature was originally created in a duality way. By the fact that the
canonical physical variables appear always in pairs, the canonical dual trans-
formation method can be used to solve many problems in natural systems. The
associated extended Lagrange duality and triality theories have profound com-
putational impacts. For any given nonlinear problem, as long as there exists a
geometrical operator Λ such that the tri-canonical forms can be characterized
correctly, the canonical dual transformation method and the associated triality
principles can be used to establish nice theories and to develop powerful al-
ternative algorithms for robust feedback control of chaotic systems. For static
three-dimensional finite deformation problems, a general analytic solution form
and associated extremality theory have been proposed (Gao, 1999, 1999b). A
general canonical dual transformation method for solving nonsmooth global op-
timization is given recently (Gao, 2000c). In general n-dimensional distributed
parameter systems, the dual algebraic equation (5.6) will be a tensor equation
 Canonical Dual Control for Nonconvex Systems 107

(a) Primal and dual solutions (b) Primal and dual actions
3 1.5

2 1

1 0.5

0 0

−1 −0.5

−2 −1

−3 −1.5
0 10 20 30 40 −4 −2 0 2 4

(c) Trajectories in phase space u−p (d) Primal and dual actions in phase space u−p
2

2
1

0
0

−1 −2
2
5
0
0
−2
−4 −2 0 2 4 −2 −5

Fig. 4. Primal and dual solutions in primal phase space

and the stability of the nonconvex system will depend on the eigenvalues of
symmetrical canonical stress tensor field ς(x, t) (see Gao, 2000d). The triality
theory can be used for studying the controllability, observability and stability
of distributed parameter control problems.
Acknowledgement. The author would like to thank the referee for the
valuable suggestions and comments.
108 Gao

(a) Trajectores in phase space ς−p,t (b) Trajectores in ς−p space

4 2

2 1

0 0

−2 −1

−4 −2
−2 −1 0 1 2 −2 −1 0 1 2

(c) Dual solution in dual phase space p−p,t (d) Dual solution in phase space ς−p−p,t
4

5
2

0
0

−2 −5
2
2
0
0
−4
−2 −1 0 1 2 −2 −2

Fig. 5. Duffing solutions in dual phase spaces

References

[1] Antoniou, I., Basios, V. and Bosco, F. (1996). Probabilistic control of chaos: The
β-adic Renyi map under control, Int. J. Bifurcation Chaos, 6, 1563-1573.
[2] Arada, N. and Raymond, J.-P. (1999). Minimax control of parabolic systems with
state constraints, SIAM J. Control Optim.[, 38, 254-271.
[3] Auchmuty, G. (1983), Duality for non-convex variational principles, J. Diff.
Equations, vol. 50, pp 80-145
[4] Auchmuty, G. (1989), Duality algorithms for nonconvex variational principles,
Numer. Funct. Anal. and Optim., 10, pp. 211-264.
 Canonical Dual Control for Nonconvex Systems 109

[5] Auchmuty, G. (1997). Min-max problems for non-potential operator equations. in

Optimization methods in partial differential equations (South Hadley, MA, 1996),
19-28, Contemp. Math., 209, Amer. Math. Soc., Providence, RI, 1997.
[6] Barron, E.N. (1990). The pontryagin maximum principle for minimax problems
of optimal control, Nonlinear Analysis, 15, 1155-1165.
[7] Bergounioux, M., Ito, K. and Kunisch, K. (1999). Primal-dual strategy for
constrained optimal control problems, SIAM J. Control Optim., 37, 1176-1194.
[8] Chan, W.L. (1985). Duality in the optimal control of non-well-posed parabolic
equations with positive controls, J. Math. Analy. Appl., 107, 509-519.
[9] Chan, W.L. and Yung, S.P. (1987). Duality theory for the linear-convex optimal
control problem with delays, IMA J. Math. Control Inform., 4, 251-262.
[10] Chen, G., Lee, E.B., Littman, W. and Markus, L. (eds) (1991). Distributed
Parameter Control Systems, Lecture Notes in Pure and Applied Math., Marcel
Dekker, New York.
[11] Chen, G. and Dong, X. (1992). On feedback control of chaotic dynamic systems,
Int. J. Bifurcation Chaos, 2, 407-411.
[12] Chen, G. and Dong, X. (1993). ¿From chaos to order-perspectives and methodolo-
gies in controlling chaotic nonlinear dynamic systems, Int. J. Bifurcation Chaos,
3, 1363-1409.
[13] Clarke, F.H. (1985), The dual action, optimal control, and generalized gradients,
Mathematical Control Theory, Banach Center Publ., 14, PWN, Warsaw, pp. 109-
119.
[14] Dolecki, S. and Russell, D.L. (1977). A general theory of observation and control,
SIAM J. Control Opt., 15, 185-220.
[15] Ekeland, I. (1990), Convexity Methods in Hamiltonian Mechanics, Springer-
Verlag, 247pp.
[16] Ekeland, I. and Temam, R. (1976), Convex Analysis and Variational Problems,
North-Holland.
[17] Favini, A., Horn, M.A., Lasiecka, I. and Tataru, D. (1996). Global existence,
uniqueness and regularity of solutions to a von Karman system with nonlinear
boundary dissipation, Diff. Integral Equations, 9, 267-294.
[18] Fowler, T.B. (1989). Application of stochastic control techniques to chaotic
nonlinear systems, IEEE Trans. Autom. Contr., 34, 201-205.
[19] Gao, D.Y. (1995) Duality theory in nonlinear buckling analysis for von Karman
equations, Studies in Appl. Math., 94, 423-444.
[20] Gao, D.Y. (1996), Post-buckling analysis of nonlinear extended beam theory and
dual variational principles, in Applied Mechanics in Americas Vol. 4, eds. L.A.
Godoy, M. Rysz and L. Suarez, Proceedings of PACAM V, San Juan, Puerto
Rico, Iowa Univ. Press.
[21] Gao, D.Y. (1997), Dual extremum principles in finite deformation theory with
applications in post-buckling analysis of nonlinear beam model, Appl. Mech.
Reviews, ASME, 50 (11), November, S64-S71.
[22] Gao, D.Y. (1998b), Duality, triality and complementary extremum principles in
nonconvex parametric variational problems with applications, IMA J. of Applied
Math., 61, 199-235.
[23] Gao, D.Y. (1999), Duality Principles in Nonconvex Systems: Theory, Methods
and Applications, Kluwer Academic Publishers, Dordrecht /Boston / London,
472pp.
110 Gao

[24] Gao, D.Y. (1999b) General analytic solutions and complementary variational
principles for large deformation nonsmooth mechanics. Meccanica, 34, 169-198.
[25] Gao, D.Y. (2000a). Finite deformation beam models and triality theory in
dynamical post-buckling analysis. Int. J. Non-Linear Mechanics, 35, 103-131.
[26] Gao, D.Y. (2000b). Analytic solutions and triality theory for nonconvex and
nonsmooth variational problems with applications, Nonlinear Analysis, to appear.
[27] Gao, D.Y. (2000c). Canonical dual transformation method and generalized triality
theory in nonsmooth global optimization, J. Global Optimization, to appear.
[28] Gao, D.Y. (2000d). Complementary extremum principles and triality theory in
non-smooth and non-convex Hamilton mechanics, to appear in Philosophical
Transactions of the Royal Society: Mathematical, Physical and Engineering
Sciences.
[29] Gao, D.Y. (2000e). Shear control against chaotic vibration of extended nonlinear
beam model, The 8th Conf. on Nonlinear Vibration, Stability and Dynamics of
Structures, Submitted to J. Vibration and Control.
[30] Gao, D.Y. and Russell, D.L. (1994). A finite element approach to optimal control
of smart beam, in Proc. of Int. Conf. Comput. Meth. Struct., Y.K. Cheung et al
eds., Vol. 1, 135-140.
[31] Gao, D.Y. and Russell, D.L. (1996). Finite deformation extended beam theory
and nonlinear buckling analysis, in Development in Mechanics and Mathematics
of Materials, Ed. by R. Batra and M. F. Beatty, CIMN Barcelona, Spain, 430-441.
[32] Gao, D.Y. and Russell, D.L. (1996). An extended beam theory for smart materials
applications, Part I. Extended beam theory, duality theory and finite element
simulations. Appl. Math. Optimization, 34, 279-298.
[33] Gao, D.Y. and Russell, D.L. (1997). Finitely deformed thick beam theory and
dual variational extremum principles in post-buckling analysis, in Proceedings of
the 14th US Army Symposium on Solid Mechanics, Edited by K.R. Iyer and Shun-
Chin Chou, Battelle Press, Columbus/Richiand, 497-506.
[34] Gao, D.Y. and Strang, G (1989), Geometric nonlinearity: Potential energy,
complementary energy, and the gap functional, Quart. Appl. Math., XLVII(3),
487-504.
[35] Gay, David M., Overton, Michael L. and Wright, Margaret H. (1998). A primal-
dual interior method for nonconvex nonlinear programming. In Advances in
nonlinear programming (Beijing, 1996), 31–56, Appl. Optim., 14, Kluwer Acad.
Publ., Dordrecht, 1998.
[36] Ghezzi, L.L. and Piccardi, C. (1997). PID control of a chaotic system: an
application to an epidemiological model, Automatica, 33, 181-191.
[37] Horn, M.A. and Lasiecka, I. (1994). Uniform decay of weak solution to a von
Karman plate with nonlinear boundary dissipation, Diff. Integral Equations, 7,
885-908.
[38] Horn, M.A. and Lasiecka, I. (1995). Global stabilization of a dynamic von Karman
plate with nonlinear boundary feedback, Appl. Math. Optimization, 31, 57-84.
[39] Koumboulis, F.N. and Mertzios, B.G. (1996). Decoupling of sigular systems via
P-D feedback, in: Proceedings of the International Conference on Control, Vol.
96, No. 1, 19-22.
[40] Koumboulis, F.N. and Mertzios, B.G. (2000). Feedback controlling against chaos,
Chaos, Solitons and Fractals, 11, 351-358.
[41] Lagnese, J. and Lions, J. (1988). Modelling Analysis and Control of Thin Plates,
 Canonical Dual Control for Nonconvex Systems 111

Masson, Paris.
[42] Lasiecka, I. (1998a). Mathematical control theory in structural acoustic problems,
Math. Models and Methods in Applied Sciences, 8, 1119-1153.
[43] Lasiecka, I. (1998b). Uniform stabilizability of a full von Karman system with
nonlinear boundary feedback. SIAM J. Control Optim., 36 (4), 1376-1422.
[44] Lasiecka, I. (1999). Mathematical Control Theory of Coupled PDE’s. CBMS
Lectures, University of Nebraska, Lincoln.
[45] Lasiecka, I. and Triggiani, R. (eds.) (1987). Control Problems for System Described
by Partial Differential Equations and Applications. Springer-Verlag, Berlin.
[46] Lasiecka, I. and Triggiani, R. (1999). Deterministic Control Theory for Infinite
Dimensional Systems, Vol. 1 and 2, Encyclopedia of Mathematics, Cambridge
University Press.
[47] Lee, R.C.H. and Yung, S.P. (1997). Optimality conditions and duality for a non-
linear time-delay control problem, Optimal Control Appl. & Methods, 18, 327-340.
[48] Lions, J.-L. (1988). Exact controllability, stabilization and perturbations for
distributed parameter systems, SIAM Review, 30, 1-68.
[49] Mossino, J. (1975). An application of duality to distributed optimal control
problems with constraints on the control and state, J. Math. Analy. Appl., 50,
223-242.
[50] Motreanu, D. and Panagiotopoulos, P. D. (1999). Minimax Theorems and
Qualitative Properties of the Solutions of Hemivariational Inequalities. Kluwer
Academic Publishers, Dordrecht.
[51] Ogorzalek, M.J., Taming chaos - Part II: control, IEEE Trans. Circuits Syst., 40,
700-706.
[52] Ott, E., Grebogi, C. and Yorke, J.A. (1990). Controlling chaos, Phys. Rev. Lett.,
64, 1196-1199.
[53] Rockafellar, R.T. and Wets, R.J.B. (1997). Variational analysis, Springer: Berlin,
New York.
[54] Russell, D.L. (1973). A unified boundary controllability theory for hyperbolic and
parabolic partial differential equations, Stud. Appl. Math., LII, 189-211.
[55] Russell, D.L. (1978). Controllability and stabilizability theory for linear partial
differential equations: recent progress and open questions, SIAM Review, 20 (4),
639-739.
[56] Russell, D.L. (1979). Mathematics of Finite Dimensional Control Systems: Theory
and Design, Marcel Dekker, Inc., New York.
[57] Russell, D.L. (1996). Observability of linear distributed-parameter systems, The
Control Handbook, W. S. Levine eds, CRC Press and IEEE Press, 1169-1175.
[58] Strang, G. (1986), Introduction to Applied Mathematics, Wellesley-Cambridge
Press, 758pp.
[59] Tanimoto, S. (1992). Duality in the optimal control of non-well-posed distributed
systems, J. Math. Analy. Appl., 171, 277-287.
[60] Toland, J.F. (1978), Duality in nonconvex optimization, J. Mathematical Analysis
and Applications, 66, 399-415.
[61] Toland, J.F. (1979), A duality principle for non-convex optimization and the
calculus of variations, Arch. Rational Mech. Anal. 71, 41-61
[62] Wright, M. H. (1998). The interior-point revolution in constrained optimization,
in High-Performance Algorithms and Software in Nonlinear Optimization (R.
DeLeone, A. Murli, P. M. Pardalos, and G. Toraldo, eds.) 359–381, Kluwer
112 Gao

Academic Publishers, Dordrecht.

[63] Zhong, W.X. (1999). Variational method and computation for H∞ control, in
APCOM-4, vol. 2, 1037-1047.
[64] Zhong, W.X., Ouyang, H.J. and Deng, Z.C. (1993). Computational Structural
Mechanics and Optimal Control. Press of Dalian University of Technology, Dalian
(in Chinese).
[65] Zuazua, E. (1996). A uniqueness result for the linear system of elasticity and its
control theoretical consequences, SIAM J. Control and Optimization, 34, 1473-
1495.
Carleman estimate for a parabolic equation in a
Sobolev space of negative order and their
applications

O. Yu. Imanuvilov, Iowa State University, Ames IA 50011-2064 USA.

E-mail: [email protected]
M. Yamamoto, The University of Tokyo, Tokyo, Japan.
E-mail: [email protected]

Abstract

We obtain a Carleman estimate for a second order parabolic equation

when the coefficients are not bounded and the right hand side is taken in
the Sobolev space L2 (0, T ; H −1(Ω)) and we apply it to
• the global exact null-controllability of a semilinear parabolic equation
whose semilinear term contains also derivatives of first order
• conditional stability in continuation of the solution
• inverse problem of determining f ∈ H −ℓ (Ω) with ℓ < 1 at the right
hand side.

1 Introduction
In this paper we first formulate Carleman estimates for a parabolic equation in a
Sobolev spaces of negative order and apply them in order to establish the exact
null-controllability for a semilinear parabolic equation, conditional stability in
the continuation and the uniqueness in determining the source term.
A weighted estimate for a partial differential equation Ly = g, which is
called a Carleman estimate, was used by Carleman [3] for proving the unique
continuation for an elliptic equation, and since then, it has been recognized
as an important technique in the theory of partial differential equations. In
particular, the Carleman estimate is very helpful for

• the unique continuation (e.g. Hörmander [11], Isakov [16], [17] and the
references therein).

• observability inequality and exact controllability (Cheng, Isakov, Ya-

mamoto and Zhou [5], Isakov and Yamamoto [18], Kazemi and Klibanov
113
114 O. Yu. Imanuvilov and M. Yamamoto

[19], Lasiecka and Triggiani [22], Lasiecka, Triggiani and Zhang [23],
Tataru [32]).

• inverse problem of determining coefficients or non-homogeneous terms in

partial differential equations (Bukhgeim and Klibanov [2], Imanuvilov and
Yamamoto [13], Isakov [15], [17], Khaı̆darov [20] Klibanov [21], Yamamoto
[34]).

Moreover for Carleman estimates and the applications to the hyperbolic

equation, the reader can consult Lavrent’ev, Romanov and Shishat·skiı̆[24], Ruiz
[28], and for the parabolic case, Chae, Imanuvilov and Kim [4], Fursikov and
Imanuvilov [10], Imanuvilov [12], Saut and Scheurer [29], Sogge [30], Tataru
[33].
Except for Ruiz [28], those papers take L2 -spaces as the space of g = Ly
(the right hand side of the partial differential equation under consideration),
and such L2 -spaces make us assume more regularity in the applications. On
the other hand, by our Carleman estimate, we can reduce the regularity
assumptions on the right hand side of the parabolic equation. Because of the
page limitation, however, we will omit the proofs of the Carleman estimate and
the exact controllability (see Imanuvilov and Yamamoto [14]), and concentrate
on the conditional stability in the continuation and the inverse problem
This paper is composed of five sections:
Section 2. Carleman estimate
Section 3. Application to the exact null-controllability
Section 4. Application to the continuation problem
Section 5. Application to the inverse problem.

Acknowledgements
The authors thank the referee for useful comments. The second named author
is partly supported by Sanwa Systems Developement Co. Ltd. (Tokyo, Japan).

2 Carleman estimate
Let (t, x) ∈ Q ≡ (0, T ) × Ω, Σ ≡ (0, T ) × ∂Ω, where Ω ⊂ Rn is a connected
bounded domain whose boundary ∂Ω is of class C 2 , ν(x) is the outward unit
∂ ∂
normal to ∂Ω, T ∈ (0, +∞) is an arbitrary moment of time, ∂t = ∂t , ∂i = ∂x i
,
1 ≤ i ≤ n, Dβ = D β0 Dβ = ∂tβ0 ∂1β1 . . . ∂nβn , β = (β0 , β ′ ) = (β0 , β1 , . . . , βn ),
′

|β| = 2β0 + β1 + · · · + βn , ∇ = (∂1 , ...., ∂n ). Let ω ⊂ Ω be an arbitrarily

fixed subdomain and let us set Qω = (0, T ) × ω. Throughout this paper,
Wpµ (Ω) = W µ,p (Ω), W2µ (Ω) = H µ (Ω), W0µ,p (Ω), p ≥ 1, µ ≥ 0, H0µ (Ω), µ ≥ 0,
denote usual Sobolev spaces (e.g. Adams [1]), and we set L2 (Ω) = W20 (Ω).
′
Moreover Wp−µ (Ω) = (W0µ,p (Ω))′ , H −µ (Ω) = (H0µ (Ω))′ : the dual, where
1 1
p + p′ = 1.
 Carleman estimate and applications 115

Let us consider a parabolic equation

n
X n
X
(2.1) Ly ≡ ∂t y − ∂i (aij (x)∂j y) + ∂i (bi (t, x)y) + c(t, x)y = g in Q,
i,j=1 i=1

with the boundary condition


(2.2) y Σ = 0.
Assume that
 1

 aij ∈ W∞ (Ω), aij = aji ,
 1 ≤ i, j ≤ n, 

∞ r
bi ∈ L (0, T ; L (Ω)), r > 2n, 1 ≤ i ≤ n,
(2.3) n o

 c ∈ L∞ (0, T ; Wr−µ (Ω)), 0 ≤ µ < 1 , r1 > max 
1 2
2n
3−2µ , 1 , 

and the coefficients aij satisfy the uniform ellipticity: There exists β > 0 such
that
Xn
(2.4) aij (x)ζi ζj ≥ β|ζ|2 , ζ = (ζ1 , ...., ζn ) ∈ Rn , (t, x) ∈ Q.
i,j=1

For the weak solution, we can show

Lemma 2.1. Let y0 ∈ L2 (Ω), g ∈ L2 (0, T ; H −1 (Ω)) and let us assume the
conditions (2.3) and (2.4). Then there exists a solution y ∈ L2 (0, T ; H01 (Ω))
to (2.1) and (2.2) with y(0, ·) = y0 . Moreover the solution is unique in L2 (Q),
and we have an estimate:
(2.5) kykL2 (0,T ;H01 (Ω))∩C([0,T ];L2 (Ω)) ≤ C0 (ky(0, ·)kL2 (Ω) + kgkL2 (0,T ;H −1 (Ω)) ),
where the constant C0 > 0 depends continuously only on
n
X n
X
kaij kW∞
1 (Ω) + kbi kL∞ (0,T ;Lr (Ω)) + kckL∞ (0,T ;Wr−µ (Ω)) .
1
i,j=1 i=1

This lemma can be proved by a usual energy method and for completeness we
will give a sketch of the proof in Appendix A.
In view of Lemma 2.1, in the succeeding arguments, we can assume the
smoothness of solutions which admit calculations such as integration by parts.
More precisely, we can use a usual density argument, i.e., we can do everything
for sufficiently smooth solutions, and then pass to the limit in the final
inequality.
In order to formulate our Carleman estimate, we need a special weight
function.
Lemma 2.2 ([4], [12]). Let ω0 ⊂ ω be an arbitrarily fixed subdomain of Ω
such that ω0 ⊂ ω. Then there exists a function ψ ∈ C 2 (Ω) such that
(2.6) ψ(x) > 0, x ∈ Ω, ψ|∂Ω = 0, |∇ψ(x)| > 0, x ∈ Ω \ ω0 .
116 O. Yu. Imanuvilov and M. Yamamoto

Now, using the function ψ constructed in Lemma 2.2, we introduce weight

functions:
exp(λψ(x))−exp(2λkψkC(Ω) )
ϕ(t, x) = exp(λψ(x))
t(T −t) , α(t, x) = t(T −t) ,
(2.7) 1
ϕ0 (t) = t(T −t)

where λ > 0 is a parameter.

We are ready to state a Carleman estimate in a Sobolev space of negative
order:
Theorem 2.1. Let (2.3) - (2.4) be fulfilled and Pnthe functions ϕ, α be
defined by (2.7). Moreover let g(t, x) = g0 (t, x) + i=1 ∂i gi (t, x) with g0 ∈
L2 (0, T ; H −1 (Ω)) and gi ∈ L2 (Q), 1 ≤ i ≤ n. Then there exists a number
b > 0 such that for an arbitrary λ ≥ λ,
λ b we can choose s0 (λ) > 0 satisfying:
there exists a constant C1 > 0 such that for each s ≥ s0 (λ) and any y ∈ L2 (Q)
satisfying (2.1) and (2.2), we have
Z  
1
(2.8) 2 2
|∇y| + sϕy e2sα dx dt
Q sϕ
n Z !
X
≤ C1 kg0 esα k2L2 (0,T ;H −1 (Ω)) + kgi esα k2L2 (Q) + sϕy 2 e2sα dxdt ,
i=1 Qω

where the constant C1 > 0 is dependent continuously on aij , bi , 1 ≤ i, j ≤ n, c,

λ, and independent of s. Moreover, if bi = 0, 1 ≤ i ≤ n, and y|Qω = 0, then we
have
Z  
1 − 12 sα 2 1
(2.9) kϕ ∂t (ye )kL2 (0,T ;H −1 (Ω)) + 2 2
|∇y| + sϕ0 y e2sα dx dt
s 0 Q sϕ 0

≤ C2 kgesα k2L2 (0,T ;H −1 (Ω)) ∀ s ≥ s0 (λ),

where the constant C2 > 0 is dependent continuously on aij , 1 ≤ i, j ≤ n, c, λ,

and independent of s.
Proof. For the Carleman estimate (2.8), we refer to [14]. We will prove the
second Carleman estimate (2.9). It is sufficient to prove (2.9) in the case of
c = 0. Indeed,

Lemma 2.3. Let c satisfy (2.3). Then we can take δ ∈ 0, 21 such that

1 1
kcyesα kL2 (0,T ;H −1 (Ω)) ≤ C3 s−δ (k(sϕ)− 2 (∇y)esα kL2 (Q) + k(sϕ) 2 yesα kL2 (Q) )

for y ∈ L2 (0, T ; H01 (Ω)) and large s > 0. Here a constant C3 > 0 is independent
of y, s, but dependent on aij , c, Ω, T .
The proof of the lemma is given in Appendix B. By taking s > 0 sufficiently
large, Lemma 2.3 and the Carleman estimate (2.9) in the case of c = 0, yield
(2.9) with c satisfying (2.3).
 Carleman estimate and applications 117

Now we proceed to the proof of (2.9) in the case of c = 0. Instead of the

b
operator L, it suffices to prove (2.9) for the operator L:
n
X
b = ∂t y −
Ly aij (x)∂i ∂j y.
i,j=1

Denote Lb s = esα Le−sα , z = yesα . We notice that the operator L

b s can be
written explicitly as follows
n
X n
X
b s z =∂t z −
L aij ∂i ∂j z + 2sλϕ aij (∂i ψ)∂j z + sλ2 ϕa(x, ∇ψ, ∇ψ)z
i,j=1 i,j=1
n
X
2 2 2
−s λ ϕ a(x, ∇ψ, ∇ψ)z + sλϕz aij ∂i ∂j ψ − s(∂t α)z,
i,j=1

P
where a(x, ζ, ζ) = ni,j=1 aij ζi ζj for ζ = (ζ1 , ..., ζn ).
We note that the function z is the solution to the initial value problem

(2.10) bs z = gesα in Q,
L z|Σ = 0, z(0, ·) = 0.

Using the partition of unity and a standard argument (see e.g., [11, p. 191]),
one can reduce the proof of the estimate (2.9) to the case when

Ω ⊂ {x; |x| < δ}

where the parameter δ can be chosen arbitrarily small.

For each t ∈ [0, T ], let w(t, x) be the solution to the boundary value problem:
n
X
− ∂i (aij (x)∂j w(t, x)) = z(t, x), x ∈ Ω, w|∂Ω = 0.
i,j=1

sw
Taking the scalar product of (2.10) with the function t(T −t) , after integration
by parts and a priori estimates, we can obtain
3
s3 kϕ 2 yesα k2L2 (0,T ;H −1 (Ω))
1 1 1
(2.11) ≤ C(skϕ 2 zk2L2 (Q) + kϕ− 2 ∇zk2L2 (Q) + kgesα k2L2 (0,T ;H −1 (Ω)) ).
s

Therefore, by (2.8) and (2.11), we have

3
(2.12) s3 kϕ 2 yesα k2L2 (0,T ;H −1 (Ω)) ≤ Ckgesα k2L2 (0,T ;H −1 (Ω)) .
118 O. Yu. Imanuvilov and M. Yamamoto

Next we have
n
X n
X
−1 − 12 −1
ϕ0 2 ∂t z = aij (∂i ∂j z)ϕ0 − 2sλϕ0 2 ϕ aij (∂i ψ)(∂j z)
i,j=1 i,j=1
−1 −1
−sλ2 ϕ0 2 ϕa(x, ∇ψ, ∇ψ)z + s2 λ2 ϕ0 2 ϕ2 a(x, ∇ψ, ∇ψ)z
Xn
−1 −1 −1
−sλϕ0 2 ϕz aij ∂i ∂j ψ + s(∂t α)ϕ0 2 z + gesα ϕ0 2 .
i,j=1

Therefore
1 − 12
(2.13) kϕ ∂t zk2L2 (0,T ;H −1 (Ω))
s 0
n
C X −1
≤ kaij ϕ0 2 (∂i ∂j z)k2L2 (0,T ;H −1 (Ω))
s
i,j=1
n
X 1 1
+Cs kaij (ϕϕ−1 2
0 ) ϕ (∂i ψ)∂j zkL2 (0,T ;H −1 (Ω))
2 2

i,j=1
1 3 −1
+Cs3 k(ϕϕ−1 2
0 ) ϕ a(x, ∇ψ, ∇ψ)zkL2 (0,T ;H −1 (Ω)) + Cskϕ0 ϕzkL2 (Q)
2 2 2 2

1
n o 3
2λ(kψkC(Ω) −ψ(x))
+Csk(2t − T )(ϕϕ−1
0 ) 2 e
−λψ(x)
− e ϕ 2 zk2L2 (0,T ;H −1 (Ω))
1 −1
+ kgesα ϕ0 2 k2L2 (0,T ;H −1 (Ω)) .
s
Noting that

(2.14) kqwkL2 (0,T ;H −1 (Ω)) ≤ CkwkL2 (0,T ;H −1 (Ω))

for q ∈ L∞ (0, T ; W∞
1 (Ω)), we have

[the third, the fifth and the sixth terms at the right hand side of (2.13)]
3
≤Cs3 kϕ 2 zk2L2 (0,T ;H −1 (Ω)) + Ckgesα k2L2 (0,T ;H −1 (Ω)) .

Moreover, by (2.3) and (2.14), we see

−1 −1
kaij ϕ0 2 (t)(∂i ∂j z)(t, ·)kH −1 (Ω) ≤ Ckϕ0 2 (t)(∂i ∂j z)(t, ·)kH −1 (Ω)
Z 
− 21  
≤Cϕ0 (t) sup  
 (∂j z)(t, x)∂i µ(x)dx
µ∈H01 (Ω),k∇µkL2 (Ω) =1 Ω

− 21
≤Cϕ0 (t)k(∂j z)(t, ·)kL2 (Ω) ,

so that we have
Z
C 1
[the first term at the right hand side of (2.13)] ≤ |∇z|2 dxdt.
s Q ϕ0
 Carleman estimate and applications 119

Finally we have
1 1 1
kaij (∂i ψ)(ϕϕ−1
0 ) (t, ·)ϕ(t, ·) (∂j z)(t, ·)kH −1 (Ω) ≤ kϕ(t, ·) (∂j z)(t, ·)kH −1 (Ω)
2 2 2
Z 
 
≤C sup  ϕ 12 (t, x)(∂j z)(t, x)µ(x)dx
1
 
µ∈H0 (Ω),k∇µkL2 (Ω) =1 Ω
1
≤Ckϕ (t, ·)z(t, ·)kL2 (Ω) ,
2

so that
Z
[the second term at the right hand side of (2.13)] ≤ C sϕ0 |z|2 dxdt.
Q

Hence (2.8) and (2.12) yield

1 − 12
kϕ ∂t zk2L2 (0,T ;H −1 (Ω)) ≤ Ckgesα k2L2 (0,T ;H −1 (Ω)) .
s 0
Thus the proof of (2.9) is complete.

3 Exact null-controllability of semilinear parabolic equations

Henceforth aij , bi , 1 ≤ i, j ≤ n and c are assumed to satisfy (2.3) and we
consider the semilinear parabolic equation
n
X n
X
(3.1) G(y) = ∂t y − ∂i (aij (x)∂j y) + bi (t, x)∂i y(t, x)
i,j=1 i=1
+c(t, x)y + f (t, x, ∇y, y)
= u+g in Q with u ∈ U(ω),

and

(3.2) y Σ = 0, y(0, x) = v0 (x), x ∈ Ω,

where v0 and g are given, and u(t, x) is a locally distributed control in the space

(3.3) U(ω) = {u(t, x) ∈ L2 (Q); supp u ⊂ Qω }.

By the exact null-controllability we mean a problem of finding a control

u ∈ U(ω) such that

(3.4) y(T, x) = 0, x ∈ Ω.

For a semilinear term f , let us assume that

(3.5) f (t, x, ζ ′ , ζ0 ) ∈ C 1 (Q × Rn+1 ), f (t, x, 0, 0) = 0, (t, x) ∈ Q

120 O. Yu. Imanuvilov and M. Yamamoto

and
(3.6)
 
 ∂f (t, x, ζ ′ , ζ0 ) 


 ≤ K,
 (t, x) ∈ Q, ζ ≡ (ζ ′ , ζ0 ) = (ζ1 , . . . , ζn , ζ0 ) ∈ Rn+1
∂ζi
for 0 ≤ i ≤ n with some constant K > 0. Set

−eλψ(x) + e2λkψkC(Ω)
(3.7) η(t, x) = ,
(T − t)ℓ(t)
where
T
(3.8) ℓ ∈ C ∞ [0, T ], ℓ(t) > 0, ℓ(t) ≥ t, t ∈ [0, T ], ℓ(t) = t, t ∈ [ , T ].
2
We set
( Z 1 )
2
L2 (Q, ρ) = y = y(t, x); kykL2 (Q,ρ) ≡ y 2 ρdxdt <∞
Q

for ρ = ρ(t, x) ≥ 0 and we recall that L is defined by (2.1).

To formulate our results we introduce the function spaces

(3.9) Xsλ (Q) = L2 (Q, (T − t)e2sη ),

(3.10)
( )
Zsλ (Q) = y = y(t, x); 2
y|Σ = 0, y, ∇y ∈ L (Q, e 2sη
), Ly ∈ Xsλ (Q)

with the norm

(3.11) kyk2Z λ (Q) = kLyk2Xsλ (Q) + kyk2L2 (Q,e2sη ) + k∇yk2L2 (Q,e2sη ) ,

s

and

(3.12) Y (Q) = {y(t, x); Ly ∈ L2 (0, T ; L2 (Ω)), y|Σ = 0, y(0, ·) ∈ H 1 (Ω)}

with the norm

(3.13) kyk2Y (Q) = kLyk2L2 (0,T ;L2 (Ω)) + ky(0, ·)k2H 1 (Ω) .

We are ready to state

Theorem 3.1. Let v0 ∈ H01 (Ω), and let the conditions (2.3)–(2.4), (3.5)
b > 0 such that for λ ≥ λ
and (3.6) be fulfilled. Then there exists λ b there exists
λ b
a constant s0 (λ) so that for g ∈ Xs (Q) with λ ≥ λ and s ≥ s0 (λ), there exists
a solution pair (y, u) ∈ Y (Q) × U(ω) to (3.1), (3.2) and (3.4).
 Carleman estimate and applications 121

As for the proof, we refer to Imanuvilov and Yamamoto [14]. This is the exact
null-controllability for a parabolic equation whose semilinear term depends also
on ∇y. The main achievement by our Carleman estimate (Theorem 2.1) is that
we can include the first order derivatives in the semilinear term.
For the approximate controllability for a parabolic equation with semilinear
term including ∇y, see Fernández and Zuazua [7]. We further refer to Fabre,
Puel and Zuazua [6] and Fernández-Cara [8].

4 Conditional stability in the continuation

Let L be defined by (2.1) and let (2.3)–(2.4) hold. Let ω ⊂ Ω be an arbitrary
subdomain. In this section, we discuss conditional stability in continuation of
solutions to a parabolic equation. By the uniqueness in the continuation, we
mean that y|(0,T )×ω = 0 implies y|Q = 0.
We can refer to Isakov [15]–[17], Lavrent’ev, Romanov and Shishat·skiı̆[24],
Saut and Scheurer [29], Sogge [30]. In fact, for the uniqueness, we can have
even sharper results: Lin [25], Poon [27]. The continuation problem is known
to be ill-posed. In other words, we cannot expect the continuity of the map
y|(0,T )×ω −→ y|Q . However, under a priori boundedness assumptions for y in
Q, we can restore stability, which is called conditional stability. In this section,
we will prove conditional stability by our Carleman estimate (Theorem 2.1).
For the proof, we follow a scheme: ”Carleman estimate” −→ ”conditional
stability”, which has been used by Hörmander [11], Isakov [15]. Thus the new
ingredient in this section is the Carleman estimate in a Sobolev space of negative
order.
In comparison with conditional stability (e.g. [11], [15]) shown by traditional
Carleman estimates, the advantages of our result are:

• less regularity of coefficients

• weaker norms of the right hand side,

thanks to our Carleman estimate.

Let us recall that ψ is defined in Lemma 2.2. We set

(4.1) Ωδ = {x ∈ Ω; ψ(x) > δ}

for sufficiently small δ > 0. Then we note that Ω0 = Ω by (2.6). We assume

(4.2) ω ⊂ Ω4δ .

This is true if δ > 0 is sufficiently small.

Theorem 4.1. Let (2.3)–(2.4) be fulfilled and y ∈ L2 (Q) satisfy

(4.3) Ly = g ∈ L2 (0, T ; H −1 (Ω)).

122 O. Yu. Imanuvilov and M. Yamamoto

Then for a given κ > 0, there exists θ ∈ (0, 1) depending on Ω, T , ω, κ, δ such

that

(4.4) kykL2 (2κ,T −2κ;H 1 (Ωδ )) ≤ C(kgkL2 (κ,T −κ;H −1(Ω)) + kykL2 (Qω ) )θ kyk1−θ
L2 (Q)
.

Here limδ→0 θ = 0 and limκ→0 θ = 0.

This estimate asserts the stability under a condition that kykL2 (Q) is bounded,
and becomes trivial as δ → 0 or κ → 0.
Proof. Let us take the cut off function χ ∈ C0∞ (Ω), 0 ≤ χ ≤ 1 on Ω such that
(
1, x ∈ Ω2δ
(4.5) χ(t, x) =
0, x ∈ Ω \ Ωδ .

We note that χ = 0 in a neighbourhood of ∂Ω. We further set

(4.6) v = χy.

Then v|∂Ω = 0 and

n
X n
X
Lv =χLy − aij (∂j y)∂i χ − ∂i (aij y)∂j χ
i,j=1 i,j=1
n
X n
X
− aij (∂i ∂j χ)y + bi y∂i χ in Q.
i,j=1 i=1

Therefore, noting that aij = aji , we have

  
 Xn n
X 
Lv =χg +  (∂j aij )∂i χ + aij ∂i ∂j χ y − 2 ∂j (aij y∂i χ)
 
i,j=1 i,j=1
(4.7)
n
X
+ bi y∂i χ in Q.
i=1

We can assume that T ≤ 1. For 0 < t0 < t1 < T , we set

(4.8) Qt0 t1 = (t0 , t1 ) × Ω,

exp(λψ(x))
ϕ = ϕt0 t1 (t, x) = ,
(t − t0 )(t1 − t)
(4.9)
exp(λψ(x)) − exp(2λkψkC(Ω) )
α = αt0 t1 (t, x) = , (t, x) ∈ Qt0 t1 .
(t − t0 )(t1 − t)
 Carleman estimate and applications 123

Therefore, applying (2.8) in Theorem 2.1 to (4.7), we have

Z
((sϕ)−1 |∇v|2 + sϕv 2 )e2sα dx dt
Q t0 t1

≤C1 kχgesα k2L2 (t0 ,t1 ;H −1 (Ω))

n
X
+C1 k((∂j aij )∂i χ + aij ∂i ∂j χ)yesα k2L2 (t0 ,t1 ;H −1 (Ω))
i,j=1
n
X n
X
+C1 kbi (∂i χ)yesα k2L2 (t0 ,t1 ;H −1 (Ω)) + C1 kaij y(∂i χ)esα k2L2 (Qt
0 t1 )
i=1 i,j=1
Z
+C1 sϕv 2 e2sα dxdt.
Qω

Here we note that C1 is independent of t0 and t1 , but dependent on t1 − t0 .

Consequently, by (4.2) and (4.5), we have
Z t1 Z
(4.10) ((sϕ)−1 |∇y|2 + sϕy 2 )e2sα dx dt
t0 Ω2δ
≤ C1 kgesα k2L2 (0,T ;H −1 (Ω))
n
X
+C1 k((∂j aij )∂i χ + aij ∂i ∂j χ)yesα k2L2 (t0 ,t1 ;L2 (Ωδ \Ω2δ ))
i,j=1
X n
+C1 kbi (∂i χ)yesα k2L2 (t0 ,t1 ;H −1 (Ω))
i=1
X n Z
+C1 kaij y(∂i χ)esα k2L2 (t0 ,t1 ;L2 (Ωδ \Ω2δ )) + C1 sϕy 2 e2sα dxdt.
i,j=1 Qω

On the other hand, in terms of bi ∈ L∞ (0, T ; Lr (Ω)) with r > 2n, we can rewrite
(4.10) as follows.
Z 
 
sα
kbi (t, ·)ye ∂i χkH −1 (Ω) ≤ sup  bi (t, x)ye (∂i χ)µdx
sα
 
µ∈H01 (Ω),k∇µkL2 (Ω) =1 Ω

≤ sup kbi (t, ·)kLn (Ω) kyesα ∂i χkL2 (Ω) kµk 2n

µ∈H01 (Ω),k∇µkL2 (Ω) =1 L n−2 (Ω)

≤Ckbi (t, ·)kLr (Ω) kyesα ∂i χkL2 (Ω) ≤ Ckbi (t, ·)kLr (Ω) kyesα kL2 (Ωδ \Ω2δ )

by the Hölder inequality, r > 2n and the Sobolev embedding. Moreover we

directly see that

sup ksϕt0 t1 e2sαt0 t1 kW∞

1 (Ω) ≤ C(λ).
t0 <t<t1 ,s>0
124 O. Yu. Imanuvilov and M. Yamamoto

We note that C(λ) is independent of t0 and t1 . Furthermore we can directly

see that
Z 
 
sα
kge (t, ·)kH −1 (Ω) ≤ sup  sα(t,x)
µ(x)dx
 g(t, x)e
µ∈H01 (Ω),k∇µkL2 (Ω)=1 Ω

≤kg(t, ·)kH −1 (Ω) kesα(t,·) µkH01 (Ω) ≤ C3 (λ)kg(t, ·)kH −1 (Ω) .

Here C3 (λ) > 0 is independent of s > 0. Therefore from (4.10) and (2.3), we
obtain
Z t1 Z
((sϕ)−1 |∇y|2 + sϕy 2 )e2sα dx dt
t0 Ω2δ
Z
(4.11) ≤ C4 kgk2L2 (t0 ,t1 ;H −1 (Ω)) + C4 y 2 dxdt + C4 kyesα k2L2 (t0 ,t1 ;L2 (Ωδ \Ω2δ )) .
Qω

We set
n o
2λkψkC(Ω) 2λkψkC(Ω)
(4.12) h(r) = e3δλ − e − (1 + 4r 2 − 4r) e2δλ − e

for 0 ≤ r ≤ 21 . Then h(0) = exp(3δλ) − exp(2δλ) > 0 and we can take

r = r(λ, δ) > 0 sufficiently small so that 0 < r < 12 and

(4.13) ε ≡ h(r) > 0.

Here we notice that λ > 0 and δ > 0 are fixed, and r can be determined
independently of t0 and t1 , so that ε > 0 is independent of t0 and t1 and
dependent on λ, δ and r. Let
 
t0 + t1 t0 + t1
t∈ − r(t1 − t0 ), + r(t1 − t0 ) ≡ It0 t1 .
2 2

Then, since r > 0 is sufficiently small, It0 t1 ⊂ (t0 , t1 ), and, by (4.13), we can
verify
 
 exp(3δλ) − exp(2λkψkC(Ω) ) 

 ≤ αt0 t1 (t, x), 


 1

2 

 2 − r (t 1 − t 0 )2 
(4.14)

 


 exp(3δλ) 


 1 ≤ ϕ t t
0 1 (t, x), t ∈ I t t
0 1 , x ∈ Ω 3δ 

(t − t )2
4 1 0

and
(4.15)
exp(2δλ) − exp(2λkψkC(Ω) )
αt0 t1 (t, x) ≤ 1 , t0 < t < t1 , x ∈ Ωδ \ Ω2δ .
4 (t1 − t0 )2
 Carleman estimate and applications 125

2 1
In fact, for t ∈ It0 t1 , we have (t−t0 )(t1 −t) ≥ 12 − r (t1 −t0 )2 , and (t−t0 )(t1 −t)
≤
1
2 . Therefore, noting that exp(λψ(x)) − exp(2λkψkC(Ω) ) < 0, we
( 12 −r) (t1 −t0 )2
obtain
exp(λψ(x)) − exp(2λkψkC(Ω) )
αt0 t1 (t, x) ≡
(t − t0 )(t1 − t)
exp(λψ(x)) − exp(2λkψkC(Ω) )
≥
2 , t ∈ It0 t1 .
1 2
2 − r (t1 − t0 )

Since x ∈ Ω3δ implies that ψ(x) > 3δ, we see the first inequality in (4.14). The
second inequality in (4.14) is straightforward. For (4.15) we have

1
(4.16) (t1 − t0 )2 ≥ (t − t0 )(t1 − t)
4
for t0 ≤ t ≤ t1 and

exp(λψ(x)) − exp(2λkψkC(Ω) )
αt0 t1 (t, x) ≤ 1
4 (t1 − t0 )2

by exp(λψ(x)) − exp(2λkψkC(Ω) ) < 0. Since x ∈ Ωδ \ Ω2δ , we obtain ψ(x) ≤ 2δ,

which implies (4.15).
Therefore by means of (4.11), (4.14) and (4.15), we have
! Z
s exp(3δλ) exp(3δλ) − exp(2λkψkC(Ω) )
1 2
exp 2s
2 × y 2 dxdt
(t − t ) 1 2
2 − r (t1 − t0 )
4 1 0 It0 t1 ×Ω3δ

≤C4 kgk2L2 (t0 ,t1 ;H −1 (Ω)) + C4 kyk2L2 (Qω )

!
exp(2δλ) − exp(2λkψkC(Ω) )
+C4 exp 2s 1 kyk2L2 (t0 ,t1 ;L2 (Ω)) .
4 (t1 − t0 )2

Consequently, taking s ≥ 1 and noting that

1 1 1
(t1 − t0 )2 ≤ T 2,
4s exp(3δλ) 4

we obtain
Z !
2 C4 T 2 exp(2λkψkC(Ω) ) − exp(3δλ)
y dxdt ≤ exp 2s
2 F
4 1 2
2 − r (t1 − t0 )
It0 t1 ×Ω3δ
!
C4 T 2 −2sε
+ exp
2 M.
4 1
− r (t1 − t0 )2
2
126 O. Yu. Imanuvilov and M. Yamamoto

Here and henceforth we set

F = kgk2L2 (t0 ,t1 ;H −1 (Ω)) + kyk2L2 (Qω ) , M = kyk2L2 (Q) ,
and C > 0 denotes a generic constant which dependent on λ > 0 and δ > 0,
but independent of s > 0.
Taking κ > 0 sufficiently small, we assume
(4.17) |t1 − t0 | ≥ κ > 0.
Therefore we see
Z
(4.18) y 2 dxdt ≤ C(e2sC F + e−2sCε M ).
It0 t1 ×Ω3δ

We can choose s > 0 so that

 
1 M
(4.19) s = max log , s0 (λ) + 1 ,
2C(1 + ε) F
in order that the Carleman estimate (2.8) holds. Thus we obtain
Z
1 ε
(4.20) y 2 dxdt ≤ CM 1+ε F 1+ε .
It0 t1 ×Ω3δ

Now we will complete the proof. Since (4.20) holds true provided
that
κ 3κ
|t1 − t0 | ≥ κ, we can apply (4.20) in the time interval (t0 , t1 ) = 2 , 2 , so that
Z (1+r)κ Z 2 ε
y 2 dxdt ≤ CkykL1+ε F
2 (Q) 0
1+ε
.
(1−r)κ Ω3δ

Here we set F0 = kgk2L2 κ ,T − κ ;H −1 (Ω) + kyk2L2 (Qω ) . Similarly for any t ∈

(2 2 )
(κ, T − κ), the estimate (4.20) yields
Z Z 2 ε
y 2 dxdt ≤ CkykL1+ε
2 (Q) F0
1+ε

It Ω3δ

where It is an interval including t with the length ≥ 2rκ. Since a finite number
of such intervals It cover (κ, T − κ), we obtain
Z T −κ Z 2 ε
(4.21) y 2 dxdt ≤ CkykL1+ε
2 (Q) F0
1+ε
.
κ Ω3δ

Now we will prove the estimate for ∇y. Set

Q = Qκ,T −κ = (κ, T − κ) × Ω,
exp(λψ(x))
ϕ = ϕκ,T −κ (t, x) = ,
(t − κ)(T − κ − t)
(4.22)
exp(λψ(x)) − exp(2λkψkC(Ω) )
α = ακ,T −κ (t, x) = , (t, x) ∈ Qκ,T −κ .
(t − κ)(T − κ − t)
 Carleman estimate and applications 127

Take

1, x ∈ Ω4δ
χ1 (t, x) =
0, x ∈ Ω \ Ω3δ ,

in place of χ defined by (4.5). Then, similarly to (4.11), by |e2sα | ≤ 1, we obtain

Z T −κ Z  Z 
−1 2 2sα
(sϕ) |∇y| e dxdt ≤ C kgk2L2 (κ,T −κ;H −1(Ω)) + 2
y dxdt
κ Ω4δ Qω
sα 2
+Ckye kL2 (κ,T −κ;L2(Ω3δ \Ω4δ )) ≤ CF0 + Ckyk2L2 (κ,T −κ;L2 (Ω3δ )) .

Hence (4.21) yields

Z T −2κ Z 2 ε
(sϕ)−1 |∇y|2 e2sα dxdt ≤ CF0 + CkykL1+ε
2 (Q) F0
1+ε
.
2κ Ω4δ

Then we fix such large s > 0. Since

(sϕ)−1 e2sα
!
1 exp(λψ(x)) − exp(2λkψkC(Ω) )
= (t − κ)(T − κ − t) exp(−λψ(x)) exp 2s >0
s (t − κ)(T − κ − t)

for (t, x) ∈ [2κ, T − 2κ] × Ω4δ , the proof is complete.

5 Inverse Source Problem

In this section, we will consider a parabolic equation with a source term:

(5.1) ∂t u + Au = f (x)R(t, x) in Q

(5.2) u|Σ = 0.

For simplicity we consider only a t-independent elliptic operator A

n
X
(5.3) (Au)(x) = − ∂i (aij (x)∂j u) + c(x)u,
i,j=1

with the homogeneous Dirichlet boundary condition, where c ∈ C ∞ (Ω), ≥ 0,

aij = aji ∈ C ∞ (Ω), 1 ≤ i, j ≤ n and there exists β > 0 such that
P n 2 n
i,j=1 aij (x)ζi ζj ≥ β|ζ| for ζ = (ζ1 , ...., ζn ) ∈ R and x ∈ Ω. We note that the
condition c ≥ 0 is not essential, because we choose large M > 0 and can consider
eM t u in place of u, if necessary. Let {ei (x)}∞ ∞
i=1 and {λi }i=1 be the sequences
128 O. Yu. Imanuvilov and M. Yamamoto

of eigenvectors and eigenvalues of the operator A. In order to formulate our

results, we introduce the following spaces
( ∞ ∞
)
X X
Xα = u = ai ei ∈ L2 (Ω); λαi a2i < ∞
i=1 i=1

and we set
∞
!1
X 2

kukX α = λαi a2i .

i=1

We note that X0 = L2 (Ω), X2

= ∩ H01 (Ω) ⊃ H02 (Ω), X −2 ⊂ H −2 (Ω)
H 2 (Ω)
and X −ℓ = H −ℓ (Ω) for 0 ≤ ℓ < 23 (e.g. Fujiwara [9]).
Let us recall that ω ⊂ Ω is an arbitrarily fixed subdomain. Assuming that
R = R(t, x) is given, we discuss
Inverse Source Problem. Let θ ∈ (0, T ) and 0 ≤ ℓ < 1 be arbitrarily fixed.
Determine f ∈ H −ℓ (Ω) and u(0, ·) ∈ X −2 (Ω) from
(5.4) u(θ, x), x∈Ω
and
(5.5) u|(0,T )×ω .
This inverse problem is closely related to the determination of the coefficient p
in ∂t u + Au = p(x)u. In fact, let ∂t v + Av = q(x)v. Setting y = u − v, f = p − q
and R = v, we have ∂t y + Ay = p(x)y + f (x)R(t, x) in Q.
In this paper, in order to concisely show the essence in applying our
Carleman estimate to the inverse problem, we assume that R is smooth:
(5.6) ∂t R, R ∈ C 0,1 (Q).

Here we set C 0,1 (Q) = {y; y, ∇y ∈ C(Q)}. Although we can establish the
stability in our inverse problem, we concentrate on the uniqueness for the
conciseness. Our main result is stated as follows:
Theorem 5.1. We assume (5.6) and
(5.7) R(θ, ·) > 0 on Ω.
Let u ∈ L2 (0, T ; X −2 ) be the weak solution to (5.1), (5.2) and u(0, ·) ∈ X −2
with f ∈ H −ℓ (Ω) where ℓ < 1. If u(θ, ·) = 0 in Ω and u = 0 in Qω , then
u(0, ·) = 0 and f = 0 in Q.
As f , we can consider δS which is a delta function concentrated R on an
(n − 1)-dimensional smooth hypersurface S ⊂ Ω: < δS , h >= S h(x)dSx for
h ∈ C0∞ (Ω).
Remark 5.1. For u(0, ·) ∈ X −2 and f ∈ H −ℓ (Ω), we can prove that there
exists a unique solution u ∈ L2 (0, T ; X −2 ) to (5.1) and (5.2) (e.g. Tanabe [31]).
 Carleman estimate and applications 129

Remark 5.2. In ususal Carleman estimates within L2 -spaces, we need the

regularity y, ∂t y, ∂i y, ∂i ∂j y ∈ L2 (Q). However, for f ∈ H −ℓ (Ω), 0 < ℓ < 1, we
have A∂t u 6∈ L2 (Q), and, in general, we cannot expect that ∂i ∂j y ∈ L2 (Q).
Proof. We regard A as an operator in L2 (Ω) with the homogeneous Dirichlet
boundary condition: D(A) = H 2 (Ω) ∩ H01 (Ω). Then the fractional power Aγ ,
γ ∈ R, can be defined and we have
(
H 2γ (Ω), 0 ≤ γ < 41
(5.8) D(Aγ ) =
{u ∈ H 2γ (Ω); u|∂Ω = 0}, 14 < γ ≤ 1, γ 6= 34

and there exists a constant Cγ > 0 such that

(5.9) Cγ−1 kukH 2γ (Ω) ≤ kAγ ukL2 (Ω) ≤ Cγ kukH 2γ (Ω) , u ∈ D(Aγ )

(e.g. Fujiwara [9]). Moreover we see

1
C −1 kukH −1 (Ω) ≤ kA− 2 ukL2 (Ω) ≤ CkukH −1 (Ω) , u ∈ H −1 (Ω),
(5.10)
C −1 kukX −2 ≤ kA−1 ukL2 (Ω) ≤ CkukX −2 , u ∈ X −2 .

Henceforth, without loss of generality, we can translate the time variable and
we discuss the whole problem in t ∈ (−δ, T ) with δ > 0 and we set
T
(5.11) θ= :
2

 
∂t u + Au = f (x)R(t, x), x ∈ Ω, −δ < t < T,
(5.12)
u(−δ, ·) ∈ X −2 , u|(−δ,T )×∂Ω = 0.

We note that −δ is considered as the initial time. We set Q = (0, T ) × Ω, not

Q = (−δ, T ) × Ω. Moreover we can take sufficiently small T > 0 if necessary,
so that we can assume that

(5.13) R(t, x) > 0, x ∈ Ω, 0 ≤ t ≤ T

by means of (5.7).
Henceforth C, C0 , Cγ , etc. denote positive constants which are independent
of s, f , (t, x) ∈ Q.
Next we examine the regularity of u in t ∈ (0, T ). By the theory of semigroup
(e.g. Tanabe [31]), we can prove
Lemma 5.1. Let f ∈ H −ℓ (Ω) with ℓ < 1, R satisfy (5.6), and a ≡ u(−δ, ·) ∈
−2
X . Then

∂t u ∈ C([0, T ]; H01 (Ω)).

130 O. Yu. Imanuvilov and M. Yamamoto

Proof of Lemma 5.1. By the semigroup theory, we can represent u by

Z t
u(t) = e−tA (e−δA a) + e−A(t−ξ) R(ξ)f dξ
−δ
Z t+δ
= e−A(t+δ) a + e−Aη R(t − η)f dη.
0
Here we write u(t) = u(t, ·), f = f (·) and R(t) = R(t, ·). In view of (5.6), we
have

Z t+δ
−A(t+δ) −A(t+δ)
∂t u(t) = −Ae a+e R(−δ)f + e−Aη (∂t R(t − η)f )dη, t > 0.
0
For γ > 0, there exists a constant Cγ > 0 such that
kAγ e−tA gkL2 (Ω) ≤ Cγ t−γ kgkL2 (Ω) , t>0
(e.g. [31]). Therefore, by (5.9), (5.10) and γ < 1, we have
1
k∂t u(t)kH01 (Ω) ≤ CkA 2 ∂t u(t)kL2 (Ω)

5
3
≤C A 2 e−A(t+δ) A−1 a + A 2 e−A(t+δ) A−1 (R(−δ)f )

Z t+δ

ℓ+1
−Aη − ℓ
+ A 2 e (A 2 (∂t R(t − η)f )dη
0 2
L (Ω)
− 52 − 32
≤C(t + δ) kA akL2 (Ω) + C(t + δ) kA−1 (R(−δ)f )kL2 (Ω)
−1
Z t+δ
ℓ+1 ℓ
+C η− 2 sup kA− 2 (∂t R(t − η)f )kL2 (Ω) dη
0 0≤η≤t+δ
− 52 3
≤Cδ kA−1 akL2 (Ω) + Cδ− 2 kA−1 (R(−δ)f )kL2 (Ω)
2C 1−ℓ ℓ
+ (t + δ) 2 sup kA− 2 (∂t R(s)f )kL2 (Ω) .
1−ℓ −δ≤s≤t

Thus the proof of Lemma 5.1 is complete.

In view of Lemma 5.1, we can justify the following calculations: Setting
z = ∂t u, we have
 
∂t z + Az = (∂t R)f in (0, T ) × Ω,
(5.14)
z(θ, ·) = R(θ, ·)f, z|(0,T )×ω = 0, z|(0,T )×∂Ω = 0.
By (2.9) we have
Z  
1 − 12 1
kϕ ∂t (yesα )k2L2 (0,T ;H −1 (Ω)) + 2 2
|∇y| + sϕ0 y e2sα dxdt
s 0 Q sϕ0
(5.15)
≤ C2 k(∂t R)f esαk2L2 (0,T ;H −1 (Ω)) , s > s0 (λ).
 Carleman estimate and applications 131

On the other hand, by (5.10) and z(0, ·)esα(0,·) = 0, we have

1
kz(θ, ·)esα(θ,·) k2H −1 (Ω) ≤ CkA− 2 (z(θ, ·)esα(θ,·) )k2L2 (Ω)
Z θ Z 
∂ − 12 sα 2
=C |A (ze )| dx dt
0 ∂t Ω
Z θ Z 
− 12 sα − 12 sα
=2C {A ∂t (ze )}{A (ze )}dx dt
0 Ω
Z θ Z 
− 12 − 12 sα − 12 1
sα
=2C {A (sϕ0 ) ∂t (ze )}{A ((sϕ0 ) 2 ze )}dx dt
0 Ω
Z T 1 −1 1 1
≤C {s−1 kA− 2 (ϕ0 2 ∂t (zesα ))k2L2 (Ω) + skA− 2 (ϕ02 zesα )k2L2 (Ω) }dt
0
C −1 1
≤ kϕ0 2 ∂t (zesα )k2L2 (0,T ;H −1 (Ω)) + Cskϕ02 zesα k2L2 (Q) .
s
Hence (5.15) yields
(5.16) kz(θ, ·)esα(θ,·) kH −1 (Ω) ≤ Ck∂t RkC 0,1 (Q) kf esα kL2 (0,T ;H −1 (Ω)) .
Next we will prove
1
(5.17) kf esα kL2 (0,T ;H −1 (Ω)) ≤ C(s − 1)− 4 kf esα(θ,·) kH −1 (Ω)
where C > 0 is independent of s > 0.
Proof of (5.17). By the mean value theorem, we can take κ = κ(t, x) such that

1 1
q(t, x) ≡ α(t, x) − α(θ, x) = (∂t2 α)(t, x)(t − θ)2 + (∂t3 α)(κ, x)(t − θ)3 .
2 6
By direct calculations, we have
−γ0
(5.18) (∂t2 α)(t, x) ≤ , (∂t3 α)(t, x)(t − θ)3 ≤ 0, (t, x) ∈ Q,
t3 (T− t)3
where γ0 > 0 is a constant independent of (t, x) ∈ Q. Hence we directly verify
 
sq(t,x) sγ0 2
(5.19) |e | ≤ exp − 3 (t − θ) , (t, x) ∈ Q
2t (T − t)3
and
|(∇q)(t, x)esq(t,x) |
   
(t − θ)2 (t − θ)3 sγ0 2
(5.20) ≤ C 3 + × exp − 3 (t − θ)
t (T − t)3 t4 (T − t)4 2t (T − t)3
 
(s − 1)γ0
≤ C(t − θ)2 exp − 3 (t − θ)2
2t (T − t)3
≤ C(t − θ)2 exp(−C0 (s − 1)(t − θ)2 )
132 O. Yu. Imanuvilov and M. Yamamoto

for large s > 1. We have

Z 
 
kf e sα(t,·)
kH −1 (Ω) = sup  f esα(θ,x) esq(t,x) µ(x)dx
 
µ∈H01 (Ω),k∇µkL2 (Ω) =1 Ω

≤ kf esα(θ,·) kH −1 (Ω) sup k∇(esq(t,·) µ)kL2 (Ω) .

µ∈H01 (Ω),k∇µkL2 (Ω) =1

By (5.19) and (5.20), we obtain

(5.21) kf esα(t,·) k2H −1 (Ω)

n 2 2
o
≤ C e−2C0 s(t−θ) + s2 (t − θ)4 e−2C0 (s−1)(t−θ) kf esα(θ,·) k2H −1 (Ω) .

Since
Z T
{exp(−2C0 s(t − θ)2 ) + s2 (t − θ)4 exp(−2C0 (s − 1)(t − θ)2 )}dt
0
Z ∞
2 2 C
≤ (e−2C0 (s−1)η + s2 η 4 e−2C0 (s−1)η )dη ≤ √ ,
−∞ s−1

the inequality (5.21) yields (5.17).

Thanks to (5.17), from (5.14) and (5.16), we obtain
1
(5.22) kR(θ, ·)f esα(θ,·) kH −1 (Ω) ≤ C(s − 1)− 4 kf esα(θ,·) kH −1 (Ω) , ∀ s ≥ s0 .

Noting the definition of the H −1 (Ω)-norm, by (5.6) and (5.7), we can prove

(5.23) kf esα(θ,·) kH −1 (Ω) ≤ CkR(θ, ·)f esα(θ,·) kH −1 (Ω) .

In (5.22), taking s > 0 sufficiently large, we see f = 0. Thus the proof of

Theorem 5.1 is complete.

A Appendix A. Sketch of Proof of Lemma 2.1

Since aij ∈ W∞ 1 (Ω), 1 ≤ i, j ≤ n, the unique existence of the solution in

L2 (0, T ; H01 (Ω)) ∩ C([0, T ]; L2 (Ω)) is seen in the case of bi = 0, 1 ≤ i ≤ n and

c = 0, for example, by Chapter 3, §1 and §4 in Lions and Magenes [26]. In the
general case with non-zero bi and c satisfying (2.3), as for the unique existence
of solution and the a priori estimate (2.5), in view of the general theorem in
Chapter 3, §4 in [26], it is sufficient to prove: for any ε > 0, there exists a
constant C = C(ε) > 0 such that
Z 
 

(A.1)  bi u∂i udx ≤ εkuk2H 1 (Ω) + C(ε)kuk2L2 (Ω) , 1 ≤ i ≤ n, u ∈ H 1 (Ω)
Ω
 Carleman estimate and applications 133

and
Z 
 
(A.2)  cu dx ≤ εkuk2 1 + C(ε)kuk2 2 ,
2
u ∈ H 1 (Ω),
  H (Ω) L (Ω)
Ω
Henceforth C > 0 denotes a generic constant independent of functions to
be estimated.
First we prove (A.1). By the Hölder inequality, we have
Z 
 
 bi u∂i udx ≤ kbi (t, ·)kLr (Ω) kuk 2r k∂i ukL2 (Ω) .
  L r−2 (Ω)
Ω
1
Since r > 2n, the Sobolev imbedding theorem (e.g. [1]) implies H 2 −δ (Ω) ⊂
2r
L r−2 (Ω) for sufficiently small δ > 0. Hence with small ε > 0 we have
Z 
 
 bi u∂i udx ≤ kuk 1 −δ kukH 1 (Ω)
  H 2 (Ω)
Ω
C
kuk2 1 −δ .
≤εkuk2H 1 (Ω) +
ε H 2 (Ω)

By the interpolation inequality (e.g. [1]), we see

(A.3) kuk2 1 ≤ δkuk2H 1 (Ω) + C(δ)kuk2L2 (Ω)
H 2 −δ (Ω)
δ
for small δ > 0. We choose sufficiently small ε > 0 and δ > 0 such that ε is
also small, so that
Xn Z 
 
(A.4)  bi u∂i udx ≤ εkuk2 1 + C(ε)kuk2 2 .
  H (Ω) L (Ω)
i=1 Ω

Next we will prove (A.2). For this, we show n o

1 2n 1 1
Lemma A.1. Let 0 < µ < 2 and r1 > max 3−2µ , 1 , r1 + r1′
= 1. Then
1
there exist constants 0 < δ < 2 and C > 0 such that
kzvkW µ′ (Ω) ≤ CkvkH 1 (Ω) kzk 1 .
r1 H 2 −δ (Ω)

As for the proof, we can refer to Lemma 2.2 in [14] for example.
By Lemma A.1, we have
Z 
 
 cu2 dx ≤ kc(t, ·)k −µ ku2 kW µ (Ω)
  Wr1 (Ω) r′
Ω 1

C
≤Ckuk 1 −δ kukH 1 (Ω) ≤ Cεkuk2H 1 (Ω) + kuk2 1 −δ
H 2 (Ω) ε H 2 (Ω)
1
with 0 < δ < 2. In view of the interpolation inequality (A.3), taking ε > 0 and
δ > 0 so small that δε is also small, we obtain
Z 
 
 cu dx ≤ εk∇uk2 2 + C(ε)kuk2 2 .
2
  L (Ω) L (Ω)
Ω
Thus the proof of (A.1) and (A.2) is complete.
134 O. Yu. Imanuvilov and M. Yamamoto

B Appendix B. Proof of Lemma 2.3

Henceforth C > 0 denotes a generic constant independent of y, and dependent
on c. First
Z T
Z 2
 
kcyesα k2L2 (0,T ;H −1 (Ω)) ≤C sup  cyesα µdx dt
 
0 kµkH 1 (Ω) =1 Ω
Z T
≤C {kck2L∞ (0,T ;W −µ (Ω)) sup ky(t, ·)esα µk2W µ (Ω) }dt
r1 ′
r1
0 kµkH 1 (Ω) =1
Z T
≤C sup ky(t, ·)µesα k2W µ (Ω) dt.
0 kµkH 1 (Ω) =1 r′
1

Then by Lemma A.1 in Appendix A, we obtain

ky(t, ·)µesα k2W µ (Ω) ≤ Ckµk2H 1 (Ω) kyesα k2 1

r′
1
H 2 −δ (Ω)

and so we see
Z T
kcyesα k2L2 (0,T ;H −1 (Ω)) ≤C kyesα k2 1 dt.
0 H 2 −δ (Ω)

Thus the proof of Lemma 2.3 is complete, when we will have proved
Z
2δ 1
(B.1) s kyesα k2 2 1 ≤C (sϕy 2 + |∇y|2 )e2sα dxdt
L (0,T ;H 2 −δ (Ω)) Q sϕ

whenever δ ∈ (0, 21 ). In the rest part of the appendix, we will verify (B.1). We
note that

(B.2) |(∂i ϕ)(t, x)|, |(∂i α)(t, x)| ≤ Cϕ(t, x), (t, x) ∈ Q

and

C −1 ϕ0 (t) ≤ ϕ(t, x) ≤ Cϕ0 (t), (t, x) ∈ Q

1
where ϕ0 (t) = t(T −t) . Therefore, by the interpolation inequality (e.g. [1], [26])
and

1 − 2δ 1−2δ
2 1 + 2δ 1+2δ
2
(B.3) |ab| ≤ |a| + |b| ,
2 2
 Carleman estimate and applications 135

we have
1 1
−δ +δ
kyesα k 1 −δ ≤ Ckyesα kH
2
1 (Ω) kye
sα 2
kL2 (Ω)
H 2 (Ω)
1 1
1 −δ 1 δ 1 +δ 1 δ
≤Cky(sϕ0 )− 2 esα kH
2
1 (Ω) (sϕ0 )
−
4 2 ky(sϕ ) 2 e
0 kL2 (Ω) (sϕ0 )− 4 − 2
sα 2

1 1
1 −δ 1 +δ
≤Cs−δ k(sϕ0 )− 2 ∇(yesα )kL2 2 (Ω) k(sϕ0 ) 2 esα ykL2 2 (Ω)
1 1
1 −δ 1 +δ 1
≤Cs−δ k(sϕ)− 2 esα ∇ykL2 2 (Ω) k(sϕ) 2 esα ykL2 2 (Ω) + Cs−δ k(sϕ) 2 esα ykL2 (Ω)
  2
−δ 1 − 2δ
1 1−2δ
− 12 sα −δ
≤Cs k(sϕ) e ∇ykL2 (Ω) 2
2
  2
−δ 1 + 2δ
1 1+2δ
1 +δ 1
+Cs 2
sα
k(sϕ) e ykL2 (Ω) 2
+ Cs−δ k(sϕ) 2 esα ykL2 (Ω) .
2

Thus the verification of (B.1) is complete.

References

[1] R. A. Adams, Sobolev Spaces, Academic Press, New York, 1975.

[2] A. L. Bukhgeim and M. V. Klibanov, Global uniqueness of a class of multidimen-
sional inverse problems, Soviet Math. Dokl., 24 (1981), pp. 244–247.
[3] T. Carleman, Sur un problème d’unicité pour les systèmes d’équations aux derivées
partielles à deux variables independentes, Ark. Mat. Astr. Fys., 2B (1939), pp. 1–9.
[4] D. Chae, O. Y. Imanuvilov, and S. M. Kim, Exact controllability for semilinear
parabolic equations with Neumann boundary conditions, J. of Dynamical and
Control Syst., 2 (1996), pp. 449–483.
[5] J. Cheng, V. Isakov, M. Yamamoto, and Q. Zhou, Lipschitz stability in the
lateral Cauchy problem for elasticity system, Preprint Series UTMS 99-33, 1999,
Graduate School of Mathematical Sciences, The University of Tokyo. 1999.
[6] C. Fabre, J.-P. Puel, and E. Zuazua, Approximate controllability of the semilinear
heat equation, Proc. Roy. Soc. Edinburgh, 125 Sect. A (1995), pp. 31–61.
[7] L. Fernández and E. Zuazua, Approximate controllability for semilinear heat
equation involving gradient terms, J. Optim. Theory Appl., 101 (1999), pp. 307–
328.
[8] E. Fernández-Cara, Null controllability of the semilinear heat equation,
ESAIM:Control, Optimization and Calculus of Variations, 2 (1997), pp. 87–103.
[9] D. Fujiwara, Concrete characterization of the domains of fractional powers of
some elliptic differential operators of the second order, Proc. Japan Acad., 43
(1967), pp. 82–86.
[10] A. V. Fursikov and O. Y. Imanuvilov, Controllability of evolution equations, vol. 34
of Lecture notes series, Seoul National University, Seoul, South Korea, 1996.
[11] L. Hörmander, Linear Partial Differential Operators, Springer-Verlag, Berlin,
1963.
[12] O. Y. Imanuvilov, Boundary controllability of parabolic equations, Sbornik Math-
ematics, 186 (1995), pp. 879–900.
136 O. Yu. Imanuvilov and M. Yamamoto

[13] O. Y. Imanuvilov and M. Yamamoto, Lipschitz stability in inverse parabolic

problems by the Carleman estimate, Inverse Problems, 14 (1998), pp. 1229–1245.
[14] , On Carleman inequalities for parabolic equations in Sobolev spaces of
negative order and exact controllability for semilinear parabolic equations, Preprint
Series UTMS 98-46, 1998, Graduate School of Mathematical Sciences, The
University of Tokyo. 1998.
[15] V. Isakov, Inverse Source Problems, American Mathematical Society, Providence,
Rhode Island, 1990.
[16] , Carleman type estimates in an anisotropic case and applications, Journal
of Differential Equations, 105 (1993), pp. 217–238.
[17] , Inverse Problems for Partial Differential Equations, Springer-Verlag,
Berlin, 1998.
[18] V. Isakov and M. Yamamoto, Carleman estimate with the Neumann boundary
condition and its applications to the observability inequality and inverse hyperbolic
problems, to appear in Contemporary Mathematics.
[19] M. A. Kazemi and M. V. Klibanov, Stability estimates for ill-posed Cauchy
problems involving hyperbolic equations and inequalities, Appl. Math., 50 (1993),
pp. 93–102.
[20] A. Khaı̆darov, Carleman estimates and inverse problems for second order hyper-
bolic equations, Math. USSR Sbornik, 58 (198), pp. 267–277.
[21] M. V. Klibanov, Inverse problems and Carleman estimates, Inverse Problems, 8
(1992), pp. 575–596.
[22] I. Lasiecka and R. Triggiani, Carleman estimates and exact boundary controlla-
bility for a system of coupled, nonconservative second-order hyperbolic equations,
in Partial Differential Equation Methods in Control and Shape Analysis, vol. 188,
Marcel Dekker, Inc., New York, 1997, pp. 215–243.
[23] I. Lasiecka, R. Triggiani, and X. Zhang, Nonconservative wave equations with
unobserved Neumann B.C.: global uniqueness and observability in one shot, to
appear in Contemporary Mathematics.
[24] M. M. Lavrent’ev, V. G. Romanov, and S. P. Shishat·skiı̆, Ill-posed Problems of
Mathematical Physics and Analysis, American Mathematical Society, Providence,
Rhode Island, 1986.
[25] F. H. Lin, A uniqueness theorem for parabolic equations, Comm. Pure Appl.
Math., 43 (1990), pp. 127–136.
[26] J. L. Lions and E. Magenes, Non-Homogeneous Boundary Value Problems and
Applications, Springer-Verlag, Berlin, 1972.
[27] C.-C. Poon, Unique continuation for parabolic equations, Comm. Partial Differ-
ential Equations, 21 (1996), pp. 521–539.
[28] A. Ruiz, Unique continuation for weak solutions of the wave equation plus a
potential, J. Math. Pures Appl., 71 (1992), pp. 455–467.
[29] J.-C. Saut and B. Scheurer, Unique continuation for some evolution equations, J.
Differential Equations, 66 (1987), pp. 118–139.
[30] C. D. Sogge, Unique continuation theorem for second order parabolic differential
operators, Ark. Mat., 28 (1990), pp. 159–182.
[31] H. Tanabe, Equations of Evolution, Pitman, London, 1979.
[32] D. Tataru, Boundary controllability for conservative PDEs, Appl. Math. Optim.
(1995), pp. 257–295.
[33] , Carleman estimates, unique continuation and controllability for anizotropic
 Carleman estimate and applications 137

PDE’s, Contemporary Math., 209 (1997), pp. 267–279.

[34] M. Yamamoto, Uniqueness and stability in multidimensional hyperbolic inverse
problems, J. Math. Pures Appl., 78 (1999), pp. 65–98.
138 O. Yu. Imanuvilov and M. Yamamoto
Bilinear control for global controllability of the
semilinear parabolic equations with superlinear
terms

Alexander Khapalov, Washington State University, Pullman, Washington

Abstract

In this paper we establish the global approximate controllability of the

semilinear heat equation with superlinear term, governed in a bounded
domain by a pair of controls: (a) the traditional internal either locally
distributed or lumped control and (b) the lumped control entering the
system as a time-dependent coefficient. The motivation for the latter is
due to the well known lack of global controllability properties for this class
of pde’s when they are steered solely by the former controls. Our approach
involves an asymptotic technique allowing us to “separate and combine”
the impacts generated by the above-mentioned two types of controls. In
particular, the addition of bilinear control allows us to reduce the use of
the additive one to the local controllability technique only.

Key words: semilinear heat equation, global approximate controllability,

bilinear controls, lumped control, asymptotic analysis.
AMS(MOS) subject classifications. 93, 35.

1. Introduction.

1.1. Problem formulation and motivation.

We consider the following Dirichlet boundary problem, governed in a

bounded domain Ω ⊂ Rn by the bilinear lumped control k = k(t) and
the additive locally distributed control v(x, t)χω (x), supported in the given
subdomain ω ⊂ Ω:
∂u
(1) = ∆u + k(t)u − f (x, t, u, ∇u) + v(x, t)χω (x) in QT = Ω × (0, T ),
∂t
u=0 in ΣT = ∂Ω × (0, T ), u |t=0 = u0 ∈ L2 (Ω),
139
140 Khapalov

k ∈ L∞ (0, T ), v ∈ L2 (QT ).
In the one space dimension we will also consider the case when both controls
are lumped, that is, they are the functions of time only: k = k(t) and v = v(t).
In this paper we are concerned with the issue of approximate controllability
of system (1) in the (phase-) space L2 (Ω). Namely, given the initial state u0 ,
we want to know whether the range of the solution mapping

(2) L∞ (0, T ) × L2 (QT ) ∋ (k, v) → u(·, T ) ∈ L2 (Ω)

is dense in L2 (Ω). (In fact, due to the possible nonuniquenes of solutions to

(1) the situation here is more complex as we discuss it below in the subsection
1.2.)
It is well known ([5], [9], [19], [8]) that a rather general semilinear parabolic
equation, governed in a bounded domain by the classical either boundary or
additive locally distributed controls only (i.e., no “changeable” bilinear control
k(·) in (1)) is globally approximately controllable in L2 (Ω), provided that the
nonlinearity is globally Lipschitz. The methods of these works make use of the
fixed point argument and the fact that such semilinear equations can be viewed
as “linear equations” with the coefficients uniformly bounded in some sense.
Alternative approach employs the global inverse function theorem – we refer in
this respect to the work [16] on the semilinear wave and plate equations.
However, the situation is principally different if nonlinear terms admit
polynomial superlinear growth at infinity. Such terms are in the focus of our
attention in this paper.
Given T > 0, we further assume that f (x, t, u, p) is Lebesgue’s measurable
in x, t, u, p, and continuous in u, p for almost all (x, t) ∈ QT , and is such
that

(3a) | f (x, t, u, p) |≤ β | u |r1 +β k p krR2n a.e. in QT for u ∈ R, p ∈ Rn ,

(3b)
Z Z Z
f (x, t, φ, ∇φ) φdx ≥ (ν − 1) k ∇φ kRn dx − ̺ (1 + φ2 )dx ∀φ ∈ H01 (Ω),
2

Ω Ω Ω

where β, ν, ̺ > 0, T ̺ ≤ β, and

4 2
(3c) r1 ∈ (1, 1 + ), r2 ∈ (1, 1 + ).
n n+2
Here and below we use the standard notations for Sobolev spaces such as
H01,0 (QT ) = {φ | φ, φxi ∈ L2 (QT ), i = 1, . . . , n, φ |ΣT = 0} and
H01 (Ω) = {φ | φ, φxi ∈ L2 (Ω), i = 1, . . . , n, φ |∂Ω = 0}. (A simple example of a
function f satisfying conditions (3a-c) is f (u) = u3 .) We refer, e.g., to [15] (p.
466), where it was shown that system (1), (3a-c) admits at least one generalized
 Bilinear control for the semilinear equations 141
T T
solution in C([0, T ]; L2 (Ω)) H01,0 (QT ) L2+4/n (QT ), while its uniqueness is
not guaranteed.
It turns out that in the superlinear case like (3a-c) the impact from the sole
additive control v(t)χω (x) does not propagate “effectively” from its support to
the rest of the space domain: regardless of how large the control applied on ω is,
the corresponding solutions remain uniformly bounded on any closed subset of
Ω\ω̄. (In other words, given the initial state u0 , there exist target states, namely,
“sufficiently large” on Ω\ω̄, which are strictly separated from the range of the
corresponding solution mapping.) This is true in any of the (phase- ) spaces
Lp (Ω), 1 ≤ p < ∞ at any positive time, e.g., for the functions f = f (x, t, u)
such that f (x, t, u)u > c1 | u |2+r −c2 for some constants c1 , c2 , r > 0 ([9],
and the references therein). (On the other hand, for certain refinements of
conditions (3a-c) a number of positive “superlinear” controllability results were
obtained in [10], [12]-[14], see Remark 1.1 for details.)
In this paper our goal is to show that the above-outlined principal difficulty
with propagation of control impact can be overcome by using an additional
bilinear lumped control k = k(t), entering the equation (1) as a coefficient and
thus affecting the qualitative behavior of system (1),(3a-c) in the entire space
domain.

Remark 1.1: Some references.

• We refer to the work [2] on controllability of the abstract infinite

dimensional bilinear system as the only known to us on this subject in the
framework of pde’s. In particular, in [2], under the additional assumption
that all the modes in the initial data are active, the global approximate
controllability of the rod equation utt + uxxxx + k(t)uxx = 0 with hinged
ends and the wave equation utt − uxx + k(t)u = 0 with Dirichlet boundary
conditions, where k is control, was shown making use of the Fourier series
approach. (To the contrary, an extensive and thorough bibliography on
controllability on bilinear ode’s is available, see, e.g., the survey [1].) We
also refer to works [4], [17] (and the references therein) on the issue of
optimal bilinear control for various pde’s.

• Some examples of physical interpretation of bilinear controls can be found

in the just-cited works. In particular, in the rod equation the control
k(t) is the axial load [2]. In turn, the equation (1) is typically used to
model the heat transfer and diffusion processes, with the semilinear term
associated, e.g., with the porous media. In this context k(t) is respectively
proportional to the heat or mass transfer coefficients. For example, in the
heat transfer models involving fluids k(t) is sensitive to their velocity.
In the diffusion processes, such as some biological models or the chain
reaction, the values of k(t) can be both positive and negative. In general,
142 Khapalov

the bilinear parameter k is linked to the physical properties of the process

at hand and the use of it as a control has a potential to be associated with
so-called “intelligent materials” that can change their physical properties
under certain conditions.

• In spite of the lack of global controllability of (1), (3a-c) discussed in

the above, it was shown in [13] that this equation is actually globally
approximately controllable at any time T > 0 solely by means of the ad-
ditive locally distributed controls in the spaces that are weaker than any
of Lp (Ω), 1 ≤ p < ∞. Moreover, under the additional assumption that
the superlinear term is locally Lipschitz (which ensures the uniqueness
of solutions in C([0, T ]; L2 (Ω))) the global finite dimensional exact con-
trollability of (1), (3a-c) (i.e., not necessarily to the equilibrium) at any
positive time T > 0 was also established in [13].

• The global approximate controllability of (1), (3a-c) with k = 0 was shown

in [10] for the static controls v = v(x) supported in the entire Ω.

• For the one dimensional version (5), (6a-c) (see below) of system (1) it was
shown in [12] that, if in (6a) β = β(t) → 0 faster than any e−ν/t , ν > 0 as
t → 0, then (5), (6a-c) with k = 0 is globally approximately controllable in
L2 (Ω) at any time only by means of the lumped control v = v(t), provided
that the endpoints of the interval (a, b) are the irrational numbers. This
result was recently extended to the case of several dimensions and locally
distributed controls in [14].

• The method of [10], [12]-[14] is based on the idea to “suppress” the

effect of nonlinearity by applying the actual control action only during
asymptotically short period of time. Similar idea is used in this article
and in [11] in the context of bilinear ode’s.

• Though in this article we discuss the global approximate controllability

of the equation (1), we would also like to mention here some related
works on the very close global exact null-controllability property (i.e., the
exact steering to the origin) by means of the additive locally distruibuted
controls only. In [7] the latter property was shown in L2 (Ω) (or
appropriate Sobolev space) with the reaction term f = f (x, t, u) only,
assuming that f can grow superlinearly at the logarithmic rate like
lim|p|→∞ f (p)/(p log | p |) = 0. Assuming the dissipativity condition, this
result was improved in [3] to the rate lim|p|→∞ f (p)/(p(log | p |)3/2 ) = 0.
Also in [3] some interesting non-global exact null-controllability results
were given.
 Bilinear control for the semilinear equations 143

1.2. Main results.

The multidimensional case with additive locally distributed controls.

Theorem 1.1. Let conditions (3a-c) hold. Then the range of the solution
mapping (2) is dense in L2 (Ω).

Note now that, since the boundary problem (1), (3a-c) admits multiple
solutions, this result is qualitatively different from the classical understanding
of the approximate controllability as steering (associated with applications in
the first place), which is as follows: (1) is said to be globally approximately
controllable in L2 (Ω) at time T if for any u0 , uT ∈ L2 (Ω), ξ > 0 there is a
control pair (k, v) such that

k u(·, T ) − uT kL2 (QT ) ≤ ξ.

Clearly, this classical definition is ill-posed in our case of possible multiple

solutions. Therefore, we will also use its adjustment which requires one to
find a control pair which steers all the possible realizations of a solution to
(1) in a uniform fashion. (This type of controllability was investigated in [10],
[12]-[14].)

Definition 1.1. We will say that the system (1), (3a-c), admitting multiple
solutions, is globally approximately controllable in L2 (Ω) at time T if for every
ξ > 0 and u0 , uT ∈ L2 (Ω) there is a control pair (k, v) ∈ L∞ (QT ) × L2 (QT )
such that for all (i.e., possibly multiple) solutions of (1), (3a-c), corresponding
to it

(4) k u(·, T ) − uT kL2 (Ω) ≤ ξ.

Theorem 1.2. System (1), (3a-c) is globally approximately controllable in

L2 (Ω) at any time T > 0 in the sense of Definition 1.1.

The 1-D case with all lumped controls.

Consider now the one dimensional version of problem (1), (3a-c) with all
lumped controls:

(5) ut = uxx + k(t)u − f (x, t, u, ux ) + v(t)χ(a,b) (x) in QT = (0, 1) × (0, T ),

u |x=0,1 = 0, u |t=0 = u0 ∈ L2 (0, 1), k ∈ L∞ (0, T ), v ∈ L2 (0, T ).

Here both k = k(t) and v = v(t) are the functions of time only.
144 Khapalov

Distinguishing the 1-D case we pursue two goals. Firstly, the positive
result for the case of lumped additive controls implies the same for the locally
distributed ones (since the former controls are a degenerate subclass of the
latter ones). Accordingly, our proof of Theorems 1.1/1.2 is given below as the
immediate consequence of the 1-D-“lumped” case. Secondly, lumped controls
are of special interest being closer to the engineering applications. Focusing on
them, we can give somewhat more “explicit” feeling of our method and of the
general conditions (3a-c), which for the equation (5) are as follows:

(6a) | f (x, t, u, p) | ≤ β | u |r1 + β | p |r2 a.e. in QT for u, p ∈ R,

Z1 Z1 Z1
(6b) f (x, t, φ, φx ) φdx ≥ (ν − 1) φ2x dx − ̺ φ2 dx ∀φ ∈ H01 (0, 1),
0 0 0

where β, ν, ̺ > 0, ̺T ≤ β, and

(6c) r1 ∈ (1, 5), r2 ∈ (1, 5/3).

Theorem 1.3. If a < b are any irrational numbers from (0,1), then system
(5), (6a-c) is globally approximately controllable in L2 (Ω) at any time T > 0 in
the sense of Definition 1.1.

Clearly the assumption that the endpoints of the interval (a, b) are the
irrational numbers makes the result of Theorem 1.1 unstable with respect to
the choice of control support (a, b). We stress however that it is well known
that this assumption is intrinsic for lumped controls even in the linear case. In
this respect one may prefer its immediate “stable” corollary – Theorem 1.2.
The paper is organized as follows. In the next two section we prove and
recall some auxiliary results for the linear version of system (5). Theorem 1.3
is proven in Section 4. Theorems 1.1/1.2 are proven in Section 5.

2. Preliminaries. Consider the boundary problem (5), (6a-c) assuming

that bilinear control k is constant, i.e., k(t) ≡ α on (0, T ):

(7) wt = wxx + αw − f (x, t, w, wx ) + v(t)χ(a,b) (x) in QT ,

w |t=0 = w0 ∈ L2 (0, 1), v ∈ L2 (0, T ).

w |x=0,1 = 0,
T
Denote B(0, T ) = C([0, T ]; L2 (Ω)) H01,0 (QT ) and
 1/2
ZT Z1
k q kB(0,T ) =  max k q(·, t) k2L2 (0,1) + 2 qx2 dxds .
t∈[0,T ]
0 0
 Bilinear control for the semilinear equations 145

We have the following a priori estimate.

Lemma 2.1. Given T > 0, α ≥ 0 and a positive number µ, any solution to

(7), (6a-c) (if there are multiple solutions) satisfies the following two estimates:
(8)

k w kB(0,T ) , k w kL6 (QT ) ≤ C(µ)e(α+̺+µ/2)T k w0 kL2 (0,1) + k v kL2 (0,T ) ,

where C(µ) does not depend on α.

Here and below we routinely use symbols c and C to denote (different)

generic positive constants or positive-valued functions.

Proof. Indeed, recall ([15]) that f (·, ·, w, wx ) ∈ L6/5 (QT ) and that the
following energy equality holds for (7) treated as a linear equation with the
source term f (x, t, w, wx ) + v(t)χ(a,b) (x), e.g., [15] (p. 142):

Zt Z1
1
k w k2L2 (0,1) |t0 + (wx2 − αw2 + f (x, s, w, wx )w
2
0 0

(9) −v(t)χ(a,b) (x)w)dxds = 0 ∀t ∈ [0, T ].

Here and everywhere below, if there exist several solutions to (5), we always
deal separately with a selected one, while noticing that all the estimates hold
uniformly.
Combining (9) and (6b) yields:

Zt Z1
k w(·, t) k2L2 (0,1) + 2ν wx2 (x, s) dxds
0 0

Zt Z1 Zt Z1
≤k w0 k2L2 (0,1) + 2(α + ̺) 2
w dxds + 2 v(t)χ(a,b) (x)wdxds
0 0 0 0
 
b−a
≤ k w0 k2L2 (0,1) + k v k2L2 (0,T )
µ
Zt
+ 2(α + ̺ + µ/2) k w(·, τ ) k2L2 (0,1) dτ
0
 
b−a
≤ k w0 k2L2 (0,1) + k v k2L2 (0,T )
µ
146 Khapalov

(10)  
Zt Zτ Z1
+ 2(α+̺+µ/2) k w(·, τ ) k2 2 wx2 (x, s) dxds dτ ∀t ∈ [0, T ].
L (0,1) +2ν
0 0 0

In the above we have used Young’s inequality

1
2wvχ(a,b) ≤ (vχ(a,b) )2 + µw2 ,
µ

which holds for any positive µ. Applying Gronwall-Bellman inequality to (10)

yields the first estimate in (8) with respect to the B(0, T )-norm. The second
estimate (with properly arranged generic constant) follows by the continuity
of the embedding B(0, T ) into L6 (QT ) (e.g., [15], pp. 467, 75). This ends the
proof of Lemma 2.1.

Remark 2.1. Note that if α < 0 in (7), then, as (10) implies, (8) holds with
no α in it.

We now intend to evaluate the difference between the solution w to (7) and
that to its truncated version

(11) yt = yxx + αy + χ(a,b) v(t) in QT , v ∈ L2 (0, T ),

y |x=0,1 = 0, y |t=0 = y0 ∈ L2 (0, 1),

assuming that w0 = y0 .
Denote z = w − y, then

zt = zxx + αz − f (x, t, w, wx ) in QT ,

z |x=0,1 = 0, z |t=0 = 0.
Similar to (9) and (10) we have,

Zt Z1
k z(·, t) k2L2 (0,1) + 2 zx2 (x, s) dxds
0 0

Zt Z1 Zt Z1
2
≤ 2α z dxds + 2 zf (x, s, w, wx )dxds
0 0 0 0

Zt Z1
≤ 2α z 2 dxds + 2 k z kL6 (Qt ) k f (·, ·, w, wx ) kL6/5 (Qt )
0 0
 Bilinear control for the semilinear equations 147

Zt Z1
≤ 2α z 2 dxds + 2c k z kB(0,t) k f (·, ·, w, wx ) kL6/5 (QT )
0 0

Zt
≤ 2α k z k2B(0,s) ds + δ k z k2B(0,t)
0

c2
(12) + k f (·, ·, w, wx ) k2L6/5 (QT ) ∀t ∈ [0, T ],
δ
where we have used Hölder’s and Young’s inequalities and the continuity of the
embedding B(0, T ) into L6 (QT ), due to which,

(13) k z kL6 (QT ) ≤ c k z kB (0, T ).

From (12), we have

Zt
max k z(·, τ ) k2L2 (0,1) ≤ 2α k z k2B(0,s) ds + δ k z k2B(0,t)
τ ∈(0,t)
0

c2
+ k f (·, ·, w, wx ) k2L6/5 (QT ) ∀t ∈ [0, T ].
δ
Hence, again from (12),

Zt
kz k2B(0,t) ≤ 4α k z k2B(0,s) ds + 2δ k z k2B(0,t)
0

2c2
+ k f (·, ·, w, wx ) k2L6/5 (QT ) ∀t ∈ [0, T ]
δ
and
(1 − 2δ) k z k2B(0,t)

Zt
2c2
(14) ≤ 4α k z k2B(0,s) ds + k f (·, ·, w, wx ) k2L6/5 (QT ) .
δ
0

Making use of Gronwall-Bellman inequality, we derive from (14) that

√
2α
T 2c
(15) k z kB(0,T ) ≤ e 1−δ √ k f (·, ·, w, wx ) kL6/5 (QT ) ,
δ
148 Khapalov

provided that
1
(16) 0<δ< .
2
Now, using (6a) and Hölder’s inequality (as in [15], p. 469; [10], p. 863),
we obtain:
5 r1
k f (·, ·, w, wx ) kL6/5 (QT ) ≤ βT 6 (1− 5 ) k w krL16 (Q
T)

5 3r2
(17) + βT 6 (1− 5
)
k wx krL22 (QT ) .

Combining (17), (15), (16) and (13) yields

Lemma 2.2. Given T > 0, α ≥ 0, δ ∈ (0, 1/2), and w0 = y0 , we

have the following two estimates for the difference z = w − y between any
corresponding solution w to (7), (6a-c) (if there are multiple ones) and the
unique corresponding solution to (11):
2αT 1 5 r1
k z kB(0,T ) , k z kL6 (QT ) ≤ Ce 1−δ √ (T 6 (1− 5 ) k w krL16 (QT )
δ

5 3r2
(18) + T 6 (1− 5
)
k wx krL22 (Q ) ) ∀δ ∈ (0, 1/2),
T

where C does not depend on α.

3. Controllability properties of the truncated linear system (11).

Here we would like to remind the reader some controllability properties of the
linear system (11). √
Denote by λk = −(πk)2 + α, ωk (x) = 2 sin πkx, k = 1, . . . the
eigenvalues and orthonormalized in L2 (0, 1) eigenfunctions of the spectral
problem: ωxx + αω = λω, ω ∈ H01 (0, 1).
It is well known that the general solution to (11) admits the following
representation:
1 
∞
X Z
y(x, t) = eλk t  y0 (r)ωk (r)dr  ωk (x) +
k=1 0

t
 
∞ Z
X Z1
(19) eλk (t−τ )  v(τ )χ(a,b) (r) ωk (r) dr  dτ ωk (x),
k=1 0 0
 Bilinear control for the semilinear equations 149

where the series converge in L2 (0, 1) uniformly over t ≥ 0.

Let {qT,k }∞k=1 be a biorthogonal sequence to {e
λk τ }∞
k=1 in the closed
subspace  
cl span{eλk τ | k = 1, . . . }

of L2 (0, T ) ([18], [6]):

ZT 
1, if k = l,
eλk τ qT,l (τ ) dτ =
0, if k 6= l,
0

where
1
(20) k qT,k kL2 (0,T ) = ,
dk (α, T )

I
X
dk (α, T ) = inf{k eλk t + bi eλi t kL2 (0,T ) | bi ∈ R, I = 1, 2, . . . }.
i=1,i6=k

R b Assume that a ± b are the irrational numbers. We need this to ensure that
a sin πkx dx 6= 0 for all k − 1, . . . .
Denote
b
√ Z
(21) vT,k (τ ) = qT,k (T − τ ) ( 2 sin πkx dx)−1 , τ ∈ (0, T ),
a

so that
ZT Z1 
λk (T −τ ) 1, if k = l,
(22) e vT,l (τ )χ(a,b) (r) ωk (r) dr dτ =
0, if k 6= l.
0 0

From (19) and (22) it follows that, given the positive integer L and the real
numbers a1 , . . . , aL , if one applies control
L
X
(23) v̂T (t) = ak vT,k (t), t ∈ (0, T )
k=1

in (11), then
 
L
X ∞
X Z1
(−(πkt)2 +α)T 
(24) y(x, T ) = ai ωk (x) + e y0 (r)ωk (r)dr  ωk (x),
k=1 k=1 0
150 Khapalov

where, by (20) and (21),

b
L
X 1 √ Z
(25) k v̂T kL2 (0,T ) ≤ γ(α, T ) = | ak ( 2 sin πkx dx)−1 | .
dk (α, T )
k=1 a

Remark 3.1. It follows from (20) that γ(α, T ) in (25) is nonincreasing in

T > 0.

Also, from (24),

L
X
(26) k y(·, T ) − ai ωk (·) kL2 (0,1) ≤ eαT k y0 kL2 (0,1) .
k=1

Now we are ready to prove Theorem 1.3.

4. Proof of Theorem 1.3. The scheme of the proof is as follows.

(1) Given the initial and target state u0 and uT , we steer the system at
hand “close” to the zero-state (equilibrium) employing the constant negative
bilinear controls only.
(2) Using a ( sort of) locally controllability technique with only additive
controls active, we steer the system “close” to a state suT for some small
parameter s > 0.
(3) Again, employing only constant positive bilinear controls, we “stretch”
the latter state to the desirable length uT .

Step 1: Approximate null-controllability. Take any T ∗ > 0. Then it

follows from the proof of Lemma 2.1 that if one applies control pair k(t) ≡
α < 0, v(t) = 0 on (0, T ∗ ), the corresponding solution(s) to (5), (6a-c) can be
made arbitrarily small in L2 (0, 1) by selecting appropriately small negative α.
Indeed, it follows from (10) that

ZT Z1
µ b−a
2(−α − ̺ − ) w2 dxds ≤ k w0 k2L2 (0,1) + k v k2L2 (0,T ) .
2 µ
0 0

As α → −∞, this implies that we can make k u(·, t∗ ) kL2 (0,1) as small as we
wish for some t∗ ∈ (0, T ) (in general, t∗ can be different for different multiple
solutions). This “smallness” is preserved on [t∗ , T ] by Remark 2.1, applied with
v = 0 and the same α on (t∗ , T ). In other words, we have the global approximate
controllability to the origin in the sense of Definition 1.1, just by using constant
bilinear controls.
 Bilinear control for the semilinear equations 151

Step 2. From Step 1 it follows that, without loss of generality, we may

further assume that the initial state u0 in (1) is arbitrarily small in L2 (0, 1).
(Otherwise, we need to apply the argument of Step 1 on some “small” time-
interval (0, T ∗ ), T ∗ < T .)
To prove Theorem 1.1, it is sufficient to show that any function like
L
X
uT (x) = ai ωk (x)
k=1

can be approached by u(·, T ) arbitrarily close in the sense of (4).

Fix any positive integer L and the real numbers a1 , . . . ,L .
Given T > 0, select a parameter s ∈ (0, 1) and also µ and δ in Lemmas 2.1
and 2.2. By Step 1, without loss of generality we may assume that

k u0 kL2 (0,1) ≤ s2 .

Consider any ε ∈ [0, T /2] and apply on the interval (0, T − ε) the control
pair (see (21)-(23) for notations)
L
X
k(t) = α = 0, vs,T −ε (t) = sv̂T −ε = s ak vk,T −ε (t), t ∈ (0, T − ε).
k=1

Then, in notations of Section 2 with α = 0, u = w = y + z, and, see (24),

(27) u(·, T − ε) = suT + (y(·, T − ε) − suT ) + z(·, T − ε),

where in view of (26) and Lemmas 2.1 and 2.2, applied with α = 0 on (0, T − ε),

(28) k y(·, T − ε) − suT kL2 (0,1) ≤ k u0 kL2 (0,1) ≤ s2 ,

k z(·, T − ε) kL2 (0,1)

min{r1 ,r2 }
(29) ≤ C(T ) k u0 kL2 (0,1) + s k v̂T −ε kL2 (0,T −ε) ) ,

as s → 0, where C(T ) does not depend on ε.

Since r1 , r2 > 1, (27)-(29) yields that

(30) u(·, T − ε) = suT + p(·, T − ε),

where, as it follows from Remark 3.1, uniformly over ε ∈ [0, T /2],

(31) k p(·, T − ε) kL2 (0,1) = o(s).

Step 3. On the interval (T − ε, T ) we apply controls

k(t) = α > 0, v(t) = 0, t ∈ (T − ε, T ).

152 Khapalov

Then, again in notations of Section 2, applied now on the interval (T − ε, T ),

(32) u(·, T ) = y(·, T ) + z(·, T ).

Here, according to (19) and (30), applied on (T − ε, T ),

1 
∞
X Z
2
y(x, T ) = e(−(πk) +α)ε  u(r, T − ε)ωk (r)dr  ωk (x)
k=1 0
 
∞
X Z1
2
= eαε e−(πk) ε  u(r, T − ε)ωk (r)dr  ωk (x)
k=1 0

(33) = eαε (suT + p(·, T − ε) + h(ε)(suT + p(·, T − ε)) ,

where h(ε) → 0 as ε → 0 (by continuity of solutions to (7) in time). In other

words, in view of (31),

(34) y(·, T ) = seαε uT + seαε g(·, s, ε),

where

(35) k g(·, s, ε) kL2 (0,1) → 0 as s, ε → 0.

On the other hand, by Lemma 2.2, applied on (T − ε, T ) with v = 0:

2αε 5 r1
k z(·, T ) kL2 (0,1) ≤ Ce 1−δ (ε 6 (1− 5 ) k w krL16 (Q
T −ε,T )

5 3r2
(36) + ε 6 (1− 5
)
k wx krL22 (QT −ε,T ) ),

where QT −ε,T = (0, 1) × (T − ε, T ). In turn by Lemma 2.1, applied on (T − ε, T )

with v = 0:
k w kL6 (QT −ε,T ) , k wx kL2 (QT −ε,T )

(37) ≤ Ce(α+̺+µ/2)ε k u(·, T − ε) kL2 (0,1) ,

for some constant C > 0.

Step 4. Summarizing the above estimates, we select parameters α, ε, and s

so that
(A) s → 0+;
1
(B) eαε = ;
s
 Bilinear control for the semilinear equations 153

(C) ε → 0, so that
2αε 5 r1 5 3r2 5 r1 5 3r2 2
)} − 1−δ
e 1−δ εmin{ 6 (1− 5 ); 6 (1− 5
)}
= εmin{ 6 (1− 5 ); 6 (1− 5 s → 0.

Under these conditions we have, firstly, that, in view of (A), (B) and (30),
(31), the right-hand side of (37) is bounded above by a constant and, secondly,
that, by (C) and (36),
k z(·, T ) kL2 (0,1) → 0.
Then from (32), (34) and (35) this yields that

(38) k u(·, T ) − uT kL2 (0,1) → 0,

which completes the proof of Theorem 1.3.

5. Proof of Theorems 1.1/1.2. This proof, in fact, is identical to that

of Theorem 1.3, with the following minor differences.

• In the proof of Theorem 1.3, in Step 2, we can select control v̂T /2 first
and then, as ε → 0, apply it only on the interval (T − ε, T − ε − 0.5),
i.e., the same additive control (but shifted in time) for all ε ∈ [0, T /2]. In
this way Remark 3.1 is not necessary to use in (31). Analogously, in the
proof of Theorem 1.1/1.2, in place of v̂t in the above, we can select any
function v = v̂(x, t), t ∈ (0, T /2), x ∈ ω. Then the argument of Theorem
1.3 will lead us to the convergence as in (38) to uT = y(·, T /2), which is
the state of the truncated multidimensional linear version of (2.5) with
α = 0 generated by the selected v̂(x, t). It remains to recall that the
latter is approximately controllable in L2 (Ω) at time T /2 (or any other
positive time, due to the dual unique continuation property from an open
set ω × (0, T /2)), i.e., the set of such y(·, T /2) is dense in L2 (Ω).

• In several space dimensions Lemmas 2.1 and 2.2 are principally no

different from the one dimensional case.

6. Concluding remarks.

• It seems quite possible that the results of this article can be extended at
no extra cost to boundary controls in place of the additive ones.

• In the proof of Theorem 1.3 we followed the Fourier series approach, which
is due to the delicate nature of the lumped additive controls involving the
Riesz’s basis properties of the sequence of exponentials (see Section 3).
However, as we showed it in the (sketch of the) proof of Theorems 1.1/1.2,
154 Khapalov

this approach can be avoided in part when we are dealing with the “stable”
locally distributed controls. From this viewpoint it seems very plausible
that these theorems can be extended to the case of more general parabolic
equations with variable coefficients.

References.
[1] A. Baciotti, Local Stabilizability of Nonlinear Control Systems, Series on
Advances in Mathematics and Applied Sciences, vol. 8, World Scientific,
Singapore, 1992.
[2] J.M. Ball, J.E. Mardsen, and M. Slemrod, Controllability for distributed
bilinear systems, SIAM J. Contr. Opt., 1982, pp. 575-597.
[3] V. Barbu, Exact controllability of the superlinear heat equation, Appl.
Math. Opt., 42 (2000), pp. 73-89.
[4] M.E. Bradley, S. Lenhart, and J. Yong, Bilinear optimal control of the
velocity term in a Kirchhoff plate equation, J. Math. Anal. Appl.,238
(1999), 451-467.
[5] C. Fabre, J.-P. Puel and E. Zuazua, Approximate controllability for the
semilinear heat equations, Proc. Royal Soc. Edinburg,125A (1995), pp.
31-61.
[6] H.O. Fattorini and D.L. Russell, Uniform bounds on biorthogonal functions
for real exponentials with an application to the control theory of parabolic
equations, Quarterly of Appl. Mathematics, April, 1974, pp. 45-69.
[7] E. Fernandez-Cara, Null controllability for semilinear heat equation,
ESAIM: Control, Optimization and Calculus of Variations, (1997), pp.
87-103.
[8] L.A. Fernandez and E. Zuazua, Approximate controllability for the semilin-
ear heat equation involving gradient terms, J. Optim. Theory Appl.,101
(1999), pp. 307-328.
[9] A. Fursikov and O. Imanuvilov, Controllability of evolution equations, Lect.
Note Series 34, Res. Inst. Math., GARC, Seoul National University, 1996.
[10] A.Y. Khapalov, Some aspects of the asymptotic behavior of the solutions
of the semilinear heat equation and approximate controllability, J. Math.
Anal. Appl.,194 (1995), pp. 858-882.
[11] A.Y. Khapalov and R.R. Mohler, Reachable sets and controllability of
bilinear time-invariant systems: A qualitative approach, IEEE Trans. on
Autom. Control, 41 (1996), pp. 1342-1346.
 Bilinear control for the semilinear equations 155

[12] A.Y. Khapalov, Approximate controllability and its well-posedness for

the semilinear reaction-diffusion equation with internal lumped controls,
ESAIM: COCV, 4 (1999), pp. 83-98.

[13] A.Y. Khapalov, Global approximate controllability properties for the semi-
linear heat equation with superlinear term, Rev. Mat. Complutense, 12
(1999), pp. 511-535.

[14] A.Y. Khapalov, A class of globally controllable semilinear heat equations

with superlinear terms, J. Math. Anal. Appl., 242 (2000), pp. 271-283.

[15] O. H. Ladyzhenskaja, V.A. Solonnikov and N.N. Ural’ceva, Linear and

Quasi-linear Equations of Parabolic Type, AMS, Providence, Rhode
Island, 1968.

[16] I. Lasiecka and R. Triggiani, Exact controllability of semilinear abstract

systems with application to waves and plates boundary control problems,
Appl. Math. Optim.,23 (1991), pp. 109-154.

[17] S. Lenhart, Optimal control of convective-diffusive fluid problem, Math.

Models and Methods in Appl. Sci., 5 (1995), pp. 225-237.

[18] W.A.J. Luxemburg, and J. Korevaar, Entire functions and Müntz-Szász

type approximation, Trans. of the AMS,157 (1971), pp. 23-37.

[19] D. Tataru, Carleman estimates, unique continuation and controllability for

anizotropic PDE’s, Contemporary Mathematics, 209 (1997), pp. 267-279.
156 Khapalov
A Nonoverlapping Domain Decomposition for
Optimal Boundary Control of the Dynamic Maxwell
System

J. E. Lagnese 1 , Georgetown University, Washington, DC 20057 USA

1 Introduction
Let Ω be a bounded, open, connected set in IR3 with piecewise smooth, Lipschitz
boundary Γ, and let T > 0. We consider the Maxwell system
(
εE ′ − rot H + σE = F
µH ′ + rot E = 0 in Q := Ω × (0, T )
(1.1)
ν ∧ E − δ ν ∧ (H ∧ ν) = J on Σ := Γ × (0, T ), δ > 0,
E(0) = E0 , H(0) = H0 in Ω

Here ′ = ∂/∂t, ν denotes the exterior pointing unit normal vector to Γ and
ε = (εjk (x)), µ = (µjk (x)) and σ = (σ jk (x)) are 3×3 Hermitian matrices with
L∞ (Ω) entries with ε and µ uniformly positive definite on Ω and σ ≥ 0. When
J = 0 the boundary condition on Σ is the so-called Silver-Müller condition. The
function F ∈ L1 (0, T ; L2 (Ω)) is given while J is a control input and is taken
from the class

U = L2τ (Σ) := {J| J ∈ L2 (0, T ; L2 (Γ)), ν · J(t) = 0

for a.a. x ∈ Γ and a.a. t ∈ (0, T )}

Function spaces of C-valued functions are denoted by capital roman letters,

while function spaces of C 3 -valued functions are denoted by capital script
P letters.
We use α · β to denote the natural scalar product in C 3 , i.e., α · β = 3j=1 αj β j ,
and write h·, ·i for the natural scalar product in various function spaces such as
L2 (Ω) and L2 (Ω). A subscript may sometimes be added to avoid confusion. The
spaces L2 (Ω) and L2 (Ω) denote the usual spaces of Lebesque square integrable
C-valued functions and C 3 -valued functions, respectively.

1
Research supported by the National Science Foundation through grant DMS-9972034

157
158 J. E. Lagnese

We set H = L2 (Ω) × L2 (Ω) with weight matrix M = diag(ε, µ). Thus

k(φ, ψ)k2H = hεφ, φi + hµψ, ψi

It will be proved below that for J ∈ U and (E0 , H0 ) ∈ H, (1.1) has a unique
solution (E, H) ∈ C([0, T ]; H). We shall consider the optimal control problem
Z
(1.2) inf |J|2 dΣ + kk(E(T ), H(T )) − (E1 , H1 )k2H , k > 0,
J∈U Σ

subject to (1.1), where (E1 , H1 ) ∈ H is given. Problem (1.1), (1.2) admits

a unique optimal control Jopt which is given through the optimality system
consisting of (1.1),
(
εP ′ − rot Q − σP = 0
µQ′ + rot P = 0 in Q
(1.3)
ν ∧ P + δ ν ∧ (Q ∧ ν) = 0 on Σ
(P (T ), Q(T )) = k((E(T ), H(T )) − (E1 , H1 )) in Ω,

and

(1.4) Jopt = Qτ := ν ∧ (Q ∧ ν)|Σ = Q|Σ − (Q|Σ · ν)ν.

The purpose of this paper is to develop a domain decomposition procedure to

approximate the solution of the above optimality system.
Remark 1.1. One may work with controls J supported on Σ1 := Γ1 ×(0, T ),
where Γ1 is a nonempty, relatively open subset of Γ. Outside of the support
of J one may replace the Silver-Müller boundary condition by (for example)
ν ∧ E = 0. The term −δ ν ∧ (H ∧ ν) represents boundary damping and is
included in order to improve the regularity of the solution of (1.1). Without it
the domain decomposition procedure becomes considerably more complicated
(see section 5 for a discussion of this case).
Prior work on domain decomposition (DD) for the dynamic Maxwell
systems seems to have been confined to the direct problem and to time harmonic
solutions; see [1], [2], [10], [11]. There is also little previous work on DD in the
context of optimal boundary control of dynamic partial differential equations
with penalization of the final state. For one-dimensional problems, we can refer
to the work of Leugering [14] - [18] dealing with networks of dynamic string or
beam equations on 1-d graphs. For higher dimensional problems, there seems
to be only the paper [13], which is concerned with transmission problems for
wave equations. On the other hand, for DD in other types of optimal control
problems let us mention the work by Benamou [3] - [8], and Benamou and
Després [9], where elliptic, parabolic and hyperbolic problems with constant
coefficients in Ω are considered together with a cost functional which involves
 Domain Decomposition for the Maxwell System 159

the entire state over space and time (in addition to the control). In these papers
the authors use an extension of P. L. Lions’ method [19], originally obtained
for elliptic problems. This same principle was employed in [13] and will also be
suitable adapted to the Maxwell system considered in the present paper.
The remainder of this paper is organized as follows. Existence, uniqueness
and regularity of solutions of (1.1) is examined in the next section. A domain
decomposition procedure for the optimality system (1.1) - (1.4) is introduced in
Section 3, and its convergence is studied in section 4. Remarks concerning the
case δ = 0, and limit of the optimality system and its domain decomposition
as δ → 0, are provided in section 5.

2 Existence and uniqueness of solution

We set

V = {φ ∈ L2 (Ω) : rot φ ∈ L2 (Ω), ν ∧ φ ∈ L2τ (Γ)},

Z Z
kφkV = (|φ| + | rot φ| )dx + |ν ∧ φ|2 dΓ.
2 2 2
Ω Γ

Let us first consider the problem

(
εφ′ − rot ψ + σφ = εf
µψ ′ + rot φ = µg in Q
(2.1)
ν ∧ φ − δ ψτ = 0 on Σ
φ(0) = φ0 , ψ(0) = ψ0 in Ω

Set
 
−1 −σ rot
A=M ,
− rot 0
Dom(A) = {(φ, ψ) ∈ V × V : ν ∧ φ − δ ψτ = 0 on Γ}.

The system (2.1) may be formally written

   
′ φ f
U = AU + F, where U = ,F =
ψ g
(2.2)  
φ0
U (0) = U0 :=
ψ0

Lemma 2.1. A is a m-dissipative operator in H.

Proof. A is densely defined and from the Green’s formula
Z
(2.3) hφ, rot ψi = hrot φ, ψi − ν · (φ ∧ ψ)dΓ, (φ, ψ) ∈ V × V,
Γ
160 J. E. Lagnese

we obtain
Z
2 √
hAU, U iH = −hσφ, φi − |ν ∧ φ|2 dΓ + 2 −1 Imhrot ψ, φi, U ∈ D(A),
δ Γ

so RehAU, U iH ≤ 0.
Let (f, g) ∈ H and let φ be the unique solution in V of the variational
equation
Z
−1 1
(2.4) h(ε + σ)φ, χi + hµ rot φ, rot χi + (ν ∧ φ) · (ν ∧ χ) dΓ
δ Γ
= hg, rot χi + hεf, χi, ∀χ ∈ V.
Set ψ = g − µ−1 rot φ ∈ L2 (Ω). Then (2.4) reads
Z
1
hψ, rot χi = h(ε + σ)φ, χi − hεf, χi + (ν ∧ φ) · (ν ∧ χ) dΓ, ∀χ ∈ V.
δ Γ
It follows that rot ψ ∈ L2 (Ω) and that
(ε + σ)φ − rot ψ = εf in Ω, δ ψτ = ν ∧ φ on Γ.
     
φ φ f
Therefore ∈ D(A) and (I − A) = .
ψ ψ g
Corollary 2.1. (1) If (φ0 , ψ0 ) ∈ H and (f, g) ∈ L1 (0, T ; H), then (2.1)
has a unique mild solution (φ, ψ) ∈ C([0, T ]; H) and


k(φ(t), ψ(t))kL∞ (0,T ;H) ≤ C k(φ0 , ψ0 )kH + k(f, g)kL1 (0,T ;H) .
(2) If (φ0 , ψ0 ) ∈ D(A) and (f, g) ∈ C 1 ([0, T ]; H), then (φ, ψ) ∈
C([0, T ]; D(A)).
Lemma 2.2. Let (φ0 , ψ0 ) ∈ H and (f, g) ∈ L1 (0, T ; H). Then the solution
of (2.1) satisfies ν ∧ φ|Σ ∈ L2τ (Σ).
Proof. First suppose that (φ0 , ψ0 ) ∈ D(A) and (f, g) ∈ C 1 ([0, T ]; H). We then
have
Z t Z Z
1 2 1 2 1 t
k(φ(t), ψ(t))kH − k(φ0 , ψ0 )kH + hσφ, φi dt + |ν ∧ φ|2 dΓdt
2 2 0 δ 0 Γ
Z t Z t
√
− 2 −1 hrot ψ, φi dt = h(f, g), (φ, ψ)iH dt.
0 0

It follows easily that

Z T Z T Z
1
k(φ, ψ)k2L∞ (0,T ;H) + hσφ, φi dt + |ν ∧ φ|2 dΓdt
0 δ 0 Γ
n o
≤ C k(φ0 , ψ0 )k2H + k(f, g)k2L1 (0,T ;H) .

The result now follows by density.

 Domain Decomposition for the Maxwell System 161

By transposition, we have
Theorem 2.1. If (E0 , H0 ) ∈ H, F ∈ L2 (0, T ; L2 (Ω)) and J ∈ L2τ (Σ), (1.1)
has a unique solution (E, H) ∈ C([0, T ]; H).
We remark that (E, H) satisfies
Z T
(2.5) h(E(T ), H(T )), (φ0 , ψ0 )iH − h(E(t), H(t)), (f (t), g(t))iH dxdt
0
Z T Z
1
= h(E0 , H0 ), (φ(0), ψ(0))iH + hF (t), φ(t)idt + J · (ν ∧ φ) dΣ
0 δ Σ
for all (φ0 , ψ0 ) ∈ H and (f, g) ∈ L1 (0, T ; H), where (φ, ψ) is the solution of
(
εφ′ − rot ψ − σφ = εf
µψ ′ + rot ψ = µg in Q
(2.6)
ν ∧ φ + δ ψτ = 0 on Σ
φ(T ) = φ0 , ψ(T ) = ψ0 in Ω

Theorem 2.2. If (E0 , H0 ) ∈ H, F ∈ L2 (0, T ; L2 (Ω)) and J ∈ L2τ (Σ), the

solution of (1.1) satisfies ν ∧ E|Σ ∈ L2τ (Σ).
Proof. If J = 0 the result follows from Lemma 2.2. Suppose that F = 0
and assume that (E0 , H0 ) and J are such that (E, H) ∈ C([0, T ]; V × V ).
This will hold if, for example, (E0 , H0 ) ∈ V × V , J = ν ∧ J| ˆ Σ , where
Jˆ ∈ C ([0, T ]; L (Ω)) ∩ C ([0, T ]; H (Ω)), and ν ∧ E0 − δ H0τ = J(0) on Γ.
2 2 1 1

As may be seen by making the change of variables Ê = E − J, ˆ Ĥ = H, one

has (Ê, Ĥ) ∈ C([0, T ]; D(A)) and therefore (E, H) ∈ C([0, T ]; V × V ). By
calculating as in Lemma 2.2 we obtain
Z t Z Z
1 2 1 t
k(E(t), H(t))kH + hσE, Ei dt + |ν ∧ E|2 dΓdt
2 0 δ 0 Γ
Z tZ
1 2 1
= k(E0 , H0 )kH + Re J · (ν ∧ E) dΓdt.
2 δ 0 Γ

From the boundary condition we have

1

Re J · (ν ∧ E) = |ν ∧ E|2 + |J|2 − δ2 |Hτ |2 ,
2
from which one obtains
Z t Z tZ
1
(2.7) k(E(t), H(t))k2H + 2 hσE, Ei dt + ( |ν ∧ E|2 + δ|Hτ |2 ) dΓdt
0 0 Γ δ
Z Z
2 1 t
= k(E0 , H0 )kH + |J|2 dΓdt.
δ 0 Γ
The result now follows by density.
162 J. E. Lagnese

3 Domain decomposition
In this section we describe an iterative domain decomposition for the optimality
system (1.1), (1.3), (1.4). Let {Ωi }m 3
i=1 be bounded domains in IR with piecewise
smooth, Lipschitz boudaries such that
m
[
Ωi ∩ Ωj = ∅, i 6= j, Ωi ⊂ Ω, i = 1, . . . , m, Ω= Ωi .
i=1

We set
[
Γij = ∂Ωi ∩ ∂Ωj = Γji , i 6= j, Γi = ∂Ωi ∩ Γ, γi = Γij .
j:Γij 6=∅

Then ∂Ωi = γi ∪ Γi . It is assumed that each Γi and Γij is either empty or has
a nonempty interior. We further set

Qi = Ωi × (0, T ), Σij = Γij × (0, T ), Σi = Γi × (0, T ), Si = γi × (0, T ).

Let (E, H), (P, Q) be the solution of the optimality system (1.1), (1.3), (1.4)
with (E0 , H0 ) ∈ H. The global optimality system may be formally expressed
as the coupled system
(
εi Ei′ − rot Hi + σi Ei = Fi
µi Hi′ + rot Ei = 0 in Qi
(3.1)
νi ∧ Ei − δHiτ = Qiτ on Σi
Ei (0) = E0i , Hi (0) = H0i in Ωi ,

(
εi Pi′ − rot Qi − σi Pi = 0
µi Q′i + rot Pi = 0 in Qi
(3.2)
νi ∧ Pi + δQiτ = 0 on Σi
(Pi (T ), Qi (T )) = k((Ei (T ), Hi (T )) − (E1i , H1i )) in Ωi ,

together with the interface conditions

(
νi ∧ Ei = −νj ∧ Ej , νi ∧ Hi = −νj ∧ Hj ,
(3.3)
νi ∧ Pi = −νj ∧ Pj , νi ∧ Qi = −νj ∧ Qj on Σij .

The subscript i on E, H, P, Q indicates restriction to Qi ; for the coefficients

and data the subscript indicates restriction to Ωi . The vector νi is the unit
exterior pointing normal vector to ∂Ωi . The subscript τ indicates the tangential
component, c.f. (1.4). The interface conditions (3.3) are realized in a weak sense
 Domain Decomposition for the Maxwell System 163

by the solution of the global optimality system and will hold in the sense of
traces if

rot E(t), rot H(t), rot P (t), rot Q(t) ∈ L2 (Ω), 0 < t < T.

Remark 3.1. Although the tangential components of E, H, P, Q are continuous

across an interface, in general there may be a discontinuity in their normal
components.
It is easy to see that (3.3) will hold if and only if

Eiτ := νi ∧ (Ei ∧ νi ) = Ei − (Ei · νi )νi = Ejτ on Σij

and similarly for the remaining three interface conditions. We further note that
(3.3) is equivalent to

νi ∧ Ei − αHiτ − βQiτ = −νj ∧ Ej − αHjτ − βQjτ

(3.4)
νi ∧ Pi + αQiτ − βHiτ = −νj ∧ Pj + αQjτ − βHjτ ,

where α, β are nonzero constants. Indeed, by interchanging i and j in (3.4)

and adding the results to (3.4) we find that E and P satisfy (3.3) and that

αHiτ + βQiτ = αHjτ + βQjτ , αQiτ − βHiτ = αQjτ − βHjτ ,

hence Hiτ = Hjτ and Qiτ = Qjτ .

Let us now consider the local iterations
(
εi (Ein+1 )′ − rot Hin+1 + σi Ein+1 = Fi
µi (Hin+1 )′ + rot Ein+1 = 0 in Qi
(3.5)
νi ∧ Ein+1 − δHiτn+1
= Qn+1iτ on Σi
Ein+1 (0) = E0i , Hin+1 (0) = H0i in Ωi ,

(
εi (Pin+1 )′ − rot Qn+1
i − σi Pi = 0
n+1 ′ n+1
µi (Qi ) + rot Pi =0 in Qi
(3.6)
νi ∧ Pin+1 + δQn+1
iτ = 0 on Σi
(Pin+1 (T ), Qn+1
i (T )) = k((Ein+1 (T ), Hin+1 (T )) − (E1i , H1i )) in Ωi ,

(
n+1
νi ∧ Ein+1 − αHiτ − βQn+1
iτ
n − βQn ,
= −νj ∧ Ejn − αHjτ jτ
(3.7) n+1 n+1 n+1 n n n
νi ∧ Pi + αQiτ − βHiτ = −νj ∧ Pj + αQjτ − βHjτ on Σij ,

Set

λnij = −νj ∧ Ejn − αHjτ

n
− βQnjτ |Σij
ρnij = −νj ∧ Pjn + αQnjτ − βHjτ
n
|Σij
164 J. E. Lagnese

and let Hi = L2 (Ωi ) × L2 (Ωi ) with weight matrix Mi = diag(εi , µi ). We denote

by h·, ·ii the natural scalar product in L2 (Ωi ).
Theorem 3.1. Assume that α > 0, β > 0. If λnij , ρnij ∈ L2τ (Σij ), ∀j :
Γij 6= ∅, problem (3.5) - (3.7) has a unique solution such that (Ein+1 , Hin+1 ) ∈
C([0, T ]; Hi ), (Pin+1 , Qn+1
i ) ∈ C([0, T ]; Hi ), and all of the traces appearing in
the boundary conditions in (3.5) - (3.7) are in L2τ .
Proof. This theorem is a consequence of the fact that (3.5) - (3.7) is the
optimality system for the local optimal control problem
Z Z
2 1 X 
|Ji | dΣ + |Jij |2 + |βHiτ + ρnij |2 dΣ
Σi β Σij
j:Γij 6=∅

+ kk(Ei (T ) − E1i , Hi (T ) − H1i )k2Hi → inf

Ji ,Jij

subject to
(
εi Ei′ − rot Hi + σi Ei = Fi
µi Hi′ + rot Ei = 0 in Qi
(3.8) νi ∧ Ei − δHiτ = Ji on Σi
νi ∧ Ei − αHiτ = Jij + λnij on Σij
Ei (0) = E0i , Hi (0) = H0i in Ωi ,

where Ji ∈ L2τ (Σi ), Jij ∈ L2τ (Σij ), as may be directly verified. Since α > 0,
problem (3.8) has the same structure as (1.1) and, from Theorems 2.1 and 2.2,
its solution satisfies (Ei , Hi ) ∈ C([0, T ]; Hi ), νi ∧ Ei , Hiτ ∈ L2τ (Σi ) ∪ L2τ (Si ).
Therefore (Pin+1 , Qn+1
i ) satisfies a system of the form
(
εi Pi′ − rot Qi − σi Pi = 0
µi Q′i + rot Pi = 0 in Qi
νi ∧ Pi + δQiτ = 0 on Σi
νi ∧ Pi + αQiτ ∈ L2τ (Σij ) on Σij
(Pi (T ), Qi (T )) ∈ Hi ,

and therefore (Pin+1 , Qn+1

i ) ∈ C([0, T ]; Hi ), νi ∧ Pi , Qiτ ∈ L2τ (Σi ) ∪ L2τ (Si ).
As a consequence of Theorem 3.1, it follows that the iteration (3.5) - (3.7)
is well defined if α > 0, β > 0, and λ0ij , ρ0ij ∈ L2τ (Σij ), ∀j : Γij 6= ∅.

4 Convergence
In this section we prove that the solutions {(Ein+1 , Hin+1 )}m i=1 ,
{(Pin+1 , Qn+1
i )}m
i=1 of the local optimality systems (3.5) - (3.7) converge
 Domain Decomposition for the Maxwell System 165

to the solution {(Ei , Hi )}m m

i=1 , {(Pi , Qi )}i=1 of the global optimality system
(3.1) - (3.3). Set
en , H
(E e n ) = (E n − Ei , H n − Hi ), (Pein , Q
en ) = (P n − Pi , Qn − Qi ), n ≥ 1.
i i i i i i i
en , H
Then (E e n ), (Pen , Qen ) satisfy
i i i i
(
εi (Ee n+1 e n+1 + σi E
)′ − rot H e n+1 = 0
i i i
µi (He n+1 )′ + rot Een+1 = 0 in Qi
i i
(4.1)
νi ∧ Ee n+1 − δH e n+1 = Qe n+1 on Σi
i iτ iτ
e n+1 e n+1
E (0) = H
i (0) = 0 in Ωi ,
i
(
e n+1 − σi Pei = 0
εi (Pein+1 )′ − rot Q i
µi (Qen+1 )′ + rot Pen+1 = 0 in Qi
i i
(4.2)
νi ∧ Pein+1 + δQ
en+1 = 0 on Σi
iτ
e n+1 e n+1 e e n+1 (T )) in Ωi ,
(Pi (T ), Qi (T )) = k(Ein+1 (T ), H i
(
e n+1 − αH
νi ∧ E e n+1 − β Q
en+1 = λ̃n ,
i iτ iτ ij
(4.3)
e
νi ∧ Pin+1 e n+1 e n+1 n
+ αQiτ − β Hiτ = ρ̃ij on Σij ,
where
en − αH
λ̃nij = −νj ∧ E e n − βQ
en ,
j jτ jτ
n e n e n e n
ρ̃ij = −νj ∧ Pj + αQjτ − β Hjτ .
Lemma 4.1. Assume that α > 0, β > 0, and that λ̃0ij , ρ̃0ij ∈ L2τ (Σij ),
∀j : Γij 6= ∅. i = 1, . . . , m. Then
E n+1 = E n − (F n+1 + F n ), n = 1, 2, . . . ,
where
m Z
1X  α2 + β 2
en+1 |2 )
e n+1 |2 + |Q
E n+1 = (|H iτ iτ
2 Si β
i=1
1
+ e n+1 |2 + |νi ∧ Pen+1 |2 ) dΣ,
(|νi ∧ Ei i
β

m 
X α

F n+1
= k+ (1−k2 ) k(E en+1 (T ), H e n+1 (T ))k2 + α k(Pen+1 (0), Q en+1 (0))k2
i i Hi i i Hi
2β 2β
i=1
Z
α T en+1 , Een+1 ii + hσi Pen+1 , Pen+1 ii ) dt
+ (hσi Ei i i i
β 0
Z 
 α n+1 2 α e n+1 2 α
+ δ |H e | + (1 + δ )|Q | + Re (H e n+1 ) dΣ .
e n+1 · Q
Σi β iτ β iτ
β iτ iτ
166 J. E. Lagnese

Proof. From Green’s formula (2.3) we have

Z T
(4.4) 0 = en+1 )′ − rot H
{hεi (E e n+1 , Pen+1 ii
e n+1 + σi E
i i i i
0
+ hµi (H en+1 , Q
e n+1 )′ + rot E en+1 ii } dt = kk(E e n+1 (T ))k2
en+1 (T ), H
i i i i i Hi
Z Z
+ |Qe n+1 | dΣ + Re {H e n+1 · (νi ∧ Pen+1 ) + (νi ∧ E en+1 } dΣ.
e n+1 ) · Q
iτ iτ i i iτ
Σi Si

Use of (4.3) in the last integral in (4.4) gives

Z
e n+1 e n+1
(4.5) 0 = kk(Ei (T ), Hi (T ))kHi + 2 en+1 | dΣ
|Q iτ
Σi
Z
+ {β(|He n+1 |2 + |Qen+1 |2 ) + Re (H e n+1 · λ̃nij )} dΣ
e n+1 · ρ̃nij ) + Re (Q
iτ iτ iτ iτ
Si

We have

2
e n+1 · ρ̃n ) = 1 |νi ∧ Pe n+1 |2 + α |Q
Re (H en+1 |2 − β |H e n+1 |2
iτ ij i
2β 2β iτ 2 iτ
1 α
− |ρ̃nij |2 + Re (νi ∧ Pein+1 ) · Q en+1
iτ
2β β
2
en+1 · λ̃nij ) = 1 |νi ∧ E
Re (Q en+1 |2 + α |He n+1 |2 − β |Q e n+1 |2
iτ i
2β 2β iτ 2 iτ
1 α e n+1 .
e n+1 ) · H
− |λ̃nij |2 − Re (νi ∧ E i iτ
2β β

Substitution into (4.5) yields

Z
en+1 (T ), H
(4.6) 0 = kk(E e n+1 (T ))k2 + en+1 | dΣ
|Q
i i Hi iτ
Σi
Z
 α2 + β 2
+ (|H e n+1 |2 ) + 1 (|νi ∧ E
e n+1 |2 + |Q en+1 |2 + |νi ∧ Pen+1 |2 )
iτ iτ i i
Si 2β 2β
α
+ Re((νi ∧ Pein+1 ) · Q e n+1 − (νi ∧ E
iτ
en+1 ) · H
i
e n+1 ) dΣ
iτ
β
Z
1 X
− (|λ̃nij |2 + |ρ̃nij |2 )dΣ.
2β Σij j:Γij 6=∅
 Domain Decomposition for the Maxwell System 167

From (4.1) and Green’s formula we calculate

Z T
0= en+1 )′ − rot H
{hεi (E e n+1 + σi E
en+1 , E
en+1 ii
i i i i
0
e n+1 )′ + rot E
+ hµi (H e n+1 ii } dt = 1 k(E
e n+1 , H en+1 (T ), H e n+1 (T ))k2
i i i i i Hi
2
Z T
√
+ {hσi E en+1 ii − 2 −1 ImhH
en+1 , E e n+1 , rot Ee n+1 i}dt
i i i i
0
Z TZ
+ He n+1 · (νi ∧ Ee n+1 ) dΣ
iτ i
0 ∂Ωi

and therefore
Z
(4.7) − Re e n+1 · (νi ∧ E
H e n+1 ) dΣ = 1 k(E
en+1 (T ), He n+1 (T ))k2H
iτ i i i
Si 2 i

Z T Z
+ hσi E en+1 ii dt +
en+1 , E {δ|He n+1 |2 + Re (He n+1 · Qe n+1 )} dΣ.
i i iτ iτ iτ
0 Σi

Similarly,
Z
(4.8) Re e n+1 · (νi ∧ Pen+1 ) dΣ = 1 k(Pen+1 (0), Q
Q en+1 (0))k2
iτ i i i Hi
Si 2
Z T Z
k2 en+1 e n+1 2 e n+1 e n+1 e n+1 |2 dΣ.
− k(Ei (T ), Hi (T ))kHi + hσi Pi , Pi ii dt + δ |Q iτ
2 0 Σi

Upon substituting (4.7) and (4.8) into (4.6) we obtain

α

(4.9) 0 = Ein+1 + k + en+1 (T ), H
(1 − k2 ) k(E i
e n+1 (T ))k2H
i
2β i

Z T
α en+1 en+1 (0))k2H + α e n+1 , E
en+1 ii + hσi Pen+1 , Pen+1 ii ) dt
+ k(Pi (0), Q i (hσi Ei i i i
2β i
β 0
Z
 α n+1 2 α en+1 2 α
δ |He e n+1 · Q e n+1 ) dΣ
+ iτ | + (1 + δ )|Qiτ | + Re (Hiτ iτ
Σi β β β
Z
1 X
− (|λ̃nij |2 + |ρ̃nij |2 )dΣ,
2β Σij j:Γij 6=∅

where
Z
1  α2 + β 2
Ein+1 = e n+1 |2 ) + 1 (|νi ∧ E
e n+1 |2 + |Q
(|H en+1 |2 + |νi ∧ Pen+1 |2 ) dΣ.
iτ iτ i i
2 Si β β
168 J. E. Lagnese

We next calculate

1 1 en |2 + |νi ∧ Pen |2 )
(4.10) (|λ̃n |2 + |ρ̃nij |2 ) = (|νi ∧ E
2β ij 2β j j

α2 + β 2 e n 2 enjτ |2 ) + α Re{(νj ∧ E ejn ) · He jτ

+ (|Hjτ | + |Q n
− (νj ∧ Pejn ) · Q
enjτ }
2β β
+ Re{(νj ∧ E ejn ) · Qe njτ + (νj ∧ Pejn ) · H
e jτ
n
}.

It follows from (4.9) and (4.10), upon summing over i, that

m 
X α

0 = E n+1 − E n + (1 − k2 ) k(E
k+ en+1 (T ), H e n+1 (T ))k2H
i i
2β i
i=1
Z T
α en+1 en+1 (0))k2 + α en+1 , Ee n+1 ii + hσi Pe n+1 , Pen+1 ii ) dt
+ k(Pi (0), Q i Hi (hσi Ei i i i
2β β 0
Z
 α n+1 2 α en+1 2 α
δ |He e n+1 · Q e n+1 ) dΣ
+ iτ | + (1 + δ )|Qiτ | + Re (Hiτ iτ
Σi β β β
Z
α ein ) · H
e iτ
− Re{(νi ∧ E n
− (νi ∧ Pein ) · Qe niτ }dΣ
β Si
Z 
− e n e n e n
Re{(νi ∧ Ei ) · Qiτ + (νi ∧ Pi ) · Hiτ }dΣ , e n
Si

that is,

(4.11) 0 = E n+1 − E n + F n+1

Xm  Z
α ein ) · H
e iτ
− Re{(νi ∧ E n
− (νi ∧ Pein ) · Q
eniτ }dΣ
β Si
i=1
Z 
+ e n e n e n e n
Re{(νi ∧ Ei ) · Qiτ + (νi ∧ Pi ) · Hiτ }dΣ .
Si

From (4.5) we have

m Z
X
(4.12) − ein ) · Q
Re{(νi ∧ E e niτ + (νi ∧ Pein ) · H
e iτ
n
}dΣ
i=1 Si
m 
X Z 
= e n e n 2
kk(Ei (T ), Hi (T ))kHi + e n 2
|Qiτ | dΣ ,
i=1 Σi
 Domain Decomposition for the Maxwell System 169

and from (4.7) and (4.8) we obtain

m
X Z
α en) · H
e n − (νi ∧ Pen ) · Q
e n }dΣ
(4.13) − Re{(νi ∧ Ei iτ i iτ
β Si
i=1
m
X 
α e in (T ))k2H + α k(Pein (0), Q
ein (T ), H eni (0))k2H
= (1 − k2 )|(E
2β i
2β i
i=1
Z
α T en, E
en ii + hσi Pen , Pen ii ) dt
+ (hσi E i i i i
β 0
Z 
α e n 2 e n 2 e n e n
+ (δ|Hiτ | + δ|Qiτ | + Re(Hiτ · Qiτ ))dΣ .
β Σi

Thus the sum of (4.12) and (4.13) equals F n so that (4.11) may be written

(4.14) E n+1 = E n − (F n+1 + F n ).

We now prove that under some additional hypotheses, (4.14) implies

convergence of the iteration scheme.
Theorem 4.1. In addition to the assumptions of Lemma 4.1, suppose that
r !
1 β β2
δ> − + + 1 , α/β < 2k/(k2 − 1) if k > 1,
2 α α2

that ε and µ are scalar valued functions with εi , µi ∈ C 2 (Ωi ). Then for
i = 1, . . . , m,

(Eein , H eni ) → 0 weakly* in L∞ (0, T ; Hi )

e in ) → 0, (Pein , Q
(Ee n (T ), He n (T )) → 0, (Pen (0), Qen (0)) → 0 strongly in Hi
i i i i
e n e n e n
νi ∧ Ei |Σi → 0, Hiτ |Σi → 0, νi ∧ Pi |Σi → 0, Q e niτ |Σ → 0 strongly in L2τ (Σi )
i

e e e e
νi ∧ Ei |Σ → 0, Hiτ |Σ → 0, νi ∧ Pi |Σ → 0, Qiτ |Σ → 0 weakly in L2τ (Σij )
n n n n
ij ij ij ij

Proof. It follows from (4.14) that

n+1
X′
n+1 1
E = E −2 F p,
p=1

n+1 n
X′ X
where cp = (c1 + cp+1 )/2 + cp . Under the stated condition on the
p=1 p=2
parameter δ, the quadratic form
α e n+1 2 α en+1 2 α e n+1 )
e n+1 · Q
δ |H iτ | + (1 + δ )|Qiτ | + Re (Hiτ iτ
β β β
170 J. E. Lagnese

is positive definite. (Note that any δ > 1/2 will satisfy the hypothesis.) If k > 1
and if we further restrict α/β so that
α

k+ (1 − k2 ) > 0.
2β
then
∞
X
F p converges and {E n }∞
n=1 is a bounded sequence.
p=1
P
The convergence of F p then implies that
en (T ), H
(E e n (T )) → 0 and (Pen (0), Q
en (0)) → 0 strongly in Hi , i = 1, . . . , m,
i i i i
e |Σ → 0 and Q
H n e |Σ → 0 strongly in L2 (Σi ), i = 1, . . . , m,
n
iτ i iτ i τ
en+1 → 0, σi Pe n+1 → 0 weakly in L2 (Qi ).
σi Ei i

From (4.1) and (4.2),

e n+1 |Σ → 0, νi ∧ Pe n+1 |Σ → 0 strongly in L2 (Σi )
νi ∧ Ei i i i τ
e n e n
(P (T ), Q (T )) → 0 strongly in Hi .
i i

The boundedness of E n implies that λ̃nij , ρ̃nij are bounded in L2τ (Σij ). Therefore

en+1 , H
(E en+1 ) are bounded in L∞ (0, T ; Hi ), i = 1, . . . , m.
e n+1 ), (Pen+1 , Q
i i i i

It follows that, on a subsequence n = nk of the positive integers,

en+1 , H
(E ei , H
e n+1 ) → (E e i ) weakly* in L∞ (0, T ; Hi )
i i
ei ) weakly* in L∞ (0, T ; Hi )
en+1 ) → (Pei , Q
(Pen+1 , Q
i i
(4.15)
en+1 → Ai ,
νi ∧ E e n+1 → Bi weakly in L2τ (Si )
H
i iτ
e
νi ∧ P n+1 → Ci , e n+1
Qiτ → Di weakly in L2τ (Si )
i

ei , H
for some Ai , Bi , Ci Di ∈ L2τ (Si ), (E e i ), (Pei , Q
ei ) ∈ L∞ (0, T ; Hi ).
∞
Let (φ, ψ) ∈ C (Ω × [0, T ]). We have
Z T
0= en+1 )′ − rot H
[hεi (E e n+1 + σi E
en+1 , φi + hµi (H e n+1 )′ + rot E en+1 , ψi] dt
i i i i i
0
Z T
en+1 (T ), H
= h(E e n+1 (T )), (φ(T ), ψ(T ))iH − en+1 , εi φ′ − rot ψ − σi φi
[hE
i i i i
0
Z
+ hHe n+1 , µi ψ ′ + rot φi] dt + e n+1 · (νi ∧ φ) + (νi ∧ E
{H e n+1 ) · ψτ }dΣ
i i i
Σi
Z
+ {H e n+1 · (νi ∧ φ) + (νi ∧ E e n+1 ) · ψτ }dΣ
i i
Si
 Domain Decomposition for the Maxwell System 171

Upon passing to the limit through the subsequence n = nk we obtain

Z T
ei , εi φ′ − rot ψi + hH
[hE e i , µi ψ ′ + rot φi] dt
0 Z
= {Bi · (νi ∧ φ) + Ai · ψτ }dΣ, ∀(φ, ψ) ∈ C ∞ (Ω × [0, T ]).
Si

Therefore
(
e′ − rot H
εi E ei = 0
i
µi H ei = 0
e ′ + rot E in Qi
i

(4.16) νi ∧ Eei = He iτ = 0 on Σi
νi ∧ Eei = Ai , He iτ = Bi on Si
ei (0) = H
E e i (0) = E
ei (T ) = H
e i (T ) = 0 in Ωi .

Similarly,
(
ei = 0
εi Pei′ − rot Q
µi Qe ′ + rot Pei = 0 in Qi
i

(4.17) νi ∧ Pei = Qe iτ = 0 on Σi
νi ∧ Pei = Ci , Qe iτ = Di on Si
Pei (0) = Q
e i (0) = Pei (T ) = Q
e i (T ) = 0 in Ωi .

We may now use a unique continuation argument to conclude that E ei =

e e e
Hi = 0 and Pi = Qi = 0 in Qi for i = 1, . . . , m, provided that (4.16) and (4.17)
hold. Indeed, suppose that Ωi is a region adjacent to Γ, that is Γi 6= ∅. From
(4.16) we have
Z t Z t
ei (t) =
εi E e i (s) ds,
rot H e i (t) = −
µi H ei (s) ds,
rot E
0 0

hence

(4.18) div(εi E e i ) = 0 in Qi ,
ei ) = div(µi H

ei ) = νi · (µi H
νi · (εi E e i ) = 0 on Σi ,

since νi · rot φ is a tangential differential operation on νi ∧ φ on Σi . Since

eiτ = E
0=E ei − (E
ei · νi )νi on Σi
172 J. E. Lagnese

we have

ei ) = (E
0 = νi · (εi E ei · νi )νi · (εi νi ) on Σi ,

ei = 0 on Σi since µi is positive definite. Therefore

which implies that νi · E

ei = νi · E
νi ∧ E ei = 0 on Σi

and, similarly, for H e i . Therefore (Eei , H

e i ) is a solution of the dynamic Maxwell
system and of (4.18) in Qi , and has zero Cauchy data on Σi . Moreover, since
(Eei , H
e i ) vanish at t = 0 and t = T , they may be continued by zero to Ω×{t < 0}
and to Ω × {t > T } as solutions of the Maxwell system satisfying (4.18) having
zero Cauchy data on Γi × (−∞, ∞). It then follows from a result of Eller [12,
Corollary 5.2] that E ei = H e i = 0 in Qi . Therefore, E ei = H
e i = 0 in Qi for every
index i such that Γi 6= ∅. The argument is the same for (Pei , Q e i ). It follows that
for such i we have Ai = Bi = Ci = Di = 0 and therefore the convergence in
(4.15) is through the entire sequence of positive integers.
Now suppose Ωj is a region adjacent to a boundary region, i.e., to a region
Ωi such that such that Γi 6= ∅. Then Σij = ∂Ωi ∩ ∂Ωj 6= ∅ and we have

(4.19) en+2 − αH
νi E e n+2 − β Q
e n+2 = −νj E
en+1 − αH
e n+1 − β Q
e n+1 ,
i iτ iτ j jτ jτ

(4.20) e n+1 − αH
νj E e n+1 − β Q
en+1 = −νi E
en − αH
e n − βQ
en on Σij .
j jτ jτ i iτ iτ

Since for the index i, convergence is through the entire sequence of positive
integers, if we pass to the weak L2τ (Σij ) limit in (4.19) and (4.20) through the
subsequence n = nk we obtain

−Aj − αBj − βDj = 0, Aj − αBj − βDj = 0 on Σij .

Similarly,

−Cj + αDj − βBj = 0, Cj + αDj − βBj = 0 on Σij ,

hence Aj = Bj = Cj = Dj = 0, that is to say

(4.21) ej = αH
νj ∧ E e jτ = νj ∧ Pej = β Q
e jτ = 0 on Σij .

The same unique continuation argument as above gives Eej = He j = Pej = Q ej = 0

in Qj . One may now proceed step-by-step into the remaining interior regions
ej = H
Ωj and conclude that E e j = Pej = Q
ej = 0 in Qj for j = 1, . . . , m.
 5. FURTHER COMMENTS 173

5 Further comments
Consider the problem (1.1) without boundary damping:
(
εE ′ − rot H + σE = F
µH ′ + rot E = 0 in Q
(5.22)
ν ∧ E = J on Σ
E(0) = E0 , H(0) = H0 in Ω

For (E0 , H0 ) ∈ H, F ∈ L1 (0, T ; L2 (Ω)) and J ∈ L2τ (Σ), (5.22) has a unique
solution which is continuous on [0, T ] into X ′ , where X ֒→ H is given by
X = {(φ, ψ) ∈ V × V : ν × φ|Γ = 0, div(µψ) ∈ L2 (Ω), ν · (µψ)|Γ = 0}
and X ′ denotes the dual space of X with respect to H. The appropriate cost
functional for (5.22) analogous to (1.2) is therefore
Z
(5.23) J (J) = |J|2 dΣ + kk(E(T ), H(T )) − (E1 , H1 )k2X ′
Σ

where (E1 , H1 ) ∈ H, and the optimality system for the problem inf J∈U J (J)
subject to (5.22) consists of (5.22),
(
εP ′ − rot Q − σP = 0
µQ′ + rot P = 0 in Q
(5.24)
ν ∧ P = 0 on Σ
(P (T ), Q(T )) = kA−1 ((E(T ), H(T )) − (E1 , H1 )) in Ω,

(5.25) Jopt = Qτ |Σ ,
where A : X 7→ X ′ is the canonical isomorphism. Note that (P (T ), Q(T )) is
determined by solving a stationary Maxwell-type system of 6 first order partial
differential equations. Therefore any DD for the optimality system (5.22), (5.24)
and (5.25) must include a DD procedure for the approximation of (P (T ), Q(T )).
This makes the overall approximation scheme considerably more complicated
than that given above for the regularized optimality system (see [13], where a
DD is given for Neumann boundary optimal control of the wave equation with
penalization of the final state, and which leads to similar considerations).
Now consider the regularized problem
(
ε(E δ )′ − rot H δ + σE δ = F
µ(H δ )′ + rot E δ = 0 in Q
(5.26)
ν ∧ E δ − δHτδ = J on Σ
E δ (0) = E0 , H δ (0) = H0 in Ω.
174 J. E. Lagnese

The solution (E δ , H δ ) is continuous on [0, T ] into H. Suppose one considers the

optimal control problem inf J∈U J (J) subject to (5.26), where J (J) is given by
(5.23). The optimality system is then given by (5.26),
(
ε(P δ )′ − rot Qδ − σP δ = 0
µ(Qδ )′ + rot P δ = 0 in Q
(5.27)
ν ∧ P δ + δQδτ = 0 on Σ
(P δ (T ), Qδ (T )) = kA−1 ((E δ (T ), H δ (T )) − (E1 , H1 )) in Ω,

δ
(5.28) Jopt = Qδτ |Σ .

It is possible to prove by compactness arguments that the solution of (5.26) -

(5.28) converges in a certain sense as δ → 0 to the solution of the optimality
system (5.22), (5.24), (5.25). However, by penalizing in the X ′ norm rather than
the H norm, a significant complication is introduced into the DD as noted above,
and the simplifications gained by regularizing the problem are lost. From this
point of view, it makes more sense to work with the cost functional (1.2) when
optimally controlling the damped system (5.26). But it cannot be expected
that the solution of the corresponding optimality system converges in any sense
as δ → 0 to the solution of (5.22), (5.24), (5.25) and, indeed, it does not.
Finally, let us remark that we have also considered the case k = ∞, that is,
the problem of minimum L2τ (Σ) exact controllability of the solution of (5.26)
to (E1 , H1 ) at time T . We have proved that, in this case, the solution of the
optimality system does converge as δ → 0 to the solution of the optimality
system for the limit problem, that is, the optimality system for the problem
of minimum L2τ (Σ) exact controllability of the solution of (5.22) to (E1 , H1 ) at
time T . Details will be published in a forthcoming paper.

Acknowledgment
The author thanks Matthias Eller for very helpful discussions regarding the
unique continuation argument used in section 4.

References

[1] A. Alonso and A. Valli, “A domain decomposition approach for heterogeneous

time-harmonic Maxwell equations,” Comput. Methods Appl. Mech. Engrg., 143
(1997), 97 - 112.
[2] A. Alonso and A. Valli, “An optimal domain decomposition preconditioner for
low frequency time-harmonic Maxwell equations,” Math. Comp., 68 (1999), 607 -
631.
 5. FURTHER COMMENTS 175

[3] J.-D. Benamou, “Décomposition de domaine pour le contrôle de systèmes gou-

vernés par des équations d’evolution,” CRAS Paris, Série 1, 324 (1997), 1065-
1070.
[4] J.-D. Benamou, “A domain decomposition method for the optimal control of
systems governed the Helmholtz equation,” Mathematical and Numerical aspects
of Wave Propagation, (G. Cohen, Ed.), SIAM 1995, 653–662.
[5] J.-D. Benamou, “A domain decomposition method with coupled transmission con-
ditions for the optimal control of systems governed by elliptic partial differential
equations,” SIAM J. Num. Anal., 36 (1995), 2401-2416.
[6] J.-D. Benamou, “A domain decomposition method for control problems,” in DD9
Proceedings 1996, Bergen, (P. Bjørstad et al., Ed.), DDM.org, 1998, 266–273.
[7] J.-D. Benamou, ”Résolution d’un cas test de contrôle optimal pour un système
gouvernè par l’équation des ondes à l’aide d’une méthode de décompostion de
domaine,” INRIA, Rapport de Rechereche No. 3095, 1997.
[8] J.-D. Benamou, ”Domain decomposition, optimal control of systems governed
by partial differential equations and synthesis of feedback laws,” J. Opt. Theory
Appl., 102 (1999), 15-36.
[9] J.-D. Benamou and B. Desprès, “A domain decomposition method for the
Helmholtz equation and related optimal control problems,” J. Comp. Physics,
136 (1997), 68-82.
[10] P. Collino, G. Delbue, P. Joly and A. Piancenti, “A new interface condition in
the non-overlapping domain decomposition method for the Maxwell equations,”
Comput. Methods Appl. Mech. Engrg., 148 (1997), 195 - 207.
[11] B. Després, P. Joly and E. Roberts, “A domain decomposition for the harmonic
Maxwell equation,” in Iterative Methods in Linear Algebra, Elsevier, Amsterdam,
1992, 475 - 484.
[12] M. Eller, “Uniqueness of continuation theorems,” in Direct and Inverse Problems
of Mathematical Physics, R. P. Gilbert, J. Kajiwara and Y. S. Xu, Eds., Kluver,
1999.
[13] J. Lagnese and G. Leugering, “Dynamic domain decomposition in approximate
and exact boundary control in problems of transmission for wave equations,”
SIAM J. Control Opt., 38 (2000), 503 - 537.
[14] G. Leugering, “On domain decomposition of controlled networks of elastic strings
with joint masses,” in Control and Estimation of Distributed Parameter Systems
(Vorau 1996), (F. Kappel, Ed.), ISNM, 126, Birkhäuser Verlag, Basel, 1998, 191-
205.
[15] G. Leugering, “On domain decomposition of optimal control problems for dynamic
networks of elastic strings,” Computational Optimization and Applications, to
appear.
[16] G. Leugering, “Dynamic domain decomposition of optimal control problems for
networks of strings and beams,” SIAM J. Control Opt., 37 (1999), 1649 - 1675.
[17] G. Leugering, “Domain decomposition of optimal control problems for multi-link
structures,” in ENUMATH 97 (Heidelberg), World. Sci. Publishing, River Edge,
NJ, 1998, 38 - 53.
[18] G. Leugering, “Dynamic domain decomposition of controlled networks of elastic
strings and joint masses,” in Control and partial differential equations (Marseille-
Luminy, 1997), ESIAM Proc., 4 (1998), 223 - 233.
[19] P.-L. Lions, “On the Schwarz alternating method 3”, The Third International
176 J. E. Lagnese

Symposium on Domain Decomposition Methods for Partial Differential Equations,

T. Chan and R. Glowinski, Eds.), Society for Industrial and Applied Mathematics,
Philadelphia, PA, 1990, 202-223.
[20] Z. Opial, “Weak convergence of the sequence of successive approximations for
nonexpansive mappings,” Bull. A.M.S., 73 (1967), 591-597.
Boundary Stabilizibility of a Nonlinear Structural
Acoustic Model Including Thermoelastic Effects

Catherine Lebiedzik, University of Virginia, Charlottesville, Virginia

Abstract

We are interested in a three-dimensional coupled PDE system arising

in problems dealing with the active control of structural acoustic systems.
A wave equation defined on a three-dimensional domain is coupled with
a nonlinear thermoelastic plate equation on a portion of the boundary.
The major issue studied here is the uniform stabilizibility of the entire
interactive model. Our principal result states that boundary nonlinear
dissipation placed on a suitable portion of the boundary suffices to stabilize
the system.

1 Introduction
Structural acoustic systems are typically modeled by a a three-dimensional
interactive system of partial differential equations(PDE) that consists of a
wave equation coupled at an interface with some sort of plate equation. An
undamped wave equation is defined on a three-dimensional bounded domain Ω
with boundary Γ. On a portion of the boundary (the interface labelled Γ0 ), the
wave equation in the chamber is coupled with a plate equation.
Though the issue of stability of wave and plate equations has attracted much
attention, there is less known about the behavior of these coupled (hybrid) PDE
structures. Only recently has significant progress been made in understanding
the nature of these interactive models [3, 1, 7, 23, 22, 16]. It becomes quickly
clear that understanding the interactions between different types of dynamics
is the key to stabilizing these systems. The works cited above deal exclusively
with linear constitutive laws corresponding to the PDE’s. This paper, on the
other hand, pertains to a nonlinear , large displacement theory in the context
of stabilization of hybrid structures. We consider the case where the interactive
portion of the boundary, Γ0 , is represented by a nonlinear thermoelastic plate
equation. Our main aim is to study the question of stability/stabilizibility of
the entire coupled model. A five-page announcement of the results of this paper
will appear in [15].
177
178 Catherine Lebiedzik

Fig. 1. Cross section of the domain Ω. The thick line denotes the area subject to
frictional damping g(zt ).

1.1 Statement of the Problem

Let Ω ∈ R3 be an open bounded domain with two dimensional boundary
Γ which is assumed sufficiently smooth. The boundary Γ consists of three
connected regions Γ = Γ̄1 ∪ Γ̄2 ∪ Γ̄0 , with Γ2 possibly empty. The pressure in
the acoustic medium is defined on Ω, whereas the displacement of the flexible
wall is defined on Γ0 . The other portions of the boundary, Γ1 and Γ2 , represent
‘hard’ walls, with Γ1 being the section subject to frictional forces.
The PDE model considered consists of the wave equation in the variable
z (where the quantity ρzt is the acoustic pressure, and ρ is the density of the
fluid)

ztt = c2 ∆z in Ω × (0, ∞)
∂
z = 0 on Γ2 × (0, ∞)
∂ν
∂
z = −g(zt ) − d z on Γ1 × (0, ∞)
∂ν
∂
(1.1) z = wt on Γ0 × (0, ∞)
∂ν

and the elastic equation representing the displacement of the wall w subject to
 Stability of a nonlinear structural acoustic model 179

thermal effects [See, e.g., [11]]:

(1.2)

wtt − γ∆wtt + ∆2 w = −∆θ − ρzt + [F(w), w] 
∆2 F(w) = −[w, w] on Γ0 × (0, ∞)

θt − ∆θ = ∆wt

w = ∂∂ν̃ w = 0; θ = 0
on ∂Γ0 × (0, ∞)
F(w) = ∂∂ν̃ F(w) = 0

Here θ is the temperature, c2 is the speed of sound as usual, and the constant
γ ≥ 0 accounts for rotational forces. The vector ν (respectively, ν̃ denotes the
outward unit normal vector to the boundary Γ, (respectively, ∂Γ0 ), and [u, v]
denotes the usual von Kármán bracket, i.e. [u, v] = uxx vyy + uyy vxx − 2uxy vxy .
The function g, a nonlinear boundary feedback control, represents frictional
damping and here is assumed continuous, monotone increasing, and zero at
the origin. The boundary conditions given are those for a clamped plate,
though hinged boundary conditions can be considered as well with no increase
in complexity. For convenience, and without loss of generality, in what follows
we will choose c = ρ = d = 1. Our goal is to show the uniform stability
of the coupled PDE system (1.1)-(1.2). To accomplish this goal we shall use
differential multipliers developed in the context of stability analysis for the wave
equation [12] together with the operator multiplier method introduced in [2].

1.2 Previous literature and the contribution of this paper.

Models of structural acoustic interactions as a coupling of wave and plate
equations go back to [24] and earlier. More recently, these models have become a
source of interest as engineers use them to try to control the noise in an acoustic
chamber [4, 5, 8]. The coupling of the wave equation in the chamber with the
equation of the wall or plate provides the essential mechanism for control of the
system. The first system thus studied was the case where the wall was modelled
by a structurally damped plate[4, 3, 1]. Since structural damping is such a
strong effect, the mathematical properties of such a system are much richer.
However, modelling of structural damping is poorly understood and structural
dampers added onto the active wall may produce a local ‘over-damping’ effect.
Thus, there is a great deal of appeal in models which do not depend on structural
damping to provide the necessary stabilizing effect.
The existing literature on interactive structural acoustic models (see [3, 1, 7,
23, 22, 14, 16] and references therein) deals with stabilization in the context of
linear plate models only. In contrast, this paper presents a model which allows
for large displacements of the active wall, thus addressing nonlinear dynamics
in the context of interactive structures. In addition, we do not assume any
source of structural damping or additional mechanical damping on the active
180 Catherine Lebiedzik

wall Γ0 . Not only is this physically appealing, but in fact it leads to interesting
mathematical difficulties. In the case where structural damping is added to the
wall, the corresponding dynamics are analytic, and the mathematical analysis
is much more straightforward [1]. In the case of mechanical damping on the
interface, we have certain regularizing effects on the traces of the wave equation.
In our situation, however, we have neither of these effects, and thus we need a
much more subtle mathematical analysis (particularly at the level of “sharp”
trace theory for waves and plates).
Hence, the main novelty of our contribution is the consideration of nonlinear
dynamics and thermoelastic effects in the context of hybrid PDE structures.
We wish to show that the addition of thermal effects on the flexible wall Γ0 ,
as well as boundary dissipation affecting a part of the hard wall of the acoustic
medium, suffices to stabilize the system. From a mathematical standpoint, the
fact that we do not assume and damping affecting Γ0 is critical. The lack of
damping on Γ0 has important implications regarding the regularity of solutions.
To appreciate this point, it is enough to notice that the presence of the damping
on the wall Γ0 provides a priori L2 regularity on the trace of the pressure zt |Γ0 ,
which is the coupling term between the wall and acoustic medium. If there is
no damping on Γ0 , the term zt |Γ0 is not even defined ( recall zt ∈ L2 (Ω)). Thus,
one of the fundamental task is to provide appropriate estimates (in appropriate
negative norms) for this term, as well as for the tangential derivatives of z on
Γ0 .

1.3 Statement of Main Results

We begin with a preliminary result that the system is well-posed. Theorem
1.1. (well-posedness) Let Ω be a bounded open domain in R3 with boundary Γ
as previously described. For all initial data
y0 = [z0 , z1 , w0 , w1 , θ0 ] ∈ Y where
Y ≡ HΓ11 (Ω) × L2 (Ω) × H02 (Γ0 ) × H0,γ
1
(Γ0 ) × L2 (Γ0 )
the solution y(t) = [z, zt , w, wt , θ] of the model (1.1)-(1.2) exists in C([0, ∞); T )
and is unique.
Here the function spaces used are defined as HΓ11 (Ω) = {f ∈
H 1 (Ω) with f |Γ1 = 0} and H0,γ 1 (Γ ) = H 1 (Γ ) f or γ > 0; L2 (Γ ) f or γ = 0
0 0 0 0
with inner product
(1.3)
1
1 (Γ ) ≡ (ω1 , ω2 )L2 (Γ ) + γ(∇ω1 , ∇ω2 )L2 (Γ )
(ω1 , ω2 )H0,γ ∀ ω1 , ω2 ∈ H0,γ (Γ0 )
0 0 0

Proof. Since the problem is a locally Lipschitz perturbation of a m-monotone

system with a priori bounds, the result follows from general theory of m-
dissipative operators(see, e.g. [6]). The full details of this argument are given
in [13].
 Stability of a nonlinear structural acoustic model 181

In order to formulate our main result on stability, we present some notation.

We define the energy functional associated with the model as
Z Z
 2 
(1.4) Eγ (t) = |zt | + |∇z|2 dΩ + z 2 dΓ1
Ω Γ1
Z  
2 2 2 1 2 2
+ |wt | + γ|∇wt | + |∆w| + |∆F(w)| + |θ| dΓ0
Γ0 2

Next, we introduce the function h(s) which is assumed concave, strictly

increasing, zero at the origin, and such that the following inequality is satisfied
for all |s| ≤ 1:

(1.5) h(s g(s)) ≥ s2 + |g(s)|2

Such a function can easily be constructed in view of the monotonicity assump-

tion imposed on g, [17]. Additionally, we will impose a geometric condition on
the ‘clamped’ portion of the boundary Γ2 . Γ2 is assumed convex (that is, the
level set function representing Γ2 has a nonnegative Hessian in the neighbor-
hood of Γ2 on the side of Ω) and there exists a point x0 ∈ R3 such that

(1.6) (x − x0 ) · ν ≤ 0, x ∈ Γ0 ∪ Γ2

Note that this condition is automatically satisfied if Γ2 is empty. In fact, one

can choose x0 to be any point in Γ0 . If Γ2 is non-empty, then x0 is a suitably
selected point in the hyperplane containing Γ0 (see Figure 1).
Our main result is the following. Theorem 1.2. (uniform stability) Let
Ω be a bounded open domain in R3 with boundary Γ as previously described.
Assume that the nonlinear function g satisfies

(1.7) m s2 ≤ g(s)s ≤ M s2 ; |s| ≥ 1

Then, with the constant γ ≥ 0, every weak (finite energy) solution of (1.1)-(1.2)
decays uniformly to zero, i.e.:

(1.8) Eγ (t) ≤ C sγ (t/T0 − 1)); t ≥ T0

where the real variable function sγ (t), which may depend on Eγ (0) and γ
converges to zero as t → ∞ and satisfies the following ordinary differential
equation (ODE):

d
(1.9) sγ (t) + q(sγ (t)) = 0, sγ (0) = Eγ (0)
dt
The nonlinear monotone increasing function q(s) is determined entirely from
the behavior at the origin of the nonlinear function g and is given by the
182 Catherine Lebiedzik

following algorithm:
(1.10) q ≡ I − (I + p)−1
 · 
(1.11) p ≡ (I + h0 )−1
K
(1.12) h0 (x) ≡ h (x/mes(0, T ) × Γ1 )
where h is given by (1.5) and the constant K > 0 may depend on Eγ (0) and γ.
The main mechanisms which allow us to stabilize the system are the thermal
effects in the plate Γ0 and the nonlinear dissipation on Γ1 . Thus, the decay rates
will be determined by the strength of the nonlinear function g(zt ). In fact, once
the behavior of g(s) at the origin is specified, the decay rates can be explicitly
solved for using the nonlinear ODE (1.9). If g is bounded from below by a linear
function, then it can be shown that the decay rates predicted are exponential.
If, instead, g has polynomial growth (or is exponentially decaying) at the origin,
then the decay rates are algebraic (or logarithmic). This can be demonstrated
by solving (1.9) (see [17]).
It is important also to note that though the decay rates in general may
depend on Eγ (0), and thus on the norm of the initial conditions, they are
independent of the profile of the initial conditions. In addition, we have not
assumed any source of boundary damping on the active wall Γ0 . This is a source
of additional mathematical difficulty, as the existence of boundary damping on
Γ0 gives rise to a priori L2 regularity on zt |Γ0 . This not only has an additional
stabilizing effect but also substantially improves the regularity of the hyperbolic
traces. In the absence of this property, a more sophisticated method is needed
to reconstruct the appropriate energy estimates, mainly at the level of treating
the traces in the case γ = 0. It is necessary to consider the two cases γ = 0 and
γ > 0 separately, and thus our estimates are not uniform as γ → 0.
We have not assumed any geometric conditions imposed on Γ1 , the portion
of the boundary subject to dissipation. This makes physical sense, since the
geometry of the boundary should only be an issue where the boundary is
‘uncontrolled’. Moreover, we have not imposed any conditions on the growth of
the nonlinearity g at the origin. This is in contrast with most of the literature
on boundary stabilization of wave and plate equations alone. (see [10] and
references therein). Finally, we have assumed Neumann, rather than Dirichlet,
boundary data on the ‘uncontrolled’ portion of the boundary Γ2 . This is a
source of technical difficulty, due to the fact that the Lopatinski condition is not
satisfied. On the other hand, in the context of the structural acoustic problem,
it is desirable to have Neumann data on Γ2 . Indeed, if we assumed Dirichlet
conditions on Γ2 , regularity for the corresponding elliptic problem would force
the assumption that Ω was not simply connected! Clearly this is not what
we want. Our techniques provide the result under the additional geometric
assumption that Γ2 is convex. It is not known if the same result can be shown
without this assumption (i.e. assuming only a “star-shaped” condition).
 Stability of a nonlinear structural acoustic model 183

We shall adopt the following notation:

Z
|w|s,Ω ≡ |w|H s (Ω) ; (u, v)Ω ≡ uvdΩ
Ω

The same notation will be used with Ω replaced by Γ etc. The negative Sobolev
spaces H −s (Ω) are defined as dual spaces to H0s (Ω). In addition, we will make
use of the following properties of Airy’s stress function.

(1.13) |[F(w), w]|−θ,Γ0 ≤ C|F(w)|3−β,Γ0 |w|2−θ+β,Γ0 , 0 < β < 1, 0≤θ≤1

(1.14) |F(w)|3−ε,Γ0 ≤ C|w|22,Γ0 , for any ε > 0

Moreover, straightforward application of (1.13) gives that

(1.15) |[F(w), w]|−1,Γ0 ≤ C|w|22,Γ0 |w|2−ε,Γ0 ε>0

2 Uniform Stabilization – proof of Theorem 1.2

Our goal is to show the uniform stability of the coupled PDE system (1.1)-(1.2).
We begin with a preliminary energy identity which illustrates the fact that the
system is dissipative.
Proposition 2.1. With respect to the system of equations (1.1)-(1.2) , the
following energy equality holds for all T > 0, s < T :
Z T Z T
(2.1) Eγ (s) = Eγ (T ) + 2 (g(zt ), zt )Γ1 dt + 2 |∇θ|20,Γ0 dt
s s

where the ‘energy’ Eγ (t) is defined by (1.4).

Proof. By running the multipliers zt on the wave equation, wt on the elastic
equation, θ on the thermal equation, and then integrating by parts, we obtain
the above equality for smooth solutions. A density argument allows us to extend
this inequality to all solutions of finite energy.
In order to prove Theorem 1.2, our strategy is to study the thermoelastic
equations on Γ0 and the wave equation on Ω separately and combine the
results. In the case of the thermoelastic plate, we will run the multiplier
A−1
D θ (introduced in [2]) on the elastic equation to yield an estimate of the
plate energy. For the wave equation, we run the multipliers z , z∇ · h , and
h · ∇z . This leads to an estimate of energy plus lower-order terms, which
are then absorbed via a standard uniqueness/compactness argument. Though
the multiplier used are “standard” by now in the context of the wave and
thermoelastic plate equations, the interactions between the two media brings
forward several technical difficulties due to the appearance of boundary traces
which are not a priori bounded by the energy. Handling of these terms
constituties the bulk of the the techincal content of this paper.
184 Catherine Lebiedzik

2.1 Thermoelastic Equations

We define the plate energy Ew,γ (t) as

1
(2.2) Ew,γ (t) = |wt |20,Γ0 + γ|∇wt |20,Γ0 + |∆w|20,Γ0 + |F(w)|20,Γ0 + |θ|20,Γ0
2
Next, we state several Lemmas which will be necessary for our estimates.
Proof of these Lemmas is in the Appendix. First, we consider the case where
γ = 0 and the dynamics of the plate are analytic:

wtt + ∆2 w = ht − ∆θ + k1 on Γ0 × (0, T )
∂
(2.3) w= w = 0 on ∂Γ0 × (0, T )
∂ν
Lemma 2.1. With reference to (2.3), where θ is the solution to the heat equation
in (1.2), we obtain the following regularity with an index α ≤ 1/2:

(2.4)
Z T Z T
2 2
[|w(t)|2+2α,Γ0 + |wt (t)|2α,Γ0 ]dt ≤ CT Ew (0) + CT [|ht |22α−2,Γ0 + |k1 |22α−2,Γ0 ]dt
0 0

where Ew (t) ≡ Ew,0 (t).

Lemma 2.2. Let γ = 0. We consider the original system given by (1.1), (1.2).
Z T
(2.5) [|w(t)|23,Γ0 + |wt (t)|21,Γ0 ]dt ≤ CT [Ew (0) + Ez (0)]
0
Z T Z T
2 2 2
+ CT [|wt |−1,Γ0 + |z|0,Γ1 + |g(zt )|0,Γ1 ]dt + C |w|42,Γ0 |w|22−ε,Γ0 dt
0 0

(2.6)
Z T Z T
2 2
[|w(t)|3−δ,Γ0 +|wt (t)|1−δ,Γ0 ]dt ≤ ε[Ew (0)+Ez (0)]+C |w|42,Γ0 |w|22−ε,Γ0 dt
0 0
Z T Z T
+ +CT,ε,δ [|wt |2−1,Γ0 + |w|21,Γ0 ]dt + CT [|g(zt )|20,Γ1 + |z|20,Γ1 ]dt
0 0

Next, we give a result for the clamped plate (for γ ≥ 0) which does not follow
from standard Sobolev trace theory. We note that the analogous result was
shown for the linear thermoelastic plate only in [2]. We have extended this
to account for the nonlinear von Kármán term and for the structural acoustic
interaction.
Lemma 2.3. With respect to the system of equations (1.1)-(1.2), the component
 Stability of a nonlinear structural acoustic model 185

w of the solution [z, zt , w, wt , θ] satisfies ∆w|∂Γ0 ∈ L2 (0, T ; L2 (∂Γ0 )) with the

estimate
(2.7)
Z T Z T

|∆w|20,∂Γ0 dt ≤ CT |∆w|20,Γ0 + |wt |20,Γ0 + γ|∇wt |2Γ0 + |∇θ|20,Γ0 + |z|20,Γ1
0 0

+|g(zt )|20,Γ1 dt + C(Ew,γ (T ) + Ew,γ (0)) + CT [Ez (0) + Eγ (0)lotγ (w, θ)]
where C, CT do not depend on the parameter γ, and lotγ (w, θ) is given by
(2.8)
h i
lotγ (w, θ) ≤ Cδ sup |w|22−δ,Γ0 + |wt |2−δ,Γ0 + γ|wt |21−δ,Γ0 + |θ|2−1/4,Γ0 , δ > 0
t∈[0,T ]

The major result of this section is the following estimate.

Lemma 2.4. With respect to the thermoelastic component of the model (1.2),
for all ε > 0, there exist CT , CEγ (0),T,ε such that the following inequality holds:

(2.9)
Z T T  Z

Ew,γ (t) dt ≤ εCT [Ew,γ (0) + Ew,γ (T ) + Ez (0)]+εCT |z|20,Γ1 + |g(zt )|20,Γ1 dt
0 0
Z T 
+ CEγ (0),T,ε |θ|21,Γ0 dt + lotγ (w, θ) + lot(z)
0

where lotγ (w, θ) is as given in (2.8),

 
(2.10) lot(z) ≤ Cδ sup |z(t)|21−δ + |zt (t)|2−δ , δ > 0
t∈[0,T ]

and the constants CT and CEγ (0),T,ε are uniformly bounded in γ ≥ 0.

Proof. We multiply the first equation in (1.2) by A−1
D θ and integrate from 0 to
T to obtain
Z T
(wtt − γ∆wtt + ∆2 w + ∆θ + zt − [F(w), w], A−1D θ)Γ0 dt = 0.
0

We deal with each part separately:

(1) Using integration by parts, substitution of boundary conditions, the second
equation of (1.2), and the fact that A−1
D is smoothing give(detailed calculations
are in [2])
(2.11)
Z 
 T Z Th i 
 −1


 wtt − γ∆wtt , AD θ L2 (Γ0 ) dt − |wt |20,Γ0 + γ|∇wt |20,Γ0 dt
 0 0 
Z Th i Z T
  2 2
≤ εC Ew,γ (0) + Ew,γ (T ) + ε |wt |0,Γ0 + γ|∇wt |0,Γ0 dt + Cε |θ|21,Γ0
0 0
186 Catherine Lebiedzik

Note that neither of the constants C and Cε depend on T or γ.

(2) Another integration by parts and application of boundary conditions gives
(2.12)
Z T Z T Z T
−1




2
∆ w, AD θ 0,Γ0 dt = − ∆w, ∇A−1
D θ 0,∂Γ0 dt + ∆w, ∆A−1
D θ 0,Γ0
dt
0 0 0

In order to estimate this term we need to use the trace regularity result from
Lemma 2.3.
(2.13)
Z 
 T  Z T
 2 −1


 ∆ w, AD θ Γ0 dt ≤ Cε |θ|21,Γ0 dt+εCT [Ew,γ (0) + Ew,γ (T ) + Ez (0)
 0  0
Z T Z T 
 2 2

+ Ew,γ (t) dt + |z|0,Γ1 + |g(zt )|0,Γ1 dt + Eγ (0)lotγ (w, θ)
0 0

(3) For the next term we just use integration by parts:

Z T Z T
−1


(2.14) ∆θ, AD θ L2 (Γ0 ) dt ≤ C |θ|20,Γ0
0 0

(4) For the term with zt |Γ0 , we will need to first integrate by parts, use trace
theory, and then substitute in the heat equation of (1.2):
Z T 
 −1


 zt , AD θ Γ0  dt

0
Z T Z T 
−1
 
≤ sup |z(t)|1/2+δ,Ω |AD θ|0,Γ0 +  (z, θ)Γ0 dt + (z, wt )Γ0 dt
t∈[0,T ] 0 0
 Z T  Z T
(2.15) ≤ CT,ε lot(z) + lotγ (w, θ) + |θ|21,Γ0 dt + ε |wt |20,Γ0 dt
0 0

(5) Finally, using (1.15) gives:

Z T  Z T
 −1




([F(w), w], AD θ)Γ0 dt ≤ C |[F(w), w]|−1,Γ0 |A−1
D θ|1,Γ0
0 0
(2.16) ≤ C|w|22,Γ0 |w|2−ε |θ|0,Γ0 ≤ C Eγ (0)lotγ (w, θ)
Combining equations (2.11) - (2.16) results in the fact that for ε small
enough there exists a constant CT > 0 so that

(2.17)
Z
T h i
(1 − 2ε) |wt |20,Γ0 + γ|∇wt |20,Γ0 dt ≤ εCT [Ew,γ (0) + Ew,γ (T ) + Ez (0)]
0
Z T Z T
2
 
+CT,ε |θ|1,Γ0 dt+CT,Eγ (0) lotγ (w, θ)+εCT E(t) + |z|20,Γ1 + |g(zt )|20,Γ1 dt
0 0
 Stability of a nonlinear structural acoustic model 187

where the non-crucial dependence of CT on ε has not been noted.

Next, we multiply the same equation of (1.2) by w and integrate from 0 to
T to obtain
T T Z T h i




wt , w 0,Γ0  + γ ∇wt , ∇w 0,Γ0  − |wt |20,Γ0 + γ|∇wt |20,Γ0 dt =
0 0 0
Z T Z T Z T Z T
− |∆w|20,Γ0 dt+ ([F(w), w], w)0,Γ0 dt+ (∇θ, ∇w)0,Γ0 dt− (zt , w)0,Γ0 dt
0 0 0 0

Taking norms and using the trace theorem gives

 T T 


 

 wt , w 0,Γ0  + γ ∇wt , ∇w 0,Γ0   ≤ ε[Ew,γ (0) + Ew,γ (T )] + Cε lotγ (w, θ)
 0 0

and
Z T
 Z T
 
 
(zt , w)0,Γ0 dt ≤ CT (lot(z) + lotγ (w, θ)) + C |wt |20,Γ0 dt

0 0

Combining these and using the divergence theorem as well as symmetricity of

the bracket on the term involving F(w) gives
Z T
 
|∆w|20,Γ0 + |∆F|20,Γ1 dt
0
h i Z Th i
2 2
≤ ε Ew,γ (0) + Ew,γ (T ) + C |wt |0,Γ0 + γ|∇wt |0,Γ0 dt
0
Z T Z T
+ε |∆w|20,Γ0 dt + Cε |θ|21,Γ0 dt + CT (lot(z) + lotγ (w, θ))
0 0

Thus, we have that there exist constants C, CT , Cε > 0 such that for ε > 0
small enough,
T  Z
  
(2.18) (1 − ε) |∆w|20,Γ0 + |∆F|20,Γ1 dt ≤ ε Ew,γ (0) + Ew,γ (T )
0
Z Th i Z T
2 2
+C |wt |0,Γ0 + γ|∇wt |0,Γ0 dt + Cε |θ|21,Γ0 dt + CT (lot(z) + lotγ (w, θ))
0 0

If the ε of equations (2.17) and (2.18) is small enough, they can be combined
to produce the inequality (2.9), which is the desired result.

2.2 Wave Equation

Let Ez (t) be the energy defined by

(2.19) Ez (t) = |zt (t)|20,Ω + |∇z(t)|20,Ω + |z(t)|20,Γ1

188 Catherine Lebiedzik

We quote here a sharp trace result for the wave equation that will be necessary
to our proof.
Lemma 2.5. Let z be a solution to ztt = ∆z in Ω×(0, T ) with interior regularity
z ∈ C(0, T ; H 1 (Ω)) ∩ C 1 (0, T ; L2 (Ω)); and the following boundary regularity

∂
z ∈ L2 ((0, T ) × Γ); zt |Γ1 ∈ L2 ((0, T ) × Γ1 );
∂ν
T
Let T > 0 be arbitrary and let α be an arbitrary small constant such that α < 2.
Then, we have:
Z T −α Z T
∂ 2 ∂
(2.20) | z|0,Γ1 dt ≤ CT,α [ [| z|20,Γ + |zt |20,Γ1 ]dt + lot(z)]
α ∂τ 0 ∂ν

Z T Z T
∂ 2
(2.21) |zt |2−4/5,Γ dt ≤ CT [Ez (0) + | z| dt]
0 0 ∂ν 0,Γ

where lot(z) is defined in (2.10).

Proof. Inequality (2.20) follows similarly as Lemma 7.1 in [18]. Though here
we are evaluating the tangential derivative and zt on Γ1 , rather than the whole
of the boundary Γ, the argument given in [18] can still be used. It is necessary
to note only that the measurements of zt |Γ1 are needed only in the nonelliptic
sector (after microlocalization), where the argument is purely local. Inequality
(2.21) is given in Theorems A and C of [20] (see also [19]).
The main result of this section is the following ”recovery” estimate for the
wave equation.
Lemma 2.6. Assume that the geometric condition (1.6) is in force. Consider
the wave equation (1.1) with finite energy solutions, where w is a finite energy
solution to (1.1) and T > 0 is arbitrary. Then, for any α < T /2 there exist
positive constants C, possibly depending on α, such that if γ > 0
Z T −α
(2.22) Ez (t) dt ≤ C [Ez (α) + Ez (T − α) + Ez (0)]
α
Z T
+ CT [|zt |20,Γ1 + |g(zt )|20,Γ1 ] dt + CT lot(z) + CT,γ lotγ (w, θ)]
0

and if γ = 0,
Z T −α
(2.23) Ez (t) dt ≤ C [Ez (α) + Ez (T − α) + Ez (0) + Ew (0)]
α
Z T
+ CT,Eγ (0) [|zt |20,Γ1 + |g(zt )|20,Γ1 + |wt |20,Γ0 ] dt + CT,Eγ (0) [lot(z) + lotγ (w, θ)]
0
 Stability of a nonlinear structural acoustic model 189

Proof. The first step of the proof involves the use of a multiplier method. As
usual, in order to apply this method it is necessary to have solutions which are
smooth enough that we can use standard differential calculus. In the nonlinear
case, our solutions may not have enough regularity, even if the initial data are
taken sufficiently smooth. We can bypass this difficulty by using a regularization
argument proposed in [17]. We will perform the necessary PDE calculus on
solutions of the ’regularized’ wave equation and obtain estimates (2.22) and
(2.23). Then we will pass through the limit, using an appropriate regularization
parameter, and in this way we will reconstruct these inequalities for the original
problem. The Lemma below states the result necessary for this limit passage.
Lemma 2.7. Given any solution of the wave equation

ztt = ∆z; in Ω × (0, T )

∂
(2.24) z = 0; on Γ2 × (0, T )
∂ν
with the following regularity properties

1. z ∈ C[0, T ; H 1 (Ω)] ∩ C 1 [0, T ; L2 (Ω)]

∂
2. ∂ν z|Γ1 , zt |Γ1 ∈ L2 (Σ1 )
∂
3. ∂ν z|Γ0 ∈ L2 (0, T ; H 1/2 (Γ0 ))

there exists a sequence z n ∈ C[0, T ; H 2 (Ω)] ∩ C 1 [0, T ; H 1 (Ω)] of solutions to the

wave equation
n
ztt = ∆z n ; in Ω × (0, T )
∂ n
(2.25) z = 0; on Γ2 × (0, T )
∂ν
such that the following convergence holds

• z n → z in C[0, T ; H 1 (Ω)] ∩ C 1 [0, T ; L2 (Ω)]

∂ n ∂
• ∂ν z |Γ1 → ∂ν z|Γ1 in L2 (Σ1 )
∂ n ∂
• ∂ν z |Γ0 → ∂ν z|Γ0 in L2 (0, T ; H 1/2 (Γ0 ))

• ztn |Γ1 → zt |Γ1 in L2 (Σ1 )

Proof. The proof of this Lemma follows from the method in [17] and is detailed
in [16]. Lemma 2.7 applies to any finite energy solutions of (1.1). From
the energy relation (2.1) and the properties of g, it can be immediately seen
that zt |Γ1 ∈ L2 (Σ1 ). Applying the boundary conditions on Γ1 and the Trace
∂ ∂
Theorem gives that ∂ν z|Γ1 ∈ L2 (Σ1 ). Finally, ∂ν z|Γ0 ∈ L2 (0, T ; H 1/2 (Γ0 )) is
1
given by the fact that w ∈ L2 (0, T ; H (Γ0 )).
190 Catherine Lebiedzik

In order to apply the result of Lemma 2.7 we will consider a sequence of

smooth solutions z n to (2.25), to which we apply the multipliers h · ∇z n and
div h z n . The vector field h here is is such that

h · ν = 0 on Γ0 ∪ Γ2 , J(h) > c0 > 0 on Ω

where J(h) denotes the Jacobian of h. Such a vector field exists (see [14, 21])
as long as the geometrical condition (1.6) is satisfied and Γ0 ∪ Γ2 is convex.
Applying these multipliers and performing familiar computations (see, e.g., [17])
gives

(2.26)
Z T Z T
∂ ∂
Ez n (t)dt ≤ C[Ez n (s)+Ez n (T )]+C [|ztn |20,Γ1 +| z n |20,Γ1 +| z n |20,Γ1 ]dt
s s ∂ν ∂τ
Z T 2 Z T Z 
 ∂ n  ∂ n ∂ n 
+C  z  dt + C  z z h · τ dΓ0 dt + CT lot(z n )
 ∂ν 
s 0,Γ0 s Γ0 ∂ν ∂τ

The main issue and difficulty here is to provide the estimates for the
tangential derivatives of z n on Γ1 and Γ0 . Indeed, these terms are not bounded
by the energy and the sharp trace regularity theory of hyperbolic solutions,
recalled in Section 2, is necessary. Tangential derivatives on Γ1 are estimated
with the help of Lemma 2.5, inequality (2.20). By applying the Trace Theorem
and Young’s Inequality we obtain estimates for the tangential derivatives on
Γ0 .
Z T Z T Z T
∂ ∂ n ∂ n2 ∂
( zn , z )Γ0 dt ≤ ε | z |−1/2,Γ0 dt + Cε | z n |21/2,Γ0 dt
s ∂ν ∂τ s ∂τ s ∂ν
Z T Z T
∂
(2.27) ≤ ε |z n |21,Ω dt + Cε | z n |21/2,Γ0 dt
s s ∂ν
Z T Z T
∂
≤ εCT [Ez n (0) + Ez n (t)dt + Cε | z n |21/2,Γ0
s 0 ∂ν

Combining (2.26),(2.20), (2.27), applied with s = α, T = T − α and taking ε

small enough so that εCT < C yields
Z T −α
(2.28) Ez n (t)dt ≤ C[Ez n (α) + Ez n (T − α) + Ez n (0)]
α
Z T Z T
n 2 ∂ n2 ∂
+ CT [|zt |0,Γ1 + | z |0,Γ ]dt + CT | z n |21/2,Γ0 dt + CT lot(z n )
0 ∂ν 0 ∂ν

Since, as previously explained, the function z in (1.1) satisfies all the require-
ments in Lemma 2.7, we may apply this Lemma and pass with the limit on all
 Stability of a nonlinear structural acoustic model 191

terms in (2.28). In addition, we apply the boundary conditions in (1.1). This

gives the estimate
Z T −α
(2.29) Ez (t)dt ≤ C[Ez (α) + Ez (T − α) + Ez (0)]
α
Z T
+ CT [ [|zt |20,Γ1 + |g(zt )|20,Γ1 + |wt |21/2,Γ0 ]dt + lot(z)]
0

In the case γ > 0, this yields

Z T −α
(2.30) Ez (t)dt ≤ C[Ez (α) + Ez (T − α) + Ez (0)]
α
Z T
+ CT [|zt |20,Γ1 + |g(zt )|20,Γ1 dt + Cγ,T lotγ (w, θ) + CT lot(z)
0

For γ = 0 , it is necessary to estimate the last integral on the RHS of (2.29),

since it no longer contributes lower order terms. At this point, the analyticity
of the w component plays a critical role. Indeed, by the second statement in
Lemma 2.2 applied with δ = 1/2 we obtain:
Z T
(2.31) CT |wt |21/2,Γ0 dt ≤ ε0 CT [Ew (0) + Ez (0)]
0
Z T
+ Cε0 ,Eγ (0),T [lotγ (w, θ) + lot(z) + |g(zt )|20,Γ1 dt]
0

Making a suitable choice of ε0 = ε0 (T ) and collecting (2.30) (for γ > 0) and

(2.29) (for γ = 0) combined with (2.31) yields desired conclusion in Lemma 2.6.

2.3 Uniform stability analysis for the coupled system

In the final analysis, we will combine the energy estimates on plate and wave
equations, and then absorb the lower order terms by means of a standard
compactness/uniqueness argument.
Proposition 2.2. With respect to the coupled PDE system (1.1)-(1.2), the
following estimate holds:
Z T

(2.32) Eγ (T ) ≤ CT,Eγ (0),γ |θ|21,Γ0 + |zt |20,Γ1 + |g(zt )|20,Γ1 dt
0
+ CT,Eγ (0),γ [lotγ (w, θ) + lot(z)]

Here the energy Eγ (t) is defined as in (1.4).

Proof. Here the argument is the same for γ > 0 and γ = 0, so we give only
the argument for γ = 0. The inequality (2.32) follows from equations (2.9)
192 Catherine Lebiedzik

and (2.22). First, we have added (2.9) and (2.22) after multiplying (2.9) by a
suitable constant AT in order to consolidate the |wt |20,Γ0 and |z|20,Γ1 terms. This
gives:
(2.33)
Z T Z T −α
AT Ez (t) dt ≤ εAT CT [Ew,γ (0) + Ew,γ (T ) + Ez (0)]
Ew,γ (t) dt +
0 α
Z T
+C [Ez (α) + Ez (T − α) + Ez (0) + Ew (0)]+(CT,Eγ (0),γ +εAT CT ) |g(zt )|20,Γ1 dt
0
Z T Z T Z T
+εAT CT |z|20,Γ1 dt+AT CT,Eγ (0) |θ|21,Γ0 dt+CT,Eγ (0),γ (|wt |20,Γ0 +|zt |20,Γ1 ) dt
0 0 0
+ (CT,Eγ (0),γ + AT CT,Eγ (0) ) (lotγ (w, θ) + lot(z))
Choosing AT > 2CT and ε, ε small enough (so that εAT CT ≤ CT ) and recalling
the definition of Eγ (t) given in (1.4) gives
Z T −α
(2.34) Eγ (t) dt ≤ CT [Eγ (0) + Eγ (T ) + Eγ (α) + Eγ (T − α)]
α
Z T


+CT,Eγ (0),γ |θ|21,Γ0 + |zt |20,Γ1 + |g(zt )|20,Γ1 dt+CT,Eγ (0),γ (lotγ (w, θ) + lot(z))
0
Next, we use the dissipativity property to eliminate the parameter α. Using
the identity (2.1) and the simple inequality
Z α Z T !
+ Eγ (t) dt ≤ 2αEγ (0)
0 (T −α)

gives
Z T
Eγ (t) dt
0
Z T

≤ CT [Eγ (0) + Eγ (T )] + CT,Eγ (0),γ |θ|21,Γ0 + |zt |20,Γ1 + |g(zt )|20,Γ1 dt
0
(2.35) +CT,Eγ (0),γ (lotγ (w, θ) + lot(z))
Again, dissipativity gives that for t < T ,Eγ (T ) ≤ Eγ (t), so that T Eγ (T ) ≤
RT
0 Eγ (t) dt. Substituting this fact and (2.1) into (2.35) and taking T > 2CT
leads to the desired conclusion in Proposition 2.2.
Our next step is to eliminate the lower order terms from equation (2.32).
Proposition 2.3. With respect to the coupled PDE system (1.1)-(1.2), there
exists a constant
CT,Eγ (0) > 0 such that
Z T


(2.36) lotγ (w, θ) + lot(z) ≤ CT,Eγ (0),γ |zt |20,Γ1 + |g(zt )|20,Γ1 + |θ|21,Γ0 dt
0
 Stability of a nonlinear structural acoustic model 193

Proof. The conclusion follows by contradiction via the usual compactness and
uniqueness argument. Since this argument is standard, we shall only point
out the main steps. The compactness of lotγ (w, θ) + lot(z), with respect to the
topology induced by the energy Eγ , γ ≥ 0, follows from the compact imbeddings
H 2−ε (Γ0 )×H 1−ε (Γ0 )×H 1−ε (Ω)×H −ε (Ω) ⊂ H 2 (Γ0 )×H 1 (Γ0 )×H 1 (Ω)×L2 (Ω);
for ε > 0. As for the uniqueness part, we deal with the following overdetermined
system (here we consider only the more difficult case γ > 0):
z̃tt = ∆z̃ on [0, T ] × Ω
∂
z̃t = 0; ∂ν z̃ + z̃ = 0 on [0, T ] × Γ1
(2.37) ∂
∂ν z̃ = w̃t ; θ̃ = 0 on [0, T ] × Γ0
∂ z̃
∂ν = 0 on [0, T ] × Γ2
∆w̃t = 0 on [0, T ] × Γ0
∂
Since z̃t = 0; ∂ν z̃t + z̃t = 0 on Γ1 × (0, T ), a version of Holmgren’s Uniqueness
Theorem applies (see Thm 3.5 in [9]) to conclude z̃t ≡ 0. Feeding back this
information into the plate equation yields the following overdetermined system
for the variable w̃.
w̃tt + ∆2 w̃ = [F(w̃), w̃] on [0, T ] × Γ0
θ̃ ≡ 0 on [0, T ] × Γ0
∆w̃t = 0 on [0, T ] × Γ0
∂
w̃ = ∂ν w̃ = 0 on [0, T ] × ∂Γ0

{z̃(0), z̃t (0), w̃(0), w̃t (0), θ̃(0)} = {z̃0 , z̃1 , w̃0 , w̃1 , θ˜0 } ∈ Y
∂
Since θ̃ ≡ 0, ∆w̃t = 0. However, we have that w̃t = ∂ν w̃t = 0 on ∂Γ0 . Thus, by
elliptic theory w̃t ≡ 0. Substituting this into (2.37) gives the following system:
∆z̃ = 0 on [0, T ] × Ω
∂
∂ν z̃ = 0 on [0, T ] × Γ0
(2.38) ∂
∂ν z̃ + z̃ = 0 on [0, T ] × Γ1
∂ z̃
∂ν = 0 on [0, T ] × Γ2

∆2 w̃ = [F(w̃), w̃] on [0, T ] × Γ0

θ̃ ≡ 0 on [0, T ] × Γ0
∂
w̃ = ∂ν w̃ = 0 on [0, T ] × ∂Γ0
By the uniqueness of elliptic solutions, we conclude that {z̃, w̃, θ̃} ≡ 0 , for all
t > t0 > 0. This allows us to assert (2.36) as desired.
To finish off the proof of Theorem 1.2, we use the inequality (2.36) to
combine terms on the right hand side of (2.32).
Z T


(2.39) Eγ (T ) ≤ CT,Eγ (0),γ |zt |20,Γ1 + |g(zt )|20,Γ1 + |θ|21,Γ0 dt
0
194 Catherine Lebiedzik

By using the assumptions imposed on the nonlinear function g and splitting

the region of integration into two: zt ≤ 1 and zt > 1 we also obtain (see [17]):
Z T
(2.40) [|zt |20,Γ1 + |g(zt )|20,Γ1 + |θ|21,Γ0 ]dt
0
Z T Z
≤ CT,m,M [I + h0 ] [ g(zt )zt dx + |∇θ|20,Γ0 ]dt
0 Γ1

where we have used Jensen’s inequality. Combining (2.39) and (2.40) and
recalling monotonicity of h0 we obtain:
Z T Z 
Eγ (T ) ≤ CT,γ,m,M [I + h0 ] g(zt )zt dΓ1 + |∇θ|20,Γ0 dt
0 Γ1
(2.41) = CT,γ,m,M,Eγ (0) [I + h0 ][Eγ (0) − Eγ (T )]

where in the last step we have used the energy relation. Since [I + h0 ] is
invertible, this gives
−1
(2.42) [I + h0 ]−1 (Cγ,T,m,M,E γ (0)
Eγ (T )) ≤ Eγ (0) − Eγ (T )

this gives
p(Eγ (T )) + Eγ (T ) ≤ Eγ (0)
with p defined by the Theorem 1.2. The final conclusion of Theorem 1.2 follows
now from application of Lemma 3.1 in [17]. The argument for γ > 0 is identical.

3 Appendix
Here we cite the necessary proofs of the Lemmas introduced in Section 2.1.
Proof. (Lemma 2.1) The proof of this result is identical to that given in [16],
though [16] deals with the case of free boundary conditions.
Proof. (Lemma 2.2) We apply Lemma 2.1 with h ≡ z, α = 1/2, k1 ≡
[F(w), w], and use (1.15).

(3.1)
Z T Z T
2 2
[|w(t)|3,Γ0 + |wt (t)|1,Γ0 ]dt ≤ CT Ew (0) + CT [|zt |2−1,Γ0 + |w|42,Γ0 |w|22−ε,Γ0 ]dt
0 0

But from (2.21) in Lemma 2.5 and the boundary conditions imposed on the
wave equation we have
Z T
(3.2) |zt |2−1,Γ0 dt
0
Z T Z T
2
≤ |zt |−4/5,Γ0 dt ≤ CT [Ez (0) + [|wt |20,Γ0 + |g(zt )|20,Γ1 + |z|20,Γ1 ]dt]
0 0
 Stability of a nonlinear structural acoustic model 195

Combining the two inequalities gives:

(3.3)
Z T Z T
[|w(t)|23,Γ0 + |wt (t)|21,Γ0 ]dt ≤ CT [Ew (0) + Ez (0) + |w|42,Γ0 |w|22−ε,Γ0 ]dt]
0 0
Z T Z T
+ε |wt |21,Γ0 dt + CT,ε [ |wt |2−1,Γ0 dt + |g(zt )|20,Γ1 + |z|2Γ1 ]dt]
0 0

where in the last step we have used interpolation inequalities.

Taking in (3.3) ε sufficiently small gives the first inequality in Lemma 2.2.
As for the second inequality, this follows from the first after an additional use
of interpolation inequalities.
Proof. (Lemma 2.3) We will multiply the first equation of (1.2) by the
quantity h · ∇w, where h(x, y) ≡ [h1 (x, y), h2 (x, y)] is a [C 2 (Γ̄0 )]2 vector field
such that h|∂Γ0 = [ν1 , ν2 ], and then integrating from 0 to T , i.e.

Z T

(3.4) wtt − γ∆wtt + ∆2 w + ∆θ + zt − [F(w), w], h · ∇w Γ0
dt = 0
0

By integration by parts, application of the divergence theorem and the estimates

in [2], we have that

(3.5)
Z T Z Th
|∆w|20,∂Γ0 dt ≤ C[Ew,γ (0)+Ew,γ (T )]+C |w|22,Γ0 +|wt |20,Γ0 +γ|∇wt |20,Γ0
0 0
i Z T Z T
2
+ |θ|1,Γ0 dt + 2 |(zt , h · ∇w)Γ0 | dt + 2 |([F(w), w], h · ∇w)Γ0 | dt
0 0

In order to estimate the last two terms we need to use (2.21) and (1.13),
respectively. First, by means of (2.21) and the boundary conditions on z, we
have that

Z T
(3.6) |(zt , h · ∇w)Γ0 | dt
0
Z T Z T
≤ C |zt |2−4/5,Γ0 dt + C |h · ∇w|24/5,Γ0 dt
0 0
 Z T  Z T
 2 2 2

≤ CT Ez (0) + |wt |0,Γ0 + |g(zt )|0,Γ1 + |z|0,Γ1 dt + C |w|22,Γ0
0 0
196 Catherine Lebiedzik

Next, (1.13) and (1.14) give

Z T Z T
|([F(w), w], h · ∇w)Γ0 | , dt ≤ C |[F(w), w]|−1/2,Γ0 |h · ∇w|1/2,Γ0
0 0
Z T
(3.7) ≤ C |F(w)|3−ε,Γ0 |w|3/2+ε,Γ0 |w|3/2,Γ0
0
Z T
≤ C |w|22,Γ0 |w|23/2+ε,Γ0
0
≤ CT sup |w|22,Γ0 |w|23/2+ε,Γ0
t∈[0,T ]
≤ CT E(0)lot(w, θ, z)

Combining estimates (3.5),(3.6), and (3.7) gives the desired result (2.7).

References

[1] G. Avalos, The exponential stability of a coupled hyperbolic/parabolic system

arising in structural acoustics, Abstr. Appl. Anal., 1 (1996), pp. 203–217.
[2] G. Avalos and I. Lasiecka, Exponential stability of a thermoelastic system without
mechanical dissipation. Dedicated to the memory of Pierre Grisvard, Rend. Istit.
Mat. Univ. Trieste, 28 (1997), pp. 1–27.
[3] , Uniform decay rates of solutions to a structural acoustic model with
nonlinear dissipation, Appl. Math. Comput. Sci., 8 (1998), pp. 287–312.
[4] H. T. Banks, R. J. Silcox, and R. C. Smith, The modeling and control of
acoustic/structure interaction problems via piezoceramic actuators: 2-d numerical
examples, ASME Journal of Vibration and Acoustics, 116 (1994), pp. 386–396.
[5] H. T. Banks, R. C. Smith, and Y. Y. Wang, The modeling of piezoceramic patch
interactions with shells, plates, and beams, Quarterly of Applied Mathematics, 53
(1995), pp. 353–381.
[6] V. Barbu, Nonlinear Semigroups and Differential Equations in Banach Spaces,
Noordhoff, Leyden, 1976.
[7] M. Camurdan, Uniform stability of a coupled structural acoustic system by
boundary dissipation, Abstract and Applied Analysis, (1999).
[8] C. R. Fuller, G. P. Gibbs, and R. J. Silcox, Simultaneous active control of flexural
and extensional power flow in beams, Journal of Intelligent Materials, Systems
and Structures, 1 (1990).
[9] V. Isakov, Inverse Problems for Partial Differential Equations, Springer Verlag,
New York, 1998.
[10] V. Komornik, Exact controllability and Stabilization – the Multipliers Method,
Masson, Paris, 1994.
[11] J. Lagnese, Boundary Stabilization of Thin Plates, SIAM, Philadelphia, PA, 1989.
[12] I. Lasiecka, Boundary stabilization of a 3-dimensional structural acoustic model,
Journal de Mathematiques Pure et Applique, 78 (1999), pp. 203–232.
[13] , Mathematical Control Theory of Coupled PDE Systems: NSF-CMBS
Lecture Notes, SIAM, 2000.
 Stability of a nonlinear structural acoustic model 197

[14] I. Lasiecka and C. Lebiedzik, Uniform stability in structural acoustic systems with
thermal effects and nonlinear boundary damping, Control and Cybernetics, 28
(1999), pp. 557–581.
[15] , Boundary stabilizibility of nonlinear acoustic models with thermal effects
on the interface, C.R Acad. Sci. Paris, 328 (2000), pp. 187–192.
[16] , Decay rates in interactive hyperbolic-parabolic pde models with thermal
effects on the interface, Appl. Math. Optim., to appear (2000).
[17] I. Lasiecka and D. Tataru, Uniform boundary stabilization of semilinear wave
equations with nonlinear boundary damping, Diff. Int. Eq., 6 (1993), pp. 507–533.
[18] I. Lasiecka and R. Triggiani, Regularity theory of hyperbolic equations under
Dirichlet boundary terms, Appl. Math. and Optim., 10 (1983), pp. 275–286.
[19] , Sharp regularity theory for second order hyperbolic equations of Neumann
type. part i. nonhomogenous data, Ann.Mat. Pura Applicata IV, CLVII (1990),
pp. 285–367.
[20] , Sharp regularity theory for second order hyperbolic equations of Neumann
type. part ii: general boundary data, J. Diff. Eq, 94 (1991), pp. 112–164.
[21] I. Lasiecka, R. Triggiani, and X. Zhang, Exact controllability and unique con-
tinuation for wave equations with Neumann uncontrolled boundary conditions,
Proceedings of the AMS, (to appear).
[22] C. Lebiedzik, Uniform stability of a coupled structural acoustic system with
thermoelastic effects, Dynamics of Continuous, Discrete, and Impulsive Systems,
(to appear).
[23] W. Littman and B. Liu, On the spectral properties and stabilization of acoustic
flow, SIAM J. Appl. Math., 59 (1999), pp. 17–34.
[24] P. Morse and K. Ingard, Theoretical Acoustics, McGraw-Hill, 1968.
198 Catherine Lebiedzik
On Modelling, Analysis and Simulation of Optimal
Control Problems for Dynamic Networks of
Euler-Bernoulli-and Rayleigh-beams

Günter Leugering, TU Darmstadt, Darmstadt, Germany

Wigand Rathmann, inuTech GmbH, Seukendorf, Germany

Abstract

We consider a network of Euler-Bernoulli- and Rayleigh-beams. For

the sake of simplicity, we concentrate on scalar displacements coupled to
torsion. We show that the model is well-posed in appropriate ramification
spaces. We then describe a dynamic nonoverlapping domain decomposi-
tion procedure of the network into its individual edges and provide a proof
of convergence. Further, we formulate typical optimal control problems,
related to exact controllability. The optimality system is solved using
conjugate gradients. Various numerical examples illustrate the method.

AMS-Classification: 49M27, 73K12, 73K50, 93C20

Key-word: Euler-Bernoulli- and Rayleigh beams, network models, nonoverlap-
ping domain decompositions, optimal controls.

1 Introduction
Beginning with the work of Chen et. al. [5], the question of controllability
and stabilizability of connected beams and more general flexible structures has
become a major field of research in the last ten years. Whereas such structures
have been investigated in the engineering literature mainly in terms of the
Finite-Element-Method (FEM), the mathematical literature is dominated by
the original continuum mechanical formulation in terms of partial differential
equations (PDE’s). It has become more and more apparent that controllability
properties as well as their counterparts in terms of stabilizability are very much
dependent on the underlying PDE-framework and can not be predicted or
analyzed on the FEM-level. Therefore, rather than to first discretize and then
apply standard optimal-control- or simulation-software (of which an abondance
is currently available), we insist on simulation- and optimal-control procedures
that are developed on the continuous level and are then discretized. In this
199
200 Leugering and Rathmann

paper we describe the continuous modelling and algorithms applied to the

continuous systems. The discretization is then based on classical FEM-tools.
Networks of strings, Timoshenko beams and combination there of have
been investigated by Lagnese, Leugering and Schmidt [13]. Networks of Euler-
Bernoulli-beams have been introduced in Leugering and Schmidt [20] and
preliminary controllability results have been shown there. About ten years later
Dekoninck and Nicaise [6] considered a scalar-displacement Euler-Bernoulli-
beam model with umbrella-type node conditions and showed controllability in
some cases where Ingham’s inequality and a certain uniqueness result apply.
Almost simultaneously Briffaut [4], in his thesis, considered in-plane-
beams for some particular configurations and provided numerical evidence for
controllability and stabilizability. His analysis and also the numerical work
heavily depends on the nodes of the system and a modal approximation to the
HUM-operator. In summary, it is fair to say that up to know no satisfactory
theory of controllability/stabilizability of networks of Euler-Bernoulli-beams, let
alone Rayleigh-beams, exists. Furthermore, very few papers deal with numerical
simulations of such networks. Of course, this has to do with the enormuous
complexity of PDE-systems describing the motion of the such networks.
Our philosophy here is to decompose the networks under consideration into
smaller pieces, in fact, into the edges of the underlying graph. This is done
by a nonoverlapping domain decomposition procedure, inspired by the work of
P. L. Lions [21] and Benamou [1],[2]. The method is not a substructering
method in the sense of Shur-complement-preconditioning. Rather it is an
iterative procedure which, in the static case, can be derived from an augmented
Lagrangian-saddle point algorithms of modified Uzawa-type, see [11] for the
Laplace-operator. For in-plane-model the appropriate algorithm has been for
Euler-Bernoulli-networks given in Leugering [16]. See also [3], [14], [15], [19],
[17], [18] for string- and Timoshenko-beam networks.
In this paper we also introduce torsional motion, while, for simplicity, we
restrict ourselves to scalar displacements. The more general case has been
considered in [23]. In contrast to [3], [14], [15],[16], [19], [17] [18], we do not
decompose the optimality system, but rather we use the domain decomposition
method (DDM) as a network solver. The optimal control problems which
we consider are related to exact controllability, but other cost functionals can
easily be considered. The method of choice here is a penalization of the final
states which acts as a Tychonov regularization of the (illposed) controllability
problem. We use a conjugate gradient (CG) approach in the spirit of R.
Glowinski and J. L. Lions [9], [10].
The numerical results are very encouraging as far as the CG-iteration is
concerned. As for the network solver based on DDM we have to say that the
convergence, which is linear by the method of choice, is strongly dependent
of the proper choice of the various penalty and relaxation parameters. This,
 Networks of Beams 201

of course, comes with no surprise. We currently perform research in the

direction of optimal interface conditions in the sense of de Sturler, Nataf
and Rogier [22]. This amounts to analyze and approximate Steklov-Poincaré-
operators of Petrovski-systems. This will be the subject of a forthcoming
publication. Analysis in this direction appears to be very important for real-
time-applicability of the algorithms.
The paper is organized as follows. We first introduce some notation (section
2). In section 3 we show that scalar non collinear networks of Euler-Bernoulli-
beams with rigrid joints but without torsion are inconsistent with mechanics.
In section 4 we develop our main model including torsion. There we also give
the existence and uniqueness result, the proof of which is shifted to section 7.
In section 5 we describe our domain decomposition procedure and state the
convergence result. The proof is shifted to section 8. Section 6 is devoted to
the numerical treatment of optimal control problems.

2 Notations
We consider a network of beams. A planar graph G = (V, E) with nodes V
and edges E is taken to describe the configuration at rest. The J-th node
and the centerline of the i-th beam are identified with the vertex vJ ∈ V
and edge ei ∈ E, respectively. The nodes (edges) are labeled 1, . . . , nv = ♯V
◦
(1, . . . , ne = ♯E).
V denotes the set of inner nodes and ∂V the set of
◦
boundary nodes. The sets V and ∂V are defined by the degree of a node

d(vJ ) := {number of the beams, which are incident at vJ },

as
◦
(2.1a) V := {vJ ∈ V : d(vJ ) > 1}

and
◦
(2.1b) ∂V := V \ V .

Clamped or free ends of single beams correspond to nodes vJ with d(vJ ) = 1

and, thus, to elements of ∂V . The position of a single beam (i = 1, . . . , ne ) is
given in the reference configuration in R3 by

(2.2) Ri (xi , t) = Ri0 + xi1 ei1 + xi2 ei2 + xi3 ei3 ,

with the local coordinate system ei . ei1 is a unit vector directed along the
centerline of the i-th beam. We assume, that the beams in the reference
configuration are straight and untwisted. The cross section area at xi1 is defined
by

Ai (xi1 ) = {(xi2 , xi3 )| − b/2 6 xi2 6 b/2, −h/2 6 xi3 6 h/2}.

202 Leugering and Rathmann

Ri0 is the offset of the point (0, 0, 0) of i-th beam in the rest configuration with
respect to the global fixed coordinate system e0 .We assume

e03 kei3 , ∀i = 1, . . . , ne ,

such that the representing planar graph lies in the e01 − e02 plane. The out-of-
the-plane displacement is taken in the ei3 direction. We denote the deformed
centerline in the following way

(2.3) R̂i (xi , t) = Ri0 + xi1 ei1 + wi (xi1 , t)ei3 .

We do not consider lateral displacement in the direction of ei2 , i.e. no in-plane

motion is considered here. The more general case will be treated in [23].
E(vJ ) denotes the set of all beams being incident at vJ ∈ V , i.e. |E(vJ )| =
d(vJ ). We distinguish the boundary nodes between Dirichlet- und Neumann-
type nodes denoted by VD and VN , respectively.
They are defined in the following way:

(2.4) VD := {vJ ∈ ∂V : wi (vJ ) = 0, wi ′ (vJ ) = 0, i ∈ E(vJ )}

(2.5) VN := {vJ ∈ ∂V : wi ′′ (vJ ) = 0, wi ′′′ (vJ ) = 0, i ∈ E(vJ )}.

Beyond the lateral displacements we introduce the angle φi (x, t) of Ai (x, t)

about ei2 with respect to the rest configuration. We assume that

(2.6) φ(x) = w′ (x),

according to the Euler-Bernuolli-hypotheses.

We use the following material constants on the i’th beam:
Ei Youngs modulus,
Gi shear modulus (not to mixed with transversal shear),
ρi density,
Ai cross section area,
Ii1 polar inertia of cross section area,
Ii2 inertia of cross section area,
Ei Ii2 bending stiffness,
Gi Ii1 torsional stiffness.
We consider the beams as onedimesional continua, denote the outer normal
by ±1 and introduce
(
−1 if R(vJ ) = Ri (0),
(2.7) εiJ =
+1 if R(vJ ) = Ri (li ).
 Networks of Beams 203

3 Scalar beam networks without torsion are inconsistent

In this section we consider a model for the motion of the network based on
conversation of the energy, as follows. The kinetic energy Ki of the i’th beam
is given by
Z li Z li
1 1
Ki = 2
ρi Ai ẇi dx + ρi Ii2 (ẇi ′ )2 dx
2 0 2 0

and the potential energy Ui by

Z li
1 2
Ui = Ei Ii2 wi′′ dx.
2 0

Therefore we obtain for the total energy Ei of i-th beam

(3.1)
Z li Z li Z li
1 1 ′ 2 1 2
Ei = Ki + Ui = 2
ρi Ai ẇi dx + ρi Ii2 (ẇi ) dx + Ei Ii2 wi′′ dx
2 0 2 0 2 0

and, hence, for the total energy E of the entire system:

nE
X
E= Ei .
i=1

The shear force F J and moment M J acting on vJ are defined as

X  
(3.2a) FJ = εiJ Ei Ii2 ẅi ′ − Ei Ii2 wi ′′′ (vJ )
i∈E(vJ )
X
(3.2b) M = J
εiJ Ei Ii2 wi ′′ (vJ )ei2 .
i∈E(vJ )
P 
The time deriviattive of the total energy is given by Ė = F J ẇ(v ) +
vJ ∈V J
P 
J ẇ′ (v ). In order to obtain the equations of motion and proper
vJ ∈V M J

compatibility conditions at the nodes, we assume that Ė = 0 (F J = 0, M J =

◦
0 ∀vJ ∈V )
nE n Z
X Z Z li
Ė = Ei Ii2 wi′′ ẇi′ dΩi − ′′′
Ei Ii2 wi ẇi dΩi + Ei Ii2 wi ′′′′ ẇi dx
i=1 ∂i ∂i 0
Z Z li Z li o
+ ρi Ii2 ẅi′ ẇi dΩ − ρi Ii2 ẅi′′ ẇi dx + ρi Ai ẅi ẇidx
∂i 0 0
= 0.
204 Leugering and Rathmann

Rewriting the boundary terms as sums over the nodes we obtain

nE
(Z )
X li  
′′′′ ′′
Ė = Ei Ii2 wi − ρi Ii2 ẅi + ρi Ai ẅi dx
i=1 0
X X  
+ εiJ Ei Ii2 ẅi ′ − Ei Ii2 wi ′′′ ẇi (vJ )
vJ ∈V i∈E(vJ )
X X
+ εiJ Ei Ii2 wi ′′ ẇi ′ (vJ ).
vJ ∈V i∈E(vJ )

Assuming continuity of the network we derive as another compatibility

condition

(3.3) ẇi ′ ei2 = ẇj ′ ej2 ∀i, j ∈ E(vJ ), vJ ∈ V .

Finally, we arrive at the system

′′
(3.4a) Ei Ii2 wi ′′′′ − ρi Ii2 ẅi + ρi Ai ẅi = 0 i = 1, . . . , ne ,
′
(3.4b) wi (vJ ) = wi (vJ ) = 0 vJ ∈ VD , i ∈ E(vJ ),
(3.4c) ẇi = ẇj i, j ∈ E(vJ ),
(3.4d) ẇi ′ ei2 = ẇj ′ ej2 i, j ∈ E(vJ ),
J
(3.4e) F = 0, J = 1, . . . , nJ ,
(3.4f) M J = 0, J = 1, . . . , nJ ,
i 0 i
(3.4g) wi (0, x ) = wi (x ) xi ∈ (0, li ),
(3.4h) ẇi (0, xi ) = wi 1 (xi ) xi ∈ (0, li ).

We find a nontrivial solution for (3.3) (ẇi ′ , ẇj ′ ) only if ei2 and ej2 are linear
independent, that means only for beams in a collinear configuration. The
problem is that this model does not account for torsion.

4 Second Model for out-of-plane dynamics

Here, we consider the displacements again in the ei3 -direction. However, now
we take torsion (denoted by θi ) into account. The kinetic and potential energy
is then given by
Z li Z li Z li
1 1 1 ′
Ki = ρi Ai ẇi 2 dx + ρi Ii2 ẇi ′2 dx + ρi Ii1 θ˙i 2 dx
2 0 2 0 2 0

and
Z li Z li
1 2 1 2
Ui = Ei Ii2 wi′′ dx + Gi Ii1 θi ′ dx.
2 0 2 0
 Networks of Beams 205

The total energy of the system is

nE
X nE
X
E= (Ki + Ui ) = Ei .
i=1 i=1

The shear force F J and the bending moment M J are now defined as
X  
(4.1a) FJ = εiJ ρi Ii2 ẅi ′ − Ei Ii2 wi ′′′ ,
i∈E(vJ )
X  
(4.1b) M =J
εiJ Gi Ii1 θi ′ ei1 + Ei Ii2 wi′′ ei2 .
i∈E(vJ )

In the next step we derive the equation of motion and the nodal conditions.
We start with the condition Ė = 0:
nE n Z
X Z Z li
Ė = Ei Ii2 wi′′ ẇi′ dΩi − Ei Ii2 wi ′′′ ẇi dΩi + Ei Ii2 wi ′′′′ ẇi dx
i=1 ∂i ∂i 0
Z Z li Z li
+ ρi Ii2 ẅi′ ẇi dΩi − ρi Ii2 ẅi′′ ẇi dx + ρi Ai ẅi ẇi dx
∂i 0 0
Z li Z li o
− Gi Ii1 θ̇i′′ θ˙i dx + ρi Ii1 θ̈i θ˙i dx
0 0
= 0.

This gives

(4.2a) ρi Ii1 θ¨i − Gi Ii1 θi ′′ = 0 (torsion)

(4.2b) ρi Ai ẅi − ρi Ii2 ẅi′′ + Ei Ii2 wi = 0 ′′
(lateral displacement)

and, at the nodes,

X X   X X  
0= εiJ ρi Ii2 ẅi ′ − Ei Ii2 wi ′′′ ẇi + εiJ Ei Ii2 wi′′ ẇi ′ + Gi Ii1 θi ′ θ˙i .
vJ ∈V i∈E(vJ ) vJ ∈V i∈E(vJ )

The transmission of moments reads like

X X  
′′ ′ ′ ˙
εiJ Ei Ii2 wi ẇi + Gi Ii1 θi θi
vJ ∈V i∈E(vJ )
X X  

= εiJ Ei Ii2 wi′′ ei2 + Gi Ii1 θi ′ ei1 θ˙i ei1 + ẇi ′ ei2 .
vJ ∈V i∈E(vJ )

Since the resulting moments M J and forces F J are zero at the node we obtain

(4.3a) θ˙i ei1 + ẇi ′ ei2 = θ̇j ej1 + ẇj′ ej2 ,

 206 Leugering and Rathmann

(4.3b) ẇi = ẇj , ∀i, j ∈ E(vJ ).

The description becomes complete with the initial data

(4.4a) θi (0, xi ) = θi 0 (xi ), θ˙i (0, xi ), = θi 1 (xi ) xi ∈ (0, li ),

(4.4b) wi (0, xi ) = wi 0 (xi ), ẇi (0, xi ) = wi 1 (xi ), xi ∈ (0, li ).

The difference of the models presented in this and the previous section
becomes obvious in the case of a carpenter square.
Theorem 4.1. A network consisting of ne Rayleigh-beams, described by
the system,

(4.5a) ρi Ii1 θ¨i − Gi Ii1 θi ′′ = 0, i = 1, . . . , ne ,

(4.5b) ρi Ai ẅi − ρi Ii2 ẅi′′ + Ei Ii2 wi ′′′′
= 0, i = 1, . . . , ne ,
(4.5c) θi ei1 + wi ′ ei2 = θj ej1 + wj′ ej2 , ∀i, j ∈ E(vJ ),
(4.5d) wi = wj , ∀i, j ∈ E(vJ ),
′
(4.5e) θi (vJ ) = wi (vJ ) = wi (vJ ), i ∈ E(VD ),

with
X  
(4.5f) εiJ ρi Ii2 ẅi ′ − Ei Ii2 wi ′′′ = 0,
i∈E(vJ )
X  
(4.5g) εiJ Gi Ii1 θi ′ ei1 + Ei Ii2 wi′′ ei2 = 0,
i∈E(vJ )

and initial data

(4.6a) θi (0, xi ) = θi 0 (xi ) ∈ H 2 (0, li ), xi ∈ (0, li ),

(4.6b) θ˙i (0, xi ) = θi 1 (xi ) ∈ H 1 (0, li ), xi ∈ (0, li ),
(4.6c) wi (0, xi ) = wi 0 (xi ) ∈ H 3 (0, li ), xi ∈ (0, li ),
(4.6d) ẇi (0, xi ) = wi 1 (xi ) ∈ H 2 (0, li ), xi ∈ (0, li ),

admits a unique strong solution (w, θ) with

θi ∈ C(R, H 1 (0, li )), θ˙i ∈ C(R, H 1 (0, li )),

wi ∈ C(R, H 2 (0, li )), ẇi ∈ C(R, H 2 (0, li )).

satisfying (4.5).
Proof. A proof of an analogous result for 3-d beams is presented in [23]. In the
statement of this theorem we do not insist on optimal regularity setups with
respect to the spaces of initial conditions. the proof is based on a generation
result a C0 -(semi)group. For the sake of simplicity we do not describe the
domain of the generator in full detail, and state sufficient conditions, only.
 Networks of Beams 207

5 Numerical Investigations
In section 4 we deduced a model for the out-of-the-plane motion of networks
of beams. In this section we first present an algorithm for dynamic domain
decomposition for networks of such beams which is implemented in a software
package, and then we present some numerical results.

5.1 Dynamic Domain Decompositon Method

The idea is, not to compute the response of the entire network at once, that is
on the global network level, but rather perform the computation on each single
edge (beam) and match the the edges in a iterative procedure. Therefore we
introduce the augmented Langrangian
Z T "X nE n Z li
1 2 2
L= −ρi Ii1 θ˙i + ρi Ai ẇi 2 − ρi Ii2 ẇi′
0 2 0
i=1
o
+ Gi Ii1 (θi ′ )2 + Ei Ii2 (wi′′ )2 dx dt
X σJ X
+ |wi (vJ ) − ζJ |2
2
J i∈E(vJ )
X νJ X
+ |θi (vJ )ei1 + wi ′ (vJ )ei2 − ηJ |2
2
J i∈E(vJ )
X X
+ µiJ (wi (vJ ) − ζJ )
J i∈E(vJ )
#
X X
′
+ ρiJ (θi (vJ )ei1 + wi (vJ )ei2 − ηJ ) dt
J i∈E(vJ )

which is of the form L = −K +U +penalty+Lagrangian multipliers for the nodal

conditions. To this Lagrangian we apply the ALG3 of Glowinski/Le Tallec [11]
to determine a saddle point:
1) ∂1 L((θ, w)n , (ζ, η)n−1 , (µ, ρ)n ) = 0,
n+ 12
2) µiJ = µniJ + σJ ∂31 L((θ, w)n , (ζ, η)n−1 , µniJ ),
n+ 21
ρiJ = ρniJ + νJ ∂32 L((θ, w)n , (ζ, η)n−1 , ρniJ ),
1
3) ∂2 L((θ, w)n , (ζ, η)n , (µ, ρ)n+ 2 ) = 0,
n+ 12
4) µn+1
iJ = µiJ + σJ ∂31 L((θ, w)n , (ζ, η)n , µniJ ),
n+ 21
ρn+1
iJ = ρiJ + νJ ∂32 L((θ, w)n , (ζ, η)n , ρniJ ).
Eliminating the Lagrange multipliers we deduce the following iteration
procedure in n for i = 1, . . . , ne
n
(5.1a) ρi Ii1 θ¨i − Gi Ii1 θi n′′ = 0,
208 Leugering and Rathmann

(5.1b) ρi Ai ẅi n − ρi Ii2 ẅin ′′ + Ei Ii2 wi n′′′′ = 0,

(5.1c) εiJ (ρi Ii2 ẅi n′′ − Ei Ii2 wi n′′′ ) + σJ wi n = giJ
n
,
n ′
(5.1d) εiJ (Gi Ii1 θi n′ ei1 + Ei Ii2 wi n′′ ei2 ) + νJ (θi n ei1 + wi n′ ei2 ) = giJ ,

with
2 X
n−1 n−1
(5.1e) giJ − 2σJ wi n−1 (vJ ) − (gjJ − 2σJ wjn−1 (vJ )) = giJ
n
,
dJ
j∈E(vJ )
′ n−1 n−1
giJ − 2ρJ (θi n−1 (vJ )ei1 + wi′ (vJ )ei2 )
2 X n−1 n−1 n
′
(5.1f) − (gjJ − 2ρJ (θi n−1 (vJ )ei1 + wj′ ′
(vJ )ei2 )) = giJ .
dJ
j∈E(vJ )

This iteration is analoguous to the one obtained in [16] with the updates
written in the format of [7]. In each single iteration we have to solve a system of
1-d hyperbolic and quasi-hyperbolic or even Petrovski-type PDE’s on each single
edge followed by an update of giJ n and g ′ n . We implement this algorithm in a
iJ
slightly different form: we formulate the Robin-boundary data and the update
for giJ on the velocity level
n
(5.2a) ρi Ii1 θ¨i − Gi Ii1 θi n′′ = 0,
(5.2b) ρi Ai ẅi n − ρi Ii2 ẅn′′ + Ei Ii2 wi n′′′′ = 0,
(5.2c) εiJ (ρi Ii2 ẅi n′′ − Ei Ii2 wi n′′′ ) + σJ ẇi n = giJ
n
,
n i n
(5.2d) εiJ (Gi Ii1 θi e1 + Ei Ii2 wi e2 ) + νJ (θ˙i e1 + ẇi e2 ) = giJ ,
n′ i n′′ i n′ i ′

with
2 X
n−1 n−1
(5.2e) giJ − 2σJ ẇi n−1 (vJ ) − (gjJ − 2σJ ẇj n−1 (vJ )) = giJ
n
,
dJ
j∈E(vJ )
n−1 n−1
′
giJ − 2ρJ (θ˙i (vJ )ei1 + ẇi ′n−1 (vJ )ei2 )
2 X n−1 n−1 ′ n
(5.2f) − ′
(gjJ − 2ρJ (θ˙i (vJ )ei1 + ẇi ′n−1 (vJ )ei2 ) = giJ .
dJ
j∈E(vJ )

The incorporation of velocities rather then displacements is motivated by the

work of [8] and [22]. Theorem 5.1. The algorithm (5.2) converges to a
solution (θ, w) in an H 1 × H 2 -sense.
Proof. The proof is presented in [23].
Some numerical results are presented in the next part of this section 5.2.

5.2 Numerical results

The dynamic domain decomposition algorithm (5.2) is been implemented using
the finite element method. We solve the equations (5.2a),(5.2b) by FEM with
 Networks of Beams 209

linear elements for the torsion and hermit-cubic elements for the wi ’s and use
a Newmark-algorithm to solve the equations in time.
The semi-discrete form of the algorithm after space discretization reads as
follows:

(5.1a) ρi Ii1 Mθi θ¨hi

n + G I K =f ,
i i1 θi θi
(5.1b)
ρi Ai Mwi + ωR ρi Ii2 IMwi w¨hi
n +E I K
i i2 wi =fwi , t ∈ [0, T ], i = 1, . . . nE ,

The right hand sides fθi , fwi include the right hand sides of the PDE,
n and g n′ . The updates for g n and g n′ are
applied moments and forces and giJ iJ iJ iJ
implemented in a relaxed form

n−1
λJ giJ − 2σJ ẇi n−1 (vJ )
(5.1c)
2 X 
n−1
− (gjJ − 2σJ ẇj n−1 (vJ )) + (1 − λJ )giJ
n−1 n
= giJ , λJ ∈ (0, 1),
dJ
j∈E(vJ )
 n−1 
n−1
′
µJ giJ − 2ρJ θ˙i (vJ )ei1 + ẇi ′n−1 (vJ )ei2
!
2 X  n−1

′ n−1
− gjJ − 2ρJ (θ˙i (vJ )ei1 + ẇi ′n−1
(vJ )ei2 )
dJ
j∈E(vJ )
′ n−1 ′ n
(5.1d) +(1 − µJ )giJ = giJ , µJ ∈ (0, 1).

For the updates we only need the velocities at the nodes, which we get from
the Newmark-algorithm. It is not necessary to compute deriviatives of θi or
wi . We mention, that the space discretization is allowed to differ on each edge.
Hence, the dimension of the matrices might be different on each edge. For time
discretization we choose the Newmark-algorithm in a predictor corrector form.
The time step size is denoted by ∆t = nTt ,

ti = i∆t, i = 0, . . . , nt .

We use the property of conservation of energy guaranteed by the Newmark

n and g n′ act as artifical forces and moments.
scheme. In the first iterates the giJ iJ
One observes
max Eh (t) − min Eh (t) → 0
ti ∈[0,T ] t∈[0,T ]

over the iterations for a closed system, but we do not use this as a stop-criteria.
In fact, we use the maximal error at the node

max max |wi − wj | 6 tol,

◦ ti ∈[0,T ]
vJ ∈V i,jE(vJ )
210 Leugering and Rathmann

and
max kXn − Xn−1 kE 6 tolE ,
t∈0:∆t:T

 n

θhi
Xn = n . We choose as norm the discrete energy norm
whi 1:nE

(5.2)
kXn (ti )kE
nE
(
X
= w˙hi
n (t )T (ρ A M
i
˙n ˙n T ˙n
i i wi + ωR ρi Ii2 IMwi )whi (ti ) + θhi (ti ) ρi Ii1 Mθi θhi (ti )
i=1
)
n
+ whi (ti )T Ei Ii2 Kwi whi
n
(ti ) + n
θhi (ti )T Gi Ii1 Kθi θhi
n
(ti ) .

The choice of the parameter σJ , νn , λJ , µJ is not obvious at all, an analytical

/ numerical treatment of optimal transmission conditions in the spirit of [22]
is under way. On the other side the dependence of material constants is
more transparent and will be examined in below. Our experiences suggest
a relaxation parameter between 0.8 and 0.9.
The implementation of a domain decomposition algorithm (DDM-
algorithm) to solve the elliptic (static) case is similar to the DDDM-algorithm
above. The only difference is with the update of traces rather of velocity-traces.
The choice of the numerical parameters can be taken from DDDM-algorithm
(and vice versa). We now discuss some examples.

Loaded clamped free beam. We consider the configuration showed in fig.

1. The beam is clamped at 0 and free at l = 0.8m. The material constants are
given in tab. 1.

F=-17.5N

q=-9.37N/m

500 300

Fig. 1. Clamped-free beam with loads.

 Networks of Beams 211
Table 1
Material constants.

density ρ = 7584 kg/m3

Young’s modulus E = 2.1 · 1011 N/m2
cross-section area 25mm A = 0.012 m2
inertia of the cross-section I2 = 4.36 · 10−9 m4

The explicit solution for the given loading is

(5.3) 
q F


 24EI ((0.8 − x)4 + 4 · 0.83 x − 0.84 ) + − x)3 + 3 · 0.52 x − 0.53 )
6EI ((0.5

 x ∈ [0, 0.5],
w(x) = q 4 3 4 F 2 3

 24EI ((0.8 − x) + 4 · 0.8 x − 0.8 ) + 6EI (3 · 0.5 x − 0.5 )


 x ∈ (0.5, 0.8].
The solution is plotted in fig. 2(a). This problem was solved with a
domain decomposition method, where we considered this configuration as a
serial network of two beams. The relative error is about 10−4 (see fig. 2(b)).
The graph is given in fig. 3. The material coefficients are equal for both beams.
We choose for the numerical parameter
σJ = νJ = 1000,
λJ = µJ = 0.9,
tol = 10−30 ,
tolE = 10−30 ,
ωR = 0.
ωR = 0 means, that we consider the Euler-Bernoulli-beam model. 266
Iterations where needed for convergence . In fig. 4(a) the total error on node 2
is plotted over the number of iterations. The relative error of about 10−4 was
reached once the error at the inner node settled at about 10−10 . Fig. 4 shows,
that we obtain linear convergence.
Vibrating beam. The lowest eigenfrequency of the beam given above is
approximately ω = 7.77719s−1 and the related lowest period is T1 = 0.128581s.
We choose as initial value the solution of the static problem 5.2 in fig. 2(a),
and ẇ1 = ẇ2 = 0.
With the numerical parameters
T = 0.26,
∆t = 0.01,
σJ = νJ = 200,
212 Leugering and Rathmann

solution
−4
x 10 relative error exact and numerical solution
0 1.115

1.11

−0.005 1.105

1.1
−0.01
1.095

1.09
−0.015
1.085

1.08
−0.02
1.075

1.07
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8

(a) Solution for clamped free beam. (b) Relativ error.

Fig. 2. Solution of the problem.

0.8

0.6

0.4

0.2

0 1 1 2 2 3
y

−0.2

−0.4

−0.6

−0.8

−1
0 0.2 0.4 0.6 0.8
x

Fig. 3. Graph and material constants of the network.

λJ = µJ = 0.9,
ωR = 0,
 Networks of Beams 213

i=266, |Xi(vJ)−Xj(vJ)|=0 0 i=266, ||Xi−Xi−1||E=1.2835e−28

0 10
10 10

−5 0
10 10
max|Xi(vJ)−Xj(vJ)|

||Xi−Xi−1||E
−10 −10
10 10

−15 −20
10 10

−20 −30
10 10
0 50 100 150 200 250 300 0 50 100 150 200 250 300
Iterationen i Iterationen i


(a) Error on node 2. − maxt |w1n (vJ ) − (b) kXn − Xn−1 kE .
′n ′n

w2n (vJ )|, - - maxt |w1 (vJ ) − w2 (vJ )|

Fig. 4. Convergence of static domain decomposition.

max ||X (t)−X (t)||

t i i−1 E
−2 2
10 10

0
10
−4
10
−2
10

−6
10 −4
10

−6
−8 10
10

−8
10
−10
10
−10
10

−12 −12
10 10
0 200 400 600 800 1000 1200 0 200 400 600 800 1000 1200
iterations n iterations n


(a) Error on node 2. − maxt |w1n (vJ ) − (b) kXn − Xn−1 kE .
′ ′

w2n (vJ )|, -. maxt |w1n (vJ ) − w2n (vJ )|

Fig. 5. Convergence of dynamic domain decomposition.

the algorithm stops after 1182 iterations with the following errors:
max kw1 (v2 ) − w2 (vJ )k = 9.99033 · 10−11 ,
t
max kw1 (v2 ) − w2 (vJ )k = 4.47642 · 10−11 ,
t
max kX1182 − X1181 kE = 3.93958 · 10−12 ,
t
max E(t) − min E(t) = 3.19250 · 10−6 .
t t
214 Leugering and Rathmann

The meshplot in fig. 6 shows that the lowest eigenfrequence with period
T1 ≈ 0.13s is recovered numerically.

0
10

−2
10 0.03

0.02

0.01
−4
10 0

−0.01

−0.02

−6 −0.03
10 0

0.5
0.25
−8 0.2
10 0.15
0.1
0 200 400 600 800 1000 1200 1 0
0.05
iterations n x time t

(a) maxt En (t) − mint En (t) (b) Dynamic responce of the beam.

Fig. 6. Variation of energy in time and dynamical response.

Carpenter square with loads. In the next example we consider the

carpenter-square with loads illustrated in fig. 7, with material parameters in
table 2. The initial value is the static solution of a domain decomposition for
w1 , w2 and θ1 , θ2 . The velocities are all equal to zero. We now observe torsional
effects in the beam, since the lateral elongation couples to the torsion angle at
the common node. With the numerical parameters
T = 1.23,
∆t = 0.025,
σJ = νJ = 160,
λJ = µJ = 0.9,
ωR = 1,
the algorithm stops after 296 iterations with the following errors:
max kw1 (v2 ) − w2 (vJ )k = 5.12849e · 10−9 ,
t
max kw1′ (v2 ) − w2′ (vJ )k = 2.63987 · 10−9 ,
t
max kX296 − X295 kE = 8.01305 · 10−11 ,
t
max E(t) − min E(t) = 3.74158 · 10−5 .
t t
 Networks of Beams 215

We choose the Rayleigh-beam model for the example. Fig. 8 shows the
convergence of the two stop-criteria. The behaviour of the algorithm is linear.
It is easy to see, that the error at the common node and the difference of two
successive iterates Xn is correlated. If we look at the loss of energy in fig. 9(a)
we observe the same. In fig. 9(b) the response in time is displayed. The state
for t = 0 is displayed at the buttom corner of the right side, while the final time
T = 1.23s upper left hand side.

F=150 N

500 N/m

2500

Fig. 7. Carpenter square with loads.

density ρ = 2690 kg/m3

Young’s modulus E = 7.3 · 1010 N/m2
shear modulus G = 2.54 · 1010 N/m2
cross-section area A = 0.000625 m2
ineratia of the cross-section I2 = 3.26 · 108 m4
polar inertia of the cross-section I1 = 6.51 · 10−8 m4

Table 2
Material constants for carpenter square.

6 Control of Networks
We now proceed to apply the domain decomposition procedure above to optimal
control problem. Since the model is time reversible it is sufficient to look at the
reachability problem. We give a CG-Algorithm to compute the control from
the adjoint system. We tested the method for serially connected beams and the
carpenter square. For the latter no exact controllabilty results are known.

6.1 The control problem

We consider a given network of Rayleigh/ Euler–Bernoulli–beams as above with
at least one boundary node clamped. We want to control at the inner (multiple)
216 Leugering and Rathmann

maxt ||Xn(t)−Xn−1(t)||E
10
10
carpenter3; Penalty: [2e+02 2e+02]; Relaxation: [0.9 0.9]; Iterationen: 296
0
10

5
10
−2
10

0
10
−4
10

−5
10
−6
10

−10
10
−8
10

−15
−10
10
10 0 50 100 150 200 250 300
0 50 100 150 200 250 300 iterations n


(a) Error on node 2. − maxt |w1n (vJ ) − (b) kXn − Xn−1 kE .
′ ′

w2n (vJ )|, -. maxt |w1n (vJ ) − w2n (vJ )|

Fig. 8. Stop criteria for DDDM.

or free nodes. The forward (in time) running system is

(6.1a) ρi Ii1 θ¨i − Gi Ii1 θi ′′ = 0,

(6.1b) ρi Ai ẅi − ρi Ii2 ẅi ′′ + Ei Ii2 wi ′′′′ = 0,

◦
(6.1c) θi ei1 + wi ′ ei2 = θj ej1 + wj′ ej2 , ∀i, j ∈ E(vJ ), vJ ∈V ,
◦
(6.1d) wi = wj , ∀i, j ∈ E(vJ ), vJ ∈V ,
(6.1e) θi (vJ ) = wi (vJ ) = wi ′ (vJ ) = 0, i ∈ E(VD ),
X
(6.1f) εiJ (ρi Ii2 ẅi ′ − Ei Ii2 wi ′′′ ) = F J , ∀vJ ∈ VC ,
i∈E(vJ )

(6.1g)
X
εiJ (Gi Ii1 θi ′ ei1 + Ei Ii2 wi ′′ ei2 ) = M J , ∀vJ ∈ VC
i∈E(vJ )

with initial conditions

(6.1h) θi (x, 0) = θi 0 , θ˙i (x, 0) = θi 1 ,

(6.1i) wi (x, 0) = wi 0 , ẇi (x, 0) = wi 1 .
 Networks of Beams 217

∆ E(n)
4
10

2
10

0
10

−2 7
10
6

1 5
−4 4
10 0

w
3
t
−1
2
−5
−6 −4
−3 1
10 −2
−1
0 50 100 150 200 250 300 0
1
2 0
iterations n

(a) maxt En (t) − mint En (t) (b) Dynamic responce of the carpenter
square.

Fig. 9. Results of DDDM for the carpenter square.

The proper function spaces are given by

(   

θi  θi
H :=  ∈ Hi0 ⊕ Hi[3]
1
:
wi i=1,...,n  wi
E

wi (vJ ) = 0 ∀i ∈ E(vJ ), vJ ∈ VD
)
wi (vJ ) = wj (vJ ) ∀i, j ∈ E(vJ ), vJ ∈ V \VD ,

and

(    

θi  θi
H1 :=  ∈ Hi1 ⊕ Hi[3]
2
,
wi i=1,...,nE
 wi
θi (vJ ) = wi (vJ ) = 0 ∀ i ∈ E(vJ ), vJ ∈ VD
θi (vJ )ei1 + wi (vJ )ei2 = θj (vJ )ej1 + wj′ (vJ )ej3 ,
′
)
wi (vJ ) = wj (vJ ), ∀i, j ∈ E(vJ ), vJ ∈ V \VD .

H′ und H1′ denote the corresponding dual spaces. Hi1 , Hi[3]

2 (i = 1, . . . , n )
E
are defined by

Hi1 := {v | v, v ′ ∈ L2 ([0, li ]; [ei1 ])},

218 Leugering and Rathmann
2
Hi[3] := {v | v, v ′ ∈ L2 ([0, li ]; [ei3 ]), v ′′ = L2 ([0, li ]; [ei3 ])}.

Further we introduce the Riesz-isomorphisms

*    +
θ φ
AH ,
w ψ
nE
X    

θ φ
(6.2a) = (ρi Ii1 θi , φi ) + (ρi Ai wi , ψi ) + (ρi Ii2 wi , ψi′ ) ′
, ∀ , ∈ H,
w ψ
i=1
*    +
θ φ
AH1 ,
w ψ
nE
X    

θ φ
(6.2b) = ′
(Gi Ii1 θi , φ′i ) + ′′
(Ei Ii2 wi , ψi′′ ) , ∀ , ∈ H1 .
w ψ
i=1

The Riesz-isomorphism between the product space

A : H1 × H → H1′ × H′

is given by
 
AH1
A= .
AH

The right hand sides of (6.2) are the scalar products of the corresponding
spaces. We consider the solutions of equation (6.1) with initial data in H′ × H1′ .
Therefore we use transposition, see [12]. Consider the solution of

(6.3a) ρi Ii1 θ̄¨i − Gi Ii1 θ̄i′′ = 0,

′′ ′′′′
(6.3b) ρi Ai w̄¨i − ρi Ii2 w̄¨i + Ei Ii2 wi = 0,

X ′
(6.3c) εiJ (ρi Ii2 w̄¨i − Ei Ii2 w̄i′′′ ) = F J ,
i∈E(vJ )
X
(6.3d) εiJ (Gi Ii1 θ̄i′ ei1 + Ei Ii2 w̄i′′ ei2 ) = M J ,
i∈E(vJ )

for fixed
   1
θ̄ 0 θ̄
0 ∈ H1 , 1 ∈ H.
w̄ w̄
 Networks of Beams 219

Thus we have

nE
(Z Z )
X T li
(ρi Ii1 θ̄¨i − Gi Ii1 θ̄i′′ )θi + (ρi Ai w̄¨i − ρi Ii2 w̄¨i + Ei Ii2 w̄i′′′′ )wi dx dt
′′
0=
i=1 0 0
nE
(Z
X li
= ρi Ii1 θ̄˙i (T )θi (T ) − ρi Ii1 θ̄i (T )θ˙i (T )
i=1 0
′
+ ρi Ai w̄˙ i (T )wi (T ) + ρi Ii2 w̄˙ i (T )wi ′ (T ) − ρi Ai w̄i (T )ẇi (T ) − ρi Ii2 w̄i′ (T )ẇi ′ (T ) dx
Z li
− ρi Ii1 θ̄i1 θi 0 − ρi Ii1 θ̄i0 θi 1 + ρi Ai w̄i1 wi 0 + ρi Ii2 w̄i1 ′ wi 0′
0
)
X Z T
0 ′ 1′
0 1
− ρi Ai w̄i wi − ρi Ii2 w̄i wi dx + F J w̄i + M J (θ̄i ei1 + w̄i′ ei2 ) dt.
vJ ∈VC 0

Upon defining
  " ˙ #! X Z S
θ̄ 0 θ̄ 1
LS , ˙1 := F J w̄i + M J (θ̄i ei1 + w̄i′ ei2 ) dt
w̄0 w̄ v∈VC 0
 1   0
(6.4) * θ θ̄ +
 w1  w̄0 
+    
 −θ 0  ,  θ̄ 1 
H′1 ×H′ ,H1 ×H
−w0 w̄1

we rewrite this identity as

   
* θ̇(T ) θ̄(T ) +
 ẇ(T )  w̄(T )
LT ((θ̄ 0 , w̄0 ), (θ̄ 1 , w̄1 )) =    
 −θ(T )  ,  θ̄˙ (T )  .
H′1 ×H′ ,H1 ×H
−w(T ) w̄˙ (T )

Definition 6.1. [(θ, w), (θ̇, ẇ)] ∈ H′ × H1′ is called a solution of (6.1) if
 
θ
w 
i)   ′ ′
 θ̇  ∈ C(R, H × H1 ),
ẇ
 
θ̄ 0
w̄0 
ii) (6.4) is fulfilled for all S ∈ R+ and all  
 θ̄ 1  ∈ H1 × H,
w̄1
220 Leugering and Rathmann

(6.5)
   
* θ̇(S) θ̄(S) +
 ẇ(S)  w̄(S)
   
 −θ(S)  ,  θ̄˙ (S) 
H′1 ×H′ ,H1 ×H
−w(S) w̄˙ (S)
   0
* θ1 θ̄ +
X Z S  w1  w̄0 
= F J w̄i + M J (θ̄i ei1 + w̄i′ ei2 ) dt +    
 −θ 0  ,  θ̄ 1  .
v∈VC 0 ′ ′
H1 ×H ,H1 ×H
−w0 w̄1

We choose as cost functional

(6.6)
Z   2   2
1 X T k θ(T )

0 k θ̇(T )

1
J= kF J k2 + kM J k2 dt + −z + −z .
2 0 2 w(T ) ′ 2 ẇ(T ) ′
J∈VC H H1

z 0 denotes the desired state at time T and z 1 the velocity at the final time.
The necessary first order optimality condition is

(6.7)
X Z T
0= F J F̂ J + M J M̂ J dt
J∈VC 0
" #
D    θ(T )   E D ˙   E
θ̂(T ) θ̂(T ) θ̇(T )
+k , A−1
H
0
−z , +k −1
, AH1 − z1 .
ŵ(T ) w(T ) ˙ )
ŵ(T ẇ(T )

[θ̂,ŵ]solves (6.1) with F J = F̂ J and M J = M̂ J and homogenous initial

φ
data. solves the equations
ψ

(6.8a) ρi Ii1 φ̈i − Gi Ii1 φ′′i = 0,

(6.8b) ρi Ai ψ̈i − ρi Ii2 ψ̈i′′ + Ei Ii2 ψi′′′′ = 0,

X
(6.8c) εiJ (ρi Ii2 ψ̈i′ − Ei Ii2 ψi′′′ ) = 0,
i∈E(vJ )
X
(6.8d) εiJ (Gi Ii1 φ′i ei1 + Ei Ii2 ψi′′ ei2 ) = 0,
i∈E(vJ )
 Networks of Beams 221

with the final data at T

    !
φ(T ) −1 θ̇(T ) 1
(6.9a) = kAH1 −z ,
ψ(T ) ẇi (T )
    !
φ̇(T ) −1 θ(T ) 0
(6.9b) = −kAH −z .
ψ̇(T ) wi (T )

From this we calculate F J and M J in equation (6.1) as

(6.10a) F J = −ψī ,
(6.10b) M J = −φī eī1 − ψī′ eī3 .

The dual problem of (6.7) is (Rockefellar-Fenchel duality)

  1 
1 ˆ ˆ −z
(6.11) hAf, f i + hΛf, f i = ,f ,
k z0
with 

φ(T )
   ψ(T ) 
f  
f = 0 =   ,
f1  φ̇(T ) 
ψ̇(T )
and Λ the so called HUM-operator associated with the problem. Λ maps f to
the final state of the forward running System:
Λ : H1 × H → H1′ × H′ ,
  
θ̇(T )
(6.12) − ẇ(T ) 
 
f 7→ Λf =    .
 θ(T ) 
w(T )

The duality h·, ·i operates between H1′ × H′ and H1 × H.

Equation (6.9) can be rewritten as

(6.13) a(f, fˆ) = l(fˆ),

where a(·, ·) is a bilinear form on (H1 × H)2 and l(·) a linear form on H1 × H
given as follows
1 1
(6.14a) a(f, fˆ) = hAf, fˆi + hΛf, fˆi = (f, fˆ) + (A−1 Λf, fˆ),
k k
  1   1 !
−z −z
(6.14b) l(fˆ) = , fˆ = A−1 , fˆ .
z0 z0
222 Leugering and Rathmann

6.2 CG-algorithm for the control problem

We want to solve (6.12) with a CG-algorithm. Glowinski and Lions suggested
in [9] and [10] a CG-Algorithmus for parabolic and hyperbolic equations for
diffenrent controls (Dirchlet or Neumann type controls). Their cost functional
is similiar to (6.6). The difference of both the final state and the desired
state, however, was taken in the L2 -norm. The CG-algorithm for the situation
considered here is given in a abstract setting.

(0) Initialization. Choose a value f0 ∈ H1 × H, calculate

(g0 , fˆ) = a(f 0 , fˆ) − l(fˆ),

or in dualities
1
hAg0 , fˆi = hAf 0 , fˆi + hΛf 0 , fˆi − l(fˆ)
k *  
1
0 ˆ 0 ˆ −z 1 ˆ
= Af , f + hΛf , f i − ,f .
k z0

For the evaluation of Λf0 we have to solve

ρi Ii1 φ̈i − Gi Ii1 φ′′i = 0,

ρi Ai ψ̈i − ρi Ii2 ψ̈i′′ + Ei Ii2 ψi′′′′ = 0,

with homogenous nodal conditions

X
εiJ (ρi Ii2 ψ̈i′ − Ei Ii2 ψi′′′ ) = 0,
i∈E(vJ )
X
εiJ (Gi Ii1 φ′i ei1 + Ei Ii2 ψi′′ ei2 ) = 0,
i∈E(vJ )

and the final data at T

 
φ(T )
(6.15a) = f00 ,
ψ(T )
 
φ̇(T )
(6.15b) = f01 .
ψ̇(T )

With the control given by

(6.16a) F J = −ψī (vJ ),

(6.16b) M J = −φī (vJ )eī1 − ψī′ (vJ )eī3 , ∀vJ ∈ VC ,
 Networks of Beams 223

we solve

ρi Ii1 θ¨i − Gi Ii1 θi ′′ = 0,

ρi Ai ẅi − ρi Ii2 ẅi ′′ + Ei Ii2 wi ′′′′ = 0,

X
εiJ (ρi Ii2 ẅi ′ − Ei Ii2 wi ′′′ ) = F J ,
i∈E(vJ )
X
εiJ (Gi Ii1 θi ′ ei1 + Ei Ii2 wi ′′ ei2 ) = M J
i∈E(vJ )

with homogenous initial conditions

θi (x, 0) = 0, θ˙i (x, 0) = 0,

wi (x, 0) = 0, ẇi (x, 0) = 0.

In order to determine g0 = (g00 , g01 ) ∈ H1 × H, the equations

 
θ̇(T )
(6.17) g00 = −A−1 −1 1 −1
H1 ẇ(T ) + AH1 z = −AH1 Ẋ(T ) + AH1 z ,
−1 1

 
−1 θ(T )
(6.18) 1
g0 = AH − AH−1 z 0 = A−1 −1 0
H X(T ) − AH z ,
w(t)

are to be solved. The Riesz-isomorphism from H to H′ is the identity. To

obtain g01 we solve

(6.19a) −Gi Ii1 gθ0i 0 ′′ = −θ̇(T ) + zθ1i ,

0 ′′′′ 1
(6.19b) Ei Ii2 gw i0
= −ẇ(T ) + zw i
,

with homogenous node conditions

X
0 ′′′
(6.19c) εiJ Ei Ii2 gw0 = 0,
i∈E(vJ )
X
0 ′ i 0 ′′ i
(6.19d) εiJ (Gi Ii1 gθ0 e1 + Ei Ii2 gw0 e2 ) = 0.
i∈E(vJ )

Hence, we obtain a solution of the equation

(g0 , fˆ) = a(f0 , fˆ) − l(fˆ).

Set
p0 = g0 .
224 Leugering and Rathmann

(1) Descent step (n ≥ 0).

Compute
1
hAḡn , vi = hApn , vi + hΛpn , vi,
k
respectively
" # !
1 −X̄˙ (T )
−1
(ḡn , v) = (pn , v) + A ,v .
k X̄(T )

Evaluate *" # +
−X̄˙ (T )
hΛpn , vi = ,v
X̄(T )
and solve the backward running system

(6.20a) ρi Ii1 φ̄¨i − Gi Ii1 φ̄′′i = 0,

′′
(6.20b) ρi Ai ψ̄¨i − ρi Ii2 ψ̄¨i + Ei Ii2 ψ̄i′′′′ = 0,

X ′
(6.20c) εiJ (ρi Ii2 ψ̄¨i − Ei Ii2 ψ̄i′′′ ) = 0,
i∈E(vJ )
X
(6.20d) εiJ (Gi Ii1 φ̄′i ei1 + Ei Ii2 ψ̄i′′ ei2 ) = 0,
i∈E(vJ )

with final data at T

 
φ̄i (T )
(6.20e) = p0n ,
ψ̄i (T )
" #
φ̄˙i (T )
(6.20f) = p1n .
ψ̄˙ (T )
i

The forward system is

(6.21a) ρi Ii1 θ̄¨i − Gi Ii1 θ̄i′′ = 0,

′′
(6.21b) ρi Ai w̄¨i − ρi Ii2 w̄¨i + Ei Ii2 wi ′′′′ = 0,

X ′
(6.21c) εiJ (ρi Ii2 w̄¨i − Ei Ii2 w̄i′′′ ) = −ψ̄ī ,
i∈E(vJ )
X
(6.21d) εiJ (Gi Ii1 θ̄i′ ei1 + Ei Ii2 w̄i′′ ei2 ) = −ψ̄ī′ eī1 − φ̄′ī eī3 ,
i∈E(vJ )
 Networks of Beams 225

with initial data

 
θ̄ 0
(6.21e) = 0,
w̄0
 1
θ̄
(6.21f) = 0.
w̄1
Now solve
 " # 
1 ˙
−X̄ (T )
(6.22) (ḡn , v) = pn + A−1 ,v ,
k X̄(T )

or
(6.23)
   
1 0 ˙ 1 1
(ḡn0 , v 0 )H1 + (ḡn1 , v 1 )H = p − A−1
H1 X̄ (T ), v
0
+ p + A−1
H X̄(T ), v
1
.
k n H1 k n H

Therefore, compute

(6.24a) 0 ′′
−Gi Ii1 ĝθn = −θ̄˙ (T ),
0 ′′′′
(6.24b) Ei Ii2 ĝwn = −w̄˙ (T ),
and set
1
(6.25) ḡn = pn + ĝn .
k
The new stepsize ρn is computed by
kgn k2H1 ×H
ρn = .
(ḡn0 , p0n )H1 + (ḡn1 , p1n )H
The inner products are defined as
(6.26a)
nE
( )
X   
0 ′ 0 ′ 0 ′′ 0 ′′
(ḡn0 , p0n )H1 = Gi Ii1 ḡθi n , pθi n + (Ei Ii2 ḡwi n , pwi n ,
i=1
(6.26b)
nE
( )
X



1 ′ 1 ′
(ḡn0 , p0n )H = ρi Ii1 ḡθ1i n , p1θi n + 1
ρi Ai ḡw in
, p1wi n + ρi Ii2 ḡw in
, pw i n ,.
i=1

Finally, we obtain the updates for the final state of the adjoint system
fn+1 and the residuum gn+1 by letting
fn+1 = fn − ρn pn ,
gn+1 = gn − ρn ḡn .
226 Leugering and Rathmann

(2) Convergence.
If
kgn+1 kH1 ×H
6 ε −→ STOP,
kg0 kH1 ×H
else compute
kgn+1 k2H1 ×H
γn =
kgn k2H1 ×H
and
pn+1 = gn+1 + γn pn ,
update n = n + 1 and go to 1).

As the error at the nodes (after domain decomposition) is not zero, we use
the averages for comuptation of the control from the backward system
1 X
(6.27a) FJ = − ψi (vJ ),
dJ
i∈E(vJ )
1 X
(6.27b) MJ = − (φ(vJ )ei1 + ψi′ (vJ )ei2 ).
dJ
i∈E(vJ )

6.3 Numerical results

The algorithm described above was applied to two serially connetcted beams
like in 5.2 and the carpenter squar from 5.2. We will present the results below.
Two serially connetcted beams. The initial state is given as an equlibrium
due to the load as shown in fig. 10. The matrial constants are displayed in
table 3. The aim is to control the system to rest. On the free end a force in
vertical direction and a moment is applied. In order to solve the backward and
forward running systems we use the same simulation routine as we described in
sec. 5.
 Networks of Beams 227

F=15oN

q=5ooN/m
25

1
0
0
125o
1
0
1
25oo
0
1
0
1
0
5ooo
1
0
1

Fig. 10. Clamped-free beam with loads.

density ρ = 2960 kg/m3

Young’s modulus E = 7.3 · 1010 N/m2
cross-section area A = 4.91 · 10−4 m2
inertia of the cross-section I2 = 1.917 · 10−8 m4

Table 3
Material constants of controlled single beam.

We started the CG-algorithm with the following parameters:

53 1
T = , ∆t = ,
64 128
σJ = νJ = 160, λJ = µJ = 0.9,
tol = 10−10 , tolE = 10−10 ,
ωR = 1, k = σJ /∆T = 20480.

The figures 11 and 12 show the results for the single beam. The same
computation was done with two serially connected beams of length 0.4m. The
results of the CG-algorithm are shown in figures 13 and 14. In this serial case we
took 10 CG-iterations. Often already 5 iterations where sufficent. The choice
for k = σJ /(∆t) appeared appropreate.
 228 Leugering and Rathmann

max. error ||w|| =2.28528e−01

final state of the adjoint system ∞
40 max. error ||w’||∞=7.99717e−02
φ(T) 0.25
φ’(T)
20
0.2

0 0.15

−20 0.1
0 1 2 3 4 5
x
20 0.05

10 0
0 1 2 3 4 5
0 max. error ||d/dt w|| = 5.69285e−01
∞
max. error ||d/dt w’||∞= 4.51186e+00
−10 d/dt φ(T)
d/dt φ’(T) 0.2
−20
0 1 2 3 4 5
control 0
40

20 −0.2

0
−0.4
−20 moment F
moment M
−40 −0.6
0 0.2 0.4 0.6 0.8 1 0 1 2 3 4 5

(a) Results of adjoint system. (b) Results of adjoint system.

Fig. 11. Results of CG-Algorithm for the single beam.

2 2
E(T)/E(0)= 0.0174 ||gn|| /||g0||
250 0
10
energy uncontrolled
energy controlled

200
−1
10

150
−2
10

100

−3
10
50

−4
10
0 1 1.5 2 2.5 3 3.5 4 4.5 5
0 0.2 0.4 0.6 0.8 1 iterations n

(a) Dissipation of energy. (b) Dissipation of energy.

Fig. 12. Convergence of CG-Algorithm.

 Networks of Beams 229

final state adjoint system max. error ||w|| =1.05017e−02

∞
40 max. error ||w’||∞=2.52588e−02
−3
φ(T) x 10
30 φ’(T) 12
20 10
10
8
0
6
−10
4
−20
0 1 2 3 4 5
2

40 0

−2
20
−4
0 1 2 3 4 5
0
max. error ||d/dt w||∞= 3.60220e−01
−20 max. error ||d/dt w’||∞= 4.55007e+00
d/dt φ(T)
d/dt φ’(T) 0.4
−40
0 1 2 3 4 5 0.3
x
controls 0.2
40
force F 0.1
2
20 moment M
2
0

0 −0.1

−0.2
−20
−0.3
−40
0 0.2 0.4 0.6 0.8 1 −0.4
t 0 1 2 3 4 5

(a) Results of adjoint system. (b) Reached final state of the sys-
tem.

Fig. 13. Results of CG-Algorithm for the serial case.

2 2
E(T)/E(0)=0.0075 ||gn|| /||g0||
250 10
0

energy uncontrolled
energy controlled

200
−1
10

150
−2
10

100

−3
10
50

−4
10
0 0 2 4 6 8 10
0 0.2 0.4 0.6 0.8 1 iterations n

(a) Dissipation of energy. (b) Convergence of CG-Algorithm.

Fig. 14. Convergence of CG-Algorithm for the serial case.

230 Leugering and Rathmann

Carpenter square. In 5.2 we described a carpenter square with loads. This

system is controlled to rest by a force and a moment on the free end. The
results are presented in figures 15 and 16. We took five CG-Iteration steps.

References

[1] J.-D. Benamou. A domain decomposition method for control problems. In

P. Bjørstad, editor, DD9 Proceedings, Bergen. John Wiley & Sons, 1996.
[2] J.-D. Benamou. A domain decomposition method with coupled transmission con-
ditions for the optimal control of systems governed by elliptic partial differential
equations. SIAM Journal Numerical Analysis, 33(6):2401–2416, 1996.
[3] U. Brauer and G. Leugering. Semi-discretization of control and observation
problems for a networks of elastic strings. Control and Cybernatics, 28(3):421–447,
1999.
[4] J.-S. Briffaut. Méthodes numérique pour le contrôle et la stabilisation rapide des
grandes structures flexibles. PhD thesis, École Nationale des Ponts et Chaussées,
1999.
[5] G. Chen, M. C. Delfour, A. M. Krall, and G. Payre. Modeling, Stabilization
and Control of Serially Connected Beams. SIAM J. Control and Optimization,
25(3):526–546, 1987.
[6] B. Dekoninck and S. Nicaise. Control of Networks of Euler-Bernoulli-Beams.
ESAIM-COCV, 4:57–81, 1999.
[7] Q. Deng. An analysis for a nonoverlapping domain decomposition iterative
procedure. SIAM J. Sci. Comput., 18(5):1517–1525, 1997.
[8] B. Despres. Methodes de decomposition de domaine pour les problemes de
propagation d’ondes en regime harmonique. Le theoreme de Borg pour l’equation
de Hill vectorielle. (Domain decomposition methods for harmonic wave-propagtion
problems. The Borg theorem for vectorial Hill equation). PhD thesis, Univ. de
Paris IX, 1991. [ISBN 2-7261-0706-0].
[9] R. Glowinski and J.L. Lions. Exact and approximate controllability for distributed
parameter systems (I). Acta Numerica, pages 269–378, 1994.
[10] R. Glowinski and J.L. Lions. Exact and approximate controllability for distributed
parameter systems (II). Acta Numerica, pages 159–333, 1995.
[11] R. Glowinski and P. Le Tallec. Augmented Lagrangian Interpretation of the
Nonoverlapping Schwarz Alternating Method. In T.F. Chan, R. Glowinski,
J. Reriaux, and O. B. Widlund, editors, The Third International Symposium on
Domain Decompostion Methods for Partial Differential Equations, pages 224–231.
SIAM, Philadelphia, 1990.
[12] V. Komornik. Exact Controllability and Stabilization, volume 36 of RAM. John
Wiley & Sons, Chichester, 1994.
[13] J. Lagnese, G. Leugering, and E.J.P.G. Schmidt. Modelling, Analysis and
Control of Multi-Link Flexible Structures. Systems & Control: Foundations &
Applications. Birkhäuser Basel, 1994.
[14] J.E. Lagnese. Domain decomposition in exact controllability of second order
hyperbolic systems on 1-d-networks. Control and Cybernetics, 28(3):531–556,
1999.
 Networks of Beams 231

[15] J.E. Lagnese and G. Leugering. Dynamic domain decomposition in approximate

and exact boundary control problems of transmission for wave equations. SIAM
Journal Control and Optimization, 38(2):503–537, 2000.
[16] G. Leugering. Domain Decomposition of Optimal Control Problems for Networks
of Euler-Bernoulli-Beams. In J.P. Puel and M. Tucsnak, editors, Control and
Partial Differential Equations, ESAIM Proc. 4, pages 223–233. ESAIM Paris,
1998.
[17] G. Leugering. Domain decomposition of optimal control problems of networks
of strings and timoshenko-beams. SIAM Journal Control and Optimization,
37(6):1649–1675, 1999.
[18] G. Leugering. On the semi-discretization of optimal control problems for networks
of elastic strings: global optimality systems and domain decomposition. Journal
of Computational and Applied Mathematics, 1999. to appear.
[19] G. Leugering. A domain decomposition of optimal control problems for dynamic
networks of elastic strings. Computational Optimization and Applications, 16:5–
27, 2000.
[20] G. Leugering and E.J.P.G. Schmidt. On the control of networks of vibrating
strings and beams. In Proceedings of the 28th IEEE Conference on Decision and
Control, volume 3, pages 2287–2290. IEEE, 1989.
[21] P. L. Lions. On the Schwarz Alternating Method III: A Variant for Nonoverlap-
ping Subdomains. In Chan, T.F., Glowinski, R., Reriaux, J. Widlund, O. B.,
editor, The Third International Symposium on Domain Decompostion Methods
for Partial Differential Equations, pages 202–223. SIAM, Philadelphia, 1990.
[22] F. Nataf, F. Rogier, and E. de Sturler. Optimal interface conditions for domain
decomposition methods. Technical report, I.R. CMAP (Ecole Polytechnique),
1994.
[23] W. Rathmann. Modellierung, Simulation und Steuerung von Netzwerken
aus schwingenden Balken mittels Bereichszerlegung. PhD thesis, Universität
Bayreuth, 2000.
 232 Leugering and Rathmann
E(T)/E(0)= 0.0221
450
energy uncontrolled
400 energy controlled

350

300

250

200
final state of adjoint system
0.4 150

0.2 100
φ(T)
0 φ’(T) 50

0
0 0.2 0.4 0.6 0.8 1 1.2 1.4
−0.2 t
0 0.5 1 1.5 2 2.5
0 x

−200

−400 d/dt φ(T)

d/dt φ’(T)
8
−600
0 0.5 1 controls 1.5 2 2.5
50 6
2

1 4
0
force F 0 t 2
moment M
−50 −1
0 0.2 0.4 0.6 0.8 1 1.2 1.4 −5 −4 −3 −2 −1 0
0 1 2 3
t

Fig. 15. Final adjoint state and control.

max. error ||w||∞=9.87220e−02
max. error ||w||∞=9.87220e−02
max. error ||w’||∞=3.51766e−02
max. error ||w’||∞=5.35614e−02
−0.03
0

−0.04
−0.02
−0.05

−0.04 −0.06

−0.07
−0.06
−0.08
−0.08
−0.09

−0.1 −0.1
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5 3

max. error ||d/dt w|| = 5.61083e−01 max. error ||d/dt w|| =8.63234e−01
∞ ∞
max. error ||d/dt w’||∞= 7.71978e−01 max. error ||d/dt ’||∞=3.36724e+00
0.7 1

0.6 0.8

0.5 0.6

0.4 0.4

0.3 0.2

0.2 0

0.1 −0.2

0 −0.4
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5 3

(a) Error on edge 1 (b) Error on edge 2

Fig. 16. Error at the edges.

Local Characterizations of Saddle Points and Their
Morse Indices

Yongxin Li , IBM Watson Research Center, Yorktown Hts, NY

Jianxin Zhou1 , Texas A&M University, College Station, Texas.

Abstract

In this paper, numerically computable bound estimates of the Morse

indices of saddle points are established through their new local minimax
type characterizations. The results provide methods for measuring
instability of unstable solutions in system design and control theory.

1 Introduction
Let H be a Hilbert space and J : H → R be a Frechet differentiable
functional. Denote by J ′ its Frechet derivative and J ′′ its second Frechet
derivative if it exists. A point û ∈ H is a critical point of J if

J ′ (û) = 0

as an operator J ′ : H → H. A number c ∈ R is called a critical value of J

if J(û) = c for some critical point û. For a critical value c, the set J −1 (c)
is called a critical level. When the second Frechet derivative J ′′ exists at a
critical point û, û is said to be non-degenerate if J ′′ (û) is invertible as a linear
operator J ′′ (û) : H → H. Otherwise û is said to be degenerate. The first
candidates for a critical point are the local maxima and minima to which the
classical critical point theory was devoted in calculus of variation. Traditional
numerical methods focus on finding such stable solutions. Critical points that
are not local extrema are unstable and called saddle points, that is, critical
points u∗ of J, for which any neighborhood of u∗ in H contains points v, w s.t.
J(v) < J(u∗ ) < J(w). In physical systems, saddle points appear as unstable
equilibria or transient excited states.

1
Supported in part by NSF Grant DMS 96-10076. E-mail: [email protected]

233
234 Li and Zhou

Assume that at a critical point u∗ , J ′′ (u∗ ) is a self-adjoint, Fredholm

operator from H → H, according to the spectral theory H has an orthogonal
spectral decomposition
H = H− ⊕ H0 ⊕ H+
where H − , H 0 and H + are respectively the maximum negative definite, the null
and the maximum positive definite subspaces of J ′′ (u∗ ) in H with dim(H 0 ) < ∞
and, H − , H 0 and H + are closed invariant subspaces under J ′′ (u∗ ). Following
the Morse theory, the Morse index of the critical point u∗ of J is MI(u∗ ) =
dim(H − ). So for a non-degenerate critical point, if its MI = 0, then it is a
local minimizer and a stable solution, and if its M I > 0, then it is a min-max
type and unstable solution. Throughout this paper, when the Morse index is
involved, we always assume that at a critical point u∗ , J ′′ (u∗ ) is a self-adjoint,
Fredholm operator from H → H. So the orthogonal spectral decomposition of
H will always be available.
When saddle points are introduced, one may mention another saddle point
approach in optimization and game theory. Note that in optimization and game
theory, saddle points are defined differently. For given vector spaces X and Y ,
let f : X × Y → R be a function. A saddle point of f is a point (x∗ , y ∗ ) ∈ X × Y
such that there exist neighborhoods N (x∗ ) of x∗ in X and N (y ∗ ) of y ∗ in Y
satisfying

f (x∗ , y) < f (x∗ , y ∗ ) < f (x, y ∗ ) ∀ x ∈ N (x∗ ), ∀ y ∈ N (y ∗ ).

Such a saddle point is usually also a critical point of f in X ×Y . The significant

difference here is that f has a splitting structure on X × Y and such a splitting
structure is known beforehand and fixed. Interactions between X and Y are
therefore very limited. While saddle points defined in critical point theory
are much more general. Such a splitting structure is in general not available.
If u∗ ∈ H is a nondegenerate critical point of a generic energy function J
with MI(u∗ )/geq1, by the spectral theory, we have H = H − ⊕ H + . Then we
have the decomposition u∗ = u− + u+ for u− ∈ H − , u+ ∈ H + and there are
neighborhoods N (u− ) of u− in H − and N (u+ ) of u+ in H + such that

J(v + u+ ) < J(u− + u+ ) < J(u− + w) ∀v ∈ N (u− ), w ∈ N (u+ ).

A formation close to a saddle point in optimization and game theory. However,

here the splitting structure H = H − ⊕ H + is only partial, interactions between
H − and H + are allowed. The most important difference here is that the
subspaces H − and H + cannot be available before we physically find the saddle
point u∗ . Therefore such an approach plays little role in finding or searching
saddle points in the critical point theory.
Multiple solutions with different performances and instabilities exist in
many nonlinear problems in natural and social sciences [22], [19], [14], [24],
 Saddle Points and Their Morse Indices 235

[13]. Stability is one of the main concerns in system design and control
theory. For instance, traveling waves have been observed to exist in suspended
bridges-a nonlinear beam equation [6] and showed as saddle points, therefore
unstable solutions, to their corresponding variational problem ((10) and (11) in
[6]). Those unstable solutions have been observed to have different instability
properties. How to mathematically measure their instability properties becomes
an interesting problem. As matter of fact, travelling waves make up an
important class of solutions to both reaction-diffusion equation and nonlinear
hyperbolic equations with “viscosity”. They are solutions of the form u =
u(x−ct) where c is a constant, the speed of the wave. Many phenomena arising in
various physical, or biological context can be modelled by travelling waves; such
as shock waves, nerve impulses, and various oscillatory chemical reactions. The
nice mathematical feature associated with such solutions is that the problem
often reduces to a nonlinear ordinary differential equations and the solutions
correspond to saddle points of their generic energy functions (ref. Chapter 24
in [21]). Stability (instability) analysis of such solutions is interesting to many
engineers and researchers and has been carried out in the literature.
Our objective is to develop some mathematical tool to measure instability
properties for saddle points. Usually stability describes certain property
possessed by a solution to a dynamic system. When one says that a solution
u∗ is a stable solution to a dynamic system, if u0 is near u∗ , the solution to
the dynamic system through u0 tends to u∗ as t → +∞. One may show that
a solution u∗ to a dynamic system is stable if the spectrum of the linearized
operator of the dynamic system at u∗ lies in the left-hand complex plane. When
the system is variational, an associated energy function is available. One may
also use the energy function to define stability. In this paper, we say that a
solution (critical point) u∗ is stable if it is a local minimizer of the associated
generic energy function. In this case MI(u∗ ) = 0. Thus any local perturbation of
a stable solution in an associated feasible function space will increase the energy
level. For an unstable solution (saddle point) u∗ , we may also use the maximum
dimension of a subspace in which a local perturbation of the unstable solution
u∗ in an associated feasible function space will always decrease the energy level
to define its instability index. Since by the definition, any local perturbation of
u∗ in H − will always decrease its energy level, for such variational problems, the
Morse index (= dim(H − )) of a solution can be used to measure its instability
(ref. [4],[21]). It is clear that the Morse index serves as a lower bound of the
instability index. On the other hand, in many applications, performance or
maneuverability is more desirable, in particular, in system design or control
of emergency or combat machineries. Usually unstable solutions have much
higher maneuverability or performance indices. For providing choice or balance
between stability and maneuverability or performance, it is important to solve
for multiple solutions and their Morse indices. When a saddle point u0 is
236 Li and Zhou

degenerate, the nullspace H 0 of J ′′ (u0 ) is nonempty. Since many different

situations may happy in H 0 , it is extremely difficult to determine the Morse
index in this case. On the other hand, whether or not a solution is degenerate
also depends on the domain of the solution. So we can not simply exclude such
situation.
When cases are variational, we will be dealing with multiple critical point
problems. So it is important for both theory and applications to numerically
solve for multiple critical points and their Morse indices in a stable way. So far,
little is known in the literature to devise such a feasible numerical algorithm.
Minimax principle is one of the most popular approaches in critical point theory.
However, most minimax theorems in the literature (See [1], [16], [17], [19],
[24]), such as the mountain pass, various linking and saddle point theorems,
require one to solve a two-level global optimization problem and therefore not
for algorithm implementation.
In [11], motivated by the numerical works of Choi-McKenna [7] and Ding-
Costa-Chen 9, the Morse theory and the idea to define a solution submanifold,
new local minimax theorems which characterize critical points as solutions to
a two-level local optimization problem are established. Based on the local
characterization, a new numerical minimax method for finding multiple critical
points is devised. The numerical method is implemented successfully to solve a
class of semilinear elliptic PDE on various domains for multiple solutions [11].
Although Morse index has been printed for each numerical solution obtained in
[11], their mathematical verifications have not been established. In [2], by using
a global minimax principle, A. Bahri and P.L. Lions established some lower
bound estimates for the Morse indices of solutions to a class of semilinear elliptic
PDE. There are also some efforts in the literature to numerically compute the
Morse index of a solution to a class of semilinear elliptic PDE. In addition to
finding a saddle point, one has to solve a corresponding linearized elliptic PDE
at the solution for its first a few eigen-values. It is very expensive and not much
success has been documented.
Since Morse index reflects local structure of a critical point, in this paper
we show that our local minimax characterization enables us to establish more
precise estimates for the Morse index of a saddle point in a more general setting.
By our results the Morse index of a saddle point based on the local minimax
method can be estimated even before we numerically compute the saddle point.
So no extra work is required in addition to computation for the saddle point.
In the last section, new local characterization of saddle points which are more
general than minimax solutions and bound estimates for their Morse indices
will be developed. In the rest of this section, we introduce some notations and
theorems from [11] for future use.
For any subspace H ′ ⊂ H, let SH ′ = {v|v ∈ H ′ , kvk = 1} be the unit
sphere in H ′ . Let L be a closed subspace in H, called a base space, and
 Saddle Points and Their Morse Indices 237
L
H = L L⊥ be the orthogonal decomposition where L⊥ is the orthogonal
complement of L in H. For each v ∈ SL⊥ , we define a closed half space
[L, v] = {tv + sw|w ∈ L, t > 0, s ∈ R}.
H
L ⊥ mapping P : SL⊥ → 2 is called the peak
Definition 1.1. A set-valued
mapping of J w.r.t. H = L L if for any v ∈ SL⊥ , P (v) is the set of all
local maximum points of J on [L, v]. A single-valued mapping p : SL⊥ → H is
a peak selection of J w.r.t. L if
p(v) ∈ P (v) ∀v ∈ SL⊥ .
For a point v ∈ SL⊥ , we say that J has a local peak selection w.r.t. L at v, if
there is a neighborhood N (v) of v and a mapping p : N (v) ∩ SL⊥ → H such that
p(u) ∈ P (u) ∀u ∈ N (v) ∩ SL⊥ .
Most minimax theorems in critical point theory require one to solve a two-
level global minimax problem and not for algorithm implementation. Our local
minimax algorithm requires one to solve only unconstrained local maximizations
at the first level. As pointed in [11], numerically it is great. However,
theoretically, it causes three major problems: (a) for some v ∈ SL⊥ , P (v)
may contain multiple local maxima in [L, v]. In particular, P may contain
multiple branches, even U-turn or bifurcation points; (b) p may not be defined
at some points in SL⊥ ; (c) the limit of a sequence of local maximum points
may not be a local maximum point. So the analysis involved becomes much
more complicated. We have been devoting great efforts to solve these three
problems. We solve (a) and (b) by using a local peak selection. Numerically it is
done by following certain negative gradient flow and developing some consistent
strategies to avoid jumps between different branches of P . As for Problem (c),
numerically we showed in [12] that as long as a sequence generated by the
algorithm converges, the limit yields a saddle point. New local characterization
of saddle points in this paper will further help us to solve those problems.
The following two local characterizations of saddle points are established in
[11] and played important role in our local theory. We then provide some
bound estimates of Morse indices of solutions based upon these two local
characterizations.
Lemma 1.1. Let vδ ∈ SL⊥ be a point. If J has a local peak selection p w.r.t.
L at vδ such that p is continuous at vδ and dis(p(vδ ), L) > α > 0 for some
α > 0, then either J ′ (p(vδ )) = 0 or for any δ > 0 with kJ ′ (p(vδ ))k > δ, there
exists s0 > 0, such that
J(p(v(s))) − J(p(vδ )) < −αδkv(s) − vδ k
for any 0 < s < s0 and
vδ + sd
v(s) = , d = −J ′ (p(vδ )).
kvδ + sdk
 238 Li and Zhou

The above result indicates that v(s) defined in the lemma represents a
direction for certain negative gradient flow of J(p(·)) from v. So it is clear
that if p(v0 ) is a local minimum point of J on any subset containing the path
p(v0 (s)) for some small s > 0 then J ′ (p(v0 )) = 0. In particular, if we define a
solution manifold n o
M = p(v) : v ∈ SL⊥ ,

we have p(v(s)) ⊂ M. A solution submanifold was first introduced by

Nehari in study of a dynamic system [15] and then applied by Ding-Ni in
study of semilinear elliptic PDE [17]. Although their definitions of a solution
submanifold are closely related to the problems that they were dealing with
and our definition of a solution submanifold is given in a quite general setting,
it is easy to check that their solution submanifold coincides with ours with
L = {0}. While for our solution submanifold, L can be any closed subspace. So
our definition of a solution submanifold generalizes their notions. Furthermore,
they prove that a global minimum point of the generic energy function J on
the solution submanifold M w.r.t. L = {0} yields a saddle point basically with
MI= 1. While our next theorem proved in [11] shows that a local minimum point
of the generic energy function J on our solution submanifold M also yields a
saddle point and the Morse index of such a saddle point can be greater than
one. Since for such an unstable saddle point u∗ , its generic energy is minimized
on the corresponding solution submanifold, the solution submanifold becomes
a stable submanifold of the unstable saddle point u∗ , in the sense that any
small perturbation of u∗ will increase the value of the generic functional J as
long as the perturbation stays on the submanifold M. Since u∗ is a saddle
point, a perturbation of u∗ off the submanifold may decrease the value of the
generic energy functional J. So the solution submanifold is also called a stable
submanifold.
Theorem 1.1. [11] Let v0 ∈ SL⊥ be a point. If J has a local peak selection
p w.r.t. L at v0 s.t.

(i) p is continuous at v0 ,

(ii) dis(p(v0 ), L) > 0 and

(iii) v0 is a local minimum point of J(p(v)) on SL⊥ .

Then p(v0 ) is a critical point of J.

The following PS condition will be used to replace the usual compact
condition.
Definition 1.2. A function J ∈ C 1 (H) is said to satisfy the Palais-Smale
(PS) condition, if any sequence {un } ∈ H with J(un ) bounded and J ′ (un ) → 0
has a convergent subsequence.
 Saddle Points and Their Morse Indices 239

2 Bound Estimates for Morse Index

The Morse index provides understanding of the local structure of a saddle point
and is used as an instability index for an unstable solution [4],[21]. It is an
important notion in stability analysis [4]. Although we have printed the Morse
index for each numerical solution computed by our minimax method in [11],
their mathematical justifications have not been verified. In this section, we
establish several bound estimates for the Morse index of a critical point based
on our minimax method.
Lemma 2.1. Let v0 ∈ SL⊥ be a point. If there exist a neighborhood
N (v0 ) of v0 and a locally defined mapping p : N (v0 ) ∩ SL⊥ → H such that
p(v) ∈ {L, v} ∀v ∈ N (v0 ) ∩ SL⊥ . If p is differentiable at v0 and u0 = p(v0 ) ∈
/ L,
then
p′ (v0 )({L, v0 }⊥ ) + {L, v0 } = H.
v0 + sw
Proof. For any w ∈ {L, v0 }⊥ , kwk = 1, denote vs = . Then there
kv0 + swk
exists s0 > 0 such that when |s| < s0 , we have vs ∈ N (v0 ) ∩ SL⊥ .
Consider the one dimensional vector function α(s) = PL⊥ (p(vs )), where
PL⊥ is the projection onto L⊥ . Since p is differentiable at v0 and vs smoothly
depends on s, α is differentiable at 0 and

∂vs
α′ (0) = PL⊥ (p′ (v0 )( )) = PL⊥ (p′ (v0 )(w)).
∂s
On the other hand, p(vs ) ∈ {L, vs }, we have α(s) = ts vs , where ts = hα(s), vs i
is differentiable. So α′ (0) = t′0 v0 + t0 w, where due to our assumption that
u0 = p(v0 ) ∈/ L, we have t0 6= 0. The two different expressions of α′ (0) imply

PL⊥ (p′ (v0 )(w)) = t′s (0)v0 + t0 w.

Then it leads to w ∈ {p′ (v0 )(w), L, v0 }. Since w is an arbitrary vector in

{L, v0 }⊥ , it follows that

{L, v0 }⊥ ⊂ {p′ (v0 )({L, v0 }⊥ ), L, v0 }

or

(2.1) p′ (v0 )({L, v0 }⊥ ) + {L, v0 } = H.

Lemma 2.2. Let v0 ∈ SL⊥ be a point. If there exist a neighborhood

N (v0 ) of v0 and a locally defined mapping p : N (v0 ) ∩ SL⊥ → H such that
p(v) ∈ {L, v} ∀v ∈ N (v0 ) ∩ SL⊥ . Assume that p is differentiable at v0 and
u0 = p(v0 ) ∈
/ L. If u0 is a critical point of J with M I(u0 ) > dim L + 1, then

p′ (v0 )({L, v0 }⊥ ) ∩ H − 6= {0}.

240 Li and Zhou

Proof. Denote H − the negative subspace of J ′′ (u0 ) and k = dim L + 1.

Then dim H − > k. By applying Lemma 2.1, there exit linearly independent
vectors e0 , e1 , . . . , ek ∈ H − which can be represented as ei = gi + fi with
gi ∈ p′ (v0 )({L, v0 }⊥ ) and fi ∈ {L, v0 }. f0 , f1 , . . . , fk have to be linearly
dependentP because k = dimP L + 1. So we can find real numbers a0 , a1 , . . . , ak
such that i=0 ai 6= 0 and ki=0 ai fi = 0. Therefore
k 2

k
X k
X
ai ei = ai gi ∈ p′ (v0 )({L, v0 }⊥ ) ∩ H − .
i=0 i=0
Pk
Because, e0 , e1 , . . . , ek are linearly independent, i=0 ai ei 6= 0. Thus, the
conclusion of the lemma is verified.
Theorem 2.1. Let v0 ∈ SL⊥ be a point. If J has a local peak selection p
w.r.t. L at v0 such that p is differentiable at v0 and u0 = p(v0 ) ∈ / L. If v0 is
a local minimum point of J ◦ p on SL⊥ , then u0 is a critical point of J with
M I(u0 ) 6 dim L + 1.
Proof. Since p is a local peak selection of J w.r.t. L at v0 , there exists a
neighborhood N (v0 ) of v0 such that p(v) ∈ {L, v}, ∀v ∈ N (v0 ) ∩ SL⊥ . By
applying Lemma 2.1, we have

p′ (v0 )({L, v0 }⊥ ) + {L, v0 } = H

or
codim(p′ (v0 )({L, v0 }⊥ )) 6 dim L + 1.
Now suppose that M I(u0 ) > dim L + 1. Denote H − the negative subspace of
J ′′ (u0 ). By Lemma 2.2, we have

(2.2) p′ (v0 )({L, v0 }⊥ ) ∩ H − 6= {0}.

Choose any w ∈ {L, v0 }⊥ , kwk = 1, such that p′ (v0 )(w) ∈ H − . Around

u0 = p(v0 ), we have the second order Taylor expansion
1
(2.3) J(u) = J(u0 ) + hJ ′′ (u0 )(u − u0 ), u − u0 i + o(ku − u0 k2 )
2
v0 + sw
Denote vs = , we have vs ∈ N (v0 ) ∩ SL⊥ for |s| small and then
kv0 + swk
dvs
|s=0 = w. So it follows
ds
(2.4) p(vs ) = u0 + sp′ (v0 )(w) + o(|s|).

Combining the above two estimates (2.3) and (2.4), we obtain

 Saddle Points and Their Morse Indices 241

J(p(vs ))
1
= J(u0 ) + hJ ′′ (u0 )(sp′ (v0 )(w) + o(|s|)), sp′ (v0 )(w) + o(|s|)i
2
+o(ksp′ (v0 )(w) + o(|s|)k2 )
1
= J(u0 ) + s2 hJ ′′ (u0 )(p′ (v0 )(w)), p′ (v0 )(w)i + o(s2 )
2
< J(u0 ),

where the last strict inequality holds for |s| sufficiently small, because
p′ (v0 )(w) ∈ H − .
Since vs ∈ N (v0 ) ∩ SL⊥ and u0 = p(v0 ), the above contradicts the
assumption that v0 is a local minimum point of J ◦ p on SL⊥ . Therefore
M I(u0 ) 6 dim L + 1.
Theorem 2.2. If p is a local peak selection of J w.r.t. L at v0 ∈ SL⊥ and
u0 = p(v0 ) is a nondegenerate critical point of J, then M I(u0 ) > dim L + 1.
Proof. Assume that k ≡ M I(u0 ) < dim L + 1. By our assumption, L u0 is
nondegenerate, i.e., J ′′ (u0 ) is invertible, we have H = H + H − where H +
is the maximum positive subspace and H − is the maximum negative subspace
corresponding to the orthogonal spectral decomposition of J ′′ (u0 ). It follows
that codim(H + ) = dim(H − ) = k < dim L + 1, so there exists a non-zero
vector v ∈ H + ∩ {L, v0 }. When v ∈ H + , for sufficient small t, we have
J(u0 + tv) > J(u0 ). But this contradicts to that u0 is a local maximum point
of J in the subspace {L, v0 }. Therefore, M I(u0 ) > dim L + 1.
Theorem 2.3. Assume that p is a local peak selection of J w.r.t. L at
v0 ∈ SL⊥ such that p is differentiable at v0 and u0 = p(v0 ) ∈ / L. If v0
′′
is a local minimum point of J ◦ p on SL⊥ , and J (u0 ) is invertible, then
M I(u0 ) = dim L + 1.
Proof. Since under the conditions, we have proved that u0 = p(v0 ) is a non-
degenerate critical point of J. The conclusion follows by combining the last two
theorems.
Theorem 2.4. Let v0 ∈ SL⊥ be a point. If there exist a neighborhood N (v0 )
of v0 and a locally defined mapping p : N (v0 )∩SL⊥ → H such that p(v) ∈ {L, v},
J ′ (p(v)) ⊥ {L, v}, ∀v ∈ N (v0 ) ∩ SL⊥ and p differentiable at v0 . If v0 ∈ SL⊥ is
a local minimum point of J ◦ p on SL⊥ with u0 = p(v0 ) ∈ / L, then u0 is a critical
point of J with M I(u0 ) 6 dim L + 1.
Proof. We first prove that u0 = p(v0 ) is a critical point of J. The second part
of the theorem follows from a similar proof of Theorem 2.1.
For any w ∈ {L, v0 }⊥ , denote

v0 + sw
v(s) = .
kv0 + swk
242 Li and Zhou
dv(s)
We have v(s) ∈ N (v0 ) ∩ SL⊥ for |s| small and ds |s=0 = w. Therefore

dv(s)
p(v(s)) = p(v0 ) + sp′ (v0 ) |s=0 + o(|s|)
ds
= u0 + sp′ (v0 )(w) + o(|s|).

It follows that

J(p(v(s))) = J(p(v0 )) + J ′ (p(v0 ))(p(v(s)) − p(v0 )) + o(kp(v(s)) − p(v0 )k)

= J(u0 ) + sJ ′ (u0 )p′ (v0 )(w) + o(|s|).

If
J ′ (u0 )p′ (v0 )(w) 6= 0
for some w ∈ {L, v0 }⊥ , then when |s| is sufficiently small, we can choose either
s > 0 or s < 0 such that

J(p(v(s))) < J(p(v0 ))

which contradicts the assumption that v0 is a local minimum point of J ◦ p on

SL⊥ . Thus
J ′ (u0 )p′ (v0 )({L, v0 }⊥ ) = 0.
Since by our assumption
J ′ (u0 )({L, v0 }) = 0
and by Lemma 2.1
p′ (v0 )({L, v0 }⊥ ) + {L, v0 } = H,
it follows that
J ′ (u0 )u = 0 ∀u ∈ H,
i.e., u0 = p(v0 ) is a critical point of J.
It is worthwhile indicating that if p is a local peak selection of J at v0 ∈ SL⊥ ,
then p(v) ∈ [L, v] and J ′ (p(v)) ⊥ [L, v] for all v ∈ N (v0 ) ∩ SL⊥ . If {vn } ⊂ SL⊥ ,
vn → v0 and un = p(vn ) → u0 , we have u0 ∈ [L, v0 ] and J ′ (u0 ) ⊥ [L, v0 ]. So
such a locally defined mapping generalized the notion of a local peak selection
and resolved the problem that a limit of a sequence of local maximum points
may not be a local maximum point. This generalization has a potential to
design a new type of local algorithm for finding multiple saddle points that are
not necessarily of a minimax type.
Theorem 2.5. If u0 ∈ / L is a non-degenerate critical point of J such that
u0 is not a local minimum point of J along any direction v ∈ {L, u0 }, then

M I(u0 ) > dim L + 1.

 Saddle Points and Their Morse Indices 243

Proof. Assume that k = M I(u0 ) < dim L + 1. By our assumption, u0 is non-

degenerate, i.e., J ′′ (u0 ) is invertible, we have H = H + ⊕ H − where H + is
the maximum positive subspace and H − is the maximum negative subspace
corresponding to the orthogonal spectral decomposition of J ′′ (u0 ). It follows
that codim(H + ) = dim(H − ) = k < dim L + 1, so there exists a non-zero vector
v ∈ H + ∩ {L, u0 }. When v ∈ H + , for sufficiently small t, we have

J(u0 + tv) > J(u0 ).

It then contradicts the assumption that u0 is not a local minimum point of J

along any direction v ∈ {L, u0 }.

3 Application to Semilinear Elliptic PDEs

Semilinear elliptic boundary-value problems (BVP) are known to be rich in
multiple solutions. Some of the solutions are stable and others are unstable.
How to measure the instability properties for unstable solutions is an interesting
problem, which is actually part of the motivation of this paper. When the profile
of a solution shown as in Figure 1 was obtained and presented to the nonlinear
partial differential equation community, it generated warm debates about the
existence and the Morse index of such a solution. The authors are happy to
know that the existence issue has been recently settled [23]. But the Morse
index of such a solution is still unsolved. In this section, we try to answer the
question.
Consider a semilinear elliptic Dirichlet BVP on a piecewise smooth bounded
domain Ω in Rn which has many applications in physics, engineering, biology,
ecology, geometry, etc

∆u(x) + f (x, u(x)) = 0, x ∈ Ω,
(3.1)
u(x) = 0, x ∈ ∂Ω,

where the function f (x, u(x)) satisfies the following standard hypothesis:

(h1) f (x, u(x)) is locally Lipschitz on Ω̄ × R;

(h2) there are positive constants a1 and a2 such that

(3.2) |f (x, ξ)| 6 a1 + a2 |ξ|s

n+2
where 0 6 s < n−2 for n > 2. If n = 2,

(3.3) |f (x, ξ)| 6 a1 exp φ(ξ)

where φ(ξ)ξ −2 → 0 as |ξ| → ∞;

(h3) f (x, ξ) = o(|ξ|) as ξ → 0;

244 Li and Zhou

(h4) there are constants µ > 2 and r > 0 such that for |ξ| > r,

(3.4) 0 < µF (x, ξ) 6 ξf (x, ξ),

Rξ
where F (x, ξ) = 0 f (x, t)dt.

(h4) says that f is superlinear, which implies that there exist positive numbers
a3 and a4 such that for all x ∈ Ω̄ and ξ ∈ R

(3.5) F (x, ξ) > a3 |ξ|µ − a4 .

The generic energy functional associated to the Dirichlet problem (3.1) is

Z Z
1
(3.6) J(u) = |∇u(x)|2 dx − F (x, u(x))dx, u ∈ H ≡ H01 (Ω),
2 Ω Ω
R
where we use an equivalent norm kuk = Ω |∇u(x)|2 dx for the Sobolev space
H = H01 (Ω). It is well known [19] that under Conditions (h1) through (h4),
J is C 1 and satisfy (PS) condition. A critical point of J is a weak solution,
and also a classical solution of (3.1). 0 is a local minimum point of J and so a
trivial solution. Moreover, in any finitely dimensional subspace of H, J goes to
negative infinity uniformly. Therefore, for any finite dimensional subspace L,
the peak mapping P of J w.r.t. L is nonempty.
We need one more hypothesis, that is
f (x,ξ)
(h5) |ξ| is increasing w.r.t. ξ, or
f (x,ξ)
(h5’) f (x, ξ) is C 1 w.r.t. ξ and fξ (x, ξ) − ξ > 0.

It is clear that (h5’) implies (h5). If f (x, ξ) is C 1 in ξ, then (h5) and (h5’)
are equivalent. All the power functions of the form f (x, ξ) = |ξ|k ξ with k > 0,
satisfies (h1) through (h5’), and so do all the positive linear combinations of
such functions. Under (h5) or (h5’), J has only one local maximum point in any
direction, or, the peak mapping P of J w.r.t. L = {0} has only one selection.
In other words, P = p. The proof can be found in [17] and [14].
Theorem 3.1. Under the hypothesis (h1) through (h5), if the peak mapping
P of J w.r.t. a finitely dimensional subspace L is singleton at v0 ∈ SL⊥ and
for any v ∈ SL⊥ around v0 , a peak selection p(v) is a global maximum point of
J in [L, v], then p is continuous at v0 .
Proof. See [11].
Theorem 3.2. Assume that Conditions (h1) – (h5’) are satisfied and that
there exist positive constants a5 and a6 s.t.

(3.7) |fξ (x, ξ)| 6 a5 + a6 |ξ|s−1

where s is specified in (h2). Then the only peak selection p of J w.r.t. L = {0}
is C 1 .
 Saddle Points and Their Morse Indices 245

Proof. See [11].

Since w0 = 0 is the local minimum point of J and along each direction
v ∈ H, J has only one maximum point p(v), we have J(p(v)) > 0, ∀v ∈ S. If
for each v ∈ SL⊥ , p(v) is a local maximum point of J in [L, v], then p(v) is the
p(v)
only local maximum point of J along the direction u = kp(v)k , Therefore we
have J(p(v)) > 0, ∀v ∈ SL⊥ . As a composite function J(p(·)) is bounded from
below by 0. So Ekeland’s variational principle can be applied. With the PS
condition, existence result can also be established.
Theorem 3.3. Under the hypothesis of (h1) through (h5), and that there
exist positive constants a5 and a6 such that

|fξ (x, ξ)| 6 a5 + a6 |ξ|s−1 ,

where s is specified in (h2), if v0 = arg minv∈SH J(p(v)) then u0 = p(v0 ) is a

critical point with MI(u0 ) = 1.
Proof. Assume u0 = p(v0 ) = t0 v0 . Consider the 1-dimensional function
Z Z
1 2 2
g(t) = J(tv0 ) = t |∇v0 (x)| dx − F (x, tv0 (x)) dx.
2 Ω Ω

We have Z Z
g′ (t) = t |∇v0 (x)|2 dx − f (x, tv0 (x))v0 (x) dx.
Ω Ω
So Z
f (x, t0 v0 (x)) 2
1= v0 (x) dx.
Ω t0 v0 (x)
Meanwhile we have
Z Z
′′ 2
g (t) = |∇v0 (x)| dx − fξ (x, tv0 (x))v02 (x) dx
Ω Ω

Z
g′′ (t0 ) = 1 − fξ (x, t0 v0 (x))v02 (x) dx
ZΩ
f (x, t0 v0 (x)) 2
< 1− v0 (x) dx (ref. (h5))
Ω t0 v0 (x)
= 0,

which implies that

H 0 ∩ {L, v0 } = {0},
where H 0 is the nullspace of the linear operator J ′′ (u0 ) and L = {0}, i.e.,
J ′′ (u0 ) may be degenerate, however its restriction to the direction of v0 will not
be degenerate, where u0 = t0 v0 is a critical point. By Theorem 2.1, we obtain

MI(u0 ) = dim(L) + 1 = 1.
246 Li and Zhou

In the rest of this section, we display a numerical solution of (3.1) where

f (x, u(x)) = u3 (x) and the domain Ω is the dumbbell shaped domain as shown
in the lower part of Figure 1. It can be checked that all the assumptions
(h1) − (h5) are satisfied. So the results in this section can be applied.
Many unstable solutions to this problem has been numerically computed and
documented in [11]. The details of the numerical minimax method and how
to compute such solutions are described in [11]. Some convergent properties of
this numerical minimax method are established in [12]. Here we only display
the profile and the contours the solution concentrated near the center of the
corridor of the dumbbell-shaped domain. This solution is also captured by
using a different numerical algorithm in [5]. It is this solution that generated
warm debates about the existence and the Morse index of such a solution. Since
L = {0}, by the results in this section, we can now conclude that the Morse
index of this solution is one.

4 Some New Saddle Point Theorems

As our convergence results in [12] indicate that our algorithm can be used to find
a non-minimax critical point, e.g., a Monkey saddle point. Thus the argument
already exceeded the scope of a minimax approach. So far the only results we
found in the critical point literature which are more general than a minimax
principle are two theorems proved by S. I. Pohozaev in [10] or [18]. The following
results are interesting generalizations. The first one is an embedding result.
It is general but lacks of characterization. The second result has potential
applications in devising a new numerical algorithm.
Theorem 4.1. Given L = span{w1 , ..., wk } in H and let

J¯(t, v, t1 , .., tk ) ≡ J(tv + t1 w1 + ... + tk wk ).

If (t∗0 , v ∗ , t∗1 , .., t∗k ) is a conditional critical point of J¯ subject to v ∈ SL⊥ with
t∗ 6= 0, then t∗0 v ∗ + t∗1 w1 + ... + t∗k wk is a critical point of J.
Proof. By the Lagrange Multiplier Theorem, there exist λ, µ, η1 , .., ηk with
λ2 + µ2 + η12 + ... + ηk2 6= 0 such that the Lagrange functional

k
X
¯ v, t1 , ..., tk ) + µkvk +
L(t, v, t1 , ..., tk ) = λJ(t, ηi hwi , vi
i=1

has a critical point at (t∗0 , v ∗ , t∗1 , .., t∗k ). So we have

 Saddle Points and Their Morse Indices 247

15

10
w−axis

−5
−1
y−

0
ax
is

1 −1 −1.5
1 0
Z 3 2
x−axis
X

1
y−axis

1.12.760.551
2.
2
5.5
3.86

0
1

6.6 1
3.3611.65
4.9

−1
−1.5 −1 0 1 2 3
x−axis

Fig. 1. The profile of a positive solution and its contours with J = 159.0 and
umax = 13.63.
248 Li and Zhou

∂L
(4.1) = 0 ⇒ λJ¯t (t∗0 , v ∗ , t∗1 , ..., t∗k ) = 0;
∂t
k
X
∂L
(4.2) = 0 ⇒ λJ¯v′ (t∗0 , v ∗ , t∗1 , ..., t∗k ) + µkv ∗ k′ + ηi wi = 0;
∂v
i=1
∂L
= 0 ⇒ λJ¯t′i (t∗0 , v ∗ , t∗1 , .., t∗k ) = 0 or
∂ti
(4.3) λhJ ′ (t∗0 v ∗ + t∗1 w1 + ... + t∗k wk ), wi i = 0, (i = 1, .., k).

From (4.2), we have

(4.4)
k
X
hλJ¯v (t∗0 , v ∗ , t∗1 , .., t∗k ), vi + µhkv ∗ k′ , vi + ηi hwi , vi = 0 ∀v ∈ H.
i=1

In particular
k
X
(4.5) λhJ¯v (t∗0 , v ∗ , t∗1 , .., t∗k ), v ∗ i + µhkv ∗ k′ , v ∗ i + ηi hwi , v ∗ i = 0.
i=1

Since hkv ∗ k′ , v ∗ i = 1 and hwi , v ∗ i = 0 for i = 1, ..., k, we obtain

(4.6) λhJ¯v (t∗0 , v ∗ , t∗1 , .., t∗k ), v ∗ i + µ = 0.

So λ = 0 will lead to µ = 0 and then ηi = 0 by choosing v = wi in (4.3). It

contradicts to λ2 + µ2 + η12 + ... + ηk2 6= 0. Therefore λ 6= 0 and (4.1) gives

J¯t′ (t∗0 , v ∗ , t∗1 , ..., t∗k ) = 0

or
hJ ′ (t∗0 v ∗ + t∗1 w1 + ... + t∗k wk ), v ∗ i = 0.
It leads to

hJ¯v′ (t∗0 , v ∗ , t∗1 , ..., t∗k ), v ∗ i = t∗0 hJ ′ (t∗0 v ∗ + t∗1 w1 + ... + t∗k wk ), v ∗ i
= 0.

(4.6) then yields µ = 0. Taking any v ⊥ [w1 , ..., wk ] in (4.4), we obtain

λhJ¯v′ (t∗0 , v ∗ , t∗1 , ..., t∗k ), vi = 0

or
t∗0 hJ ′ (t∗0 v ∗ + t∗1 w1 + ... + t∗k wk ), vi = 0.
 Saddle Points and Their Morse Indices 249

Since t∗0 6= 0, it leads to

hJ ′ (t∗0 v ∗ + t∗1 w1 + ... + t∗k wk ), vi = 0 ∀v ⊥ [w1 , ..., wk ].

Taking (4.3) into account, we have

hJ ′ (t∗0 v ∗ + t∗1 w1 + ... + t∗k wk ), ui = 0 ∀u ∈ H

or
J ′ (t∗0 v ∗ + t∗1 w1 + ... + t∗k wk ) = 0.
So u∗ = t∗0 v ∗ + t∗1 w1 + ... + t∗k wk is a critical point.
It is clear that Theorem 4.1 reduces to Pohozaev’s embedding result in [10]
or [18] if we set L = {0}, the trivial subspace.
Theorem 4.2. Let v ∗ ∈ SL⊥ be a point. If J has a local peak selection p
w.r.t. L at v ∗ and u∗ = p(v ∗ ) such that

(i) p is Lipschitz continuous at v ∗ ,

(ii) dis(u∗ , L) > 0,

(iv) u∗ is a conditional critical point of J on p(SL⊥ ),

then u∗ is a critical point of J.

Proof. Suppose that kJ ′ (u∗ )k > 0. Set δ = 21 kJ ′ (u∗ )k. By Lemma 1.1, there
exists s0 > 0 such that

J(p(v ∗ (s))) − J(p(v ∗ )) < −δdis(u∗ , L)kv ∗ (s) − v ∗ k ∀0 < s < s0

where
v ∗ + sd
v ∗ (s) = ∈ N (v ∗ ) ∩ SL⊥ , d = −J ′ (p(v ∗ )).
kv ∗ + sdk
Then we have
J(p(v ∗ (s))) − J(p(v ∗ )) kp(v ∗ (s)) − u∗ k
(4.7) < −δdis(u∗ , L), ∀ 0 < s < s0 ,
kp(v ∗ (s)) − u∗ k kv ∗ (s) − v ∗ k

where N (v ∗ ) is a neighborhood of v ∗ in which the local peak selection p is

defined. Since p is Lipschitz continuous at v ∗ and u∗ is a conditional critical
point of J on p(SL⊥ ),
kp(v ∗ (s)) − u∗ k
kv ∗ (s) − v ∗ k
is bounded and
J(p(v ∗ (s))) − J(p(v ∗ ))
→ 0 as s → 0.
kp(v ∗ (s)) − u∗ k
So the left hand side of (4.7) goes to zero, which leads to a contradiction.
250 Li and Zhou

Theorem 4.3. Let v0 ∈ SL⊥ be a point. If J has a local peak selection

p w.r.t. L such that p is differentiable at v0 and u0 = p(v0 ) ∈ / L. If u0 is a
conditional critical point of J on p(SL⊥ ) and v0 is not a local maximum point
of J ◦ p along the projection of any direction v on SL⊥ , then u0 is a critical
point of J with M I(u0 ) 6 dim L + 1. In addition, if u0 is nondegenerate, then
M I(u0 ) = dim L + 1.
Proof. By Theorem 4.2, we obtain that u0 is a critical point of J. Then following
a similar argument in the proof of Theorem 2.1, until we have
1
J(p(vs )) = J(u0 ) + hJ ′′ (u0 )(sp′ (v0 )(w) + o(|s|)), sp′ (v0 )(w) + o(|s|)i
2
+o(ksp′ (v0 )(w) + o(|s|)k2 )
1
= J(u0 ) + s2 hJ ′′ (u0 )(p′ (v0 )(w)), p′ (v0 )(w)i + o(s2 )
2
< J(u0 ),

where
v0 + sw
vs = ∈ N (v0 ) ∩ SL⊥ , w ∈ [L, v0 ]⊥ , kwk = 1, and p′ (v0 )(w) ∈ H − .
kv0 + swk

So the last strict inequality contradicts to our assumption that v0 is not a

local maximum point of J ◦ p along the projection of any direction v on SL⊥ .
Thus M I(u0 ) 6 dim L + 1. If in addition, u0 is nondegenerate, we can apply
Theorem 2.5 to conclude that M I(u0 ) = dim L + 1.
Condition (iv) in Theorem 4.2 is clearly satisfied if w∗ is a local minimum
point of J on the solution (stable) manifold M = p(SL⊥ ). So it is clear that
Theorems 4.1 and 4.2 are indeed more general than a minimax principle. As
matter of fact, Condition (iv) in Theorems 4.2 can be weekend as that w∗ is
a conditional critical point of J on any subset containing the path p(v ∗ (s)) for
small s > 0.
If we set L = {0}, the trivial subspace and assume p(v) is the global
maximum point of J on [L, v] = {tv : t > 0} for each v ∈ SL⊥ = SH and
p is C 1 , then Theorem 4.2 reduces to a result of Pohozaev in [18]. As we did
in our algorithm, by gradually expanding the subspace L, Theorem 4.2 can be
used to provide us with information on the Morse index of the critical point.
For an example, when we solve a semilinear elliptic equation as shown in [11],
we start from the trivial solution and set L = {0} to approximate a solution w1
with MI= 1; then we set L = {w1 } to search for a solution w2 with MI= 2,....
This is the advantage of our approach.
As we can see in the above theorems, it becomes very important to check
whether or not the local peak selection p is continuous or differentiable at v ∗ .
This is very difficult at this stage, since an explicit expression of the local peak
selection is not available. We are happy that we have found a solution. We
 Saddle Points and Their Morse Indices 251

can embed a local peak selection into a more general local selection and then
use the implicit function theorem to check whether or not the generalized local
selection is differentiable at v ∗ . Finally we prove that if a local peak selection p
coincides with the more general local selection at v ∗ , then p is also differentiable
at v ∗ . To apply the implicit function theorem, we only have to check whether or
not the determinant of an nxn matrix, where n = dim(L), is equal to zero. This
can be numerically carried out. This study has led to new approach, details
will be presented in a future paper.

References

[1] A. Ambrosetti and P. Rabinowitz, Dual variational methods in critical point

theory and applications, J. Funct. Anal. 14(1973), 349-381.
[2] A. Bahri and P.L. Lions, Morse index of some min-max critical points. I.
Application to multiplicity results, Comm. Pure Appl. Math., Vol. XLI (1988)
1027-1037.
[3] H. Brezis and L. Nirenberg, Remarks on Finding Critical Points, Comm. Pure
Appl. Math., Vol. XLIV, (1991), 939-963.
[4] K.C. Chang, Infinite Dimensional Morse Theory and Multiple Solution Problems,
Birkhäuser, Boston, 1993.
[5] G. Chen, W. Ni and J. Zhou, Algorithms and Visualization for Solutions of
Nonlinear Elliptic Equations Part I: Dirichlet Problems, Int. J. Bifurcation &
Chaos, to appear.
[6] Y. Chen and P. J. McKenna, Traveling waves in a nonlinearly suspended beam:
Theoretical results and numerical observations, J. Diff. Equ., 136(1997), 325-355.
[7] Y. S. Choi and P. J. McKenna, A mountain pass method for the numerical
solution of semilinear elliptic problems, Nonlinear Analysis, Theory, Methods and
Applications, 20(1993), 417-437.
[8] W.Y. Ding and W.M. Ni, On the existence of positive entire solutions of a
semilinear elliptic equation, Arch. Rational Mech. Anal., 91(1986) 238-308.
[9] Z. Ding, D. Costa and G. Chen, A high linking method for sign changing solutions
for semilinear elliptic equations, Nonlinear Analysis, 38(1999) 151-172.
[10] I. Kuzin and S. I. Pohozaev, Entire Solutions of Semilinear Elliptic Equations,
Birkhauser, Boston, 1997.
[11] Y. Li and J. Zhou, A minimax method for finding multiple critical points and its
applications to semilinear PDE, SIAM J. on Scientic Computing, to appear.
[12] Y. Li and J. Zhou, Convergence Results of a minimax method for finding multiple
critical points, submitted.
[13] F. Lin and T. Lin, “Minimax solutions of the Ginzburg-Landau equations”, Slecta
Math. (N.S.), 3(1997) no. 1, 99-113.
[14] J. Mawhin and M. Willem, Critical Point Theory and Hamiltonian Systems,
Springer-Verlag, New York, 1989.
[15] Z. Nehari, On a class of nonlinear second-order differential equations, Trans.
Amer. Math. Soc., 95(19960), 101-123.
[16] W.M. Ni, Some Aspects of Semilinear Elliptic Equations, Dept. of Math. National
Tsing Hua Univ., Hsinchu, Taiwan, Rep. of China, 1987.
252 Li and Zhou

[17] W.M. Ni, Recent progress in semilinear elliptic equations, in RIMS Kokyuroku
679, Kyoto University, Kyoto, Japan, 1989, 1-39.
[18] S. I. Pohozaev, On an approach to nonlinear equations, Dokl. Akad. Nauk SSSR,
247(1979), 1327-1331;
[19] P. Rabinowitz, Minimax Method in Critical Point Theory with Applications to
Differential Equations, CBMS Regional Conf. Series in Math., No. 65, AMS,
Providence, 1986.
[20] M. Schechter, Linking Methods in Critical Point Theory, Birkhauser, Boston,
1999.
[21] J. Smoller, Shock Waves and Reaction-Diffusion Equations, Springer-Verlag, New
York, 1982.
[22] M. Struwe, Variational Methods, Springer, 1996.
[23] J. Wei and L. Zhang, “On the effect of the domain shape on the existence of large
solutions of some superlinear problems”, preprint.
[24] M. Willem, Minimax Theorems, Birkhauser, Boston, 1996.
Static Buckling in a Supported Nonlinear Elastic
Beam

David L. Russell, Virginia Polytechnic Institute and State University,

Blacksburg, VA
Luther W. White, University of Oklahoma at Norman, Norman, OK

Abstract

In the present paper we study finite static buckling effects in a nonlin-

ear elastic beam supported on a flat, rigid, inelastic surface and subject to
a gravitational force. We obtain necessary conditions characterizing equi-
librium states as minima of the corresponding potential energy expression
and we show that we can obtain closed form expressions for the displace-
ment of the beam above the supporting surface on intervals where that
displacement is positive.

AMS-MOS Classifications: 35J25, 35J50, 35Q72, 49L10; Keywords: nonlinear

beam, supported beam, buckling, constrained equilibrium

1 Introduction
In a recent article [6] the authors have studied buckling phenomena in the
context of a nonlinear beam model originally introduced by Lagnese [2]. Our
purpose in the present article is to revisit this model in a constrained situation
wherein the beam is supported on a flat, rigid, inelastic surface – so that the only
permissible transverse displacements are positive – and is subject to a uniform
negative force, which can be interpreted as gravity. A variety of applications
occur in circumstances where a strip of material, e.g., a track or a roadbed, is
laid out over a supporting surface and may be subject to buckling away from the
supporting surface as a consequence of temperature-induced horizontal stresses,
fast moving loads, etc..
We consider, then, an elastic beam of length L, with uniform cross section,
in a two dimensional geometric context. The longitudinal extent of the beam
corresponds to the interval 0 ≤ x ≤ L and the beam is assumed to have
thickness 2h. It will be convenient to suppose that in equilibrium the elastic
axis coincides with the x-axis, even though, strictly speaking, that violates
the constraints described in the preceding paragraph. The displaced elastic
253
254 Russell and White

axis, or ”neutral curve”, admitting both transverse and lateral displacements,

is described by    
x x + ξ(x)
−→ , x ∈ [0, L],
y η(x)
where, minimally, ξ ∈ H 1 [0, L] and η ∈ H 2 [0, L]; in many cases we will need
to assume more smoothness than this. The support constraint corresponds to
the condition η(x) ≥ 0. Assuming that linear material elements orthogonal to
the elastic axis in equilibrium remain so under admitted displacements, such
a deformation results in an infinitesimal stretching or contraction of material
filaments parallel to the elastic axis having vertical coordinate y in equilibrium,
with resultant arclength increment, using ′ to denote differentiation with respect
to x,
q
(1.1) ds = (1 + ξ ′ )2 + (y η ′ )2 dx.
′
With the basic assumption, described in [6], to the effect that ξ and ξ are of
′ ′′
the same order as η 2 , (η )2 and (η )2 , and, further, that these three are of the
same order as the thickness of the beam, discarding terms of higher order one
obtains, again as in [6], the potential energy expression for the beam in the
form Z Z
A L ′ 
′
2 2 Bh2 L ′′
2
V(ξ, η) = 2ξ + η dx + η dx
2 0 6 0

Z L
(1.2) +g η(x) dx − κ ξ(L).
0

Here g > 0 is the constant gravitational force acting in the negative y direction
and κ > 0 represents a horizontal compressional force acting at the right
(horizontally free) end of the beam; the left end of the beam is assumed
horizontally fixed so that ξ(0) = 0. The beam is assumed “transversely
′
clamped” at the endpoints x = 0 and x = L in the sense that η(x) and η (x)
vanish there; we formalize these requirements in §2 to follow. The positive
constants A and B involve the elastic constants of the beam material and its
width (in the third dimension which does not concern us here). All constants
shown here include, implicitly, a factor 2h corresponding to beam thickness
(for example, the gravitational force is, for constant material mass density,
proportional to beam thickness).
In §2, to follow, we establish the existence of a minimizer of the potential
energy functional (1.2) under certain restrictions on the parameters of the
problem, as made precise in §6. In §3 we continue to develop necessary
conditions characterizing such a minimum. The purpose of §4 is to present
two detailed examples to clarify the results presented in §3. In §5 we carry
 Static Buckling in a Supported Nonlinear Elastic Beam 255

out explicit computations resulting in analytic expressions for the solution in

certain restricted circumstances. Finally, §6 serves as an appendix, establishing
certain bounds required in the work of §2.

2 Existence of a Potential Energy Minimizer

In this section we study the static equilibrium problem for the supported elastic
beam introduced in §1. Such an equilibrium is characterized as a minimizer of
the potential energy expression V(ξ, η) given by (1.2) subject to the positivity
constraint

(2.1) η(x) ≥ 0, x ∈ [0, L].

We further suppose that the beam is “clamped” at the endpoints x = 0 and

′
x = L; i.e., we have the boundary conditions (using to indicate differentiation
with respect to x)
′ ′
(2.2) η(0) = η (0) = η(L) = η (L) = 0.

We assume the left endpoint of the beam is horizontally fixed, corresponding

to

(2.3) ξ(0) = 0

but we place no constraint on the horizontal right hand endpoint displacement;

thus ξ(L) is free. These general boundary conditions are distinct from certain
free boundary conditions which we will need to introduce later in order to
characterize intervals in which the minimizer component η(x) > 0.
A rigorous study of the potential energy minimization problem requires,
first of all, a definition of appropriate “state spaces”. We let

(2.4) V = H02 [0, L]

and define the subset corresponding to the imposed constraint as

V+ = { η ∈ V | η(x) ≥ 0 } .

We further define
 
(2.5) U = ξ ∈ H 1 [0, L]  ξ(0) = 0 }

and
   
ξ
(2.6) X = U×V = χ ≡ ξ ∈ U, η ∈ V ;
η

X+ = U × V+ .
256 Russell and White

Expanding the potential energy in the form

Z L  ′ 2  Z L  2
Bh2  ′′ 2 ′ ′
V(ξ, η) = 2A ξ + η dx + 2A ξ η dx
0 6 0

Z L  ′ 4 Z L
A
(2.7) + η dx + g η dx + κ ξ(L),
2 0 0

we designate the first term as the bilinear form a : X × X 7−→ R defined by

Z L 
′ ′ Bh2 ′′ ′′
(2.8) a(χ1 , χ2 ) = 4A ξ1 ξ2 + η η dx.
0 3 1 2
It is clear that a satisfies, for some positive numbers γ0 , γ1 , the coercivity and
boundedness relations

γ0 kχk2X ≤ a(χ, χ), |a(χ1 , χ2 )| ≤ γ1 kχ1 kX kχ2 kX .

We further define the nonlinear functional G : X 7−→ R by

Z L  2
′ ′
(2.9) G(χ) = 2A ξ η dx.
0
 1
R L  ′ 4 4
Defining W = W01,4 [0, L] with norm kηk = 0 η dx ; we observe that

′
|G(χ)| ≤ 2A kξ kL2 kηk2W ≤ 2A kξkU kηk2W .

Using the definitions of the previous paragraph, we may express the

potential energy as
Z L
1 A 4
V(χ) = a(χ, χ) + G(χ) + kηkW + g η dx + κ ξ(L).
2 2 0

For ε > 0 we introduce the perturbed form

(2.10)
Z L
1 A
Vε (χ) = (1 + ε) a(χ, χ) + G(χ) + kηk4W + g η dx + κ ξ(L),
2 2 0

for which we have the lower bound

Z L  ′ 2 Bh2  ′′ 2 
1
Vε (χ) ≥ (1 + ε) 4A ξ + η dx
2 0 3
Z L 2  12 Z L  ′ 4  21 Z L
′ A 4
− 2A ξ dx η dx + kηkW + g η dx + κ ξ(L)
0 0 2 0
 Static Buckling in a Supported Nonlinear Elastic Beam 257
Z L 
 ′ 2 Bh2  ′′ 2   Z L  1 
1 1 ′ 2
≥ (1 + ε) 4A ξ + η dx − A 2 4A ξ kηk2W
2 0 3 0
Z L
A 1
+ kηk4W + g η dx + κ ξ(L) ≥ (1 + ε) a(χ, χ)
2 0 2
1 A 1 1 1
− A 2 a(χ, χ) 2 kηk2W +kηk4W − β0 g L 2 kηkV − κ L 2 kξkU .
2
Rx ′
Here we have used, for ξ(x) = 0 ξ (x) dx,
Z x Z x  1 Z x 1
′ 2 ′ 2
2
| ξ(x)| ≤ | ξ (s)| dx ≤ dx | ξ (s)| dx
0 0 0

1
(2.11) ⇒ | ξ(x)| ≤ x 2 kξkU , x ∈ [0, L],

and, also, the estimates

Z L Z L  21 Z L  12
1
2
η dx ≤ dx η dx ≤ β0 L 2 kηkV ,
0 0 0

kηkL2 [0,l] ≤ β0 kηkV , η ∈ V.

As a consequence we have
ε 1 1 ε
Vε ≥ a(χ, χ) − β0 g L 2 kηkV − κ L 2 kξkU ≥ a(χ, χ)
2 2

1
1 ε γ0 1
1
(2.12) − L 2 β0 g2 + κ2 2
kχkX ≥ kχk2X − L 2 β0 g2 + κ2 2 kχkX .
2
We now consider the regularized problem

Minimize Vε (χ);
Pε : .
χ ∈ X+ .

From (2.12) we conclude that there exists an M0 such that

(2.13) Vε (χ) ≥ M0

and, for χ satisfying Vε (χ) ≤ M1 there exists M2 such that

(2.14) kχkX ≤ M2 .

Let {χn } be a minimizing sequence for Vε in X+ ; thus

Vε (χn ) −→ d ≡ inf Vε (χ).

+
χ∈X
258 Russell and White

From (2.14) it is clear that {χn } is bounded in X. Thus there is a

subsequence, which we still denote by { χn }, converging weakly in X to an
element χ ∈ X+ , since X+ is weakly closed. In particular the corresponding
subsequence { ηn } converges weakly to η ∈ V. Since V embeds compactly
in W we may extract a further subsequence, which we continue to call { ηn },
converging strongly to η in W which, further, implies

kηn kW −→ kηkW .
n ′ o
We may now conclude, since the corresponding ξn converges weakly to ξ in
    ′ 2
′ 2
2
L [0, L] and ηn converges strongly to η in L2 [0, L], that

Z L  ′ 2 Z L  ′ 2
′ ′
G(χn ) = 2A ξn ηn dx −→ 2A ξ η dx = G(χ).
0 0

It follows that
d = lim inf Vε (χn ) ≥ Vε (χ)
and we conclude that there exists an element, which we will now call χε , such
that
Vε (χε ) = min Vε (χ).
+
χ∈X

Now let us suppose that, possibly with some restrictions

 on the parameters

of our problem, we can show that the set of ε-minima χε  ε > 0 ⊂ X+ is
bounded in X. Then, as ε −→ 0, a subsequence { ηεk } converges weakly in
V to some η 0 ∈ V+ and, passing to a further subsequence if necessary, { ξεk }
converges weakly in U to an element ξ0 , which implies

lim inf a(χε , χε ) ≥ a(χ0 , χ0 ).

Let χ ∈ X+ . Then
1 A
Vε (χ) = (1 + ε)a(χ, χ) + G(χ) + kηk4W + (g, η)L2 [0,L] + κ ξ(L)
2 2
1 A
≥ Vε (χε ) ≥ (1 + ε) a(χε , χε ) + G(χε ) + kηε k4W + (g, ηε )L2 [0,L] + κ ξε (L).
2 2
In the limit as ε −→ 0 we have
1 A
V(χ) ≥ a(χ0 , χ0 ) + G(χ0 ) + kη 0 k4W + (g, η 0 )L2 [0,L] + κ ξ0 (L),
2 2
since it is clear that limε → 0 Vε (χ) = V(χ). Then we have the desired result:

V(χ) ≥ V(χ0 ), χ ∈ X+ .
 Static Buckling in a Supported Nonlinear Elastic Beam 259

Thus the existence of a (not necessarily unique) minimizer for V is established.

The proof of the boundedness of the χε is somewhat complicated; we leave it
for §6, which serves as an appendix, at the end of this paper. That proof requires
some restrictions on the parameters of the problem which may or may not be
realistic in some cases. The proof given above is, of course, valid whenever we
have a uniform bound on the χε , whether or not that boundedness is obtained
as in §6.

3 General Conditions Characterizing a Minimizing Element

We begin this section with a choice of some minimizing element for V, as shown
to exist in the previous section and there identified as χ0 . Here we will refer
to such an element as χ and we will primarily be dealing with its individual
components ξ and η. We have the potential energy form (1.2) and the general
boundary conditions (2.2) and (2.3). We impose the positivity constraint (2.1)
and we continue to use the spaces U, V and X as in (2.5), (2.4) and (2.6),
respectively.
Let Y = H0p [0, L] for 12 < p ≤ 2; we define Z = U × Y and the mapping
F : Z 7−→ Y by
F (ξ, η) = − η.
The restriction on p guarantees H0p [0, L] ֒→ C 0 [0, L] and thus that Y has
a positive cone Y+ , corresponding to η(x) ≥ 0, x ∈ [0, L] with non-empty
interior corresponding to η(x) > 0, x ∈ [0, L].
Applying an infinite dimensional version of the Kuhn-Tucker theorem [4],[5]
there is an element λ ∈ Y∗ = H −p [0, L] such that the Lagrangian, defined by
L(χ, λ) = V(χ) + hλ, F (χ)i ,
has a stationary point at the solution of the minimization problem
min V(χ).
F (χ) ≤ 0

Thus, using the symbol DL to denote the Gateaux derivative with respect to
the χ variable in X, we have
DL(χ, λ) χ̂ = 0, ∀χ̂ ∈ X.
Hence, for all ξ̂ ∈ U and η̂ ∈ V we have
(3.1)
Z L   ′ 2   ′  Bh2 ′′ ′′ 
′ ′ ′
0 = A 2ξ + η 2ξ + 2η η̂ + ˆ
η η̂ + g η̂ dx + κ ξ(L) − hλ, η̂i .
0 3
Let η̂ = 0. Then for all ξˆ ∈ U we have
Z L   ′ 2  ′ 
′
(3.2) 2A 2ξ + η ˆ
ξ dx + κ ξ(L) = 0.
0
260 Russell and White

From (3.2) we conclude

  ′ 2 ′  ′ 2
′ ′ κ
(3.3) 2ξ + η = 0 ⇒ 2ξ + η = − .
2A

Setting ξˆ = 0 in (3.1) we have

Z L   ′ 2  ′ 
′ ′ Bh2 ′′ ′′
(3.4) 0 = 2A η 2 ξ + η η̂ + η η̂ + g η̂ dx − hλ, η̂i .
0 3

and then substituting (3.3) into (3.4) we obtain

Z L 
′′ ′′ ′ ′
(3.5) α η η̂ − κ η η̂ + g η̂ dx − hλ, η̂i = 0, ∀ η̂ ∈ V,
0
2
where α = Bh 3 .
1
From the cited Kuhn - Tucker conditions λ ∈ H −p [0, L], 2 < p ≤ 2,
satisfies the complementary slackness condition

(3.6) hλ, ηi = 0, λ ≥ 0.
1
Letting p = 1
2 + ε, 0 < ε ≤ 2, (3.5) yields the equality, in H −( 2 +ε) ,
′′′′ ′′
(3.7) αη + κ η = − g + λ.

Using the essential boundary conditions (2.2) and applying the Sobolev
7
embedding theorems we conclude from (3.7) that η ∈ H02 ∩ H 2 −ε and, further,
that
′′′ 1 3
(3.8) η ∈ C 2 [0, L], η ∈ L ε [0, L], 0 < ε ≤ .
2
We now take Y = H01 [0, L] so that λ ∈ H −1 [0, L] and conclude there is an
element λ̂ ∈ H01 [0, L] such that, for all φ ∈ H01 [0, L],
Z L
′ ′
(3.9) λ̂ φ dx = hλ, φi .
0

From (3.5) we see that for all φ ∈ H02 [0, L] we have

Z Ln o Z L
′′ ′′ ′ ′ ′ ′
(3.10) α η φ − κ η φ + g φ dx − λ̂ φ dx = 0.
0 0

We define O to be the open set

n  o

(3.11) O = x ∈ (0, L)  η(x) > 0 .
 Static Buckling in a Supported Nonlinear Elastic Beam 261

Then λ must vanish on O in the sense that if supp φ ⊂ O and φ(x) ≥ 0, x ∈

[0, L], then
Z L
′ ′
(3.12) hλ, φi = 0 = λ̂ φ dx.
0

From (3.10), for all φ ∈ H02 [0, L]

such that supp φ ⊂ O,
Z Ln o
′′ ′′ ′ ′
(3.13) α η φ − κ η φ + g φ dx = 0.
0

Let F = [0, L] \ O. Then η(x) = 0, x ∈ F. Suppose φ ∈ H02 [0, L], supp φ ⊂

Int F. Then, from (3.10),
Z L 
′ ′
(3.14) g φ − λ̂ φ dx = 0.
0

The conclusion η ∈ C 2 [0, L] in (3.8) implies, in particular, that

′′
(3.15) η (xb ) = 0
at any boundary point xb of O which is also a cluster point for the set F.
This constitutes an important free boundary condition characterizing boundary
points of isolated open intervals included in O.

4 Illustrative Examples
Since the implications of the results obtained in the previous section may not
be completely obvious, we present two examples in this section showing how
those conditions apply and, in particular, the role and limitations of the free
′′
boundary condition η (xb ) = 0 applying at certain boundary points xb of the
set O as described in (3.11).
Example 1 Let us consider a situation wherein, for 0 < x1 < x2 < L, we
have
η(x) > 0, x ∈ (0, x1 ) ∪ (x2 , L); η(x) = 0, x ∈ [x1 , x2 ].
We assume V(χ) = V(ξ, η) is minimized, subject to the constraints described
earlier by the pair (ξ, η), that λ, is the corresponding Lagrange multiplier and
that λ̂ is related to λ by (3.9). From the regularity results of the preceding
section we conclude that
′ ′′
η(x1 ) = η (x1 ) = η (x1 ) = 0;
′ ′′
η(x2 ) = η (x2 ) = η (x2 ) = 0.
From the minimality of V(χ) = V(ξ, η) and the inactivity of the constraint
η(x) ≥ 0 on (0, x1 ) ∪ (x2 , L) we see that if φ ∈ H02 [0, L], supp φ ⊂
(0, x1 ) ∪ (x2 , L), then, in particular,
Z x1  
′′ ′′ ′ ′
0 = α η φ − κ η φ + g φ dx
0
262 Russell and White
 Z x1 h ′′′  ′ i
′′ ′ x1 ′
= αη φ  − α η + κ η φ − g φ dx
0 0
 ′′′   x1 Z x1  
′  ′′′′ ′′
= − α η + κ η φ + αη + κ η + g φ dx
0 0
Z x1  ′′′′ 
′′
= α η + κ η + g φ dx = 0;
0
in each case the properties of φ show that the boundary terms vanish. We
conclude from the last identity that
′′′′ ′′
(4.1) αη + κ η + g = 0, in (0, x1 ) .

Comparing with (3.7) we conclude λ = 0 on (0, x1 ). Similarly we conclude

that λ = 0 on (x2 , L) and
′′′′ ′′
(4.2) αη + κ η + g = 0, in (x2 , L) .

From the above it follows, using (3.9), that for φ ∈ H02 [0, L], supp φ ⊂
(0, x1 ) ∪ (x2 , L) we have
Z L
′ ′
(4.3) λ̂ φ dx = hλ, φi = 0.
0

Since λ̂ ∈ H01 [0, L] we have the endpoint conditions

(4.4) λ̂(0) = 0, λ̂(L) = 0.

Further, since we can require supp φ ⊂ (0, x1 ), we can use the argument of
′
the classical du Bois - Reymond lemma [3] to see that λ̂ is constant on (0, x1 );
equivalently
′′
(4.5) λ̂ = 0, x ∈ (0, x1 ).

Now let supp φ ⊂ (0, x2 ). From (3.10) we see that

Z x2   Z x2
′′ ′′ ′ ′ ′ ′
0 = α η φ − κ η φ + g φ dx − λ̂ φ dx
0 0
Z x1  ′′ ′′  Z x2 Z x2
′ ′ ′ ′
= α η φ − κ η φ + g φ dx + g φ dx − λ̂ φ dx
0 x1 x1
 Z x1  ′  Z x2 Z x2
′′ ′ x1 ′′ ′ ′ ′
= αη φ  − αη + κη φ − g φ dx + g φ dx − λ̂ φ dx
0 0 x1 x1
Z x1  ′ ′  Z x2  
′′ ′ ′
= − α η + κ η φ − g φ dx + g φ dx − λ̂ φ dx = 0,
0 x1
 Static Buckling in a Supported Nonlinear Elastic Beam 263
′ ′′
the last equality being valid because φ (0) = 0 and η (x1 ) = 0. Integrating
by parts again we have
 ′′ ′ x1 Z x1  ′′ 
 ′′
− α η + κ η φ + αη + κη + g φ dx
0 0
 x2 Z x2     ′ 
′  ′′ ′′ ′
− λ̂ φ  + λ̂ + g φ dx = − αη + κη (x1 −) + λ̂ (x1 +) φ(x1 )
x1 x1
Z x1  ′′  Z x2  ′′ 
′′
+ αη + κη + g dx + λ̂ + g φ dx = 0.
0 x1

This being true for all φ ∈ H02 [0, L]

with supp φ ⊂ (0, x2 ) we once again have
the earlier result (4.1) for the interval (0, x1 ) together with the “point support
condition” at x1 ,
′ ′′′
(4.6) λ̂ (x1 +) − α η (x1 −) = 0,

and the adjoint equation in the interval (x1 , x2 ),

′′
(4.7) λ̂ (x) + g = 0, x ∈ (x1 , x2 ).

Next passing to φ ∈ H02 [0, L] with supp φ ⊂ (0, L) and carrying out
computations similar to those performed above we again obtain (4.2), the
“support condition” at x2 ,
′ ′′′
(4.8) λ̂ (x2 −) − α η (x2 +) = 0,

and the adjoint equation

′
(4.9) λ̂ (x) = 0, x ∈ (x2 , L).

Combining the endpoint conditions (4.4), the support conditions (4.6) and
(4.8) with the adjoint equations (4.5), (4.9) and (4.7) we have a complete set
of equations determining λ̂ in terms of η and ξ. The equation (4.3) identifies
′′
λ, in the distributional sense, as − λ̂ . Thus λ ≡ g on the interval (x1 , x2 )
′′′ ′′′
and includes Dirac delta components of magnitude − α η (x1 −) and α η (x2 )
at the points x1 and x2 , respectively. On (x1 , x2 ) the multiplier λ corresponds
to the constraint force required to support the beam on that interval. Support
for the beam on the intervals (0, x1 ) and (x2 , L) corresponds to the vertically
oriented distributional forces just identified at x1 and x2 together with the
constraint forces exerted at x = 0 and x = L, which are not included in the
analysis because the conditions η(0) = η(L) = 0 are given a priori. Further
analysis shows that
 ′′′

 α η (x1 ) x, x ∈ (0, x1 ),
g x 2
λ̂(x) = − 2 + c1 x + c2 , x ∈ (x1 , x2 ),

 α η ′′′ (x )(x − L),
2 x ∈ (x2 , L),
264 Russell and White

with c1 and c2 determined by the requirement that λ̂ should be continuous at

x1 and x2 .
Example 2 The situation studied here differs from that in Example 1 in
that the middle interval [x1 , x2 ] is collapsed to a single point which we will call
x1 . Using (3.10) and integrating by parts we have, for φ ∈ H02 [0, L],
Z L  Z L
′′ ′′ ′ ′ ′ ′
0 = α η φ − κ η φ + g φ dx − λ̂ φ dx
0 0
  Z L  ′ ′  Z L
′′ ′  x1 ′′ ′ L ′′ ′ ′
αη φ  + αη φ  − α η + κ η φ − g φ dx − λ̂ φ dx = 0.
0 x1 0 0
Collecting boundary terms and integrating by parts again we have
h ′′ i ′ ′ x
 ′′ ′ x1
′′ 
α η (x1 −) − η (x1 +) φ (x1 ) − λ̂ φ0 1 − α η + κ η φ
0
 ′′ ′ L Z L  ′′  Z L
 ′′ ′′
− α η + κ η φ + αη + κη + g φ dx + λ̂ φ dx = 0.
x1 0 0
′
Restricting first to φ such that φ(x1 ) = φ (x1 ) = 0 and arguing as we did to
obtain (4.1) and (4.2) in Example 1, we obtain those equations again and also
′′ ′
conclude that λ̂ vanishes identically, i.e., λ̂ is a constant, on (0, x1 ) and x1 , L).
′
Then considering φ with φ (x1 ) = 0 we obtain
′ ′
 ′′′ ′′′

(4.10) λ̂ (x1 +) − λ̂ (x1 −) = α η (x1 +) − η (x1 −) .
′
Combining (4.10), the result that λ̂ is constant on (0, x1 ) and on (x1 , L)
and the conditions λ̂(0) = λ̂(L) = 0 following from the original definition of λ̂
as an element of H01 [0, L] we have all that is needed to determine λ̂ in terms of
ξ and η (actually, only η is involved). The discussion of the role of the original
Lagrange multiplier λ is much the same as in Example 1 except that in this
case λ is zero except for its single distributional component,
 ′′′ a multiple ofthe
′′′
Dirac distribution with support at x1 of strength α η (x1 −) − η (x1 +) . It
′′
should be noted that in this example there is no requirement on η (x1 ) other
′′ ′′
than the continuity requirement η (x1 −) = η (x1 +).

5 Analytic Discussion of Static Equilibrium States

This section serves as as a somewhat extended third example, supplementary to
Examples 1 and 2 of the preceding section. We want to consider the case where
O, the support set for η(x), consists of a single interval (x1 , x2 ) ⊂ Int [0, L].
Throughout the interval [0, L] we have the partial differential equation
  ′ 2 ′
′
(5.1) 2ξ + η = 0.
 Static Buckling in a Supported Nonlinear Elastic Beam 265

We have η(x) ≡ 0, x ∈ [0, x1 ]∪[x2 , L] and, in the interval (x1 , x2 ), η(x) satisfies,
2
recalling the abbreviation α = B3h introduced in §3, the partial differential
equation
   ′ 2 ′
′′′′ ′ ′
(5.2) αη − 2A η 2ξ + η + g = 0.

From the first of these equations we have

′
 ′ 2
2ξ + η = C, x ∈ [0, L],

′
for some constant C. Since η vanishes on [0, x1 ] ∪ [x2 , L], (3.3) gives
′
4A ξ + κ = 0, x = 0, L,

on that set and we must, in fact, have

 ′ ′
2
 κ
(5.3) A 2ξ + η ≡ − , x ∈ [0, L].
2
Then setting
κ g
(5.4) K = , G = ,
α α
the equation (5.2) can be rewritten in the form
′′′′ ′′
(5.5) η + K η + G = 0.

Since this differential equation is autonomous solutions are invariant under

translation of the support interval (x1 , x2 ). It will be convenient to assume
that this interval is centered at x = L2 . Looking for solutions symmetric with
respect to x = L2 , we note that (5.5) is equivalent to

2 !′′
′′ G x − L/2
η + Kη + = 0
2

from which, taking the symmetry into account again, there is a constant d such
that

2
′′ G x − L/2
(5.6) η + Kη + + d = 0.
2

2
b x−L/2
Trying for a particular solution of the form η(x) = 2 + c we find that
G G
b = −d − Kc and Kb = −G. Thus b = − K , d = −b − Kc = K − Kc.
266 Russell and White

Including the solutions of the homogeneous equation symmetric with respect to

x = L2 we have

2

G x − L/2
η(x) = a cos ω(x − L/2) − + c,
2K
1
where ω = K 2 . Let ρ = L/2 − x1 = x2 − L/2. Then the “essential”
′
boundary conditions η(x2 ) = η (x2 ) = 0 yield

Gρ2 Gρ
(5.7) a cos ωρ − 2
+ c = 0, − aω sin ωρ − 2 = 0,
2ω ω
′′
while the “free boundary” condition η (x2 ) = 0 gives

G
(5.8) − aω 2 cos ωρ − = 0.
ω2
The second equation of (5.7) together with (5.8) give

(5.9) tan ωρ = ωρ.

We initially take ρ to be the smallest positive solution of this equation; the one
such that σ ≡ ωρ lies in (π, 3π/2). Then the third equation yields

G
a = − ,
ω 4 cos ωρ

which is positive since cos ωρ < 0. Finally, from the first equation of (5.7) we
have  
Gρ2 G Gρ2 G σ2
c = − a cos ωρ + = 4 + = 2 1+ .
2ω 2 ω 2ω 2 K 2


Recalling (5.4) the solution η(x) is thus given, in the interval L2 − ρ, L2 + ρ
wherein it is positive, as
 
gα cos ω(x − L/2) σ2 g
(5.10) η(x) = 2 1 − + − (x − L/2)2 .
κ cos ωρ 2 2κ

The amplitude at the mid-point x = L/2 is

   
gα 1 σ2 gα 1 σ2
η(L/2) = 2 1 − + = 2 1+ +
κ cos ωρ 2 κ cos(σ − π) 2

and the second derivative there is

′′ g α ω2 g
η (L/2) = 2
− .
κ cos σ κ
 Static Buckling in a Supported Nonlinear Elastic Beam 267

0.25

0.2

0.15
beam displacement eta(x)

0.1

0.05

−0.05

−0.1

0 0.5 1 1.5 2 2.5 3 3.5 4

longitudinal variable x

Fig. 1. Basic Equilibrium Form for η(x); j = 1.

In Figure 1 we show a typical plot of η(x) obtained using the formula (5.10).
Computing the third derivative of η(x) in the interval where η(x) > 0 we
obtain the expression
 
′′′ g α ω3 L L L
η (x) = − 2 sin ω x − , −ρ < x < + ρ.
κ cos ωρ 2 2 2

Since, as
we have noted earlier, cos ωρ < 0 this is a positive multiple of
′′′
sin x − L2 . Since η(x) and hence η (x) vanish to the left of L2 − ρ and to
the right  of L2 + ρ, the lateral forces experienced by the beam at these points
′′′ L

 
L


are B η 2 − ρ + and − B 2 + ρ − , respectively. Using (5.9) we see
that these both have the positive value

g B α ω3 g α ω3 g ρ B α K2
tan ωρ = ωρ =
κ2 κ2 κ2
and represent point forces exerted on the beam by the supporting surface at the
points indicated. If the supporting surface were endowed with elastic qualities
these would be replaced by distributed forces, of course.
All of the above assumes that L ≥ 2ρ. If this is not the case then the free
boundary condition (5.8) cannot be achieved; we have η(x) > 0 throughout
′′
the open interval (0, L) with η (x) > 0 at both of the points x = 0 and
x = L. There is, of course, the critical case where L = 2ρ for which η(x) > 0
268 Russell and White

1.4

1.2

0.8
j=4
beam displacement eta(x)

0.6

j=3
0.4

0.2
j=2

−0.2

−0.4

−0.6

0 0.5 1 1.5 2 2.5 3 3.5 4

longitudinal variable x

Fig. 2. Equilibrium Forms for η(x), j = 2, 3, 4.

′′ ′′
throughout (0, L) and η (0) = η (L) = 0. It is clear that the equation (5.9)
has infinitely many solutions ρj tending asymptotically to (2j+1)π
2 as j → ∞,
the one just discussed corresponding to j = 1. In Figure 2 we show the
equilibrium forms obtained from (5.10) for the cases j = 2, 3, 4. Our conjecture
is that the case j = 1 corresponds to a stable equilibrium associated with a
minimum of the potential energy form whereas the cases j > 1 are unstable,
corresponding to stationary points of the potential energy functional which are
not minima of that functional.

Appendix: Proof of Boundedness of the χε

We retain the definitions of spaces, etc., introduced in §2. For ξ ∈ U, η ∈ V
and for a variation δ to ξ, also in U, we have (cf. (2.10))
Z L  ′ 2 
′ ′ ′
Dξ Vε (ξ, η) δ = 4A(1 + ε) ξ δ dx + 2A δ η dx + κ δ(L).
0

Applied at the point χε = (ξε , ηε ) and with δ replaced by ξε , this gives

Z L  Z L  
′ 2 ′ ′ 2
(A.1) 4(1 + ε) ξε dx = − 2A ξε ηε dx − κ ξε (L)
0 0

which yields the estimate, uniform for ε > 0,

1
4A kξε k2U ≤ 2A kξε kU kηε k2W + κ L 2 kξε kU
 Static Buckling in a Supported Nonlinear Elastic Beam 269
1
1 κ L2
(A.2) ⇒ kξε k ≤ kηε k2W + .
2 4A
Applying integration by parts to (A.1) we obtain
  ′ 2  
′
4A(1 + ε)ξε + 2A ηε (L) + κ δ(L)
Z L   ′ 2 ′
′
− 4A(1 + ε)ξε + 2A ηε δ dx = 0
0

Restricting attention to perturbations δ(x) such that δ(0) = δ(L) = 0 we

obtain the equation
  ′ 2 ′
′
(A.3) −2A 2(1 + ε)ξε + ηε = 0.

Our assumptions on ξ give ξε (0) = 0; using (A.3) with variable δ(L) we

conclude that
  ′ 2 
′
(A.4) 4A(1 + ε)ξε + 2A ηε (L) + κ = 0.

Since (A.3) implies

  ′ 2    ′ 2 
′ ′
2(1 + ε)ξε + ηε (L) = 2(1 + ε)ξε + ηε (x), x ∈ [0, L),

we obtain
  ′ 2 
′ κ
(A.5) 2(1 + ε)ξε + ηε (x) ≡ − .
2A

Integration and the condition ξε (0) = 0 then give

 Z x  
κx 1 ′ 2
ξε (x) = − + ηε ds
4A(1 + ε) 2(1 + ε) 0
so that, uniformly for ε > 0,
Z x 2
κx 1 ′
(A.6) |ξε (x)| ≤ + ηε ds.
4A 2 0

Now considering perturbations υ ∈ V to η ∈ V+ such that η + t υ ∈ V+

2
for small t > 0 and recalling the abbreviation α = B3h introduced in §3, we
have the inequality Dη Vε (ξ, η)υ =
Z L   ′ 2  ′ ′  Z L
′ ′′ ′′ ′′ ′′
2A 2 ξ + η η υ + α η υ + g υ dx + ε α η υ dx ≥ 0.
0 0
270 Russell and White

Applying this to ξε and ηε we obtain

Z L Z L  ′ 2  ′ ′ Z L
′′ ′′ ′
(1 + ε)α ηε υ dx + 2A 2ξε + ηε ηε υ dx + g υ dx ≥ 0.
0 0 0
′
Using (A.5) to solve for ξε we have
′ 1  ′ 2 κ
ξε (x) ≡ − ηε (x) −
2(1 + ε) 4A(1 + ε)
so that
Z Z   Z L
L
′′ ′′
L
2Aε  ′ 2 κ ′ ′
(1 + ε)α ηε υ dx + η − η υ dx + g υdx ≥ 0.
0 0 (1 + ε) ε (1 + ε) ε 0

Using (2.11) we now have

1
1 κ L2
k ξε kU ≤ kηε k2W +
2 4A

  21 1
1 L3 κ L2
2
(A.7) ⇒ k ξε kU ≤ kηε kV + .
2 3 4A

We clearly have (cf. (2.10))

1+ε A
0 = Vε (0, 0) ≥ a (χε , χε ) + G (χε ) + kηε k4W + (g, ηε ) + κ ξε (L)
2 2
which implies, using (2.8) and (2.9), that
Z L   ′ 2  ′′ 2 
1+ε
− (g, ηε ) − κ ξε (L) ≥ 4A ξε + α ηε dx
2 0
Z L  ′ 2 Z L 4
′ A ′
+ 2A ξε ηε dx + ηε dx ≥ 2A(1 + ε)k ξε k2U
0 2 0
α(1 + ε) A
+ kηε k2V + kηε k4W − 2A k ξε kU kηε k2W .
2 2
α(1 + ε) A A
≥ 2A(1 + ε)k ξε k2U + kηε k2V + kηε k4W − 2A k ξε k2U − kηε k4W
2 2 2
α(1 + ε) α(1 + ε)
= 2A ε k ξε k2U + kηε k2V ≥ kηε k2V .
2 2
From this we conclude, using (2.11) again, that

α(1 + ε) g L2 1
kηε k2V ≤ kηε kV + κ L 2 k ξε kU .
2 2
 Static Buckling in a Supported Nonlinear Elastic Beam 271

Using (A.7) and dropping ε on the left hand side we see that
"  1 1
#
α g L 2 1 1 L 3 2 κ L 2
kηε k2V ≤ kηε kV + κ L 2 kηε k2V + .
2 2 2 3 4A

As a result we see that

 
κ L2 α g L2 κ2 L
(A.8) √ − kηε k2V + kηε kV + ≥ 0.
2 3 2 2 4A

If
κ L2 α
(A.9) √ − < 0,
2 3 2

so that the parabola described by the equation obtained from (A.8) by changing
≥ to = opens downward, then the inequality is valid only to the left of the
largest root of that quadratic equation, leading to the conclusion
r   
2
g L2 g L2 κ2 L L2
κ√ α
2 + 2 − A 2 3
− 2
kηε kV ≤   ≡ K0 .
L2
κ√ α
− 2 3
− 2
√
3α
Thus if (A.9) is true, i.e., if κ < L2
, we have the estimate, independent of
ε > 0,

(A.10) 0 ≤ kηε kV ≤ K0 .

Using this in (A.7) we obtain

  21 1
1 L3κ L2
(A.11) k ξε kU ≤ + . K02
2 3 4A
n  o

Combining (A.10) with (A.11) we see that the set χε  ε > 0 is bounded in
X, as required to complete the existence argument in §2.

References

[1] Adams, R. A.: Sobolev Spaces, Academic Press, New York, 1975
[2] Lagnese, J. E.: Recent progress in exact boundary controllability and uniform
stabilizability of thin beams and plates, in Distributed Parameter Control Systems,
G. Chen, E. B. Lee, W. Littman and L. Markus, Eds., Marcel Dekker, New York,
1991, pp. 61-111
272 Russell and White

[3] Ewing, G. M.: Calculus of Variations with Applications, W. W. Norton & Co.,
Inc., New York, 1969
[4] Luenberger, David G.: Optimization by Vector Space Methods, John Wiley &
Sons, Inc., New York, 1969
[5] Russell, D. L.: The Kuhn - Tucker conditions in Banach space with an application
to control theory. J. Math. Anal. Appl. (1966).
[6] Russell, D. L., and L. W. White: An elementary nonlinear beam theory with finite
buckling deformation properties, to appear in SIAM J. Appl. Math.
Optimal control of a nonlinearly viscoelastic rod

Thomas I. Seidman, University of Maryland Baltimore County, Baltimore,

MD. E-mail: [email protected]
Stuart S. Antman1 , University of Maryland College Park, College Park, MD.
E-mail: [email protected]
Abstract
We consider some typical optimal control problems for a nonlinear
model of longitudinal vibrations in a viscoelastic rod. In trying to
follow the usual pattern of showing that every infimizing sequence of
controls contains a subsequence suitably converging to an optimal control,
we confront the severe technical difficulty that the constitutive function
cannot be uniformly Lipschitzian in its arguments — e.g., it blows up at
a ‘total compression.’ One needs to make careful use of the structure of
the system to overcome this difficulty.

1 Introduction
We consider a PDE model for the longitudinal motion of a uniform2 viscoelastic
rod:
(1.1) wtt = νs + f, ν = n (ws , wst )
holding on Q := (0, ℓ)×(0, T ). Here w = w(s, t) is the position at time t ∈ [0, T ]
of the material point with reference position s ∈ (0, ℓ) so ws gives the strain
and wst is the strain rate; ν = ν(s, t) is then the contact force (given by the
constitutive function n) and f = f (s, t) is the external body force. One natural
set of boundary conditions for this problem consists of the specification of the
contact forces at the endpoints. It is plausible to consider either the body
force f as a distributed control or the contact force ν̂ at s = ℓ as a boundary
control. We take homogeneous boundary conditions at one end, for simplicity,
so that
 
 
(1.2) ν ≡ 0, ν = ν̂(t).
s=0 s=ℓ

1
Supported in part by a MURI Grant from the ARO.
2
This uniformity is purely for expository simplicity. There would be no difficulty in
permitting the density ρ, here normalized to 1, and the constitutive function n(··) to depend
explicitly (piecewise continuously) on s as well.
273
274 T.I. Seidman and S.S. Antman

It is convenient to introduce u := ws , v := wt and then to rewrite (1.1) in the

form

i) ut = vs ,
(1.3)
ii) hω, vt i + hωs , n(u, ut )i = ω(ℓ)ν̂ + hω, f i,

with (1.3-ii) holding for t ∈ (0, T ) and for all suitable test functions ω ∈
H 1 (0, ℓ). Here and below we use h·, ·i to denote the L2 (0, ℓ) inner product
and comparable duality products.
This model was considered in [1] and [3] (and is generalized in the
forthcoming paper [2] to a full vector model which considers transverse motion,
shear, and torsion as well as longitudinal motion). Suitable assumptions there
on the constitutive function n(··) (permitting fully nonlinear dependence on the
strain rate) ensure, for suitable data, both well-posedness and the preclusion
of ‘total compression’, i.e., u = ws is pointwise bounded away from 0. In
Section 3 we adapt those hypotheses, with particular attention to weakening
the conditions imposed on the data f, ν̂ which we take as possible controls.
Our primary objective in this paper is to prove the existence of optimal con-
trols for three closely related and reasonably typical optimal control problems:

Problem 1: Taking the boundary condition ν̂ (contact force at: s = ℓ) as

control, track a target trajectory w̄(··) over [0, T ],
Problem 2: Taking the external force f as control, approximate a target
state [w̄, w̄t ] at the fixed time T > 0,
Problem 3: Again with boundary control ν̂, attain a target state in
minimum time, subject to constraints.
These problems are closely related here in that essentially the same a priori
estimates suffice to provide the compactness needed for the arguments. Note
that in order to take limits through the nonlinearity of the constitutive
relation, the compactness obtained must necessarily involve pointwise bounds
and convergence for u, ut , since we impose no growth condition restricting the
dependence of n(y, z) on z and since n(y, z) must blow up as y → 0 to penalize
‘total compression’. This lack of regularity in the constitutive function is a
principal technical difficulty for our analysis.

2 Compactness and optimal control

Perhaps the principal technical point of this paper is that the same estimates
and compactness needed to obtain existence (and well-posedness) for the direct
problem (with f, ν̂ specified) also suffice to give existence of optimal controls
for the particular problems we consider.
 Optimal control of a viscoelastic rod 275

We do not attempt to seek especially weak conditions on the initial data

but we impose, as sufficient for our arguments, the requirement that the initial
data satisfy:
◦ ◦ ◦
(2.1) w, wt ∈ H 2 (0, ℓ) with w (s) > c̄ > 0.

On the other hand, we weaken slightly the assumption (used in [1], e.g.,)
that ft , ν̂t have L2 bounds. Let us define


(2.2) U := W 1,p [0, T ] → L2 (0, ℓ) , V := W 1,4/3 (0, T ).

where p is an arbitrary

fixed number with p > 1. [Note that then U ⊂
2
C [0, T ] → L (0, ℓ) with a fixed modulus of continuity (depending on the
choice of p) and that V ⊂ C[0, T ] with compact embedding.] Our key hypothesis
on the cost functionals considered for optimization is an appropriate coercivity
condition:

f is bounded in U and ν̂ is bounded in V (at least when the cost

functional is bounded)

While deferring the detailed proofs, which constitute the next two sections,
we now assert two lemmas which will be fundamental to our arguments:
Lemma 2.1. Under the hypotheses of Section 3, the equation ( 1.1 ) , equiv-
alently (1.3), has a unique solution corresponding to any choice of data [f, ν̂]
in U × V and any choice of initial data satisfying ( 2.1 ) .
Lemma 2.2. Under the hypotheses of Section 3, the solutions of ( 1.1 ) ,
( 1.2 ) corresponding to [f, ν̂] bounded in U × V for fixed or suitably bounded
initial data as in ( 2.1 ) all lie in a fixed compact set: In particular, there is a
compact K such that

(2.3) ut = vs ∈ K ⊂ C(Q)

and, of course, w, u = ws , and v = wt also lie similarly in compact subsets of

C(Q). Further, u is uniformly bounded away from 0 and vt = wtt is bounded
in L∞ ((0, T ) → L2 (0, ℓ)).
We now show how these lemmas may be used to prove the existence of
optimal controls for the problems under consideration, beginning with the
following corollary.
Corollary 2.1. Suppose that {f k , ν̂ k } are bounded in the sense of the
coercivity condition and converge weakly in L2 (Q), ×L2 (0, T ): f k ⇀ f¯, ν̂ k ⇀ ν̄.
Then the corresponding solutions converge on Q (i.e., wk , uk = [wk ]s , v k =
wtk , ukt = vsk = wst
k converge uniformly and v k = wk converges weakly in L2 (Q))
t tt
to the solution corresponding to the limit data f¯, ν̄.
276 T.I. Seidman and S.S. Antman

Proof. By Lemma 2.1 there are indeed corresponding solutions wk and by

Lemma 2.2 there is a subsequence for which these corresponding solutions
converge uniformly with vtk = wtt k converging weakly in L2 (Q). Since uk , uk
t
k
are bounded with u bounded away from 0 by Lemma 2.2, the corresponding
functions ν k = n(uk , ukt ) then also converge uniformly to the appropriate limit;
we note, in particular, that the limit boundary condition is then necessarily
satisfied. With smooth test functions we easily see that the limit of the solutions
satisfies (1.3-ii) with the limit data, as desired, and such ω are dense. Since
this solution is unique by Lemma 2.1, we have convergence for the full sequence.

Problem 1: Suppose that f ∈ U and initial data as in (2.1) are already

specified. We may then consider the determination of the boundary contact
force ν̂(·) in (1.1), (1.2) as an optimal control problem once we have specified
a cost functional J for optimization. We take J = J1 (ν̂) + J2 (w) (subject to
(1.1), (1.2)) with, for example,
Z T
(2.4) J1 (ν̂) := max{|ν̂(t)} + a |ν̂t (t)|4/3 dt
[0,T ] 0

and, tracking a target trajectory w̄,

J2 (w) := b supQ {|wtt |} + χ∗ (|w − w̄|/c)

(2.5)

where 
0 if ω 6 1 on Q,
χ∗ (ω) :=
+∞ otherwise.
Thus, we wish to control the rod ‘gently’ (so that ν̂ stays small and does not
change too abruptly) and to keep the rod from accelerating too violently (so
that wtt stays small) while demanding that the rod match the specified target
trajectory w̄ = w̄(s, t) to within a tolerance c. We do not expect that J is finite
for arbitrary boundary data ν̂ ∈ V, but assume a priori — presumably as a
condition on the target w̄ under consideration — that the set Vad of admissible
controls is nonempty: there is at least one ν̂ ∈ V for which (1.1), (1.2) gives a
solution w with wtt bounded and |w − w̄| 6 c everywhere on Q.
Theorem 2.1. With f, w̄ and the initial data as above and under the
hypotheses on the constitutive function n(··) of Section 3, there is an optimal
control ν̄ for Problem 1, i.e., the cost functional J given by ( 2.4 ) , ( 2.5 )
attains its minimum.
Proof. Given Lemmas 2.1, 2.2 and Corollary 2.1, the argument has a fairly
standard pattern. Let (ν̂ k ) be an infimizing sequence for J with corresponding
 Optimal control of a viscoelastic rod 277

solutions (wk ) so
J k := J1 (ν̂ k ) + J2 (wk ) ց J∗ := inf{J }
Note that the assumption Vad 6= ∅ means that J∗ < ∞ and we can assume
that J k < [bound] iy for each k. Since this bounds {ν̂ k } in V, we may assume,
without loss of generality, that (ν̂ k ) converges uniformly on [0, T ] with weak
convergence in L4/3 (0, T ) of (ν̂tk ), i.e., ν̂ k ⇀ ν̄, ν̂tk ⇀ ν̄t . Since J1 (·) is lower
semicontinuous in this topology, we have J1 (ν̄) 6 lim inf J1 (ν̂ k ).
Next, using Corollary 2.1, we have convergence of (wk ) to the solution
corresponding to this control ν̄. Since χ∗ = 0 for each wk , the uniform
convergence ensures that this also holds in the limit. We only have weak L2 (Q)
convergence for the accelerations wtt k , but we note that ω 7→ sup {|ω|} is lower
Q
semicontinuous with respect to the weak L2 topology (since {ω ∈ L2 (Q) :
|ω| 6 α} is convex and strongly closed for each α). Thus, in the limit we
have J2 6 lim inf J2 (wk ) so J 6 J∗ and the minimum is attained at ν̄.
Some characterization of this optimal control ν̄ through (formal) computa-
tion of first-order optimality conditions (expressed in terms of a linear adjoint
equation) would certainly be possible, if rather messy for the particular cost
functional we have treated here, but we do not pursue this.

Problem 2: Now suppose that ν̂ and the initial data are specified and that
we seek an optimal distributed control f ∈ U. We consider two variants of this
problem: we may insist on minimizing the U-norm while exactly matching the
target state at t = T or we may penalize deviation from the target as measured
in some norm. Thus, we either consider


(2.6) J = J1 (f ) := kf kU subject to: [w, wt ] = [w̄, w̄t ]
t=T

or, for example, topologizing deviation in C 2 [0, ℓ]

× C 1 [0, ℓ] for the second
variant, we may take J = J1 (f ) + J2 (w) with J1 (·) as in (2.6) and
J2 (w) := sup{ |w(s, T ) − w̄(s)|, |wss (s, T ) − w̄′′ (s)|,
(2.7)
|wt (s, T ) − w̄t (s)|, |wts (s, T ) − [w̄t ]′ (s)| : s ∈ [0, ℓ]}.
[Note that (2.6) is equivalent to introducing J2 (w) := 0 if there is such a match,
and J2 (w) := +∞ otherwise.] For neither variant do we expect J to be finite for
arbitrary f ∈ U (for (2.7), because our estimates do not bound wss at all), but
we do assume a priori that the admissible control set Uad := {f ∈ U : J < ∞}
is nonempty for whichever variant is under consideration. This restricts our
choice of target states to consider and, in particular, for the first variant it
means requiring that the target be exactly reachable.
Theorem 2.2. With ν̂, w̄ and the initial data as above and under the
hypotheses on the constitutive function n(··) of Section 3, there is an optimal
278 T.I. Seidman and S.S. Antman

control ν̄ for Problem 2, i.e., the cost functional J , given by ( 2.6 ) or ( 2.7 )
as appropriate, attains its minimum.
Proof. The proof is similar enough to that given for Theorem 2.1 that we only
comment on it briefly. We can now begin by finding an infimizing
sequence (f k )
for which fk ⇀ f¯ in L2 (Q) and ftk ⇀ f¯t in Lp [0, T ] → L2 (0, ℓ) , the topology
for which J1 (·) is both coercive and lower semicontinuous. As before, we now
extract, if necessary, a subsequence for which Lemma 2.2 gives convergence
of the corresponding solutions (wk ). For the first variant we then need only
note that Corollary 2.1 ensures that the terminal condition in (2.6) holds in
the limit since it holds for each wk . For (2.7) we note, much as in the proof of
Theorem 2.1, that the convexity of the set {w : J2 (w) 6 α} for each α gives
the needed lower semicontinuity of J2 (·). In any case, we have J 6 J∗ in the
limit and the minimum is attained at f¯.

Problem 3: Finally, we suppose that f ∈ U and the initial data (subject

to (2.1)) are specified and that we have also specified a target state [w̄, w̄t ] ∈
C[0, ℓ] × C[0, ℓ] and constraints. These are to include (the possibility of)
both control constraints (for which the boundary control ν̂ lies in a specified
subset Vad ⊂ V) and state constraints (for which the trajectory t 7→ w
determined by using this ν̂ in (1.2) lies in a specified subset W ⊂ C(Q̄)).
While one could consider more general examples, we assume, for expository
simplicity, that Vad is convex, closed, and bounded in V and that W is defined
(almost) pointwise so that


(2.8) W = {w(·) : [w, wt ] ∈ Ŵ(t) for 0 6 t 6 τ = τ (ν̂)}
t

where each Ŵ(t) is closed in [C 1 (0, ℓ)]2 . We also ask, of course, that there be
some control ν̂ ∈ Vad for which w(·) is not only inW, but also matches the
target, i.e.,


(2.9) [w, wt ] = [w̄, w̄t ]
t=τ (ν̂)

Our goal is to minimize the control time τ for this match.

Theorem 2.3. With f, w̄, the initial data, and the constraint sets Vad , W
as above and under the hypotheses on the constitutive function n(··) of Section 3,
there is an optimal control ν̄ for Problem 3, i.e., there is a control for which the
corresponding solution reaches the target state at the minimum control time τ∗ .
Proof. As before, we consider an infimizing sequence (ν̂ k ) with corresponding
solutions wk for which

τ k = τ (ν̂ k ) ց τ∗ := inf{τ (ν̂) : ν̂ is admissible }

 Optimal control of a viscoelastic rod 279

where ‘admissible’ means that ν̂ ∈ Vad , w(·) ∈ W, and ( 2.9 ) . Without

loss of generality, since Vad is bounded, we may assume that (ν̂ k ) is weakly
convergent: ν̂ k ⇀ ν̄ and our assumptions on Vad ensure that we also have ν̄ ∈

Vad . The admissibility of each ν̂ k ensures that [wk , wtk ] ∈ Ŵ(t) for 0 6 t 6 τ∗
t
since that τ k > τ∗ . Since Lemma 2.2 ensures (for a subsequence, to which we
now restrict our attention) that wk and uk = wsk converge uniformly on Q̄, we
have wk (·, t) converging in C 1 (0, ℓ) and, similarly we get uniform convergence
of v k = wtk and of wtsk , so that wk (·, t) converges in C 1 (0, ℓ). Hence, the limit
t
∗
solution w , whose existence is given by Corollary 2.1, also satisfies the defining
condition of W on [0, τ∗ ] and so will be in W, as desired, once we show that w∗
satisfies (2.9) at t = τ∗ , making τ (ν̄) = τ∗ . To verify (2.9), we need only note
that  
 
[wk , wtk ] k −→ [w∗ , wt∗ ]
τ τ∗

is ensured by the uniform convergence provided by Lemma 2.2.

3 Hypotheses and first estimates

In this section and the next we provide the estimates giving the compactness of
Lemma 2.2 and underlying the existence proof (which we will not present here)
of Lemma 2.1.
We split the constitutive function n(y, z) for the contact force into an elastic
part n(y, 0) =: ϕ′ (y) and a dissipative part σ(y, z) (with σ(y, 0) ≡ 0):

(3.1) n(y, z) = ϕ′ (y) + σ(y, z) for y > 0 and all z.

We assume that ϕ, σ are smooth where defined, e.g., ϕ′ , σ are C 1 .

We impose two hypotheses on the constitutive function n(··), i.e., on ϕ, σ:
the first is a rather natural assumption that there be some minimal dissipation
while the second, formulated in terms of an auxiliary function ψ, is rather
technical. Our first hypothesis is that

(3.2) σz (y, z) > µ (for some fixed µ > 0).

Note that this gives µ|z|2 6 z σ(y, z). Next, we introduce ψ : (0, ∞) → IR+ ,
requiring that ψ(y) → ∞ as y → 0 to enforce our prohibition against total
compression, and then impose our second hypothesis:
(a) For each c > 0 there is a constant λ = λ(c) such that

(3.3) [z ny (y, z)]2 6 λσz (y, z) [1 + z σ(y, z) + ϕ(y)] when ψ(y) 6 c

while
(b) There are β, c̄ > 0 such that
280 T.I. Seidman and S.S. Antman

(3.4) z ψ ′ (y) 6 n(y, z) + β when ψ(y) > c̄.

We include in this section some estimation which requires only L2 bounds

for f, ν̂, beginning with the usual energy estimate. Taking ω = v in (1.3-ii) and
noting (2.1)) gives
Z t
1 2
2 kvk + hϕ(u)i + hvs σ(u, vs )i
(3.5) 0
◦ ◦ Rt Rt
= 21 k v k2 + hϕ(u)i + 0 v(ℓ, ·)ν̂ + 0 hv, f i
Rℓ
where hϕ(u)i = 0 ϕ(u(s, t)) ds, etc. We have used (3.1) and the fact that
vs ϕ′ (u) = [ϕ(u)]t since vs = ut . Estimating |v(ℓ, t)|2 6 εkvs k2 + Cε kvk2 and
recalling that µ|vs |2 6 vs σ(u, vs ) so we may choose ε > 0 small and absorb this
term εkvs k2 on the left. We apply the Gronwall Inequality to obtain
Z t
2


(3.6) kvk , hϕ(u)i, hvs σi 6 C 1 + kν̂k2 + kf k2 6 C
0

with C denoting a positive constant depending only on T, µ and bounds on the

data, as indicated.
Next we show that our hypotheses, as hoped, ensure the impossibility of
total compression: we will bound u away from 0. We first integrate (1.1)
over (0, s̄) to obtain
 Z s̄

(3.7) n(u, ut ) =: ν(s̄, t) = [vt − f ] ds.
(s̄,t) 0
◦
We may assume in (3.4) that c̄ > maxs {ψ(u)} so if ψ(u(s̄, t̄)) > c̄, then there is
some τ > 0 with ψ(u(s̄, τ )) = c̄ and ψ(u(s̄, ·)) > c̄ on (τ, t̄). Thus (3.4) applies
to give ut ψ ′ (u) 6 β + n(u, ut ) there, and use (3.7); integrating in t over (τ, t̄)
then gives Z s̄  Z t̄ Z s̄
t̄
ψ(u(s̄, t̄)) 6 c̄ + (t̄ − τ )β + v  − f.
0 τ τ 0

The terms on the right can all be estimated, by using an L2 (Q) bound on f for
the last and (3.6) for the penultimate term; thus, ψ(u) 6 C and we will later be
able to apply (3.3). Further, since ψ(y) → ∞ as y → 0, this shows that u = wt
is uniformly bounded away from 0.

4 Further estimates and compactness

We continue our estimation, now assuming U, V bounds for f, ν̂ and using (3.3)
with the appropriate λ.
◦ ◦
We begin by setting ζ := vt = wtt . Note, first, that (2.1) gives u=ws
◦
in H 1 (0, ℓ), so that u is continuous (and bounded pointwise on [0, ℓ]) as well
 Optimal control of a viscoelastic rod 281
◦
as bounded away from 0; similarly, ut = [wt ]s is bounded. These bounds
ensure that the arguments remain in a compact set for which we have bounds
◦
on ny and nz = σz . This will enable us to verify the regularity of ζ:= ζ(·, 0).
Using (1.1) at t = 0, we have
◦ ◦ ◦ ◦ ◦
ζ= [n(u, v s )]s + f (·, 0) = ny wss +σz [w t ]ss + f

which, with ny , σz bounded, is in L2 (0, ℓ).

Differentiating (1.3-ii) in t one obtains

hω, ζt i + hωs , σz ζs i = ω(ℓ)ν̂t + hω, ft i − hωs , ny ut i.

Taking ω = ζ and integrating over (0, t) then gives

1 2
Rt
2 kζ(t)k + 0 hζs , σz ζs i hR i1/4
◦
t
(4.1) 6 21 k ζ k2 + C 0 |ζ(ℓ, ·)|4 + C z̄(t)
Rt 
+ 0 εhζs , σz ζs i + λ4 ε (1 + hvs σ(u, vs )i + hϕ(u)i)

where we have set z̄(t) := max{kζ(t̄)k : 0 6 t̄ 6 t}, have applied (3.3) after
noting that
|ζs ny ut | 6 εσz (ζs )2 + (1/4εσz )|ny ut |2 ,
have used the relevant bounds on ft , ν̂t , and have noted that the final set of
terms were already bounded by (3.6). Since |ζ(ℓ, t)|2 6 (1/ℓ)kζk2 + 2z̄kζs k
we have Z t 1/4
4
|ζ(ℓ, ·)| 6 εkζ(t)k2 + εkζs k2 + Cε [1 + z̄(t)] .
0

Using this in (4.1) with ε chosen to absorb those terms on the left then gives

kz(t)k2 6 C[1 + z̄(t)] 6 C[1 + z̄(t̄)]

with t̄ arbitrarily fixed so t 6 t̄. Taking the maxt over [0, tb] then gives z̄ 2 6
C[1 + z̄] (uniformly for t̄ ∈ [0, T ]) and we have bounded3 ζ = vt = wtt
in L∞ ((0, T ) → L2 (0, ℓ)).
In view of (3.7), this gives us also a uniform pointwise bound for ν.
Returning to (4.1) we see that we have also bounded the integral on the left
and so have bounded ζs = wstt = utt in L2 (Q); integrating this over (0, t),
using (2.1), bounds ut = vs = wst in L∞ ((0, T ) → L2 (0, ℓ)). Summarizing, we
have by now shown

(4.2) kwtt (·, t)k 6 C, |ν(s, t)| 6 C, kwstt kQ 6 C, kwst (·, t)k 6 C

3
This and a density argument show that ζ is actually in C([0, T ] → L2 (0, ℓ)).
282 T.I. Seidman and S.S. Antman

as well as (3.6) and the bound away from 0 obtained for u = ws .

Now consider ϕ′ (u) + σ(u, ut ) =: n(u, ut ) = ν(s̄, ·) with s = s̄ arbitrarily
fixed in (0, ℓ). Multiplying this by ut and integrating the product over [0, T ]
gives, using the bound on ν,
Z T Z T
2
µkut k[0,T ] 6 ϕ(u) + ut σ(u, ut ) dt = ut ν dt 6 Ckut k[0,T ]
0 0

which bounds ut = vs = wst in L2 (0, T )

on each s = s̄. Integrating over (0, t)
and noting (2.1) then uniformly bounds u pointwise on Q and so, with our
previous bound away from 0, restricts the argument of ϕ(·) to a compact
subinterval of its domain, whence ϕ(u), ϕ′ (u) are necessarily also bounded. Now
we apply |z| 6 |σ(y, z)|/µ to bound ut pointwise, noting that σ(u, ut ) = ν−ϕ′ (u)
has been bounded.
This, of course, only bounds u = ws and ut = vs = wst in L∞ (Q) and,
although a density argument would suffice to show that these remain within
the closed subspace C(Q̄), we wish to show that they
lie in a compact subset of
that. [Our bound on ft is in Lp [0, T ] → L2 (0, ℓ) , by (2.2), but we note that
it would have been sufficient, up to this point, only to have required a bound
on ft in the dual space of C([0, T ] → L2 (0, ℓ)).]
We prepare for the next step by recalling a compactness lemma appearing
in a footnote in [4] (note also [5]):
Lemma 4.1. Let X ֒→ Y ֒→ Z with the embedding X ֒→ Y compact.
Suppose that F is a set of functions on [0, T ] which is bounded in L∞ ([0, T ] →
X ) and for which {ψt : ψ ∈ F} is bounded in Lp ([0, T ] → Z) for some p > 1.
Then F is a precompact subset of C([0, T ] → Y).
We now apply this to F = {ν} with X = H 1 (0, ℓ), Y = C[0, ℓ], Z = L2 (0, ℓ),
which satisfy the embedding requirements. From (3.7) we already have νs =
ζ − f bounded in L2 (0, ℓ) so ν is bounded in H 1 (0, ℓ). On the other hand,
(4.3) νt = [n(u, ut )]t = ny ut + σz ζs
on noting that utt = vst = ζs . At this point, since the arguments of n(u, ut ) have
been restricted to a fixed compact subset of the domain (0, ∞) × IR we have
uniform bounds on the derivatives ny , nz = σz . As we already have bounds
on ut in L∞ (Q) and on ζs in L2 (Q), the right hand side of (4.3) is bounded
in L2 ([0, T ] → Z) as desired and we have shown that ν lies in a fixed compact
subset of C([0, T ] → Y) = C(Q̄).
Next we notice that the ordinary differential equation: n(u, ut ) = ν (with
◦
specified initial data u∈ C[0, ℓ]) defines a continuous solution map4
S : ν 7→ ut : C(Q̄) → C(Q̄)

4
Since n(y, z) is not globally smooth and is not defined on all of IR2 (which has forced
 Optimal control of a viscoelastic rod 283

and the set K of (2.3) is the image under S of the compact subset of C(Q̄)
obtained to contain ν. The compactness of K follows immediately, of course,
from the continuity of S.
This effectively completes the proof of Lemma 2.2. The estimates here
provide the core of the existence proof but we omit the actual proof, referring
the reader to [1] for further detail.

References

[1] S.S. Antman and T.I. Seidman, Quasilinear hyperbolic-parabolic equations of one-
dimensional viscoelasticity, J. Diff. Eqns., 124 (1996), pp. 132–185.
[2] S.S. Antman and T.I. Seidman, The spatial motion of nonlinearly viscoelastic rods,
in preparation.
[3] D. French, S. Jensen, and T.I. Seidman, A space-time finite element method for a
class of nonlinear hyperbolic-parabolic equations, Appl. Numer. Math., 31 (1999),
pp. 429–450.
[4] T.I. Seidman, The transient semiconductor problem with generation terms, II,
in Nonlinear Semigroups, PDE, and Attractors (LNM #1394; T.E. Gill,
W.W. Zachary, eds.), Springer-Verlag, New York, 1989, pp. 185–198.
[5] J. Simon, Compact sets in the space Lp (0, T ; B), Ann. Mat. Pura Appl., 146
(1987), pp. 65–96.

us to use the restriction of the arguments which we have obtained), one must redefine S.
E.g., suppose the rectangle D is the compact subset of IR2 to which we have restricted values
of [u, ut ]. We redefine n(y, z) when |z| is too large for D so as to be smooth (still), uniformly
Lipschitzian in z, and still satisfying nz > µ; then further redefine n (now for y outside the
relevant range) to coincide with the values for the nearest admissible y-value. There is then
no obstruction to the global (re)definition of S and we observe that the results must coincide
with the original results for all relevant inputs so the redefinition is nugatory.
284 T.I. Seidman and S.S. Antman
Mathematical Modeling and Analysis for Robotic
Control

S. Tsui, Oakland University, Rochester, MI.

Abstract
We present the results of investigation of the torsional elastic robot
beams. Next, we study the geometry of the joint space of a multi-joint
robot. This opens doors to a new horizon of future research for torsional
elastic multi-joint robots.

1 Introduction
With the rapid development of robotics in engineering, the coupled bending and
torsional vibrations of elastomers appear frequently in application. Therefore,
in this article, we summarize the recent results of the research on two topics: 1.
the design of control for a loong and thin flexible robot arm (see Sections 2 and
3); 2. the analysis of joint space of multi-joint robots (see Section 4). In Section
5, we propose to study mathematical modeling and analysis of multi-joint robots
with flexible arms. We first describe the flexible robot system as an evolution
equation in an appropriate Hilbert space, and then apply functional analysis,
spectral theory of linear operators and semigroup theory of linear operators to
investigate stability. Then, we design a controller so that the considered system
is exponentially stable under this control, and the tip of the arm of the robot
can reach any designated point. Related works on the control of beams can be
found in [[2], [3], [7], [8], [9], [10], [11], [12], [13], [14], [15], [16]]. We also present
the results of the study of the geometry of the joint space of multi-joint robots.
Selected references for this topics are from [[1], [4], [5], [6]].

2 The model of a beam with a tip body

Consider a long and thin flexible beam which is rotated by a motor in a
horizontal plane. The beam is clamped on a vertical shaft of the motor at
one end and has a tip body rigidly attached at the free end as shown in Fig.
1. The beam is of length l and with a uniform mass density ρ per unit length,
uniform flexural rigidity EI, and uniform torsional rigidity GJ. Let X0 , Y0 , Z0
be the inertial Cartesian coordinate axes, where X0 , Y0 axes span a horizontal
285
286 Tsui

plane, and Z0 axis is the axle of rotation of the motor. Let X1 , Y1 , Z1 with
Z1 = Z0 denote coordinate axes rotating with the motor and θ(t) be the angle
of rotation of the motor. Let Q be the mass center of the rigid tip body, and P
be the intersection of the beam tip’s tangent with a perpendicular plan passing
through the Q. Let C denote the distance between the beam’s tip point and P ,
and C is assumed to be small. It is also assumed that P and Q never coincide
and lie on the same vertical line in the equilibrium state. Let e be the distance
between P and Q.
We take another coordinate axes, X2 , Y2 , Z2 attached to the tip body, where
X2 is the beam’s tip tangent and is obtained by rotation X1 axis by θ1 due to
the bending of the beam. During the motion the tip body oscillates about
a shear-center axis P X2 like a pendulum. Let Φ be the angle of rotation of
the tip body about P X2 . The axes Y2 , Z2 also oscillates together with the tip
body. Since the tip body is a rigid body, it is characterized by mass m, and two
moments of inertia J0 and JE , where J0 is 0 with respect to the line passing
through Q and parallel to the axis P Z2 and JE is with respect to the line
passing through Q and parallel to the axis P X2 .

Now let y(t, x) and φ(t, x) be the transverse displacement of the beam in
the rotating frame X1 , Y1 and the angle of twist of the beam, respectively, at
position x, 0 < x < l, and at time t. For the transverse vibration we use the
Euler-Bernoulli model with internal viscous damping of the Voigt type [14]
( 2
∂ y(t,x) ∂ 5 y(t,x) EI ∂ 4 y(t,x)
2 + 2δ EI
ρ 4 + ρ ∂x4
= −xθ̈(t)
(2.1) ∂t
′
∂t∂x
y(t, 0) = y (t, 0) = 0,

where δ > 0 is a small damping constant of the beam material. The initial
conditions are due to the fact that the beam is clamped at x = 0.
We assume that the beam material is isotropic and the internal damping
constant for the torsional vibration is equal to that of the transverse vibration.
Therefore the torsional vibration is governed by
( 2
∂ φ(t,x) GJ ∂ 3 φ(t,x) GJ ∂ 2 φ(t,x)
2 − 2δ ρk 2 · ∂t∂x2 − ρk 2 · ∂x2
=0
(2.2) ∂t
φ(t, 0) = 0,

where ρk2 is the polar momentum of inertia mass for per length of beam.
Obviously φ(t, l) = Φ(t), yx (t, l) = θ1 (t).
Neglecting some nonlinear small quantities, we obtain the total kinetic
energy of end body by
1 1
T = JE [φ̇(t, l)]2 + J0 [θ̇(t) + ẏ ′ (t, l)]2
2 2
1
+ m[(l + c)θ̇(t) + ẏ(t, l) + cẏ ′ (t, l) + eϕ̇(t, l)]2 ,
2
Mathematical Modeling and Analysis for Robotic Control 287

Y1 Y0 Y2

X2
eΦ
Q θ1 (a)

P
φ
y
X1
L x
θ
X0

Z0

(b)

Fig. 1. Bending and torsion of a flexible beam with a tip body.

288 Tsui

where “.” denotes the time derivative, and “′ ” denotes the spatial derivative.
We choose y(t, l), y ′ (t, l), φ(t, l) as the generalized coordinates, and f1 , f2 and
f3 as the corresponding generalized forces defined by

f1 = EIy ′′′ (t, l) + 2δEI ẏ ′′′ (t, l)

f2 = −EIy ′′ (t, l) − 2δEI ẏ ′′ (t, l)
f3 = −GIφ′ (t, l) − 2δGI ϕ̇′ (t, l).

¿From the second class Lagrange’s equation we have

 
d ∂T ∂T
− = fi (i = 1, 2, 3),
dt ∂ q̇i ∂qi
where q1 = y(t, l), q2 = y ′ (t, l), q3 = φ(t, l). We can derive the following
boundary equations of coupled bending and torsional vibrations of flexible beam
as follows.
 
(2.3) m (l + c)θ(t) + ÿ(t, l) + cÿ ′ (t, l) + eϕ̈(t, l) = EIy ′′′ (t, l) + 2δEI ẏ ′′′ (t, l)

h i h i
(2.4) mc (l + c)θ̈(t) + ÿ(t, l) + cÿ ′ (t, l) + eϕ̈(t, l) + J0 θ̈ ′ (t) + ÿ ′ (t, l)
= −EIy ′′ (t, l) − 2δEI ẏ ′′ (t, l)

h i
(2.5) me (l + c)θ̈(t) + ÿ(t, l) + cÿ ′ (t, l) + eϕ̈(t, l) + JE ϕ̈(t, l)
= −GJφ′ (t, l) − 2δGJ ϕ̇′ (t, l).

The rigid turning angle θ(t) of the beam is described by:


Jm θ̈(t) + µθ̇(t) = τc (t) + EIy ′′ (t, 0)
(2.6)
θ(0) = θ (1) , θ̇(0) = θ (2) ,
where Jm is the inertia moment of the electrical motor, µ is the viscous-friction
coefficient, EIy ′′ (t, 0) is the bending moment of the flexible beam, and τc (t)
is the torque of the motor. Unlike in equations (2.3), (2.4), and (2.5), there
is no damping term for the angle of turning θ(t) at the shaft end in (2.6), for
it is negligible in comparison with the damping at the end of the tip body.
The motion differential equations of the robot system is described by (2.1)–
(2.6). The turning angle θ(t, y(t)) depends on time variable, t, and the bending
moment, y ′′ (t, 0).
We shall choose the space H = L2 (0, l) × L2 (0, l) × R3 as a state space,
which is a Hilbert space equipped with the inner product defined as
Z lh i 5
X
hu, viH = ρ u1 (x)v1 (x) + k2 u2 (x)v2 (x) dx + ui v̄i ,
0 i=3
Mathematical Modeling and Analysis for Robotic Control 289

where u = (u1 , u2 , . . . , u5 )T , v = (v1 , v2 , . . . , v5 )T , u, v ∈ H, and (. . . )T means

the transpose of (. . . ). Let V be a subspace of H defined by

V = {u = (u1 , u2 , . . . , u5 )T | u1 (x) ∈ H 2 (0, l), u2 (x) ∈ H 1 (0, l), u3 = u1 (l),

u4 = u′1 (l), u5 = u2 (l), u1 (0) = 0, u′1 (0) = 0, u2 (0) = 0},

where H m (0, l) = {f ∈ L2 (0, l) = f ′ , f ′′ , . . . , f (m) ∈ L2 (0, l)} is the m-th degree

Sobolev space, m = 1, 2.
We now define the inner product on V by
Z lh i 5
X
hu, viV = u′′1 (x)v1′′ (x) + u′2 (x)v2′ (x) dx + ui v̄i .
0 i=3

It is easy to see that V with the inner product h·, ·iV is a Hilbert space.
Define an operator ∧ : H → H as follows.
 
1 0 |
 | 0 
 
 0 1 | 
 
∧u = 
 −− | −−  u,
 (u ∈ H)
 | 
 
 0 | M 
|

where  
m mc me
M =  mc J0 + mc2 mce .
me mce JE + me2

It is obvious that ∧ and M are symmetric positive operators.

Due to the positivity of the operator ∧, we can define another inner product
as
Rlh i
hu, viH′ = h∧u, viH = ρ 0 u1 (x)v1 (x) + k2 u2 (x)v2 (x) dx
+(u3 , u4 , u5 )M (v̄3 , v̄4 , v̄5 )T .

We denote the space (H, h·, ·iH′ ) by H′ . It is apparent that there are two
constants c1 and c2 such that

c1 kukH 6 kukH′ 6 c2 kukH .

Thus, H′ is also a Hilbert space. Furthermore, we define the operator B :

D(B) → H by
 
EI d4 GJ d2 d3 d d
Bu = diag , − 2 2 , −EI 3 , EI , GJ u, u ∈ D(B).
ρ dx4 ρk dx dx dx dx
290 Tsui

Here D(B) = {u = (u1 , u2 , . . . , u5 )T | u ∈ V, u′′1 ∈ H 2 (0, l), u′2 ∈ H 1 (0, l)} is

the domain of B.
In the systems (2.1)–(2.6), the turn angle θ is related to time t and the
bending vibration displacement y of the beam,i.e., θ = θ(t, y). If we introduce
the notation Ω = −(x, 0, m(l + c), J0 + mc(l + c), me(l + c))T , and then the
system (2.1)–(2.5) can be described as the following second order homogeneous
evolution equation.

(2.7) ∧ü(t) + 2δB u̇(t) + Bu(t) = Ωθ̈(t, y(t)).

Let A = ∧−1 B, D(A) = D(B). Then (2.7) becomes

(2.8) ü(t) + 2δAu̇(t) + Au(t) = θ̈(t, y(t)) ∧−1 (Ω).

The corresponding second order homogeneous evolution equation is as follows

(2.9) ü(t) + 2δAu̇(t) + Au(t) = 0.

The above setup first appeared in [14].  

⇀ du 0 I
Set u= (u(1) , u(2) )T , u(1) = u(t), u(2) = dt , A = ,
−A −2δA
⇀ ⇀
D(A) = D(A) × D(A), and F (t, u) = (0, ∧−1 (Ω)θ̈(t, y(t)))T . Then (2.8)
becomes
⇀
d u (t) ⇀ ⇀ ⇀
(2.10) = A u (t)+ u (t, F ).
dt
The corresponding homogeneous evolution equation is
⇀
d u (t) ⇀
(2.11) = A u (t).
dt
The equations (2.10) and (2.11) first appeared in [9]. Now, we list four theorems,
a lemma and a corollary whose proofs can be found in [9].
Theorem 2.1. The operator A is a densely defined, self-adjoint, positive
definite operator on V .
Theorem 2.2. The inverse of A exists and it is compact.
Theorem 2.3. The spectrum σ(A) of A consists of only countable eigen-
values {λn } with finite multiplicity, so that 0 < λ1 < λ2 < · · · < λn < . . . and
λn → ∞(n → ∞).
Let orthogonal unital eigenvectors of A corresponding to the eigenvalue
λn be φnj where (j = 1, 2, . . . , nk ; nk is finite) such that Aφnj = λn φnj ,
kφnj kH′ = 1. It is known that {φk1 , . . . , φnnj }∞
k=1 form an orthonormal basis
′
for H .
Mathematical Modeling and Analysis for Robotic Control 291

Next, we shall discuss the spectral properties of the main operator A in the
1
evolution equation (2.10). Let’s consider a dense subspace E = D(A 2 ) × H′
with a new inner product defined by
⇀ ⇀ 1 1
(2.12) h u , v iE = hA 2 u(1) , A 2 v(1) iH′ + hu(2) , v(2) iH′ ,

⇀ ⇀
where u= (u(1) , u(2) )T , v = (v(1) , v(2) )T ∈ E
Lemma 2.1. The space E with the inner product defined in (2.12) is a
Hilbert space.
It is easy to see that E has an orthonormal basis consisting of the following
vectors        ∞
φk1 0 φknk 0
, ,..., , .
0 φk1 0 φknk k=1

Theorem 2.4. Denote the spectrum of A by σ(A), the point spectrum of

A by σp (A), the resolvent of A by ρ(A), then we have the following results:
1
1. σ(A) = σp (A) ∪ {− 2δ }, σp (A) = {ξk , ηk }∞
k=1 ,where

p p
ξk = −δλk + (δλk )2 − λk , ηk = −δλk − (δλk )2 − λk
λ
= pk
−δλk − (δλk )2 − λk
1 1
→ √ = , as k → ∞
−δ − δ2 −2δ

and the eigenvectors of A corresponding to ξk and ηk are, respectively, as

follows
   
⇀ 1 φkj ⇀ 1 φkj
φ kj = p , ψ kj = p
λk + | ξk |2 ξk φkj λk + | ηk |2 ηk φkj

⇀ ⇀
with k φ kj kE = | u ψkj kE = 1.

2. if µ ∈ ρ(A), then

(µI − A)−1
 
µ2 + 2δµA + A)−1 (µ + 2δA) (µ2 + 2δµA + A)−1
= .
−I + (µ + 2δµA + A) (µ + 2δµA) µ(µ2 + 2δµA + A)−1
2 −1 2

Corollary 2.1. The operator A is a closed linear operator.

292 Tsui

3 Stability and Control of the System

In this section we show that the system in the previous section is asymptotically
stable, and we design a control for the system. Again, the proofs for the
theorems in this section can be found in [9]. First, we show that the real
parts of the eigenvalues of A are bounded above, and there exists a constant
ω1 > 0 such that

sup{Reµn : µn ∈ σp (A)} = −ω1 (ω1 > 0).

In fact, it is easy to see from Theorem 2.5 that Reξk < 0, Reηk < 0. Since
1 1
limk→∞ ξk = − 2δ , limk→∞ ηk = −∞, it follows that limk→∞ Reξk = − 2δ and
limk→∞ Reηk = −∞. Hence we have the following theorem.
Theorem 3.1. The operator A in (2.10) or (2.11) is the infinitesimal
generator of a C0 -semigroup T (t) on Hilbert Space H′ ⊕ H′ , and there are
constants M > 0 and ω > 0 such that kT (t)k 6 M e−ωT (t > 0).
Theorem 3.2. The first order homogeneous evolution equation (2.11) has
⇀
a unique solution u (t).
Theorem 3.3. The solution u(t) of the second order evolution equation
(2.9) is asymptotically stable.
Theorem 3.4. Suppose for every T > 0, θ̈ : [0, T ] × L2 (0, l) → L2 (0, l)
is Lipschitz continuous (with constant N ) in y on L2 (0, l), then nonlinear
⇀
evolution equation (2.10) has a unique weak solution u∈ C([0, T ]; H).
Theorem 3.5. Let T > 0, θ̈ : [0, T ] × L2 (0, l) → L2 (0, l) be continuously
⇀
differentiable, then u 0 = (u(0), u̇(0))T ∈ D(A), and nonlinear evolution
equation (2.10) has a unique strong solution.
In order to investigate the properties of the solution to (2.10), we denote
C([0, +∞)) = {f : f is continuous on [0, ∞) and kf k∞ = supt>0 |f (t)| < +∞}.
It is clear that the space C([0, +∞)) with norm k · k∞ is a Banach space.
We define an operator on C([0, +∞)) by
Z t
Kg(t) = e−ω(t−s) g(s) ds, g ∈ C([0, +∞)),
0

where ω can be found in Theorem 3.1.

Lemma 3.1. The operator K is a linear bounded operator on C([0, +∞))
and kKk∞ 6 1/ω.
Theorem 3.6. Suppose θ̈ : [0, T ] × L2 (0, l) → L2 (0, l) is uniformly Lipshitz
continuous in y on L2 (0, l) for any T > 0 with a Lipschitz constant N <
√ ⇀
cρ ω/a0 M k ∧−1 k. Then the solution u (t) to the nonlinear evolution (2.10)
decays exponentially so that the solution u(t) to original system (2.1) − −(2.5)
is asymptotically stable in exponential form.
Theorem 3.7. If we design the following controller

τc (t) = −EIy ′′ (0, t) − η(θ(t) − θ0 ),

Mathematical Modeling and Analysis for Robotic Control 293

where θ0 ∈ [0, 2π], 0 6 η 6 µ2 /4Jm . µ and Jm can be found in (2.6), for the
system (2.1)–(2.6), then the bending vibration y(t, x) and torsional vibration
φ(t, x) of the robot arm can be suppressed to be exponentially stable, and the
elastic arm of robot can be arrived at any designated position, that is,

lim θ(t) = θ0 .
t→∞

4 Joint spaces of robots with joints

In the case of robots with joints the dynamics on the robot is known to
be derived through one of the following ways; (1) Lagrangian formulation,
(2) Newton-Euler formulation, (3) Kane’s partial velocity, and (4) compact
dynamic formulation. In the following figure, we demonstrate how to derive
an n-dimensional manifold from an n-joint robot. In order to position the end-
effector of a robot arm, one needs to specify a position and an orientation for the
end-effector, each of which has three degrees of freedom. In general, a complete
robot arm has six joints for six degree of freedoms. A joint can be represented as
a rotational angle or a length of a segment. Therefore, the motion of the robot
arm depends on how one manipulates these six joints, each joint represented by
a variable xi , for i = 1, 2, ..., 6. The set of all (x1 , x2 , . . . , x6 ) form a manifold
in a Euclidean space of higher dimension. This is called the joint space of the
robot. This invisible joint space dictates the dynamics as well as the kinematics
of the robot arm. It is the curvature tensor of this joint space that determines
the dynamics of a robot arm.

On this joint space we define a metric by the Hamiltonian of the system

through the Newton-Euler formulation, and calculate the curvature. In case of
n = 2, we can determine the Gaussian curvature, and in higher dimensional
cases, we find sectional curvatures. With the aid of metric and curvature, one
can find the shortest path between two fixed points on the joint space, which is
called the geodesic. The geodesics on a joint space are the paths of minimum
energy between two positions, and hence the best path for the robot arm to
move. Therefore the most efficient robot design should always try to move the
arm on a geodesic. The apparent difficulty in this design is that the visible arm
with its end-effector give a false image of the position for the robot arm in the
invisible joint space.
In the case of robots with two planar joints which is described in the
following figure.
In this case, we can calculate the first and the second fundamental forms of
the joint space. By the Rigidity Theorem for hypersurfaces the joint space is
locally (metrically) diffeomorphic to a 2-torus. [5], [6].
Based on the classical Chasler’s Theorem, every rigid body motion can
be decomposed into two decoupled portion: one is the translation of its mass
294 Tsui

Fig. 2. Six-joint robot

center and the other one is the rotation with respect to the mass center. In other
words, this 6-dimensional configuration manifold M 6 can be diffeomorphic to
the compact topological space T (3) × SO(3), where T (3) is a compact subspace
in Euclidean 3-space ℜ3 , representing the bounded translation, and SO(3) is the
3-dimensional special orthogonal Lie group, representing rotation. Clearly, any
T (3) is already a compact Euclidean 3-space, while any subspace of SO(3) has
to be smoothly and isometrically embedded into Euclidean 9-space in general.
Therefore, the configuration manifold M 6 of a rigid body should have up to 12-
dimensional ambient Euclidean space for its smooth and isometric embedding.
Since an n-joint robot arm has n links that form a serial open-chain with the
first one fixed on the base, basically the n-dimensional combined configuration
manifold Mcn of the entire robotic system may need up to 12n-dimensional
Euclidean space for its smooth and isometric embedding. Furthermore, because
the joint positions and velocities of a robot are always the most preferable
variables for any robotic task/path planning and other robotic applications, all
the n joint positions (q 1 · · · q n )T become the best choice of the n-tuple local
coordinate system to be defined at each point on the n-dimensional combined
configuration manifold Mcn . If we denote the coordinates of the m-dimensional
Mathematical Modeling and Analysis for Robotic Control 295

Fig. 3. Two-joint planar robot

ambient Euclidean space by z = (z 1 · · · z m )T , then the embedding can be

written by z = ζ(q). Based on differential geometry, all the n columns g1 , · · · , gn
of the Jacobian matrix G = ∂ζ/∂q of Mcn span a tangent space Tq (Mcn ) of
dimension n at each point on Mcn . The normal space Nq (Mcn ) should be
the orthogonal complementary space of Tq (Mcn ) in ℜm , and thus (m − n)-
dimensional, which can usually be determined by the second fundamental form
of the manifold.
Furthermore, it was shown in [9],[10] that the topological structure of the
combined configuration manifold Mcn for an n-joint robot is solely determined
by the robotic kinematics. The robotic dynamics, however, determines all
geometrical details of the manifold. In other words, the embedding equation
can be rewritten as

(4.1) z = ζ(q) = Z(q)ξ,

where Z(q) is the structure matrix that depends only on the local coordinates
q, i.e., the joint positions of the robot, and determined directly by the robotic
kinematic transformation, while ξ is a column vector of all the dynamic
parameters of the robot. Based on this central idea, we will be able to
create a number of interesting concepts and reveal new properties for robot
dynamic model reduction and adaptive control algorithm development. Based
on equation (4.1), all the dynamic parameters with their uncertainty can be
296 Tsui

viewed as a “coat” wrapping around the structure of the configuration manifold

[11]. We may therefore extend such a parametric coat from its originally
lumped nature to a new distributed version. In other words, the investigation
of dynamic modeling and control of robots with flexible links can be carried
out directly via a new study on geometry and topology of the configuration
manifolds with distributed parametrization. This leads to a new challenging
research topic: partial differential equations on configuration manifolds.

5 Future research problems

Now, we propose the following basic research subjects and further to focus on
their feasible algorithms development towards potential industrial applications:
1. Geometrical and topological analysis of the configuration manifold for
a robotic system to reveal more intimate dynamic behavior and new
characteristics hidden behind the robot dynamic system;

2. Develop new models and adaptive control algorithms based on the concept
of configuration manifolds with their embeddings and isometrizations for
robotic systems;

3. Structural stability analysis of robot dynamic systems to be adaptively

controlled by means of the singularity and catastrophe theories;

4. Modeling, analysis and control design of robot arms with flexible links
and/or flexible joints by developing partial differential equations and
solutions on their differentiable configuration manifolds; and

5. Computer simulation studies and industrial tests for the new models and
control algorithms developed.

References

[1] H. Asada and J-J. E.Slotine, Foundations of Mechanics, The Ben-

jamin/Cummings, 1978.
[2] G. Chen and J. Zhou, Vibration and Damping in Distributed Systems, Vol. 1,
CRC, 1993.
[3] , Vibration and Damping in Distributed Systems, Vol. 2, CRC, 1993.
[4] K. S. Fu, R. C. Gonzalez and C. S. G. Lee, Robotics: Control, Sensing, Vision,
and Intelligence, McGraw-Hill, 1987.
[5] E. Y. L. Gu, Dynamic Systems Analysis and Control Based on a Configuration
Manifold Model, International Journal of Nonlinear Dynamics, 19 (1999), pp. 111-
132.
[6] , A Configuration Manifold Embedding Model for Dynamic Control of
Redundant Robots, International Journal of Robotics Research, (1999).
Mathematical Modeling and Analysis for Robotic Control 297

[7] K. Hannsgen, O. J. Staffans and R. L. Wheeler, Rational approximations of

transfer functions of some viscoelastic rods, with applications to robust control,
preprint, New York.
[8] X. Hou and S. -K. Tsui, A control theory for Cartesian flexible robot arms, Journal
of Mathematical Analysis and Applications, 225 (1998), pp. 265–288.
[9] Control and stability of a torsional elastic robot arm, Journal of Mathemat-
ical Analysis and Applications, 243 (2000), pp. 140–162.
[10] Z. -H. Luo, Direct strain feedback control of flexible robot arms: new theoretical
and experimental results, IEEE Transactions on Automatic Control, 38 (Novem-
ber 1993), no. 11.
[11] F. Matsuno, T. Asano and Y. Sakawa, Modeling and quasi-static hybrid po-
sition/force control of constrained planar two-ling flexible manipulators, IEEE
Transactions on Robotics and Automation, 10 (June 1994), no. 3.
[12] F. Matsuno, T. Murachi and Y. Sakawa, Feedback control of decoupled bending
and torsional vibrations of flexible beams, Journal of Robotic Systems, 11 (1994),
no. 5, pp. 341–353.
[13] K. A. Morris and K. J. Taylor, A variational calculus approach to the modelling
of flexible manipulators, SIAM Review, 38 (June 1996), no. 2, pp. 294–305.
[14] Y. Sakawa and Z. H. Lou, Modeling and Control of Coupled Bending and Torsional
Vibrations of Flexible Beams, IEEE Trans. Auto. Contr., 34 (1989), pp. 970–977.
[15] M. W. Spong and M. Vidyasagar, Robot Dynamics and Control, John Wiley &
Sons, New York, 1989.
[16] S. Timoshenko, D. H. Young and W. Weaver, Jr., Vibration problems in
Engineering, Wiley, New York, 1974.
298 Tsui
Optimal Control and Synthesis of Nonlinear Infinite
Dimensional Systems

Yuncheng You, University of South Florida, Tampa, Florida, USA.

E-mail: [email protected]

Abstract

In this paper, optimal control problems of nonlinear distributed sys-

tems in Banach spaces with a general criterion functional are considered.
The maximum principle as a necessary condition for an optimal process is
proved forthright by the approach of analysis of the oscillating variations
of control. It is shown that value functions are locally Lipschitz continu-
ous. By using lower Dini derivatives, a general synthesis result on optimal
feedback control via solving a differential inclusion is obtained.

1 Introduction
In this paper, we shall consider a nonlinear control system in a real, separable
reflexive Banach space X, whose norm is denoted by | · |. After making certain
assumptions, we shall also deal with a dense subspace W of X, with a stronger
topology, and the norm of W will be denoted by k · k. The nonlinear control
system is governed by an abstract evolutionary equation as follows,

dx
= Ax + f (t, x, u), t > 0,
(1.1) dt
x(0) = x0 ,

where A : D(A)(⊂ X) → X generates a C0 semigroup of linear operators eAt ,

t > 0, on X. X is the state-valued space of x0 and x(t). Let Y be another real
Banach space and U ⊂ Y be an arbitrary, given subset called the control-valued
set. Let T > 0 be a fixed constant. We define

(1.2) U = {u : [0, T ] → Y | u(t) ∈ U and u is strongly measurable}.

We shall make some assumptions on the nonlinear term f and more assumptions
on A later. Since (1.1) is usually formulated from many initial-boundary value
299
300 You

problems of PDEs, it can be called a distributed control system. As usual, we

shall adopt the mild solution of (1.1), i.e.,

Zt
At
(1.3) x(t) = x (t, x0 , u) = e x0 + eA(t−s) f (s, x(s), u(s)) ds,
0

as the state function or called trajectory, where the integral is Bochner integral
in X. For a nonlinear system, there is no guarantee in general that for each
u ∈ U, the mild solution exists globally on [0, T ]. So we have to define a class
of control functions as follows,

(1.4) Uad = {u(·) ∈ U | x (t, x0 , u) exists on [0, T ], for any x0 ∈ W } .

Any u(·) ∈ Uad is an admissible control. Set a cost functional as a criterion,

ZT
(1.5) J (x0 , u) = Q(t, x(t), u(t)) dt,
0

where Q(t, x, u) is a nonlinear functional defined on [0, T ] × X × Y and to be

specified in detail later. An optimal control problem is proposed: For any given
x0 ∈ W , find a control u∗ (·) ∈ Uad such that

(1.6) J (x0 , u∗ ) = min J (x0 , u) .

u∈U
ad

We shall call u∗ an optimal control and x∗ (·) = x (·, x0 , u∗ ) an optimal

trajectory.
We may refer to this optimal control problem described by (1.1), (1.5) and
(1.6) briefly as the (OCP).
Now we make some specific assumptions on A, f and Q that will enable us
to analyze the described optimal control problem concretely.
Hypothesis I. Let −A : D(A) → X be a positive sectorial operator.
Then eAt is an analytic semigroup. Let W 2α = D ((−A)α ) be the family of
interpolation Banach spaces with the norm kxk2α = |(−A)α x|, for α > 0.
The basic properties of eAt under Hypothesis I are listed below, cf. [15, 16]:

(a) For any α > 0 and t > 0, the operator eAt maps X into W 2α and it is
strongly continuous in t > 0.

(b) For any α > 0, there are constants a > 0 and Mα > 0 such that
At
(1.7) e
L(X,W 2α )
= (−A)α eAt L(X) 6 Mα e−at t−α , for t > 0.
 Optimal Control and Synthesis of Nonlinear Systems 301

(c) For 0 < α 6 1, there is a constant Cα > 0 such that

 At 
(1.8) e w − w 6 Cα tα |(−A)α w| , for t > 0, w ∈ W 2α .

We shall fix an α > 0 in Hypothesis III. For that fixed α, let W = W 2α .

We now introduce some notation. Let E be any real Banach space. We shall
use CLip (E) = CLip ([0, T ] × W × Y ; E) to denote the collection of all strongly
continuous functions g : [0, T ] × W × Y → E that are uniformly bounded and
uniformly Lipschitz continuous with respect to x ∈ W for any bounded set
B ⊂ W . That means there exist constants K0 (B) and K1 (B) such that

(1.9) kg(t, x, u)kE 6 K0 (B), for any t ∈ [0, T ] and (x, u) ∈ B × U,

and
kg (t, x1 , u) − g (t, x2 , u)kE 6 K1 (B) kx1 − x2 kW , for any t ∈ [0, T ],
(1.10)
and (xi , u) ∈ B × U, i = 1, 2.

We then define C 1 (E) = C 1 ([0, T ] × W × Y ; E) to be the collection of all

functions g : [0, T ] × W × Y → E that are continuously Fréchet differentiable in
x ∈ W . The first-order Fréchet derivative of g with respect to x will be denoted
by gx or ∂x g which is valued in L(W, E). Let
1
CLip (E) = CLip (E) ∩ C 1 (E).

Hypothesis II. Assume that f ∈ CLip 1 (X) and Q ∈ C 1 (R). Moreover,

Lip
the following assumptions are satisfied:
(i) For any x(·) ∈ C([0, T ], W ) and any u ∈ U, one has

(1.11) |f (t, x(t), u(t))| ∈ Lp [0, T ],

and

(1.12) |Q(t, x(t), u(t))| ∈ L1 [0, T ].

(ii) For any bounded set B ⊂ W and each u(·) ∈ U, there exist scalar functions
β(·) ∈ Lp [0, T ] and γ(·) ∈ L1 [0, T ], both may depend on B and u, such
that

(1.13) kfx (t, x(t), u(t))kL(W,X) 6 β(t),

and

(1.14) kQx (t, x(t), u(t))kL(W,R)=W ′ 6 γ(t).

for all t ∈ [0, T ], provided that x(t) ∈ B, 0 6 t 6 T .

302 You

Hypothesis III. Let the exponent α in Hypothesis I and the exponent p in

Hypothesis II be fixed and satisfy the following condition,
p−1
(1.15) either p > 1 and 0 6 α < ,
p
(1.16) or p = 1 and α = 0.

From now on, W = W 2α is accordingly fixed and its norm is denoted by k · k.

From these hypotheses, any constant control u(t) ≡ u0 ∈ U belongs to Uad .
For any x(·) ∈ C([0, T ], W ) and u(·) ∈ U, the integral
Zt
Mα
|f (s, x(s), u(s))| ds
(t − s)α
0

is convergent for any t ∈ [0, T ].

We have more notational remarks. For simplicity, we shall write the norm
k · kC([0,T ],W ) as k · kC . For any Banach spaces E1 and E2 , the Banach space
L (E1 , E2 ) consists of all bounded linear operators from E1 into E2 with the
operator norm. It is denoted by L (E1 ) when E1 = E2 . For any Banach space
E, its dual space is written as E ′ . The dual product of the functional action
between a ϕ ∈ E ′ and a e ∈ E is denoted by hϕ, ei. For any densely defined,
closed linear operator K, we use K † to denote its adjoint operator. Since
X is a
reflexive Banach space, by [15, Corollary 10.6], we know that A† : D A† → X ′
†
also generates a C0 -semigroup eA t , t > 0.
Let us consider a linearized evolutionary equation,
dw
(1.17) = Aw + fx (t, x̂(t), û(t)) w,
dt
where û ∈ Uad is given and x̂(t) = x (t, x0 , û). Let F (t) = fx (t, x̂(t), û(t)),
then by Hypothesis II one has F ∈ Lp ([0, T ], L(W, X)). By the perturbation
theory of semigroups, cf. [16], there exists a family of bounded linear operators
G(t, s) defined on ΩT = {(t, s) : 0 6 s 6 t 6 T } and valued in L(X) ∩ L(W ),
which is strongly continuous in (t, s), uniformly bounded in L(X) and L(W ),
and satisfies the following:

G(t, t) = IX , 0 6 t 6 T;
G(t, s) = G(t, τ )G(τ, s), 0 6 s 6 τ 6 t 6 T;

and for any w0 ∈ X and s ∈ [0, T ],

Zt
A(t−s)
G(t, s)w0 = e w0 + eA(t−σ)
(1.18)
s
· fx (σ, x̂(σ), û(σ))G(σ, s)w0 dσ, t ∈ [s, T ].
 Optimal Control and Synthesis of Nonlinear Systems 303

The rest of the paper is outlined as follows. The first part is devoted to
proving the maximum principle via a straightforward approach featuring the
oscillating variations of control. This part includes Sections 2, 3 and 4. Then
the second part is a contribution of a general synthesis result of optimal control
processes, based on the proof of Lipschitz continuity of the value functions and
utilizing a technique of differential inclusions. That part consists of Sections 5
and 6.

2 Oscillating Variations of Control

There have been many results on the maximum principle for infinite dimensional
optimal controls in various cases and shown by various methods. Notably, [3, 14]
and [17] make use of vector measures and the Eidelheit separation theorem; [6]
and [9] use Ekeland’s variational principle and directional derivatives; [7] takes
the approach of nonlinear programming by using the Kuhn-Tucker technique;
[8] exploits the relaxation of control and trajectories; [11, 12, 13] cover the area
of quadratic optimal control; [19] and [20] directly treat spike perturbations
and provide results on nonlinear and nonquadratic applications.
Here we shall present a forthright proof of the maximum principle, a
necessary condition satisfied by optimal control if it exists, for the described
(OCP) by the approach of oscillating variations of control. Since this approach
can reach the goal based on essentially two tool lemmas, the perturbation and
robustness lemma for Volterra integral equations and the generalized Riemann-
Lebesgue lemma, it seems mathematically simpler in comparison with some
other approaches.
In the first part, that consists of Sections 2 through 4, all the lemmas and
theorems are valid under Hypotheses I, II and III. Sometimes these hypotheses
will not be mentioned repeatedly.
We start with a given admissible control u∗ (·) ∈ Uad , that may not be an
optimal control at this moment. Let x∗ (t) = x (t, x0 , u∗ ), 0 6 t 6 T . For any
u ∈ U and any δ ∈ [0, 1], we define a variation of control uδn (t) as follows,
(
u(t), for t ∈ Enδ ,
(2.1) uδn (t) =
u∗ (t), for t ∈ [0, T ]\Enδ ,

where the sets Enδ are determined by

[
n−1
kT (k + δ)T

(2.2) Enδ = , , n = 1, 2, . . . .
n n
k=0

Note that the variation of control certainly depends on u ∈ U, which may not
be in Uad . This type of variation of control is called an oscillating variation.
304 You

In order to analyze the implications of the oscillating variations of control,

let us define
 
(2.3) Γδn = Γδn (t, s, x) = eA(t−s) f s, x, uδn (s) ,

where uδn is defined by (2.1). This Γδn is a mapping such that

Γδn : ΩT × W → X and Γδn : ΩT \{t = s} × W → W.

The two important properties of Γδn are stated in the following lemma.
Lemma 2.1. The family of mappings Γδn for n = 1, 2, . . . and δ ∈ [0, 1] has
the following properties:
(P1) Γδn is a strongly measurable function and it satisfies the local Lipschitz
condition in the following sense: for any x0 ∈ W , there exist a constant
η > 0 and a nonnegative scalar function ρ(·) ∈ L1 [0, T ] such that

δ
(2.4) Γn (t, s, x1 ) − Γδn (t, s, x2 ) 6 ρ(s) kx1 − x2 k ,

for (t, s) ∈ ΩT \{t = s}, for any x1 , x2 ∈ Nη (x0 ) =

{x ∈ W : kx − x0 k 6 η} and uniformly for all δ ∈ [0, 1] and all n > 1.

(P2) For any given x(·) ∈ C([0, T ], W ), one has Γδn (t, ·, x(·)) ∈ L1 ([0, t], W )
Rt
and Γδn (t, s, x(s)) ds is strongly continuous in t ∈ [0, T ]. And
0

Zt Zt
(2.5) lim Γδn (t, s, x(s)) ds = Γ0n (t, s, x(s)) ds,
δ→0+
0 0

in which the convergence is uniform with respect to 0 6 t 6 T and all

n > 1.
Proof. By the hypothesis, obviously Γδn (t, s, x) is strongly measurable. For
x1 , x2 in Nη (x0 ), we have

δ δ
Γ
n (t, s, x1 ) − Γ n (t, s, x )
2
    
 
6 eA(t−s) f s, x ,
1 nuδ
(s) − f s, x ,
2 nuδ
(s) 
L(X,W )
Mα
6 K1 (Nη (x0 )) kx1 − x2 k ,
(t − s)α
1 and (1.10) here for this f . By taking ρ(s) = M K (N ) (t−
because of f ∈ CLip α 1 η
s)−α if 0 6 s < t and = 0 elsewhere, we see that (2.4) is valid and (P1) is
satisfied. Note that the L1 norm of ρ is independent of t in [0, T ].
In order to verify (P2), we make the observations as follows:
 Optimal Control and Synthesis of Nonlinear Systems 305

i) Let 0 6 t 6 T be relatively fixed. By the strong continuity of f in u, we

have
 

Γδn (t, s, x(s)) = eA(t−s) f s, x(s), uδn (s) → eA(t−s) f s, x(s), u0n (s)

as δ → 0+ , for almost every s ∈ [0, t].

ii) Define a family of periodic functions θnδ (·) by
 h 
(k+δ)T
1 − δ, if s ∈ kT n , n ,
θnδ (s) = h 
(2.6) −δ, if s ∈ (k+δ)T , (k+1)T
∪ {T },
n n

for k = 0, 1, . . . , n − 1.
It is easy to see that the following identity holds,
 
Γδn (t, s, x) = eA(t−s) f s, x, uδn (s)
= eA(t−s) [δf (s, x, u(s)) + (1 − δ)f (s, x, u∗ (s))
(2.7) i
+θnδ (s) (f (s, x, u(s)) − f (s, x, u∗ (s))) ,
s ∈ [0, t], x ∈ W.
By Hypothesis I and (1.11), it follows from (2.7) that
(2.8)

δ
Γn (t, s, x(s)) 6 eA(t−s) (|f (s, x(s), u(s))| + |f (s, x(s), u∗ (s))|)
L(X,W )
Mα
6 (|f (s, x(s), u(s))| + |f (s, x(s), u∗ (s))|)
(t − s)α
, b (s; x, u, u∗ ) , for 0 6 s < t.

Extend b(·) by letting b (s; x, u, u∗ ) = 0 for t 6 s 6 T . Then b (·; x, u, u∗ ) ∈

L1 [0, T ] and kb(·)kL1 [0,T ] is independent of δ ∈ [0, 1], n > 1 and t ∈ [0, T ].
The facts shown in i) and ii) above allow us to apply the Lebesgue
Dominated Convergence Theorem to get the uniform convergence relation (2.5)
with respect to 0 6 t 6 T and n > 1. The rest is easy to check. Therefore,
(P2) is satisfied. The proof is completed.
Lemma 2.2. Let uδn and Γδn be defined as above. Then there exists a positive
δ0 , 0 < δ0 6 1, such that for every δ ∈ [0, δ0 ], one has
(2.9) uδn (·) ∈ Uad ,


and the corresponding mild solution xδn (t) = x t, x0 , uδn exists on [0, T ] and
satisfies


(2.10) sup xδn (·) − x∗ (·) → 0, as δ → 0+ .
>
n 1 C
306 You

Proof. We shall use the fixed point theorem to show that
for δ > 0
sufficiently small, there exists a unique global solution x t, x0 , uδn of Eq. (1.3)
on [0, T ].
By (P1) and the compactness argument, it follows that for any given
x̂(·) ∈ C([0, T ], W ) and its closed neighborhood

(2.11) Nλ (x̂(·)) = {g ∈ C([0, T ], W ) | kg − x̂kC 6 λ} , λ > 0,

there exists a nonnegative scalar function µ(·) ∈ L1 [0, T ] depending on x̂ and λ

only, such that

δ
Γn (t, s, x1 (s)) − Γδn (t, s, x2 (s))
(2.12)
6 µ(s) kx1 (s) − x2 (s)k , a.e. on [0, t],

for any x1 (·), x2 (·) ∈ Nλ (x̂(·)) and all δ ∈ [0, 1] and n > 1.
Now let x∗ (·) = x (·, x0 , u∗ ) take the place of x̂(·) in (2.11) and (2.12).
Then µ ∈ L1 [0, T ] depending on x∗ and λ is determined. Define mappings
Tnδ : Nλ (x∗ (·)) → C([0, T ], W ) by

  Zt
(2.13) δ At
Tn x (t) = e x0 + Γδn (t, s, x(s)) ds, t ∈ [0, T ],
0

where Γδn is defined by (2.3), for δ ∈ [0, 1] and integers n > 1. By definition, we
have

(2.14) x∗ (·) = Tn0 x∗ , for n = 1, 2, . . . .

For any x(·) ∈ Nλ (x∗ (·)), by (2.12) we have

   
δ
Tn x (t) − Tnδ x∗ (t)
Zt
6 µ(s) kx(s) − x∗ (s)k ds
0
Zt Rs Rs
2 µ(τ )dτ −2 µ(τ )dτ
= µ(s)e 0 e 0 kx(s) − x∗ (s)k ds
0
 
Zt Rs  Rs 
2 µ(τ )dτ −2 µ(τ )dτ
(2.15) 6 µ(s)e 0 max e 0 kx(s) − x∗ (s)k ds
s∈[0,T ]  
0
s=t  
Rs   −2 Rs µ(τ )dτ 
1 2 0 µ(τ )dτ 
= e  max e 0 kx(s) − x∗ (s)k
2  s∈[0,T ]  
s=0
 Optimal Control and Synthesis of Nonlinear Systems 307
 
Rt  Rs 
1 02 µ(τ )dτ −2 µ(τ )dτ
(2.16) 6 e max e 0 kx(s) − x∗ (s)k .
2 s∈[0,T ]  

We can define k · kC,µ by

 
 Rt 
−2 µ(τ )dτ
(2.17) kgkC,µ = max e 0 kg(t)k , for g ∈ C([0, T ], W ).
t∈[0,T ]  

Note that k · kC,µ can be called a Bielecki norm and it is equivalent to k · kC on

the space C([0, T ], W ) because
RT Rt
−2 µ(τ )dτ −2 µ(τ )dτ
0 < c0 (T ) = e 0 6e 0 6 1.

Let Nλ,µ (x∗ (·)) be the closed λ-neighborhood of x∗ (·) in C([0, T ], W ) with
respect to the new norm. Then Nλ,µ (x∗ (·)) ⊂ Nλ (x∗ (·)) for any λ > 0. From
(2.15), we obtain
1 λ
δ
(2.18) Tn x − Tnδ x∗ 6 kx − x∗ kC,µ 6 ,
C,µ 2 2
for any x(·) ∈ Nλ,µ (x∗ (·)).
By the property (P2) we have shown in Lemma 2.1, there exists a constant
δ0 , 0 < δ0 6 1, such that
λ
δ ∗
(2.19) Tn x − x∗ = Tnδ x∗ − Tn0 x∗ < ,
C,µ C,µ 2
for all δ ∈ [0, δ0 ] and n > 1. Combining (2.18) and (2.19), we can assert that

(2.20) x(·) ∈ Nλ,µ (x∗ (·)) implies Tnδ x ∈ Nλ,µ (x∗ (·)) .

for any δ ∈ [0, δ0 ] and any integer n > 1. Therefore, Tnδ maps Nλ,µ (x∗ (·)) into
itself.
Similar to (2.15), one can prove that for any x1 (·), x2 (·) ∈ Nλ,µ (x∗ (·)), the
following inequality holds,
1
δ
(2.21) Tn x1 − Tnδ x2 6 kx1 − x2 kC,µ .
C,µ 2

Hence the mapping Tnδ : Nλ,µ (x∗ (·)) → Nλ,µ (x∗ (·)) is a contraction. According
to the fixed point theorem, there exists a unique fixed point of Tnδ in Nλ,µ (x∗ (·)).
By the definitions (2.13) and (2.3), this fixed point must be
 
(2.22) xδn (t) = x t, x0 , uδn , t ∈ [0, T ].
308 You

Therefore, we have proved that for δ ∈ [0, δ0 ], (2.9) is true.

Finally, let us prove (2.10) by estimating

δ
(2.23) xn − x∗ 6 Tnδ xδn − Tnδ x∗ + Tnδ x∗ − Tn0 x∗
C,µ C,µ C

By (P2) and (2.5), we have

δ ∗ 0 ∗
Tn x − T n x → 0, as δ → 0+ ,
C

where the convergence is uniform with respect to n > 1. Thus, for any ε > 0,
there is a δ1 = δ1 (ε) ∈ (0, 1] such that for any 0 6 δ 6 δ1 , one has

δ ∗ 0 ∗ δ ∗ ∗
(2.24) Tn x − Tn x = Tn x − x < ε/2.
C C

Then, using (2.21), we obtain

1
δ δ δ
(2.25) Tn xn − Tnδ x∗ 6 xn − x∗ ,
C,µ 2 C,µ

for any δ ∈ [0, min {δ0 , δ1 }], and uniformly with respect to n > 1. Substituting
(2.24) and (2.25) into (2.23), we find that
1 ε
δ δ
xn − x∗ < xn − x∗ + ,
C,µ 2 C,µ 2
which implies that for all n > 1,

δ
(2.26) xn − x∗ < ε, for 0 6 δ 6 min {δ0 , δ1 } .
C,µ

Since k · kC,µ and k · kC are equivalent, (2.26) leads to the conclusion (2.10).
The proof is completed.
The argument we made in Lemma 2.1 and Lemma 2.2 can be generalized
to prove a result on perturbation and robustness of nonlinear Volterra integral
equations (NVIE). We state this result as the following lemma. Its proof is
omitted to avoid any substantial duplication.
Lemma 2.3 (Perturbation and Robustness of NVIE). Consider a
nonlinear Volterra integral equation in a Banach space E, with parameters
δ ∈ [0, δ1 ] and n = 1, 2, . . . ,

Zt
δ
(2.27) x(t) = h (t) + Hnδ (t, s, x(s)) ds, t ∈ [0, T ].
0

Suppose the following conditions are satisfied:

 Optimal Control and Synthesis of Nonlinear Systems 309

(P0) For each δ ∈ [0, δ1 ], hδ (·) ∈ C([0, T ], E), and hδ − h0 C([0,T ],E) → 0, as
δ → 0+ .
(P1) Hnδ (t, s, x) : ΩT \{t = s} × E → E satisfies the corresponding property
(P1) in Lemma 2.1 with Γδn , W replaced by Hnδ , E.
(P2) Hnδ (t, s, x) also satisfies the corresponding property (P2) in Lemma 2.1
with Γδn , W replaced by Hnδ , E.

(P3) For δ = 0, there exists a common solution x∗ (·) ∈ C([0, T ], E) of Eq.

(2.27) for all n > 1.
Then there exists a positive δ0 , 0 < δ0 6 δ1 , such that for any δ ∈ [0, δ0 ] and for
any n > 1, there exists a unique solution xδn ∈ C([0, T ], E) of Eq. (2.27) and


(2.28) sup xδn (·) − x∗ → 0, as δ → 0+ .
>
n 1 C([0,T ],E)

This lemma will be used in the next section.

3 First Order Variations of Trajectory

Consider the mild solution of an initial value problem of the following variational
equation
dw
= Aw + fx (t, x∗ (t), u∗ (t)) w, t ∈ [τ, T ],
(3.1) dt
w(τ ) = w0 ∈ W,

where u∗ (·) ∈ Uad is given and x∗ (·) = x (·, x0 , u∗ ). Since Eq. (3.1) is exactly
Eq. (1.17) with {x̂, û} replaced by {x∗ , u∗ }, we shall use the same notation
G(t, s) associated with (1.17) to denote the corresponding evolution operators
associated with (3.1). Specifically, by the uniqueness of mild solution w(·) of
the initial value problem, it can be expressed by

(3.2) w(t) = G(t, τ )w0 , t ∈ [τ, T ],

where G(t, s) : ΩT → L(X) ∩ L(W ) is strongly continuous, uniformly bounded,

and satisfies the following equations in the strong sense, for (t, s) ∈ ΩT ,
Zt
(3.3) A(t−s)
G(t, s) = e + eA(t−σ) fx (σ, x∗ (σ), u∗ (σ)) G(σ, s) dσ,
s

and
Zt
(3.4) G(t, s) = e A(t−s)
+ G(t, σ)fx (σ, x∗ (σ), u∗ (σ)) eA(σ−s) dσ.
s
310 You

Notice that (3.3) and (3.4) are valid both in X and W . Consequently, we get
the commutative formula of the variational perturbation, i.e.,
Zt
(3.5) eA(t−σ) fx (σ, x∗ (σ), u∗ (σ)) G(σ, s) ds
s
Zt
= G(t, σ)fx (σ, x∗ (σ), u∗ (σ)) eA(σ−s) dσ, (t, s) ∈ ΩT .
s
Let us define a function q (t, u, u∗ ) by
q(t) = q (t, u, u∗ )
Zt
(3.6)
= G(t, τ ) [f (τ, x∗ (τ ), u(τ )) − f (τ, x∗ (τ ), u∗ (τ ))] dτ,
0
where u(·) ∈ U and u∗ (·) ∈ Uad are arbitrarily given. Here, x∗ (t) = x (t, x0 , u∗ ),
but it is not required in the proof of the next lemma.
Lemma 3.1. For any fixed u∗ ∈ Uad and any u ∈ U, the function q(t) is
a unique, strongly continuous solution of the following linear Volterra integral
equation,
Zt
y(t) = eA(t−s) fx (s, x∗ (s), u∗ (s)) y(s) ds
0
(3.7)
Zt
+ eA(t−s) [f (s, x∗ (s), u(s)) − f (s, x∗ (s), u∗ (s))] ds, t ∈ [0, T ].
0
Proof. Since the uniqueness is relatively easy to show, it suffices to prove
that q(·) defined by (3.6) satisfies Eq. (3.7). The verification is as follows,
Zt
eA(t−s) fx (s, x∗ (s), u∗ (s)) q(s) ds
0
ZtZs
= eA(t−s) fx (s, x∗ (s), u∗ (s)) G(s, τ )
0 0
· [f (τ, x∗ (τ ), u(τ )) − f (τ, x∗ (τ ), u∗ (τ ))] dτ ds
(by the Fubini-Tonelli theorem and Hypotheses to interchange the order of
integration)
ZtZt
= eA(t−s) fx (s, x∗ (s), u∗ (s)) G(s, τ ) ds
0 τ
 Optimal Control and Synthesis of Nonlinear Systems 311

· [f (τ, x∗ (τ ), u(τ )) − f (τ, x∗ (τ ), u∗ (τ ))] dτ

Zt h i
= G(t, τ ) − eA(t−τ ) [f (τ, x∗ (τ ), u(τ )) − f (τ, x∗ (τ ), u∗ (τ ))] dτ
0
Zt
= q(t) − eA(t−s) [f (s, x∗ (s), u(s)) − f (s, x∗ (s), u∗ (s))] ds,
0

where (3.3) is used. The proof is completed.

Before investigating the first order variations of trajectory, we now present
another key lemma in this oscillating variation approach, which can be called
a generalized Riemann-Lebesgue lemma. Let θnδ be given by (2.6). Let uδn and
xδn be the same as defined in (2.1) and Lemma 2.2.
Lemma 3.2 (Generalized Riemann-Lebesgue Lemma). Let h(·) ∈
L1 ([0, T ], E) with E being any real Banach space and let θnδ be defined by (2.6).
Then,

ZT
(3.8) lim h(t)θnδ (t) dt = 0,
n→∞
0

for each δ ∈ [0, 1].

Proof. By definition, we have
 t 
Z 
 
(3.9)  θnδ (s) ds 6 δ(1 − δ)T 6 T → 0, as n → ∞,
  n n
 
0

uniformly with respect to t ∈ [0, T ] and δ ∈ [0, 1]. Then for any subinterval
(a, b) ⊂ [0, T ] and any h0 ∈ E, let h(t) = χ(a,b) (t)h0 where χ∆ is the
characteristic function of subset ∆. Then by (3.9), we have
 b 
ZT Z
χ(a,b) (t)h0 θnδ (t) dt =  θnδ (t) dt h0 → 0, as n → ∞.
0 a

It follows that for any finitely many disjoint open intervals (ai , bi ) ⊂ [0, T ],
i = 1, . . . , m, and for any hi ∈ E, i = 1, . . . , m, we have

ZT X
m
χ(ai ,bi ) hi θnδ dt → 0, as n → ∞.
0 i=1
312 You
∞
P
Then for any countably valued function h ∈ L1 ([0, T ]; E), h(t) = χ∆i (t)hi
i=1
with hi ∈ E and ∆i ’s being mutually disjoint measurable sets in [0, T ], we have
ZT
(3.10) h(t)θnδ (t) dt → 0, as n → ∞.
0

Finally, for any general h ∈ L1 ([0, T ], E), by definition of the Bochner integral,
there is a sequence of countably valued functions hj ∈ L1 ([0, T ], E) such that

ZT
(3.11) kh(t) − hj (t)kE dt → 0, as j → ∞.
0

We can write
T T
Z ZT Z

h(t)θ δ (t) dt 6 kh(t) − hj (t)k dt + hj (t)θ δ (t) dt .
n E n

0 E 0 0 E

By (3.11) and (3.10), for any ε > 0, there is an integer j0 such that
RT
kh(t) − hj0 (t)kE dt < ε/2 and then for that j0 there is a positive integer
0
RT

n0 = n0 (j0 , ε) such that for n > n0 we have hj0 (t)θnδ (t) dt < ε/2. It
0
E
follows that T
Z

h(t)θnδ (t) dt < ε, for any n > n0 .


0 E
Thus (3.8) is proved. The proof is completed.
Let δ0 > 0 be the constant stated in Lemma 2.2 and fixed.
Lemma 3.3. For every δ ∈ [0, δ0 ], there is a positive integer n(δ) such that
1h δ i
(3.12) lim xn(δ) (t) − x∗ (t) = q(t),
δ→0+ δ

where the convergence is uniform in t ∈ [0, T ] and q(t) is given by (3.6).

Moreover,
ZT
1
(3.13) lim [f (t, x∗ (t), u(t)) − f (t, x∗ (t), u∗ (t))] θn(δ)
δ
(t) dt = 0,
δ→0 δ
+
0

where u ∈ U and u∗ ∈ Uad are the same as in Lemma 2.1, and x∗ (t) =
x (t, x0 , u∗ ).
 Optimal Control and Synthesis of Nonlinear Systems 313

Proof. By Hypothesis II, we know that

f (t, x∗ (t), u(t)) − f (t, x∗ (t), u∗ (t)) ∈ L1 ([0, T ], X).

According to Lemma 3.2, it follows that for each δ ∈ [0, 1],

ZT
(3.14) lim [f (t, x∗ (t), u(t)) − f (t, x∗ (t), u∗ (t))] θnδ (t) dt = 0.
n→∞
0

Hence for each 0 < δ 6 δ0 , there exists a positive integer n1 (δ) such that

(3.15)
 T 
Z 
 
 [f (t, x∗ (t), u(t)) − f (t, x∗ (t), u∗ (t))] tδn (t) dt < δ2 , for n > n1 (δ).
 
 
0

Now consider
1h δ i
xn (t) − x∗ (t)
δ
Zt h    i
1
= eA(t−s) f s, xδn (s), uδn (s) − f s, x∗ (s), uδn (s) ds
δ
0
Zt h   i
1
(3.16) + eA(t−s) f s, x∗ (s), uδn (s) − f (s, x∗ (s), u∗ (s)) ds,
δ
0

where the two integral terms on the right side of (3.16) are denoted by I1 (n)
and I2 (n), respectively. Then we have the expression

Zt  
1
(3.17) I2 (n) = F (t, s) δ + θnδ (s) ds,
δ
0

where

(3.18) F (t, s) = eA(t−s) [f (s, x∗ (s), u(s)) − f (s, x∗ (s), u∗ (s))] .

By Lemma 3.2, for every 0 < δ 6 δ0 , it holds that

Zt
(3.19) lim F (t, s) θnδ (s) ds = 0
n→∞
0
314 You

uniformly in t ∈ [0, T ]. Hence for every such δ, there exists a positive integer
n2 (δ) such that
t
Z

(3.20) sup δ 2
F (t, s) θn (s) ds < δ , for n > n2 (δ).
t∈[0,T ]
0

Let n(δ) = max {n1 (δ), n2 (δ)} and denote uδn(δ) = uδ , xδn(δ) = xδ . Define

1h δ i
(3.21) ∆x(t, δ) = x (t) − x∗ (t) , for 0 < δ 6 δ0 ,
δ
and

(3.22) xδ (t, λ) = λxδ (t) + (1 − λ)x∗ (t), for 0 6 λ 6 1.

Note that Lemma 2.2 and (2.10) imply that

(3.23) sup xδ (·, λ) − x∗ (·) → 0, as δ → 0+ .
66
0 λ 1 C

Hence there is δ2 ∈ (0, δ0 ] such that xδ (·, λ) ∈ N1 (x∗ (·)) for 0 6 δ 6 δ2 .

Therefore,
n o

sup xδ (s, λ) : 0 6 s 6 T, 0 6 λ 6 1, 0 6 δ 6 δ2
(3.24)
6 kx∗ (·)kC + 1 (a constant depending on x∗ only).

On the other hand, we have

Zt Z1  
A(t−s)
(3.25) I1 (n(δ)) = e fx s, xδ (s, λ), uδ (s) ∆x(s, δ) dλ ds.
0 0

Assembling together (3.16) and (3.17) with n = n(δ), (3.20) and (3.25), we get
 
Zt Z1  
∆x(t, δ) = eA(t−s) fx s, xδ (s, λ), uδ (s) dλ ∆x(s, δ) ds
0 0
(3.26)
Zt
+ F (t, s) ds + R(δ),
0

where R(δ) is a term satisfying kR(δ)k < δ, so that R(δ) → 0 uniformly.

 Optimal Control and Synthesis of Nonlinear Systems 315

Here we have a Volterra integral equation with a parameter δ ∈ [0, δ2 ], with

unknown y, that is
Zt
δ
(3.27) y(t) = h (t) + H δ (t, s, y(s)) ds, t ∈ [0, T ],
0

where
Zt
δ
h (t) = F (t, s) ds + R(δ),
0
 
Z1  
H δ (t, s, x) = eA(t−s) fx s, xδ (s, λ), uδ (s) dλ x.
0

Indeed, Eq. (3.27) is a particular case of Eq. (2.27), where δ ∈ [0, δ1 ] is replaced
by δ ∈ [0, δ2 ] and no parameter n. We can check that, in this case, all four
conditions in Lemma 2.3 are satisfied.
(P0) is satisfied because hδ − h0 C = kR(δ)kC 6 δ.
(P1) can be verified as follows. Since H δ (t, s,
x) is linear in x, and using an
identity similar to (2.7) for fx s, xδ (s, λ), uδ (s) , one has


(3.28) H δ (t, s, x1 ) − H δ (t, s, x2 )
Z1    
Mα δ δ ∗
6 f
x s, x (s, λ), u(s) + f
x (s, x (s, λ), u (s) dλ kx1 − x2 k
(t − s)α
0
Z1
Mα
6 (β(s, u) + β (s, u∗ )) dλ kx1 − x2 k = ρ(s) kx1 − x2 k
(t − s)α
0

where β(·, u) and β (·, u∗ ) come from (1.13) and depend on the bounded set B
which is the closed ball in W of radius kx∗ kC + 1 from (3.24), and
(
Mα (t − s)−α [β(s, u) + β (s, u∗ )] , for 0 6 s < t,
(3.29) ρ(s) =
0, for t 6 s 6 T.

(P2) can be verified as follows. First, by (3.23), (1.13) and β(·, u), β (·, u∗ )
as in (3.28), one can use the Lebesgue Dominated Convergence Theorem to get

Z1  
(3.30) lim fx s, xδ (s, λ), uδ (s) dλ = fx (s, x∗ (s), u∗ (s)) ,
δ→0+
0
316 You

uniformly in s ∈ [0, T ]. Eq. (3.30) shows that for each s ∈ [0, T ],

(3.31) H δ (t, s, x(s)) → H 0 (t, s, x(s)), as δ → 0+ .

Besides, one has H δ (t, s, x(s) 6 ρ(s)kx(·)kC ∈ L1 [0, T ] and the L1 norm of
ρ is independent of t ∈ [0, T ], where ρ(·) is the same function given in (3.29).
Thus, one can use the Lebesgue Dominated Convergence Theorem again to
obtain the uniform convergence relation (2.5) for H δ .
(P3) is also satisfied. For δ = 0, Eq. (3.27) becomes exactly Eq. (3.7) and
that equation has a strongly continuous solution q(t) given by (3.6), as shown
by Lemma 3.1.
Therefore, we can apply Lemma 2.3 to conclude that there is a δ3 ∈ (0, δ2 ]
such that for any 0 < δ 6 δ3 , there exists a unique solution of Eq. (3.27), which
is exactly ∆x(·, δ) ∈ C([0, T ], W ), such that
h i
1 δ ∗

(3.32) k∆x(·, δ) − q(·)kC = x (·) − x (·) − q(·)
+
→ 0, as δ → 0 .
δ C

Thus (3.12) is proved. Finally, (3.13) follows from (3.15). The proof is
completed.
Lemma 3.3 yields an expression of perturbation,
(3.33) xδ (t) = x∗ (t) + δq(t) + o(δ),
that indicates q(t) is the first order variation of the trajectory. We emphasize
that only two essential tool lemmas are used so far: the lemma of perturbation
and robustness of NVIE (Lemma 2.3) and the generalized Riemann-Lebesgue
lemma (Lemma 3.2).

4 Maximum Principle for Optimal Control

The maximum principle is a necessary condition satisfied by an optimal control,
if it exists. The usual description of the maximum principle features a local
condition that is satisfied pointwise by an optimal control function u∗ (t) except
possibly for a set of Lebesgue measure zero. A bridge toward the pointwise
result is a global maximality result, or called a variational inequality, as stated
in the following theorem.
Theorem 4.1. Under Hypotheses I, II and III, if there exists an optimal
control u∗ (·) ∈ Uad and let x∗ (·) = x (·, x0 , u∗ ), then the following inequality is
satisfied,
(4.1)
ZT ZT
Qx (t, x∗ (t), u∗ (t)) q(t) dt > [Q (t, x∗ (t), u∗ (t)) − Q (t, x∗ (t), u(t))] dt,
0 0

where q(·) is defined by (3.6).

 Optimal Control and Synthesis of Nonlinear Systems 317

Proof. By the optimality of the control process {u∗ (·), x∗ (·)}, one has
1h   i
(4.2) 06 J x0 , uδ − J (x0 , u∗ ) = J1 + J2 ,
δ
where uδ and xδ are the same as described in the proof of Lemma 3.3, just
before (3.21), and
ZT h    i
1
(4.3) J1 = Q t, xδ (t), uδ (t) − Q t, x∗ (t), uδ (t) dt,
δ
0
ZT
1 h  ∗  i
(4.4) J2 = Q t, x (t), uδ (t) − Q (t, x∗ (t), u∗ (t)) dt.
δ
0

For J1 part, we have

 
 ZT 
 

(4.5) J1 − Qx (t, x (t), u (t)) q(t) dt
∗ ∗
 
0
Z  h   i 
1 ∗ ∗ ∗

6  δ 
 δ Q t, x (t), u(t) − Q (t, x (t), u(t)) − Qx (t, x (t), u (t)) q(t) dt
δ (δ)
En
Z  h   i
1
+  Q t, xδ (t), u∗ (t) − Q (t, x∗ (t), u∗ (t))
δ
δ (δ)
[0,T ]\En

∗ ∗

−Qx (t, x (t), u (t)) q(t) dt.

1 and by (3.12), we have

Since Q ∈ CLip
 h   i
1
lim  Q t, xδ (t), u∗ (t) − Q (t, x∗ (t), u∗ (t))
δ→0+ δ
(4.6) 

−Qx (t, x (t), u (t)) q(t) = 0.
∗ ∗

On the other hand, by utilizing (3.24), (1.14) and (3.12) we can get a constant
δ4 , 0 < δ4 6 δ3 (6 δ2 6 δ0 6 1), such that
 h   i
1 ∗ ∗ ∗
 Q t, x (t), u (t) − Q (t, x (t), u (t)) 
δ
δ 
 1 
Z   δ ∗

(4.7)  ∗ x (t) − x (t) 
 δ
=  Qx t, x (t, λ), u (t) dλ ·
δ 
 
0
6 g1 (t) (kq(·)kC + 1) , t ∈ [0, T ], for any δ ∈ (0, δ4 ] ,
318 You

where g1 (·) ∈ L1 [0, T ] depends on N1 (x∗ (·)), cf. (3.24). Based on (4.6) and
(4.7), we can apply the Lebesgue Dominated Convergence Theorem to obtain
Z  h   i
1
lim  Q t, xδ (t), u∗ (t) − Q (t, x∗ (t), u∗ (t))
δ→0+
δ
δ
[0,T ]\En(δ)
(4.8)

∗ ∗

−Qx (t, x (t), u (t)) q(t) dt = 0

Moreover, similarly we can get a δ5 , 0 < δ5 6 δ3 , such that

 h   i 
1 
 Q t, xδ (t), u(t) − Q (t, x∗ (t), u(t)) − Qx (t, x∗ (t), u∗ (t)) q(t)
(4.9)  δ 
6 g2 (t), t ∈ [0, T ], for any δ ∈ (0, δ5 ] ,

where g2 (·) ∈ L1 [0, T ] depends on N1 (x∗ (·)) and u as well. Equation (4.9) and
the fact that
δ
meas En(δ) = δT → 0, as δ → 0+ ,
imply the first integral term on the right side of inequality (4.5) also converges
to zero as δ → 0+ . Therefore, we have proved that
ZT
(4.10) lim J1 = Qx (t, x∗ (t), u∗ (t)) q(t) dt.
δ→0+
0

Next we treat the I2 part. In fact, we have

Z
1
J2 = [Q (t, x∗ (t), u(t)) − Q (t, x∗ (t), u∗ (t))] dt
δ
δ
En(δ)

ZT
= [Q (t, x∗ (t), u(t)) − Q (t, x∗ (t), u∗ (t))] dt
0
(4.11)
ZT
1
+ [Q (t, x∗ (t), u(t)) − Q (t, x∗ (t), u∗ (t))] θn(δ)
δ
dt
δ
0
ZT
→ [Q (t, x∗ (t), u(t)) − Q (t, x∗ (t), u∗ (t))] dt text, asδ → 0+ ,
0

because
ZT
1
(4.12) lim [Q (t, x∗ (t), u(t)) − Q (t, x∗ (t), u∗ (t))] θ δ (t) dt = 0,
δ→0+ δ
0
 Optimal Control and Synthesis of Nonlinear Systems 319

that can be shown similarly as we did for (3.13).

Finally, we substitute (4.10) and (4.11) into (4.2) and then obtain

ZT ZT
∗ ∗
06 Qx (t, x (t), u (t)) q(t) dt − [Q (t, x∗ (t), u∗ (t)) − Q (t, x∗ (t), u(t))] dt,
0 0

that immediately leads to (4.1). The proof is completed.

Now we are going to present the main results in the first part, the maximum
principle. Associated with the (OCP), we define a functional H(t, x, ψ, u) to be

(4.13) H(t, x, ψ, u) = hψ, f (t, x, u)i − Q(t, x, u),

where (t, x, ψ, u) ∈ [0, T ] × W × W ′ × Y and h·, ·i stands for the dual product
of W ′ versus W .
Theorem 4.2 (Maximum Principle). Let Hypotheses I, II and III be
satisfied. If u∗ (·) ∈ Uad is an optimal control of the (OCP ) and x∗ (·) =
x (·, x0 , u∗ ), then it holds that

H (t, x∗ (t), ψ(t), u∗ (t)) = max H (t, x∗ (t), ψ(t), v) ,

(4.14) v∈U
for a.e. t ∈ [0, T ],

where the dual function ψ ∈ C ([0, T ], W ′ ) is given by

ZT
(4.15) ψ(t) = − Qx (s, x∗ (s), u∗ (s)) G(s, t) ds, t ∈ [0, T ],
t

in which G(·, ·) is defined by (3.2) and satisfies (3.3) and (3.4).

320 You

Proof. We can substitute (3.6) for q(t) in (4.1) and then interchange the
order of integration by the Fubini-Tonelli theorem. It yields

(4.16)
ZT Zt
∗ ∗
Qx (t, x (t), u (t)) G(t, τ ) · [f (τ, u∗ (τ ), u(τ )) − f (τ, x∗ (τ ), u∗ (τ ))] dτ dt
0 0
 T 
ZT Z
=  Qx (t, x∗ (t), u∗ (t)) G(t, τ )dt
0 τ
· [f (τ, u (τ ), u(τ )) − f (τ, x∗ (τ ), u∗ (τ ))] dτ
∗

ZT
=− hψ(τ ), f (τ, u∗ (τ ), u(τ )) − f (τ, x∗ (τ ), u∗ (τ ))i dτ
0
ZT
> [Q (t, x∗ (t), u∗ (t)) − Q (t, x∗ (t), u(t))] dt.
0

It follows that for any u(·) ∈ U, the following maximality inequality holds,

ZT ZT
∗ ∗
(4.17) H (t, x (t), ψ(t), u (t)) dt > H (t, x∗ (t), ψ(t), u(t)) dt.
0 0

Recall that starting from Lemma 2.1 in Section 2 and Lemma 3.1 in Section 3,
we only require u(·) ∈ U, that may or may not be in Uad . Therefore, for any
v ∈ U and any nontrivial subinterval [a, b] ⊂ [0, T ], the control function u(·)
defined by
(
v, if t ∈ [a, b],
(4.18) u(t) = ∗
u (t), if t ∈ [0, T ]\[a, b],

is in U. Then (4.17) is valid for this particular control u(·), which implies that

Zb Zb
∗ ∗
(4.19) H (t, x (t), ψ(t), u (t)) dt > H (t, x∗ (t), ψ(t), v) dt.
a a

Since a and b, with 0 6 a < b 6 T , are arbitrary, by the Lebesgue differentiation

theorem, we can assert that
Zt+r
1
(4.20) lim H (s, x∗ (s), ψ(s), u∗ (s)) ds = H (t, x∗ (t), ψ(t), u∗ (t))
r→0 2r
+
t−r
 Optimal Control and Synthesis of Nonlinear Systems 321

for almost every t ∈ [0, T ] (called Lebesgue points), and

Zt+r
1
(4.21) lim H (s, x∗ (s), ψ(s), v) ds = H (t, x∗ (t), ψ(t), v)
r→0 2r
+
t−r

for any t ∈ [0, T ], by taking a = t − r and b = t + r. From (4.19), (4.20) and

(4.21) it follows that for any v ∈ U ,
(4.22) H (t, x∗ (t), ψ(t), u∗ (t)) > H (t, x∗ (t), ψ(t), v) , for a.e. t ∈ [0, T ].
Thus (4.14) is proved. The proof is completed.
Note that in the presentation of Theorem 4.2, the functional H(t, x, ψ, u)
defined by (4.13) is in terms of the dual product h·, ·i of W ′ versus W . If we
adopt the dual product h·, ·i of W versus W ′ , with this notation unchanged,
but the first component is in W and the second component is in W ′ , then the
same maximum principle takes the adjoint version, which is the next theorem.
Theorem 4.3. Let Hypotheses I, II and III be satisfied. If u∗ (·) ∈ Uad is
an optimal control of the (OCP ) and x∗ (·) = x (·, x0 , u∗ ), then it holds that
H (t, x∗ (t), ϕ(t), u∗ (t)) = max H (t, x∗ (t), ϕ(t), v) ,
(4.23) v∈U
for a.e. t ∈ [0, T ]
where
(4.24) H(t, x, ϕ, u) = hf (t, x, u), ϕi − Q(t, x, u),
with the dual product being W versus W ′ , and
ZT
(4.25) ϕ(t) = − G† (s, t)Q†x (s, x∗ (s), u∗ (s)) ds, t ∈ [0, T ],
t

that is the mild solution of the “backward” evolution equation in W ′ ,

dϕ
= −A† ϕ − fx† (t, x∗ (t), u∗ (t)) ϕ + Q†x (t, x∗ (t), u∗ (t)) ,
(4.26) dt
t ∈ [0, T ], ϕ(T ) = 0,
and G(·, ·) is the same as in Theorem 4.2.
Proof. As we mentioned in Section 1, the superscript † is used to denote
the adjoint operator. Eq. (4.26) is a linear evolution equation, whose backward
mild solution satisfying ϕ(T ) = 0 is given by
ZT
† (s−t)
ϕ(t) = eA fx† (s, x∗ (s), u∗ (s)) ϕ(s) ds
t
(4.27)
ZT
† (s−t)
− eA Q†x (s, x∗ (s), u∗ (s)) ds, t ∈ [0, T ].
t
322 You

From (4.15) and (4.25), we see that exactly ψ † (t) = ϕ(t) for t ∈ [0, T ]. Then
(4.23) follows from (4.14). It suffices to show that ϕ(t) given by (4.25) satisfies
the equation (4.27). In fact, by (3.4) we have
Zs
A† (s−t) † (σ−t)
(4.28) G† (s, t) = e + eA fx† (σ, x∗ (σ), u∗ (σ)) G† (s, σ) dσ.
t

Substituting (4.28) into (4.25), we find

ZT
ϕ(t) = − G† (s, t)Q†x (s, x∗ (s), u∗ (s)) ds
t
ZT
† (s−t)
= − eA Q†x (s, x∗ (s), u∗ (s)) ds
t
ZTZs
† (σ−t)
(4.29) − eA fx† (σ, x∗ (σ), u∗ (σ)) G† (s, σ) · Q†x (s, x∗ (s), u∗ (s)) dσ ds
t t
ZT
† (s−t)
= − eA Q†x (s, x∗ (s), u∗ (s)) ds
t
ZT ZT
†
− eA (σ−t) fx† (σ, x∗ (σ), u∗ (σ)) G† (s, σ) · Q†x (s, x∗ (s), u∗ (s)) ds dσ
t σ
ZT
† (s−t)
= − eA Q†x (s, x∗ (s), u∗ (s)) ds
t
ZT
† (σ−t)
+ eA fx† (σ, x∗ (σ), u∗ (σ)) ϕ(σ) dσ, t ∈ [0, T ].
t

Therefore, ϕ(·) satisfies (4.27). The proof is completed.

Inheriting the terminology of finite dimensional optimal control theory, one
can call the function ϕ(·) as the co-state function and call Eq. (4.26) as the
adjoint state equation with a terminal value condition.
Remark 4.1. The maximum principle is an implicit, nonlinear equality
relation satisfied by any optimal control, if it exists. The maximum principle is
a necessary condition only. In its descriptions, the dual function ψ in Theorem
4.2 and the co-state function ϕ in Theorem 4.3 depend on the optimal control
u∗ and the optimal state trajectory x∗ . Therefore, except for some simpler
cases, the maximum principle alone is not enough for solving an optimal control
problem.
 Optimal Control and Synthesis of Nonlinear Systems 323

5 Lipschitz Continuity of Value Functions

For infinite dimensional nonlinear optimal control, a weak solution or a viscosity
solution of the Hamilton-Jacobi equation associated with the value functions is
closely related to the synthesis of an optimal control under certain conditions,
cf. [1, 2], and other references on Hamilton-Jacobi equations. However, the
value functions may fail to be Gâteaux differentiable almost everywhere, and in
the current theory of viscosity solutions the uniqueness and regularity aspects
are better established than the existence aspect. The latter issue is usually
addressed case-by-case and usually requires sophisticated techniques, even in
finite dimensional cases. In short, the optimal feedback control for nonlinear
distributed systems remains a largely open issue.
In this section, we shall prove the Lipschitz continuity of the value functions
for the (OCP) described in Section 1 and the affiliated problems. Based on this
crucial property, the next section will provide a general result on the state
feedback implementation of an optimal control, assuming it exists.
First, let us imbed the (OCP) over the given time interval [0, T ] into a family
of optimal control problems over subintervals Iτ = [τ, T ] ⊂ [0, T ], described by

dx
= Ax + f (t, x, u), t ∈ [τ, T ],
(5.1) dt
x(τ ) = x0 ,

 
 ZT 
(5.2) minτ Jτ (x0 , u) = Q(t, x(t), u(t)) dt ,
u(·)∈Uad  
τ

with
τ
Uad = {u : [τ, T ] → U | u is strongly measurable and
(5.3)
x (t, τ, x0 , u) exists over [τ, T ]}

Here, the mild solution of the initial value problem (5.1) is denoted by x(t) =
x (t, τ, x0 , u). We shall refer to this optimal control problem described by (5.1),
(5.2) and (5.3) briefly as the (OCP)τ . Since we do not address the existence of
optimal control in this paper, we have to make certain assumptions.
Hypothesis IV. Assume that for each x0 ∈ W , there exists an optimal
control process denoted by {u∗τ , x∗τ }, where x∗τ (t) = x (t, τ, x0 , u∗τ ), of the
(OCP )τ , for τ ∈ [0, T ]. Moreover, for any given bounded set B ⊂ W , there
is a constant ρ = ρ(B) > 0 such that kx∗τ kC([τ,T ],W ) 6 ρ(B) for any x0 ∈ B.
Below, the norm of C([τ, T ], W ) for any τ ∈ [0, T ] will be denoted by k · kC
for simplicity. For convenience, the notation of initial state x0 will be replaced
by z in the sequel. We shall denote by Sr (E) the closed ball in a Banach space
E centered at the origin and of radius r.
324 You

Define the value function V (τ, z) to be

τ
(5.4) V (τ, z) = min {Jτ (z, u) | u ∈ Uad }.

We are going to investigate the properties of the value function.

Lemma 5.1. Let Hypotheses I–IV be satisfied. Then for any r > 0, there is
a constant L1 = L1 (r) > 0 such that

(5.5) |V (τ, z1 ) − V (τ, z2 )| 6 L1 (r) kz1 − z2 k ,

for any τ ∈ [0, T ] and any z1 , z2 in Sr (W ).

Proof. By Hypothesis IV, let {u1 , x1 } and {u2 , x2 } be optimal control
processes of the (OCP)τ with the initial state being z1 and z2 , respectively.
Here we drop the optimal indication, superscript ∗. The optimality condition
tells us

V (τ, z1 ) − V (τ, z2 ) 6 Jτ (z1 , u2 ) − Jτ (z2 , u2 ) ,

(5.6)
V (τ, z2 ) − V (τ, z1 ) 6 Jτ (z2 , u1 ) − Jτ (z1 , u1 )

Hence we have

|V (τ, z1 ) − V (τ, z2 )|
(5.7)
6 max {|Jτ (z1 , u2 ) − Jτ (z2 , u2 )| , |Jτ (z2 , u1 ) − Jτ (z1 , u1 )|} .

By Hypothesis IV, for any given r > 0, there exists a constant ρ(r) > 0 such
that

(5.8) kx (·, τ, z, u∗ )kC 6 ρ(r), for any z ∈ Sr (W ).

By Hypothesis II, using the mild solution formula and the Henry-Gronwall
formula, cf. [16], one can show that there exists a constant K(r) > 0 such that

kx (·, τ, z1 , u) − x (·, τ, z2 , u)kC 6 K(r) kz1 − z2 k ,

(5.9)
for any z1 , z2 ∈ Sr (W ),

where u = either u1 or u2 , the corresponding optimal controls mentioned above.

Then by Hypothesis II again, specifically Q ∈ CLip1 , (5.8) and (5.9) imply that

there exists a constant L1 = L1 (r) > 0 such that for any τ ∈ [0, T ],

|Jτ (z1 , u) − Jτ (z2 , u)| 6 L1 (r) kz1 − z2 k ,

(5.10)
for any z1 , z2 ∈ Sr (W ),

with u = either u1 or u2 as above. Finally, substituting (5.10) into (5.7), we

obtain (5.5). The proof is finished.
 Optimal Control and Synthesis of Nonlinear Systems 325

Lemma 5.2. Let Hypotheses I–IV be satisfied. Then for any r > 0, there is
a constant L2 = L2 (r) > 0 such that

(5.11) |V (τ1 , z) − V (τ2 , z)| 6 L2 (r) |τ1 − τ2 | ,

for any τ1 , τ2 ∈ [0, T ] and any z ∈ Sr W 2 = Sr (D(A)).
Proof. Let 0 6 τ1 < τ2 6 T and let u2 denote an optimal control
corresponding to the (OCP)τ2 over [τ2 , T ] with the initial state x (τ2 ) = z.
Define u ∈ Uadτ by


u0 , for t ∈ [τ1 , τ2 ) ,
(5.12) u(t) =
u2 (t) for t ∈ [τ2 , T ] ,

where u0 ∈ U is arbitrarily given. Then one has

(5.13) V (τ1 , z) − V (τ2 , z)

6 Jτ1 (z, u) − Jτ2 (z, u2 ) 6 |Jτ1 (z, u) − Jτ2 (z, u2 )|
Zτ2
6 |Q (t, x (t, τ1 , z, u0 ) , u0 )| dt
τ1
ZT



+ Q t, x1 (t), u2 (t) − Q t, x2 (t), u2 (t)  dt,
τ2

where

x1 (t) = x (t, τ2 , x (τ2 , τ1 , z, u0 ) , u2 ) , x2 (t) = x (t, τ2 , z, u2 ) .

Let the two integral terms on the right side of the last inequality in (5.13) be
1
I1 and I2 , respectively. Since Q ∈ CLip and by (1.9), there exists a constant
K0Q = K0 (r, u0 , Q) such that


|Q (t, x (t, τ1 , z, u0 ) , u0 )| 6 K0Q , for t ∈ [τ1 , τ2 ] and z ∈ Sr W 2 ,

so that

(5.14) I1 6 K0Q |τ1 − τ2 | .

Next, by (1.10), we can get a Lipschitz constant K1Q = K1 (r, u0 , u2 , Q) such

that
ZT

(5.15) I2 6 K1Q x1 (t) − x2 (t) dt.
τ2
326 You

From the formula of mild solution and by Hypothesis I, for τ2 < t 6 T , we have
1
x (t) − x2 (t)
Zτ2
A(τ2 −τ1 ) A(t−τ2 ) A(t−τ2 )
6 e e z−e z + eA(t−s) f (s, x (s, τ1 , z, u0 ) , u0 ) ds
τ1
Zt
A(t−s) 


+ e f s, x1 (s), u2 (s) − f s, x2 (s), u2 (s)  ds
L(X,W )
τ2
 
 
6 C1 |τ1 − τ2 | (−A)1+α eA(t−τ2 ) z 
Zτ2
A(t−τ2 )
(5.16) + eA(τ2 −s) e |f (s, x (s, τ1 , z, u0 ) , u0 )| ds
L(W ) L(X,W )
τ1
Zt
Mα 


+ α
f s, x1 (s), u2 (s) − f s, x2 (s), u2 (s)  ds
(t − s)
τ2
Zτ2
C1 Mα M0 Mα
6 |(−A)z| |τ1 − τ2 | + |f (s, x (s, τ1 , z, u0 ) , u0 )| ds
(t − τ2 )α (t − τ2 )α
τ1
Zt
Mα 


+ α
f s, x1 (s), u2 (s) − f s, x2 (s), u2 (s)  ds.
(t − s)
τ2

Let u0 ∈ U be relatively fixed. Then there is a constant ρ1 (r) > 0 such that


kx (s, τ1 , z, u0 )k 6 ρ1 (r), for s ∈ [τ1 , τ2 ] and z ∈ Sr W 2 .

Hence, there exist constants K0f = K0 (r, u0 , f ) and K1f = K1 (r, u0 , u2 , f ) such
that


|f (s, x (s, τ1 , z, u0 ) , u0 )| 6 K0f , for s ∈ [τ1 , τ2 ] and z ∈ Sr W 2 ,

and



f s, x1 (s), u2 (s) − f s, x2 (s), u2 (s)  6 K f x1 (s) − x2 (s) ,
1


for s ∈ [τ2 , T ] and z ∈ Sr W 2 . Then, from (5.16) and the aforementioned, it
follows that
1
(5.17) x (t) − x2 (t)
C1 Mα r M0 Mα K0f
6 |τ 1 − τ 2 | + |τ1 − τ2 |
(t − τ2 )α (t − τ2 )α
 Optimal Control and Synthesis of Nonlinear Systems 327

Zt
Mα K1f
x1 (s) − x2 (s) ds
+ α
(t − s)
τ2
Zt
Mα K1f
x1 (s) − x2 (s) ds,
= h(t) |τ1 − τ2 | + α
t ∈ (τ2 , T ] ,
(t − s)
τ2

where
C1 Mα r + M0 Mα K0f
h(t) = .
(t − τ2 )α
By the Henry-Gronwall inequality, cf. [16, Appendix D], (5.17) implies that
1
(5.18) x (t) − x2 (t) 6 h(t) |τ1 − τ2 | E (µε , t) ,

where E(µ, t) is a positive, continuous, nondecreasing scalar function of t,

defined on [0, ∞) and E(µ, 0) = 1, that satisfies
1
lim log E(µ, t) = µ,
t→∞ t
and here µε = Mα K1f Γ(ε), with ε > 0 arbitrarily small and Γ(·) being the
Gamma function. Substituting (5.18) into (5.15), we then get
1  
(5.19) I2 6 K1Q E (µε , T ) Mα C1 r + M0 K0f T 1−α |τ1 − τ2 | .
1−α
From (5.13), (5.14) and (5.19), there exists a constant ℓ1 (r) = ℓ1 (r, u0 , u2 ) > 0
such that
(5.20) V (τ1 , z) − V (τ2 , z) 6 ℓ1 (r) |τ1 − τ2 | ,


for any τ1 < τ2 in [0, T ] and z ∈ Sr W 2 .
On the other hand, let u1 be an optimal
control of the (OCP)τ1 over [τ1 , T ]
with the initial state x (τ1 ) = z ∈ Sr W 2 . Then
V (τ2 , z) − V (τ1 , z) 6 Jτ2 (z, u1 ) − Jτ1 (z, u1 ) 6 |Jτ2 (z, u1 ) − Jτ1 (z, u1 )|
Zτ2
6 |Q (t, x (t, τ1 , z, u1 ) , u1 (t))| dt
τ1
ZT
+ |Q (t, x (t, τ2 , z, u1 ) , u1 (t)) − Q (t, x (t, τ1 , z, u1 ) , u1 (t))| dt.
τ2

A parallel argument of what we have done from (5.13) through (5.20) shows
that there exists a constant ℓ2 (r) = ℓ2 (r, u1 ) such that
(5.21) V (τ2 , z) − V (τ1 , z) 6 ℓ2 (r) |τ1 − τ2 | ,
328 You


for any τ1 < τ2 in [0, T ] and z ∈ Sr W 2 . Finally, let L2 (r) = max {ℓ1 (r), ℓ2 (r)}
and (5.11) is proved. The proof is completed.
Theorem 5.1. Under Hypotheses I–IV, the value function V (τ, z) : [0, T ] ×
W 2 → R is locally Lipschitz continuous in (τ, z), that is, for any given r > 0,
there exists a constant L(r) > 0 such that

(5.22) |V (τ2 , z) − V (τ1 , z)| 6 L(r) (|τ1 − τ2 | + kz1 − z2 k) ,

for any (τi , zi ) ∈ [0, T ] × Sr W 2 , i = 1, 2.
Proof. By Hypothesis I, W 2 has a stronger topology than W . Thus, there
is a constant C > 0 such that kwk 6 CkwkW 2 = C|(−A)w|. From Lemmas 5.1
and 5.2, (5.22) is valid by taking

L(r) = max {L1 (Cr), L2 (r)} .

Certainly in (5.22) one can replace kz1 − z2 k by C kz1 − z2 kW 2 .

In order to simplify the presentation of a synthesis result in the next section,
we now convert the original (OCP)τ of Lagrange form to its Meyer form, by
introducing a new state variable y. Let y(t) = y(t, τ, z, u) be the solution of the
following ODE with the initial condtion,
dy
= Q(t, x(t, τ, z, u), u(t)), t ∈ [τ, T ],
(5.23) dt
y(τ ) = ζ.

Then define the augmented state function and its initial data by
 
x(t)
(5.24) ξ(t) = col(x(t), y(t)) = ,
y(t)
 
z
(5.25) ξ(0) = Z = col(z, ζ) = .
ζ

The augmented value function is defined to be

(5.26) Φ(τ, Z) = minτ {y(T, τ, Z, u)},

u∈Uad

and (5.22) is translated to the Lipschitz property of Φ as follows,

|Φ (τ1 , Z1 ) − Φ (τ2 , Z2 )| 6 L(r) |τ1 − τ2 | + kZ1 − Z2 kW ×R ,
(5.27)

for (τi , Zi ) ∈ [0, T ] × Sr W 2 × R.

6 Synthesis of Optimal Control

As commented earlier, except for quadratic optimal controls, cf. [11, 12, 13],
and some nonquadratic optimal controls, cf. [1, 2, 19, 20, 18], etc., the synthesis
 Optimal Control and Synthesis of Nonlinear Systems 329

of a nonlinear distributed optimal control in general is a widely-open issue.

Here, synthesis means an implementation of an optimal control by using a
state feedback. In this section, a theoretical approach is provided to reach a
fairly general result on the synthesis of an optimal control, if it exists, for the
described (OCP). This result can be regarded as a generalization of the results
on optimal feedback controls in [19] and [20] related to systems governed by
specific PDEs. We should also mention that the idea of synthesis via solving
a multivalued differential equation is rooted in the finite dimensional theory of
optimal control, cf. [4] and [5].
Let us define a set-valued function in X × R,
 
Az + f (t, z, u)
(6.1) Π(t, z) = , for (t, z) ∈ [0, T ] × W 2 .
Q(t, z, u)
For each (t, z), the value Π(t, z) is a set in general neither convex nor compact.
Definition

6.1. The lower Dini derivative of a function g(t, Z) : [0, T ] ×
W 2 × R → R at a point (t, Z) in the direction (1, η) with η ∈ W 2 × R is
denoted and defined by
1
(6.2) D− g(t, Z; 1, η) = lim inf [g(t + δ, Z + δη) − g(t, Z)].
δ→0 + δ
By Theorem 5.1, since Φ has the shown Lipschitz property, we know
D− Φ(t, Z; 1, η) exists as a finite number. We make more assumptions in this
section.
Hypothesis V. Assume that for any (t, x) ∈ [0, T ]×W , f (t, x, U ) ⊂ X and
Q(t, x, U ) ⊂ R are closed sets respectively. Also assume that for any bounded
set B ⊂ W , there exist constants K2f (B) > 0 and K2Q (B) > 0 such that

|f (t1 , x1 , u) − f (t2 , x2 , u)| 6 K2f (B) (|t1 − t2 | + kx1 − x2 k) ,

(6.3)
|Q (t1 , x1 , u) − Q (t2 , x2 , u)| 6 K2Q (B) (|t1 − t2 | + kx1 − x2 k) ,
for any t1 , t2 ∈ [0, T ] and any x1 , x2 ∈ B, u ∈ U .
For a control function u(·) valued in U , the minimal requirement is just for
any x(·) ∈ C([0, T ], W ), both f (t, x(t), u(t)) and Q(t, x(t), u(t)) are measurable.
So we adopt this definition for U accordingly. By [16, Lemma 47.2], under
Hypotheses I–IV, for any τ ∈ [0, T ], every x0 ∈ W 2 and every constant control
function u, the mild solution x(·) is a strong solution in W and satisfies

 
0,1−θ
(6.4) x(·) ∈ C [0, T ], W 2 ∩ Cloc (0, T ], W 2θ , 0 6 θ 6 1.

Now we use the method of dynamic programming to
prove a key lemma.

Lemma 6.1. For any (τ, Z) ∈ [0, T ] × W 2 × R , one has

(6.5) min D− Φ(τ, Z; 1, η) = 0.

η∈Π(τ,z)
330 You

Proof. Take any η ∈ Π(τ, z), then there is a û ∈ U such that

 
Az + f (τ, z, û)
(6.6) η= , p (τ, z, û) .
Q (τ, z, û)

Let u(t) ≡ û, t ∈ [τ, T ], and let ξ(t) be the corresponding augmented trajectory
with this u(·) and the initial data ξ(τ ) = Z = col(z, ζ) in W 2 × R. Then for
0 6 δ 6 T − τ , one has
τ +δ
Z
(6.7) ξ(τ + δ) = Z + p (s, x(s), û) ds,
τ

where x(·) = x (·, τ, z, û). As we mentioned earlier, this x(·) is a strong solution
in W 2 and it satisfies (6.4). Hence, p (s, x(s), û) is strongly continuous on [τ, T ].
It follows that

p (s, x(s), û) = p (τ, z, û) + o(1), as δ → 0+ ,

and

(6.8) ξ(τ + δ) = Z + δp (τ, z, û) + o(δ), as δ → 0+ .

By the optimality condition, we have

Φ(τ + δ, ξ(τ + δ)) − Φ(τ, Z)

ZT



= y(τ + δ) + Q t, x t, τ + δ, x(τ + δ), u∗τ +δ , u∗τ +δ (t) dt − Φ(τ, Z)
τ +δ
τ +δ
Z
= y(τ ) + Q (t, x (t, τ, z, û) , û) dt
τ
ZT



(6.9) + Q t, x t, τ + δ, x(τ + δ), u∗τ +δ , u∗τ +δ (t) dt − Φ(τ, Z)
τ +δ
ZT
= ζ+ Q (t, x (t, τ, z, uτ ) , uτ (t)) dt − Φ(τ, Z) > 0.
τ

where u∗τ +δ stands for an optimal control of the (OCP)τ +δ with the initial data
x (τ + δ, τ, z, û), and uτ is given by
(
û, for τ 6 t 6 τ + δ,
uτ (t) = ∗
uτ +δ (t), for τ + δ < t 6 T,
 Optimal Control and Synthesis of Nonlinear Systems 331

τ . From (6.9), (6.8), and the Lipschitz continuity of Φ shown

which belongs to Uad
at the end of Section 5, it follows that
1
[Φ(τ + δ, ξ(τ + δ)) − Φ(τ, Z)]
δ
1
(6.10) = [Φ (τ + δ, Z + δp (τ, z, û) + o(δ)) − Φ(τ, Z)]
δ
1
= [Φ (τ + δ, Z + δp (τ, z, û)) − Φ(τ, Z)] + o(1)
δ
1
= [Φ(τ + δ, Z + δη) − Φ(τ, Z)] + o(1) > 0.
δ
By taking the infimum limit in (6.10) as δ → 0+ , we obtain

(6.11) inf D− Φ(τ, Z; 1, η) > 0.

η∈Π(τ,z)

On the other hand, let {u∗τ , x∗τ } be an optimal process of the (OCP)τ
with the initial data x(τ ) = z and Z = col(z, ζ) again. Let ξ ∗ (·) be the
corresponding augmented trajectory with ξ ∗ (τ ) = Z. Since f (s, x∗τ (t), u∗τ (t))
and Q (s, x∗τ (t), u∗τ (t)) are Bochner and Lebesgue integrable, respectively, almost
every s ∈ [τ, T ] is a Lebesgue point. Thus, from (6.7), we get
τ +δ

Z
1
ξ ∗ (τ + δ) = Z + δ  p (s, x∗τ (s), u∗τ (s)) ds
(6.12) δ
τ
= Z + δp (τ, z, u∗τ (τ )) + o(δ), as δ → 0+ .

According to dynamic programming, we have

(6.13) Φ (τ + δ, ξ ∗ (τ + δ)) − Φ(τ, Z) = 0

Repeating steps in (6.10) and then taking the infimum limit as δ → 0+ , now
with ξ(τ + δ) replaced by ξ ∗ (τ + δ) and (6.12), we end up with

(6.14) D− Φ (τ, Z; 1, p (τ, z, u∗τ (τ ))) = 0,

where u∗τ (τ ) ∈ Π(τ, z). Finally, (6.11) and (6.14) imply (6.5).
As a corollary of Lemma 6.1, let {u∗ , x∗ } be an optimal process of the
original (OCP) with x0 ∈ W 2 , and let ξ ∗ be the corresponding augmented state
trajectory. Then

(6.15) D− Φ (t, ξ ∗ (t); 1, ∂t ξ ∗ (t)) = min D− Φ (t, ξ ∗ (t); 1, η) = 0,

η∈Π(t,x∗ (t))

for almost every t ∈ [0, T ], where ∂t represents a strong derivative in time t.

332 You

Note that Φ(t, Z) satisfies the following nonlinear equation,

(6.16) min D− Φ(t, Z; 1, η) = 0, for a.e. t ∈ [0, T ],

η∈Π(t,z)

(6.17) Φ(T, Z) = ζ.

Define a set-valued function P (t, Z) by

P (t, Z) = argη {D− Φ(t, Z; 1, η) = 0}

(6.18)
= {η ∈ Π(t, z) : D− Φ(t, Z; 1, η) = 0} .

Then consider a differential inclusion or a multivalued differential equation with

the initial value condition as follows,

dξ
∈ P (t, ξ(t)), t ∈ [0, T ],
(6.19) dt
ξ(0) = Z0 = col (x0 , 0) , with x0 ∈ W 2 .


Definition 6.2. A function ξ(·) ∈ C [0, T ], W 2 × R is called a strong
solution of the initial value problem (6.19) if its strong derivative exists a.e.,
it satisfies the differential inclusion in (6.19) a.e., and the given initial value
condition is satisfied.
The following theorem is the main result in the second part of this paper,
and it provides a synthesis of an optimal control under the made assumptions.
Theorem 6.1. Let Hypotheses I–V be satisfied. Let x0 ∈ W 2 . A strong
solution ξ(t) = col(x(t), y(t)), t ∈ [0, T ], of the initial value problem of the
differential inclusions (6.19) provides a synthesis (closed-loop) solution to the
original optimal control problem in the following sense:

1) its first component x(t) is an optimal trajectory with x(0) = x0 ;

2) there exists a strongly measurable selection b(t) = b(t, x(t)) ∈ P (t, ξ(t)),
whose affiliated u(t), t ∈ [0, T ], is a corresponding optimal feedback
control;

3) the terminal value of its second component gives the optimum of the
criterion functional, that is, y(T ) = V (0, x0 ) = min J (x0 , u).
u∈Uad

Proof. Let ξ(t) = col(x(t), y(t)) be a strong solution of the IVP (6.19). It
will be shown later in Lemma 6.3 that there exists a control function u ∈ Uad
such that the following equation is satisfied for almost every t ∈ [0, T ],
 
dξ Ax(t) + f (t, x(t), u(t))
(6.20) = .
dt Q(t, x(t), u(t))
 Optimal Control and Synthesis of Nonlinear Systems 333

We now show that this control process {u, x} is optimal in comparison with
any other admissible control process denoted by {v, g}, where v(·) ∈ Uad and
g(·) = x (·, x0 , v) and whose augmented state trajectory is denoted by λ(·).
By the Hypotheses, for any x0 ∈ W 2 and any admissible control, the
mild solution must be absolutely strongly continuous. Then, by the shown
Lipschitz continuous property of Φ, we can assert that Φ(t, ξ(t)) and Φ(t, λ(t))
are absolutely continuous on [0, T ]. Hence there derivatives in t exist almost
everywhere and are Lebesgue integrable over [0, T ]. Thus, by the Newton-
Leibniz formula, we have
Φ(T, λ(T )) − Φ (0, Z0 )
ZT
d
= Φ(t, λ(t)) dt
dt
0
ZT
1
= lim [Φ(t + δ, λ(t + δ)) − Φ(t, λ(t))] dt (by (6.8))
δ→0+ δ
0
ZT
1
(6.21) = lim [Φ(t + δ, λ(t) + δp(t, g(t), v(t)) + o(δ)) − Φ(t, λ(t))] dt
δ→0+ δ
0
ZT
1
= lim [Φ(t + δ, λ(t) + δp(t, g(t), v(t))) − Φ(t, λ(t))] dt
δ→0+ δ
0
ZT
= D− Φ(t, λ(t); 1, p(t, g(t), v(t))) dt > 0,
0

where the penultimate equality follows from the Lipschitz property of Φ and
the last inequality comes from Lemma 6.1 and its corollary (6.15).
On the other hand, through the steps parallel to (6.21) and by the definition
in (6.18), one can show that
ZT
(6.22) Φ(T, ξ(T )) − Φ (0, Z0 ) = D− Φ (t, ξ(t); 1, ∂t ξ(t)) dt = 0,
0

where ∂t ξ(t) = dξ(t)/dt is given by the selection in (6.20). Therefore, we find

that by (6.21) and (6.22),
(6.23) J (x0 , v) = Φ(T, λ(T )) > Φ (0, Z0 ) = Φ(T, ξ(T )) = J (x0 , u) ,
for any v(·) ∈ Uad . This proves that u(·) ∈ Uad is an optimal control. Thus,
statements 1) and 3) have been proved. Statement 2) will be shown in Lemma
6.3.
334 You

The following lemma, whose proof can be found in [10, Theorem 4.6], is a
generalization of the famous Filippov lemma.
Lemma 6.2. Let Λ be a measure space with a complete σ-finite nonnegative
measure. Let Θ be a Banach space such that there exists a countable subset S
in its dual space, which separates points of Θ. Let M be a separable complete
metric space. Assume that h : Λ → M is a measurable closed-set-valued
function and ρ : Λ × M → Θ is a function such that
i) ρ(·, m) is strongly continuous for each m ∈ M, and

ii) ρ(t, ·) is demicontinuous for a.e. t ∈ Λ.

If κ : Λ → Θ is a strongly measurable function satisfying

(6.24) κ(t) ∈ ρ(t, h(t)), a.e. in Λ,

then there exists a strongly measurable selection m(t) ∈ h(t) such that

(6.25) κ(t) = ρ(t, m(t)), a.e. in Λ.

Lemma 6.3. Under the same assumptions as in Theorem 6.1, for any strong
solution ξ(·) of the differential inclusion in (6.19), thre exists a control function
u ∈ Uad such that (6.20) is satisfied a.e. on [0, T ].
Proof. Since P (t, Z) ⊂ Π(t, z), any strong solution ξ(t) of the differential
inclusion in (6.19) can be viewed as a strong solution of the following differential
inclusion,
dξ
(6.26) ∈ Π(t, x(t)), a.e. t ∈ [0, T ], ξ(0) = Z0 = col (x0 , 0) ,
dt
in which the set-valued function Π(t, x(t)) is given by
  
Ax(t) + f (t, x(t), u)
(6.27) Π(t, x(t)) = :u∈U
Q(t, x(t), u)

In order to apply Lemma 6.2, let

Λ = [0, T ], Θ = X × R, and M = X × R.

Note that Θ = M is a separable, reflexive Banach space. Define ρ : Λ×M → Θ

by
 
Ax(t)
(6.28) ρ(t, m) = + m,
0

that satisfies the two conditions in Lemma 6.2. Let h : Λ → M be defined as

  
f (t, x(t), u)
(6.29) h(t) = :u∈U ,
Q(t, x(t), u)
 Optimal Control and Synthesis of Nonlinear Systems 335

which is a measurable and closed-set-valued function, by Hypothesis V. In this

dξ(t)
case, κ(t) = : Λ → M is strongly measurable as a strong derivative of an
dt
almost everywhere differentiable function, and κ(t) satisfies Eq. (6.24) with ρ
and h given by (6.28) and (6.29).
Therefore, all the conditions in Lemma 6.2 are satisfied. We can apply
Lemma 6.2 to this case and conclude that there exists a strongly measurable
selection m(t, u) ∈ h(t), which means there exists a control function u(·), such
that
 
dξ Ax(t) + f (t, x(t), u(t))
(6.30) = ρ(t, m(t, u)) =
dt Q(t, x(t), u(t))

for a.e. t ∈ [0, T ]. It is seen that u(·) ∈ Uad and Eq. (6.20) is satisfied almost
everywhere on [0, T ]. The proof is completed.
This also completes the proof of Theorem 6.1.
Remark 6.1. The results of this paper also demonstrate that it does not
require the convexity of the cost functional to establish the maximum principle
and the general synthesis for an optimal control. However, nonconvex criteria
of optimality may affect the existence and uniqueness theory of optimal control.
An example of an optimal control problem related to the diffusion of epidemics
that has a nonconvex cost functional can be found in [3, Section 6.1].
Two examples of finding concrete optimal feedback controls by this ap-
proach can be found in [19] and [20]. Even though the style of presenting and
proving this synthesis result seems abstract, there is actually a potentiality to
develop a scheme of approximation for constructing optimal feedback controls
based on this method.

References

[1] V. Barbu, Optimal feedback controls for a class of nonlinear distributed parameter
systems, SIAM Control & Optim., 21 (1983), pp. 871–894.
[2] V. Barbu and G. D. Prato, Hamilton-Jacobi equations and synthesis of nonlinear
control processes in Hilbert spaces, J. Diff. Eqns., 48 (1983), pp. 350–372.
[3] N. Basile and M. Mininni, An extension of the maximum principle for a class
of optimal control problems in infinite dimensional spaces, SIAM J. Control &
Optim., 28 (1990), pp. 1113–1135.
[4] L. Berkovitz, Optimal feedback control, SIAM J. Control & Optim., 27 (1989),
pp. 991–1007.
[5] V. Dzhafanov, Multivalued synthesis for one class of controllable systems, J. Appl.
Math. Mech., 56 (1992), pp. 581–583.
[6] H. Fattorini, A unified theory of necessary conditions for nonlinear nonconvex
control systems, Appl. Math. & Optim., 15 (1987), pp. 141–185.
[7] , Optimal control problems for distributed parameter systems in Banach
spaces, Appl. Math. & Optim., 28 (1993), pp. 225–257.
336 You

[8] , Existence theory and the maximum principle for relaxed infinite dimen-
sional optimal control problems, SIAM J. Control & Optim., 32 (1994), pp. 311–
331.
[9] , Optimal control problems with state constraints for semilinear distributed
parameter systems, J. Optim. Theory & Appl., 88 (1996), pp. 25–59.
[10] S. Hou, Implicit function theorem in topological spaces, Applicable Analysis, 13
(1982), pp. 209–217.
[11] I. Lasiecka and R. Triggiani, Differential and Algebraic Riccati Equations With
Applications to Boundary/Point Control Problems: Continuous Theory and
Approximation Theory, Springer, Berlin, 1991.
[12] , Control Theory for Partial Differential Equations: Continuous and Ap-
proximation Theories, vol. I, Cambridge University Press, 2000.
[13] , Control Theory for Partial Differential Equations: Continuous and Ap-
proximation Theories, vol. II, Cambridge University Press, 2000.
[14] X. Li and Y. Yao, Maximum principle of distributed parameter systems with time
lag, in Lecture Notes in Control and Information Science, F. Kappel, K. Kunish,
and W. Schappacher, eds., vol. 75, Springer, New York, 1985, pp. 410–427.
[15] A. Pazy, Semigroups of Linear Operators and Applications to Partial Differential
Equations, Springer, New York, 1983.
[16] G. Sell and Y. You, Dynamics of Evolutionary Equations, Springer, New York,
2000. (to be published).
[17] Y. Yao, Vector measure and maximum principle of distributed parameter systems,
Scientia Sinica (Ser. A), 26 (1988), pp. 102–112.
[18] Y. You, A nonquadratic Bolza problem and a quasi-Riccati equation for distributed
parameter systems, SIAM J. Control & Optim., 25 (1987), pp. 904–920.
[19] , Nonlinear optimal control and synthesis of thermal nuclear reactors, in
Distributed Parameter Control Systems: New Trends and Applications, C. Chen,
E. Lee, W. Littman, and L. Marcus, eds., Marcel Dekker, New York, 1991,
pp. 445–474.
[20] , Optimal feedback control of Ginzburg-Landau equation for superconductivity
via differential inclusion, Discussions Mathematicae Differential Inclusions, 16
(1996), pp. 5–41.
Forced Oscillation of The Korteweg-De
Vries-Burgers Equation and Its Stability

Bing-Yu Zhang1 , University of Cincinnati, Cincinnati, Ohio

Abstract

This paper studies an infinite-dimensional dynamic system described

by the Korteweg-de Vries-Burgers equation posed on a finite domain with
an external excitation. It shows that if the external excitation is time
periodic with small amplitude, then the system admits a unique time
periodic solution which, as a limit cycle, forms an inertial manifold for
the system.

1 Introduction

An initial and boundary-value problem (IBVP) for a model equation for uni-
directional propagation of waves is investigated here. The equation in question,
which incorporates nonlinear, dispersive and dissipative effects, is the forced
Korteweg-de Vries-Burgers equation

(1.1) ut + ux + uux − uxx + uxxx = f

for x ∈ [0, 1] and t ≥ 0. The equation is subjected to the initial condition

(1.2) u(x, 0) = φ(x), x ∈ (0, 1),

and the boundary conditions

(1.3) u(0, t) = 0, u(1, t) = 0, ux (1, t) = 0

Here f ≡ f (x, t) is a given function. The IBVP (1.1)-(1.3) may be viewed

as an infinite-dimensional dynamic system with the forcing f as an external
excitation.

1
Supported in part by the Charles P. Taft Memorial Fund. E-mail: [email protected]

337
338 Zhang

Our main concern in this paper is the following two questions.

Question 1. If the external excitation f is time periodic with period ω, does it
force the equation (1.1) produce a time periodic solution (forced oscillation) of
periodic ω satisfying the boundary conditions (1.3)?
If such a time periodic solution exists, it forms a limit cycle in the phase
space of the dynamic system (1.1)-(1.3).
Question 2. What stability does this limit cycle possess?
There have been many studies on time periodic solutions of partial differen-
tial equations in the literature. For early works on this subject, see Brézis [6],
Vejvoda et al. [17], Keller and Ting [9], Rabinowitz [12, 13] and the references
therein. For recent works, see Bourgain [5], Craig and Wayne [8], Wayne [19].
In particular, see Wayne [18] for a recent review on time periodic solutions of
nonlinear partial differential equations. While there have been many results for
parabolic and hyperbolic equations, there are few discussions on time-periodic
solutions of nonlinear dispersive wave equations such as the KdV-Burgers equa-
tion discussed in this paper. Especially, there is very few discussion on stability
of time periodic solutions.
In this paper it will be shown that for given time periodic forcing f (of
period ω) with small amplitude, the equation (1.1) does possess a unique
time periodic solution u∗ (x, t) of period ω satisfying the boundary conditions
(1.3). Furthermore, it will be shown that this unique time periodic solution
u∗ (x, t), as a limit cycle for the system (1.1)-(1.3), is also a global attractor
in the space H j (0, 1) (j = 0, 3); for any φ ∈ H j (0, 1) with φ satisfying
φ(0) = φ(1) = φ′ (1) = 0 if j = 3, the corresponding solution u(x, t) of (1.1)-
(1.3) satisfies
(1.4) ku(·, t) − u∗ (·, t)kH j (0,1) ≤ Ce−µt
for any t ≥ 0 where µ > 0 is independent of φ. In other words, this unique time
periodic solution, as a limit cycle, forms an inertial manifold for the dynamic
system (1.1)-(1.3).
The paper is organized as follows.
In section 2 we discuss the global well-posedness of the IBVP (1.1)-
(1.3). It will be shown that for given T > 0, φ ∈ H j (0, 1) and f ∈
H j/3 (0, T ; L2 (0, 1)) ∩ L2 (0, T ; H j/3 (0, 1)), (1.1)-(1.3) admits a unique solution
u ∈ C(0, T ; H j (0, 1)) ∩ L2 (0, T ; H j+1 (0, 1)) with j = 0, 3. The proof is
standard; the local well-posedness is established first by using contraction
mapping principle and then the global well-posedness is obtained by finding
the needed global a priori estimate.
In section 3 we discuss large time behavior of the system (1.1)-(1.3) without
assuming time periodicity of the forcing. The asymptotic estimates of (1.1)-
(1.3) in the space H j (0, 1) when j = 0 and j = 3 are established directly as
 Forced Oscillation of the KdV-Burgers Equation 339

usual via energy estimate method. In particular, the solution of the system
(1.1)-(1.3) tends to zero in the space H j (0, 1) as t → ∞ if the forcing f tends
to zero and the solution decays exponentially if the forcing decay exponentially
as t → ∞. The results presented in this section are crucial in discussing time
periodic solutions of the system (1.1)-(1.3).
The existence of the forced oscillation and its stability analysis will be
discussed in section 4.
The similar results can be also obtained by the same approach for the system
with the external excitations acted on the boundaries:


 ut + ux + uux − uxx + uxxx = 0,



(1.5) u(x, 0) = φ(x),





u(0, t) = h1 (t), u(1, t) = h2 (t), ux (1, t) = h3 (t)
for x ∈ (0, 1) and t ≥ 0.

2 Well-posedness

Consideration is first directed to the homogeneous linear problem



 ut + ux + uxxx − uxx = 0, for x ∈ (0, 1) and t ≥ 0,



(2.1) u(x, 0) = φ(x),





u(0, t) = 0, u(1, t) = 0, ux (1, t) = 0.
By semigroup theory, its solution is given by
u(t) = W (t)φ,
where the spatial variable is suppressed and W (t) is the C0 −semigroup in the
space L2 (0, 1) generated by the operator
Aψ = −ψ ′′′ − ψ ′ + ψ ′′
with the domain of
D(A) = {ψ ∈ H 3 (0, 1), ψ(0) = ψ(1) = ψ ′ (1) = 0}.
By d’Alembert’s formula, one may use the semigroup W (t) to formally write
the solution of the inhomogeneous linear problem


 ut + ux + uxxx − uxx = f, for x ∈ (0, 1), t ≥ 0,



(2.2) u(x, 0) = 0,





u(0, t) = 0, u(1, t) = 0, ux (1, t) = 0
340 Zhang

in the form Z t
u(t) = W (t − τ )f (·, τ )dτ.
0

For T > τ ≥ 0, let XTj (j = 0, 3) be the collection of all

(φ, f ) ∈ L2 (0, 1) × H j/3 (0, T ; L2 (0, 1)) ∩ L2 (0, T ; H j/3 (0, 1))
with φ satisfying the compatibility condition
(2.3) φ(0) = 0, φ(1) = 0, φ′ (1) = 0.
if j = 3 and let
j
Yτ,T = C([τ, T ]; H j (0, 1)) ∩ L2 (τ, T ; H j+1 (0, 1)).

For (φ, f ) ∈ XTj , its norm k(φ, f )kX j is defined by

T

 1/2
k(φ, f )kX j = kφk2H j (0,1) + kf k2H j/3 (0,T ;L2 (0,1)) + kf k2L2 (0,T ;H j/3 (0,1)
T

j
and for v ∈ Yτ,T , its norm kvkY j is defined by
τ,T

kvkY j = kvkC([τ,T ];H j (0,1)) + kvkL2 (τ,T ;H j+1 (0,1)) .

τ,T

j
If τ = 0, the space Yτ,T will be abbreviated simply YTj .
The following two lemmas reveal some smoothing effects for solutions of
(2.1) and (2.2), which will also play important roles later in studying stability
of time periodic solutions of the nonlinear system (1.1)-(1.3).
Lemma 2.1. Let T > 0 be given and u be a solution of (2.1). Then there
exists a constant C independent of φ such that
3
(2.4) kukY 0 ≤ kφkL2 (0,1)
T 2
if φ ∈ L2 (0, 1) and
(2.5) kukY 3 + kut kY 0 ≤ CkφkH 3 (0,1)
T T

if φ ∈ H 3 (0, 1)
with φ(0) = φ(1) = = 0. φ′ (1)
Lemma 2.2. Let T > 0 be given and u be a solution u of (2.2). Then there
exists a constant C independent of f such that
(2.6) kukY 0 ≤ Ckf kL2 (0,T ;(0,1))
T

if f ∈ L2 (0, T ; L2 (0, 1)) and

(2.7) kukY 3 + kut kY 0 ≤ C kf kL2 (0,T ;H 1 (0,1)) + kft kL2 (0,T ;L2 (0,1))
T T

if f ∈ L2 (0, T ; H 1 (0, 1)) ∩ H 1 (0, T ; L2 (0, 1)).

 Forced Oscillation of the KdV-Burgers Equation 341

We only present a proof for Lemma 2.1. The proof of Lemma 2.2 is similar.
Proof of Lemma 2.1: Multiply both sides of the equation in (2.1) by 2u and
integrate over (0, 1) with respect to x. Integration by parts leads to the equality
Z 1 Z 1
d 2
u (x, t)dx + u2x (0, t) +2 u2x (x, t)dx = 0
dt 0 0

for any t ≥ 0, from which (2.5) follows easily if one notes that
Z 1 Z 1
u2 (x, t)dx ≤ u2x (x, t)dx.
0 0

To prove (2.5), let v = ut , which solves



 vt + vx + vxxx − vxx = 0,



v(x, 0) = φ∗ (x),





v(0, t) = 0, v(1, t) = 0, vx (1, t) = 0

with φ∗ (x) = φ′′ (x) − φ′′′ (x) − φ′ (x). By (2.4),

3
(2.8) kut kY 0 ≤ kφ∗ kL2 (0,1) ≤ CkφkH 3 (0,1) .
T 2
It follows from the equation uxxx = −ut − ux + uxx that

kuxx kL2 (0,T ;L2 (0,1))

≤ Ckuxxx kL2 (0,T ;H −1 (0,1))
≤ Ckut − ux + uxx kL2 (0,T ;H −1 (0,1))


≤ C kut kL2 (0,T ;L2 (0,1)) + kukL2 (0,T ;L2 (0,1)) + kux kL2 (0,T ;L2 (0,1))
≤ CkφkH 3 (0,1) .

One infers that

kuxxx kL2 (0,T ;L2 (0,1)) ≤ Ckut − ux + uxx kL2 (0,T ;L2 (0,1)) ≤ CkφkH 3 (0,1) ,
kuxxxx kL2 (0,T ;L2 (0,1)) ≤ Ck(ut − ux + uxx )x kL2 (0,T ;L2 (0,1)) ≤ CkφkH 3 (0,1)

and
sup ku(·, t)kH 3 (0,1) ≤ CkφkH 3 (0,1) .
0≤t≤T

The proof is complete.

Attention now is turn to the well-posedness of the nonlinear problem (1.1)-
(1.3).
342 Zhang

Theorem 2.3. Let T > 0 be given. Then for any (φ, f ) ∈ XTj , j = 0, 3,
with φ satisfies (2.3) if j = 3, the IBVP (1.1)-(1.3) admits a unique solution
u ∈ YTj satisfying

(2.9) kukY j ≤ γ(k(φ, f )kX 0 )k(φ, f )kX j ,

T T T

where γ : R+ → R+ is a nondecreasing continuous function. Moreover, the

corresponding solution map is continuous from the space XTj to the space YTj .
Remark 2.4. The solution map is in fact real analytic from the space XTj
to the space YTj , j = 0, 3 (cf. [20]).
Proof: Using the notation of the semigroup W (t), rewrite (1.1)-(1.3) in its
integral form:
Z t Z t
(2.10) u(t) = W (t)φ + W (t − τ )f (·, τ )dτ − W (t − τ )(uux )(·, τ )dτ.
0 0

Let r > 0 and θ > 0 be two constants to be determined and let S denote the
set
S = {v ∈ Yθ0 ; kvkY 0 ≤ r}.
θ

For given r and θ, S is a complete metric space. For given (φ, f ) ∈ XT0 , define
a map Γ on S:
Z t Z t
Γ(v) = W (t)φ + W (t − τ )f (·, τ )dτ − W (t − τ )(vvx )(·, τ )dτ
0 0

for any v ∈ S. By Lemma 2.1 and Lemma 2.2,

Z θ
kΓ(v)kY 0 ≤ Ck(φ, f )kX 0 + kvvx (·, τ )kL2 (0,1) dτ
θ θ
0
Z θ
≤ Ck(φ, f )kX 0 + sup |v(x, τ )|kvx (·, τ )kL2 (0,1) dτ
θ
0 x∈(0,1)
Z θ
1/2 1/2
≤ Ck(φ, f )kX 0 + C kvkL2 (0,1) kvx kL2 (0,1) kvx kL2 (0,1) dτ
θ
0
Z θ
1/2 3/2
≤ Ck(φ, f )kX 0 + C sup kv(·, τ )kL2 (0,1) kvx (·, τ )kL2 (0,1) dτ
θ
0≤τ ≤θ 0

≤ Ck(φ, f )kX 0 + Cθ 1/4 kvk2Y 0 .

θ θ

If one chooses

(2.11) r = 2Ck(φ, f )kX 0

θ
 Forced Oscillation of the KdV-Burgers Equation 343

and

(2.12) Cθ 1/4 r ≤ 1/2,

then
kΓ(v)kY 0 ≤ r
θ

for any v ∈ S. Thus Γ maps S into S. Similarly, one can show that for r and
θ chosen as in (2.11)-(2.12),
1
kΓ(v1 ) − Γ(v2 )kY 0 ≤ kv1 − v2 kY 0 .
θ 2 θ

In other words, the map Γ is a contraction. Its unique fixed point

u = Γ(u)

is the unique solution of (1.1)-(1.3) in the space S. We have thus established

the local well-posedness of (1.1)-(1.3) in the space XT0 . To obtain the global
well-posedness, it suffices now to show that estimate (2.9) with j = 0 holds for
any smooth solution u of (1.1)-(1.3). Multiply both sides of the equation (1.1)
by 2u and integrate over (0, 1) with respect to x. Integration by parts leads to
Z Z 1 Z 1
d 1 2 2 2
u( (x, t)dx + ux (0, t) + 2 ux (x, t)dx = 2 f (x, t)u(x, t)dx.
dt 0 0 0

It yields that
Z 1 Z 1 Z 1
d
u2 (x, t)dx + u2x (x, t)dx ≤ f 2 (x, t)dx,
dt 0 0 0

from which (2.9) (j = 0) follows.

To see that the IBVP (1.1)-(1.3) is well-posed in the space XT3 , let S ∗ denote
the set
S ∗ = {v ∈ Yθ3 ; vt ∈ Yθ0 , kvkY 3 + kvt kY 0 ≤ r}
θ θ

for given r and θ. A similar argument shows Γ is a contraction mapping from S ∗

to S ∗ if θ and r are chosen appropriately. Thus (1.1)-(1.3) is locally well-posed
in the space XT3 . For its globally well-posedness, one needs to show that (2.9)
with j = 3 holds for any smooth solution of (1.1)-(1.3). To this end, let h = ut .
Then h solves


 ht + hx + (uh)x − hxx + hxxxx = ft ,



h(x, 0) = φ∗ (x, 0)





h(0, t) = 0, h(1, t) = 0, hx (1, t) = 0
344 Zhang

with φ∗ (x) = f (x, 0) − φ′ (x) − φ′′′ (x) + φ′′ (x) − φ(x)φ′ (x). As before, it follows
that
Z Z 1 Z 1 Z 1
d 1 2
h (x, t)dx+ h2x (x, t)dx ≤ ft2 (x, t)dx+2ku(·, t)kL2 (0,1) h2x (x, t)dx,
dt 0 0 0 0

which yields that

kut kY 0 = khkY 0 ≤ C0 (kukY 0 )k(φ, f )kX 3

T T T T

where C0 : R+
→ R+
is a nondecreasing continuous function. The estimate
(2.9) (j = 3) then follows from the equation

uxxx = f − ut − uux − ux + uxx ,

the above inequality and the estimate (2.9) with j = 0. The proof is complete.

3 Large time behavior

In this section we view (1.1)-(1.3) as a dynamic system with external forcing.

Our main concern is large time behavior of its solutions. First we investigate
its large time behavior in the space L2 (0, 1).
Theorem 3.1. Let T > 0 and 0 < ε < 1 be given. For f ∈
L2loc (R+ ; L2 (0, 1)) and φ ∈ L2 (0, 1), the solution u of (1.1)-(1.3) satisfies

ku(·, t)kL2 (0,1)

r Z 1/2
s
−(1−ε)t 2 −(1−ε)(t−s)
≤ e kφkL2 (0,1) + e e−2(1−ε)(s−τ ) kf (·, τ )k2L2 (0,1) dτ
ε 0
r Z t 1/2
2
(3.1) + e−2(1−ε)(t−τ ) kf (·, τ )k2L2 (0,1) dτ
ε s
and

(3.2) kukL2 (t,t+T ;H 1 (0,1)) ≤ ku(·, t)kL2 (0,1) + kf kL2 (t,t+T ;L2 (0,1))

for any 0 ≤ s ≤ t < +∞. Consequently, assuming f ∈ C(R+ ; L2 (0, 1)),

(i) if lim kf (·, t)kL2 (0,1) = 0, then
t→+∞

lim ku(·, t)kL2 (0,1) + kukL2 (t,t+T ;H 1 (0,1)) = 0;

t→+∞

(ii) if kf (·, t)kL2 (0,1) ≤ Ce−αt for some α > 0, then

ku(·, t)kL2 (0,1) +kukL2 (t,t+T ;H 1 (0,1)) ≤ 2e−(1−ε)t kφkL2 (0,1) +Cε,α Ce− min{1−ε,α}t
 Forced Oscillation of the KdV-Burgers Equation 345

for any t ≥ 0, where

 q q

 1
+ 1
if α 6= 1 − ε,
 ε|1−ε−α| α
Cε,α = q q


 2t
+ 1
if α = 1 − ε.
ε α

Proof: For given φ and f , the solution u of (1.1)-(1.3) satisfies the identity
Z Z 1 Z 1
d 1 2 2 2
u (x, t)dx + ux (0, t) + 2 ux (x, t)dx = 2 f (x, t)u(x, t)dx.
dt 0 0 0

for any t ≥ 0. Since

Z 1 Z 1
2
u (x, t)dx ≤ u2x (x, t)dx,
0 0

one obtains
Z 1 Z 1 Z 1
d 2
u2 (x, t)dx + 2(1 − ε) u2x (x, t)dx ≤ f 2 (x, t)dx
dt 0 0 ε 0

and by Gronwall’s inequality,

Z 1 Z
2 t −2(1−ε)(t−τ )
u2 (x, t)dx ≤ e−2(1−ε)t kφk2L2 (0,1) + e kf (·, τ )k2L2 (0,1) dτ
0 ε 0

for any t ≥ 0. In particular, for any 0 ≤ s ≤ t

Z 1
u2 (x, t)dx
0
Z
−2(1−ε)(t−s) 2 2 t −2(1−ε)(t−τ )
≤ e ku(·, s)kL2 (0,1) + e kf (·, τ )k2L2 (0,1) dτ
ε s
Z
−2(1−ε)t 2 2 −2(1−ε)(t−s) s −2(1−ε)(s−τ )
≤ e kφkL2 (0,1) + e e kf (·, τ )k2L2 (0,1) dτ
ε 0
Z
2 t −2(1−ε)(t−τ )
+ e kf (·, τ )k2L2 (0,1) dτ
ε s
and
Z t+T Z 1 Z t+T
u2x (x, τ )dxdτ ≤ ku(·, t)k2L2 (0,1) + kf (·, τ )k2L2 (0,1) dτ
t 0 t

for any t ≥ 0 and T > 0. Combining the above inequalities yields the estimate
(3.1) and (3.2). The proof is complete.
346 Zhang

Next we describes large time behavior of solutions of (1.1)-(1.3) in the space

H 3 (0, 1). Theorem 3.2. Let T > 0 be given. Suppose f ∈ C 1 (R+ ; L2 (0, 1)) ∩
L2loc (R+ ; H 1 (0, 1)) and φ ∈ H 3 (0, 1). If φ satisfies (2.3) and f satisfies the
condition
lim kf (·, t)kL2 (0,1) < 1/2,
t→+∞

then for any η with 0 < η < 1 − 2 lim kf (·, t)kL2 (0,1) , there exists s1 > 0
t→+∞
depending only on kφkL2 (0,1) + kf kCb (R+ ;L2 (0,1)) such that the corresponding
solution u of (1.1)-(1.3) satisfies

ku(·, t)kH 3 (0,1)

"
   
≤ γ kφkL2 (0,1) + kf kCb (R+ ;L2 (0,1)) ku(·, s1 )kL2 (0,1) + kut (·, s1 )kL2 (0,1)
Z t   1/2
−η(t−s1 ) −2η(t−τ )
·e + e kf (·, τ )k2L2 (0,1) + kft (·, τ )k2L2 (0,1) dτ
s
+kf kH 1 (t,t+T ;L2 (0,1))

Z #
s   1/2
+e−η(t−s) e−2η(s−τ ) 2 2
kf (·, τ )kL2 (0,1) + kft (·, τ )kL2 (0,1) dτ
s1

and

kukL2 (t,t+T ;H 4 (0,1))





≤ γ kφkL2 (0,1) + kf kCb (R+ ;L2 (0,1)) ku(·, s1 )kL2 (0,1) + kut (·, s1 )kL2 (0,1)
Z t   1/2
−η(t−s1 ) −2η(t−τ ) 2 2
·e + e kf (·, τ )kL2 (0,1) + kft (·, τ )kL2 (0,1) dτ
s
+kf kH 1 (t,t+T ;L2 (0,1))

Z s   1/2 
+e−η(t−s) e−2η(s−τ ) kf (·, τ )k2L2 (0,1) + kft (·, τ )k2L2 (0,1) dτ
s1
+kf kL2 (t,t+T ;H 1 (0,1)) .

for any s1 ≤ s ≤ t, where γ : R+ → R+ is a nondecreasing continuous function.

Consequently,
 Forced Oscillation of the KdV-Burgers Equation 347


(i) if lim kf (·, t)kL2 (0,1) + kft (·, t)kL2 (0,1) + kf kL2 (t,t+T ;H 1 (0,1)) = 0, then
t→+∞

lim kukY 3 = 0;
t→+∞ t,t+T

(ii) if kf (·, t)kL2 (0,1) + kft (·, t)kL2 (0,1) + kf kL2 (t,t+T ;H 1 (0,1)) < Ce−αt , for some
α > 0 and any t ≥ 0, then


kukY 3 ≤ γ kφkL2 (0,1) + kf kCb (R+ ;L2 (0,1)) ku(·, s1 )kL2 (0,1) +
t,t+T

i
+kut (·, s1 )kL2 (0,1) e−η(t−s1 ) + Cα,η Ce− min{α,η}t

for any t ≥ s1 , where

 q
 1 √2
 |1−ε−α| + α
if α 6= η,
Cη,α =

 √t − s +
1 √2 if α = η.
α

Proof: Let h = ut . Then h solves



 ht + hx + (uh)x − hxx + hxxxx = ft ,



h(x, 0) = φ∗ (x, 0)





h(0, t) = 0, h(1, t) = 0, hx (1, t) = 0

with
φ∗ (x) = f (x, 0) − φ′ (x) − φ′′′ (x) + φ′′ (x) − φ(x)φ′ (x).
It holds that
Z 1 Z 1
d 2
h (x, t)dx + 2 h2x (x, t)dx + h2x (0, t)
dt 0 0
Z 1 Z 1
=2 ft (x, t)h(x, t)dx + 2 u(x, t)hx (x, t)h(x, t)dx.
0 0

Thus
d 2
kh(·, t)k2L2 (0,1) + 2(1 − ε − ku(·, t)kL2 (0,1) )khx (·, t)k2L2 (0,1) ≤ kf (·, t)k2L2 (0,1) ,
dt ε
from which one obtains
Z t
−2ηs,t (t−s) 2
kh(·, t)k2L2 (0,1) ≤e kh(·, s)k2L2 (0,1) + e−2ηs,t (t−τ ) kft (·, τ )k2L2 (0,1) dτ
ε s
348 Zhang

and
Z t+T Z t+T
1 1
khx (·, τ )k2L2 (0,1) dτ ≤ kf (·, τ )k2L2 (0,1) dτ + kh(·, t)k2L2 (0,1)
t εηs,t t 2ηs,t
for 0 ≤ s ≤ t where

ηs,t = 1 − ε − sup ku(·, τ )kL2 (0,1) .

s≤τ ≤t

Since
lim ku(·, t)kL2 (0,1) ≤ 2 lim kf (·, t)kL2 (0,1) < 1
t→+∞ t→+∞

for given 0 < η < 1 − 2 lim kf (·, t)kL2 (0,1) , if let

t→+∞

1
ε = (1 − η − 2 lim kf (·, t)kL2 (0,1) ),
2 t→+∞

then according to Theorem 3.1, there exists s1 > 0 depending only on

kφkL2 (0,1) + kft kCb (R+ ;L2 (0,1)) such that for any s > s1 ,

inf ηs,t ≥ η.
s1 ≤s≤t<+∞

Thus

kh(·, t)kL2 (0,1)

r Z 1/2
t
−η(t−s) 2
≤ e kh(·, s)kL2 (0,1) + e−2η(t−τ ) kft (·, τ )k2L2 (0,1) dτ
ε s
for any t ≥ s > s1 . In particular,

kh(·, s)kL2 (0,1)

r Z 1/2
s
−η(s−s1 ) 2 −2η(s−τ ) 2
≤ e kh(·, s1 )kL2 (0,1) + e kft (·, τ )kL2 (0,1) dτ
ε s1
for any s > s1 and

kh(·, t)kL2 (0,1)

r Z 1/2
s
−η(t−s1 ) 2 −η(t−s1 ) −2η(s−τ )
≤ e kh(·, s1 )kL2 (0,1) + e e kft (·, τ )k2L2 (0,1) dτ
ε s1
r Z s 1/2
2
+ e−2η(s−τ ) kft (·, τ )k2L2 (0,1) dτ
ε s1
for any s1 < s < t. By Theorem 2.3, there exists a γ = γ(k(φ, f )kXs0 ) such
1
that
kh(·, s1 )kL2 (0,1) ≤ γ(k(φ, f )kXs0 )k(φ, f )kXs3 .
1 1
 Forced Oscillation of the KdV-Burgers Equation 349

One arrives that

r Z s 1/2
2 −η(t−s) −η(t−s1 ) −2η(s−τ ) 2
kut (·, t)kL2 (0,1) ≤ e e e kft (·, τ )kL2 (0,1) dτ +
ε s1
r Z s 1/2
−ηt 2 −2η(s−τ ) 2
(3.3)
+γ(k(φ, f )kXs0 )k(φ, f )kXs3 e + e kft (·, τ )kL2 (0,1) dτ
1 1 ε s1

and
r r
1 1
(3.4) kut kL2 t,t+T ;H 1 (0,1) ≤ kut (·, t)kL2 (0,1) + kft kL2 (t,t+T ;L2 (0,1))
2η εη
for any s1 ≤ s ≤ t. Recall that

(3.5) uxxx = f − ut − ux − uux + uxx .

There exists a constant C such that

kuxx (·, t)kL2 (0,1) ≤ Ckuxxx (·, t)kH −1 (0,1)

 
≤ C kf (·, t)kL2 (0,1) + kut (·, t)kL2 (0,1) + ku2 (·, t)kL2 (0,1) + kux (·, t)kL2 (0,1)
 
≤ C kf (·, t)kL2 (0,1) + kut (·, t)kL2 (0,1) + (1 + ku(·, t)kL2 (0,1) )kux (·, t)kL2 (0,1) .

Thus, for given T > 0,


kuxx kL2 (t,t+T ;L2 (0,1)) ≤ C kf kL2 (t,t+T ;L2 (0,1)) + kut kL2 (t,t+T ;L2 (0,1))
  
+ sup 1 + ku(·, τ )kL2 (0,1) kux kL2 (t,t+T ;L2 (0,1)) .
t≤τ ≤t+T

It follows from the equation (3.5) again that

kuxxx kL2 (t,t+T ;L2 (0,1)

≤ kf − ut − ux kL2 (t,t+T ;L2 (0,1)) + kuxx − uux kL2 (t,t+T ;L2 (0,1))
 
≤ kf − ut − ux kL2 (t,t+T ;L2 (0,1)) + sup 1 + ku(·, τ )kL2 (0,1) kuxx kL2 (t,t+T ;L2 (0,1))
6
t τ ≤t+T
 2 
≤ C sup 1 + ku(·, τ )kL2 (0,1) kf kL2 (t,t+T ;L2 (0,1)) + kut kL2 (t,t+T ;L2 (0,1))
6
t τ ≤t+T

+kux kL2 (t,t+T ;L2 (0,1)) .

As u ∈ L2 (t, t + T ; H 3 (0, 1)) and ut ∈ L2 (t, t + T ; H 1 (0, 1))), it yields that

u ∈ C(t, t + T ; H 2 (0, 1)) by interpolation and that
 
sup kuxx (·, τ )kL2 (0,1)) ≤ C kukL2 (t,t+T ;H 3 (0,1)) + kut kL2 (t,t+T ;H 1 (0,1))
t≤τ ≤t+T
350 Zhang

for some constant C independent of t, T and u. Consequently, using (3.5) yields

that

kuxxx (·, t)kL2 (0,1))

≤ kf (·, t) − ut (·, t)kL2 (0,1)) + kuxx (·, t) − ux (·, t) − u(·, t)ux (·, t)kL2 (0,1))
 
≤ kf (·, t) − ut (·, t)kL2 (0,1)) + 2 + ku(·, t)kL2 (0,1)) kuxx (·, t)kL2 (0,1))
 
≤ kf (·, t) − ut (·, t)kL2 (0,1)) + C 2 + ku(·, t)kL2 (0,1)) kukL2 (t,t+T ;H 3 (0,1))

+kut kL2 (t,t+T ;H 1 (0,1))
 3
≤ kf (·, t)kL2 (0,1)) + kut (·, t)kL2 (0,1)) + C sup 2 + ku(·, τ )kL2 (0,1)
t6τ ≤t+T
 
· kf kL2 (t,t+T ;L2 (0,1)) + kut kL2 (t,t+T ;H 1 (0,1)) + kux kL2 (t,t+T ;L2 (0,1))
 3 
≤ C 1 + ku(·, t)kL2 (0,1) + kf kL2 (t,t+T ;L2 (0,1)) kf kL2 (t,t+T ;L2 (0,1)) +

+kut kL2 (t,t+T ;H 1 (0,1)) + kux kL2 (t,t+T ;L2 (0,1))
+kf (·, t)kL2 (0,1)) + kut (·, t)kL2 (0,1))

and

kuxxxx kL2 (t,t+T ;L2 (0,1))

≤ kf − ut kL2 (t,t+T ;H 1 (0,1)) + k(ux + uux )x kL2 (t,t+T ;L2 (0,1))
+kuxxx kL2 (t,t+T ;L2 (0,1))
≤ kf − ut kL2 (t,t+T ;H 1 (0,1)) + kuxx kL2 (t,t+T ;L2 (0,1)) + kuxxx kL2 (t,t+T ;L2 (0,1))
+2kux kL2 (t,t+T ;L2 (0,1)) sup kuxx (·, τ )kL2 (0,1)
t6τ 6t+T
 
≤ kf kL2 (t,t+T ;H 1 (0,1)) + C 1 + kux kL2 (t,t+T ;L2 (0,1)) kukL2 (t,t+T ;H 3 (0,1))

+kut kL2 (t,t+T ;H 1 (0,1)) .

Thus there exists a constant Cε,η depending only on η and ε such that for any
t ≥ s1 ,

ku(·, t)kH 3 (0,1)

 
≤ Cε,η 1 + ku(·, t)kL2 (0,1) + kf kL2 (t,t+T ;L2 (0,1)) ) kf kH 1 (t,t+T ;L2 (0,1))

+ku(·, t)kL2 (0,1) + kut (·, t)kL2 (0,1)

and

kukL2 (t,t+T ;H 4 (0,1))

 Forced Oscillation of the KdV-Burgers Equation 351
 
≤ kf kL2 (t,t+T ;H 1 (0,1)) + Cε,η 1 + ku(·, t)kL2 (0,1) + kf kL2 (t,t+T ;L2 (0,1)) )


· kf kH 1 (t,t+T ;L2 (0,1)) + ku(·, t)kL2 (0,1) + kut (·, t)kL2 (0,1) .

Theorem 3.2 follows consequently. The proof is complete.

4 Forced oscillation and its global stability

In this section we assume the forcing f is a time-periodic function of period

ω > 0 and study if it generates a time periodic solution for the equation (1.1)
with the boundary condition (1.3).
Theorem 4.1. If f ∈ Cb1 (R+ ; L2 (0, 1)∩L2loc (R+ , H 1 (0, 1)) is a time periodic
function of period ω satisfying

1
(4.1) sup kf (·, t)kL2 (0,1) < ,
0≤t≤ω 4

then the equation (1.1) admits a unique time periodic solution u∗ ∈

Cb (R+ ; H 3 (0, 1)) of period ω satisfying the boundary conditions (1.3).
Proof: For the given forcing f satisfying (4.1), choose φ ∈ H 3 (0, 1) satisfying
the compatibility condition (2.3). Let u(x, t) be the corresponding solution of
the IBVP (1.1)-(1.3). By Theorem 3.2, the set {ku(·, t)kH 3 (0,1) }+∞
t=0 is uniformly
bounded. Let tk be a sequence with tk → ∞ as k → ∞ such that u(·, tk )
converges to a function ψ ∈ H 3 (0, 1) weakly in H 3 (0, 1) and strongly in L2 (0, 1)
as k → ∞. If one takes ψ as an initial data in the IBVP (1.1)-(1.3) with the
given forcing f , then the corresponding solution, named as u∗ (x, t), is a time
periodic function of period ω. To see this is true, let v(x, t) = u(x, t+ω)−u(x, t).
Because of the periodicity of f , v(x, t) solves the following linear problem with
the variable coefficient b(x, t) = u(x, t + ω) + u(x, t):


 vt + vx + (bv)x − vxx + vxxx = 0,



(4.2) v(x, 0) = φ∗ (x),





v(0, t) = 0, v(1, t) = 0, vx (1, t) = 0,

where φ∗ (x) = u(x, ω) − u(x, 0). It leads to

Z 1 Z 1 Z 1
d
v 2 (x, t)dx + vx2 (0, t) + 2 vx2 (x, t)dx = 2 b(x, t)vx (x, t)v(x, t)dx
dt 0 0 0

or Z Z
1 1
d 2
v (x, t)dx + 2(1 − kb(·)kL2 (0,1) ) v 2 (x, t)dx ≤ 0
dt 0 0
352 Zhang

for any t ≥ 0. By Gronwall’s inequality,

Rt
−2(1−kb(·,s)kL2 (0,1) )ds
kv(·, t)kL2 (0,1) ≤ kv(·, τ )kL2 (0,1) e τ

for any t ≥ τ ≥ 0. By Theorem 3.1,

lim kb(·, t)kL2 (0,1) ≤ 2 lim ku(·, t)kL2 (0,1) ≤ 4 lim kf (·, t)kL2 (0,1) < 1.
t→∞ t→∞ t→∞

One can choose τ > 0 large enough such that

 
2γ = 2 1 − sup kb(·, t)kL2 (0,1) > 0.
t≥τ

Consequently,
kv(·, t)kL2 (0,1) ≤ kv(·, τ )kL2 (0,1) e−γ(t−τ )
for any t ≥ τ . In particular,

ku(·, tk + ω) − u(·, tk )kL2 (0,1) ≤ kv(·, τ )kL2 (0,1) e−γ(tk −τ )

for any tk ≥ τ . Note that u(·, tk ) converges to ψ = u∗ (·, 0) strongly in L2 (0, 1)

and that u(·, tk + ω) converges to u∗ (·, ω) strongly in L2 (0, 1) as k → ∞. Since

ku∗ (·, ω) − u∗ (·, 0)kL2 (0,1) ≤ ku∗ (·, ω) − u(·, tk + ω)kL2 (0,1) +

+ku(·, tk + ω) − u(·, tk )kL2 (0,1) + ku(·, tk ) − u∗ (·, 0)kL2 (0,1)

for any tk ≥ τ , we conclude that u∗ (x, ω) = u∗ (x, 0) for x ∈ (0, 1) a.e. and
that u∗ (x, t) is a time periodic function of period ω. To show the uniqueness,
let u1 and u2 be such two time periodic solutions with the given forcing f . Let
v = u1 − u2 . Then v solves the linear problem (4.2) with b = u1 + u2 and
φ∗ (x) = u1 (x, 0) − u2 (x, 0). By Theorem 3.1,

lim kb(·, t)kL2 (0,1) ≤ lim ku1 (·, t)kL2 (0,1) + lim ku2 (·, t)kL2 (0,1)
t→∞ t→∞ t→∞

≤ 4 lim kf (·, t)kL2 (0,1)

t→∞

< 1.

Consequently, v decays to zero exponentially in the space L2 (0, 1). Therefore

u1 (x, t) ≡ u2 (x, t) for any x ∈ (0, 1) and t ≥ 0 because they are time periodic
functions. The proof is complete.
The time periodic solution u∗ forms a limit circle for the dynamic system
described by (1.1)-(1.3). To investigate its stability, let φ be a s−compatible
 Forced Oscillation of the KdV-Burgers Equation 353

function in the phase space H s (0, 1) and u be the corresponding solution of the
IBVP (1.1)-(1.3). If let w = u − u∗ , then w solves the equation

wt + wx + wwx + (u∗ w)x − wxx + wxxx = 0

and satisfies the initial condition w(x, 0) = φ − u∗ (x, 0) and the homogeneous
boundary condition (1.3). This leads us to considering the large time behavior
of the following initial-boundary-value problem


 ut + ux + uux + (vu)x + uxxx − uxx = 0,



(4.3) u(x, 0) = φ(x),





u(0, t) = 0, u(1, t) = 0, ux (1, t) = 0

for x ∈ (0, 1) and t ≥ 0 where v ≡ v(x, t) is a given function

with v ∈ C(R+ ; H 3 (0, 1)) ∩ L2loc (R+ ; H 4 (0, 1)) and vt ∈ C(R+ ; L2 (0, 1)) ∩
L2loc (R+ ; H 1 (0, 1)). By the same arguments used in the section 2 one can show
that (4.3) is globally well-posed in the space H j (0, 1) for j = 0, 3. As for its
large time behavior, we have the following result.
Proposition 4.2. Let T > 0 and j = 0 or 3 be given. There exist constants
η1 ∈ (0, 1) and η2 > 0 such that if

(4.4) lim kvkY 0 < η1 ,

t→∞ t,t+T

then for φ ∈ H j (0, 1) with φ satisfying (2.3) if j = 3, the unique solution u of

(4.3) satisfies


(4.5) ku(·, t)kH j (0,1) ≤ γ kφkL2 (0,1) e−η2 t kφkH j (0,1)

for any t ≥ 0, where γ : R+ → R+ is a nondecreasing continuous function

depending on j, T and v.
The following stability result for the forced oscillation u∗ (x, t) then follows.
Theorem 4.3. There exists an η ∈ (0, 1) such that if f ∈ Cb1 (R+ ; L2 (0, 1) ∩
L2loc (R+ , H 1 (0, 1)) is a time periodic function of period ω satisfying

sup kf (·, t)kL2 (0,1) < η,

0≤t≤ω

then the equation (1.1) admits a unique time periodic solution u∗ ∈

Cb (R+ ; H 3 (0, 1)) of period ω satisfying the boundary conditions (1.3). More-
over, for given j = 0, 3, there exists a δ > 0 such that for any compatible
φ ∈ H j (0, 1) with φ satisfying (2.3) if j = 3, the corresponding solution u of
(1.1)-(1.3) satisfies

ku(·, t) − u∗ (·, t)kH j (0,1) ≤ Ce−δt

354 Zhang

for any t ≥ 0. In other words, the set {u∗ (·, t), 0 ≤ t ≤ ω}, as a limit circle,
forms an inertial manifold in the space H j (0, 1) for the dynamic system (1.1)-
(1.3).
Indeed, for given compatible φ ∈ H j (0, 1), let u(x, t) be the corresponding
solution. Then w(x, t) = u(x, t) − u∗ ((x, t) solves the system (4.3) with
v(x, t) = u∗ (x, t) and φ∗ (x) = φ(x) − u∗ (x, 0). By Theorem 3.1, one may
choose η small enough such that

lim ku∗ kY 0 < η1 ,

t→∞ t,t+T

where η1 is as given in Proposition 4.1. If Proposition 4.2 holds, then there

exist τ > 0 and δ > 0 such that

kw(·, t)kH s (0,1) ≤ γ(kw(·, τ )kL2 (0,1) )kw(·, τ )kH j (0,1) e−δ(t−τ )

for any t ≥ τ , which yields Theorem 4.3. Thus it remains to prove Proposition
4.2.
Proof of Proposition 4.2: For the solution u of (4.3) it holds that
Z Z 1 Z 1
d 1 2
u (x, t)dx + u2x (0, t) + 2 u2x (x, t)dx = 2 ux (x, t)u(x, t)v(x, t)dx
dt 0 0 0

for any t ≥ 0, which implies that

Z Z 1
d 1 2
u (x, t)dx + 2(1 − kv(·, t)kL2 (0,1) ) u2x (x, t)dx ≤ 0
dt 0 0

for any t ≥ 0. By the assumption (4.4), there exists s > 0 such that

sup (1 − kv(·, t)kL2 (0,1) ) = η > 0.

s≤t<∞

Thus Z 1
u2 (x, t)dx ≤ e−2η(t−s) ku(·, s)k2(0,1)
0
and Z Z
t+T 1
1
u2x (x, τ )dxdτ ≤ u2 (x, t)dx
t η 0
for any t ≥ s. The estimate (4.5) with j = 0 follows consequently.
To see Proposition 4.2 holds for j = 3, let u be the solution of (4.3) and
h = ut . Then h solves


 ht + ((v + u)h)x + hxxx − hxx = −(uvt )x ,



h(x, 0) = φ∗ (x),





h(0, t) = 0, h(1, t) = 0, hx (1, t) = 0
 Forced Oscillation of the KdV-Burgers Equation 355

for x ∈ (0, 1) and t ≥ 0 where

φ∗ (x) = −φ′′′ (x) − v(x, 0)φ′ (x) − vx (x, 0)φ(x) + φ′′ (x).

Its integral form is given by

Z t Z t
h(t) = W (t−s)h(s)− W (t−τ )(h(v +u))x (·, τ )dτ − W (t−τ )(uvt )x (·, τ )dτ
s s

for any 0 ≤ s ≤ t < +∞. For a given T > 0, let Q = W (T + s), qk = h(kT + s),
Z (k+1)T +s
fk = − W ((k + 1)T + s − τ )(uvt )x (·, τ )dτ
kT +s

and Z (k+1)T +s
g(qk ) = − W ((k + 1)T + s − τ )((u + v)h)x (·, τ )dτ
kT +s

for k = 0, 1, 2, · · · . We have then

qk+1 = Qqk + g(qk ) + fk , k = 0, 1, 2, · · · .

Note that there exists a constant C1 depending only on T , v and kφkL2 (0,1) such
that √
kfk kL2 (0,1) ≤ 2 T kvt kY 0 kukY 0
kT +s,(k+1)T +s kT +s,(k+1)T +s

and
kg(qk )kL2 (0,1) ≤ CT ak kqk kL2 (0,1) ,
where ak = ku + vkY 0 . Since kQk < 1 and
kT +s,(k+1)T +s

lim ku + vkY 0 ≤ η1
t→∞ kT +s,(k+1)T +s

by the estimate (4.5) with j = 0 we have just proved, one can choose η1 and s
such that, for any k ≥ 0,

kQk + CT ak = α < 1

and
kfk k ≤ C2 ku(·, s)kL2 (0,1) β k ,
where 0 ≤ β < 1 and the constant C2 depends only on kvt kCb (R+ ;L2 (0,1)) and T .
As a result, for such chosen η1 and s,

kqk+1 kL2 (0,1) ≤ αkqk kL2 (0,1) + C2 β k

356 Zhang

for any k ≥ 0, which leads to

k
X
kqk+1 kL2 (0,1) ≤ αk kq0 kL2 (0,1) + C2 ku(·, s)kL2 (0,1) αj β k−j
j=0

or

kqk+1 kL2 (0,1) ≤ αk kq0 kL2 (0,1) + C2 kη k ku(·, s)kL2 (0,1)

= αk kut (·, s)kL2 (0,1) + C2 kη k ku(·, s)kL2 (0,1)

for any k ≥ 0 where η = min{α, β}.

Hence there exists a η2 > 0 such that


khkY 0 = kut kY 0 ≤ γ1 kφkL2 (0,1) kφkH 3 (0,1) e−η2 t
t,t+T t,t+T

for any t ≥ 0 where γ1 : R+ → R+ is a nondecreasing function depending only

on v and T. It follows from the equation

uxxx = −uux − ux + uxx − ut − (vu)x

by the same argument as that used in the proof of Theorem 3.2 that


kukY 3 ≤ γ kφkL2 (0,1) e−η2 t kφkH 3 (0,1)
t,T +t

for any t ≥ 0. The proof is complete.

References

[1] C. J. Amick, J. L. Bona and M. E. Schonbek, Decay of solutions of some nonlinear

wave equations. J. Differential Equations 81 (1989), 1–49.
[2] J. L. Bona and L. Luo, Decay of solutions to nonlinear, dispersive wave equations,
Diff. & Integral Equ. 6 (1993), 961–980.
[3] J. L. Bona, V. A. Dougalis, An initial and boundary value problem for a model
equation for propagation of long waves, J. Math. Anal. Appl. 75 (1980), 503–522.
[4] J. L. Bona and L. Luo, Initial-boundary-value problems for model equations for
the propagation of long waves, Evolution equations (Baton Rouge, LA, 1992), 63,
65–94, Lecture Notes in Pure and Appl. Math., 168, Dekker, New York, 1995.
[5] J. Bourgain, Construction of periodic solutions of nonlinear wave equations in
higher dimension, Geom. Funct. Anal., 5 (1995), 105 – 140.
[6] H. Brézis, Periodic solutions of nonlinear vibrating strings and duality principles,
Bull. Amer. Math. Soc., 8 (1983), 409 – 425.
[7] P. Constantin, C. Foiacs, B. Nicolaenko and R. Temam, Integral manifolds
and inertial manifolds for dissipative partial differential equations, Applied
Mathematical Sciences, 70 (1989), Springer-Verlag, New York-Berlin
 Forced Oscillation of the KdV-Burgers Equation 357

[8] W. Craig and C. E. Wayne, Newton’s method and periodic solutions of nonlinear
wave equations, Comm. Pure Appl. Math., 46 (1993), 1409–1498.
[9] J. B. Keller and L. Ting, Periodic vibrations of systems governed by nonlinear
partial differential equations, Comm. Pure and Appl. Math. (1966), 371 – 420.
[10] J. U. Kim, Forced vibration of an aero-elastic plate, J. Math. Anal. Appl.
113 (1986), 454–467.
[11] R. M. Miura, The Korteweg-de Vries equation: A survey of results, SIAM review
18 (1976), 412 – 459.
[12] P. H. Rabinowitz, Periodic solutions of nonlinear hyperbolic differential equations,
Comm. Pure Appl. Math. 20 (1967), 145 – 205.
[13] P. H. Rabinowitz, Free vibrations for a semi-linear wave equation, Comm. Pure
Appl. Math. 31 (1978), 31 – 68.
[14] D. L. Russell and B.-Y. Zhang, Controllability and stabilizability of the third-
order linear dispersion equation on a periodic domain, SIAM J. Contr. Optim.
31 (1993), 659–676.
[15] D. L. Russell and B.-Y. Zhang, Exact controllability and stabilizability of the
Korteweg-de Vries equation, Trans. Amer. Math. Soc. 348 (1996), 3643–3672
[16] D. L. Russell and B.-Y. Zhang, Smoothing and decay properties of solutions of
the Korteweg-de Vries equation on a periodic domain with point dissipation, J.
Math. Anal. Appl. 190 (1995), 449–488.
[17] O. Vejvoda, partial differential equations: time-periodic solutions, Mrtinus Nijhoff
Publishers, 1982.
[18] C. E. Wayne, Periodic solutions of nonlinear partial differential equations, Notices
Amer. Math. Soc. 44 (1997), 895–902.
[19] C. E. Wayne, Periodic and quasi-periodic solutions of nonlinear wave equations
via KAM theory, Comm. Math. Phys. 127 (1990), 479 – 528.
[20] B.-Y. Zhang, Analyticity of solutions for the generalized Korteweg-de Vries
equation with respect to their initial datum, SIAM J. Math. Anal. 26 (1995),
1488–1513.