DOI: 10.2478/v10006-010-0016-4
SENSORS AND BOUNDARY STATE RECONSTRUCTION OF HYPERBOLIC SYSTEMS
E
LH
ASSANZERRIK
∗, H
AMIDBOURRAY
∗∗, S
AMIRBEN HADID
∗∗∗∗
MACS Group, Faculty of Sciences
University of Moulay Ismaïl, Bni Mhamed 4011, Meknes, Morocco e-mail:
zerrik3@yahoo.fr∗∗
MACS Group, Faculty of Poly Disciplinary
University of Moulay Ismaïl, Boutalamine I, Errachidia, Morocco e-mail:
hbourrayh@yahoo.fr∗∗∗
Department of Mathematics, Faculty of Exact Sciences Mentouri University, Constantine, Algeria
e-mail:
ihebmaths@yahoo.frThis paper deals with the problem of regional observability of hyperbolic systems in the case where the subregion of interest is a boundary part of the system evolution domain. We give a definition and establish characterizations in connection with the sensor structure. Then we show that it is possible to reconstruct the system state on a subregion of the boundary.
The developed approach, based on the Hilbert uniqueness method (Lions, 1988), leads to a reconstruction algorithm. The obtained results are illustrated with numerical examples and simulations.
Keywords: distributed systems, hyperbolic systems, regional observability, boundary reconstruction, strategic sensor.
1. Introduction
The study of dynamical spatiotemporal systems has ge- nerated wide literature with applications in fields such as ecology, pollution control, population dynamics as well as many others. A wide portion of the literature is devo- ted to the problem of the analysis and control partial dif- ferential equations, and many notions have been studied and explored. Numerous works have been devoted to the observation problem in the whole domain, see (Gilliam and Martin, 1988; Kobayashi, 1980). But some delicate problems need to be studied only in some subregion of the system evolution domain. This is the subject of the regional control theory of distributed parameter systems (DPSs), which was pioneered by El Jai and his co-workers since the 1990s and consists in studying notions related to control and observation only on a subregion of the system evolution domain. The reader may find interesting deve- lopments of these topics for parabolic and hyperbolic sys- tems, see (Amouroux et al., 1994), when the subregion is interior to the system domain; one also finds examples for a system which is not observable (controllable) within the
whole domain Ω but observable (controllable) in a subre- gion ω ⊂ Ω. An extension of these results to a bounda- ry subregion for parabolic systems was then discussed in (Zerrik et al., 2002).
There are many applications of these notions and an interesting one may be the problem of determining lami- nar boundary flux conditions developed in a steady-state by a vertical heated plate and consists of the study of the thermal transfer by natural convection generated by a uni- formly heated plate located in a small enclosure. Inside that enclosure, differences in the wall surface produce na- tural convection movements. The heat exchanger mainta- ins a prescribed temperature on the back face of the plate by means of hot water circulation. All the faces of this ac- tive wall are insulated except for the front face. The objec- tive is to find the unknown boundary convective condition on a part of the front face of the active plate using measu- rements given by internal thermocouples, see (Aparron, 1963) for more details.
There is an extensive literature on the exact and ap-
proximate regional observability problem for linear pa-
228
rabolic systems, but very little has been done for hyper- bolic ones. Recently, regional observability for hyperbo- lic systems has been introduced in the internal case and the developed theory leads to interesting results perfor- med through numerical examples and simulations (Zerrik et al., 2007).
Here we are interested in the regional observabili- ty of hyperbolic systems where the subregion target is a part of the boundary of the system evolution domain. We establish results which are extensions of those given in (Zerrik et al., 2002) to a class of hyperbolic systems. This is the aim of this paper, which is organized as follows:
In Section 2 we introduce definitions and properties of regional observability and show that regional observabi- lity implies boundary one. In Section 3 we characterize the sensor which ensures regional boundary observability.
In Section 4 we give two approaches for regional recon- struction: the first one is direct and based on pseudoinver- se techniques and the second one uses an extension of the Hilbert uniqueness method. In the last section we give an example of hyperbolic systems in a two-dimensional ca- se which illustrates the obtained results through numerical simulations.
2. Regional observability
Let Ω be an open bounded subset of R
n, with a boun- dary ∂Ω which is regular enough. For T > 0, we set Q = Ω×]0, T [, Σ = ∂Ω×]0, T [ and we consider a sys- tem described by the equation
⎧ ⎪
⎪ ⎪
⎪ ⎨
⎪ ⎪
⎪ ⎪
⎩
∂
2y
∂t
2(x, t) + Ay(x, t) = 0 in Q, y(x, 0) = y
0(x), ∂y
∂t (x, 0) = y
1(x) in Ω,
∂y
∂ν
A(ζ, t) = 0 on Σ,
(1)
where A is the elliptic differential operator of the second order given by
A = −
ni,j=1
∂
∂x
ia
ij∂
∂x
jwith domain D(A) =
ϕ ∈ H
2(Ω), ∂ϕ
∂ν
A= 0 on ∂Ω
.
We assume that
a
ij= a
ji∈ C
1(Ω) and there exists α > 0 such that
n i,j=1a
ijζ
iζ
j≥ α
nj=1
| ζ
j|
2, ∀ζ = (ζ
1, . . . , ζ
n) ∈ R
n,
where
∂y
∂ν
A=
n i,j=1a
ij∂y
∂x
jη
iis the conormal derivative of the operator A, and η
istands for the i-th component of the conormal η to ∂Ω.
We consider the state space F = H
2(Ω) × H
1(Ω) and O = L
2(0, T ; R
q) as the observation space. The sys- tem (1) is augmented with the output
z(t) = Cy(t), (2)
where C : H
2(Ω) −→ R
qis the observation operator, q indicates the number of the sensors considered, see (El Jai and Pritchard, 1988).
Let the operator
A =
0 I A 0
be defined by
A(z
1, z
2) = (z
2, Az
1), for all
(z
1, z
2) ∈ D(A) = D(A) × H
1(Ω).
This operator generates a semigroup (S(t))
t≥0.
Denoting ¯ y = (y, ∂y/∂t), the system (1) may be written in the following form:
∂ ¯ y
∂t = A¯ y in Q,
¯
y(0) = ¯ y
0in Ω,
(3)
where ¯ y
0= (y
0, y
1) and the output function (2) takes the form
¯
z(t) = C ¯ y(t), (4)
for C = (C, 0). Then the system (3) admits a unique solu- tion given by
¯
y(t) = S(t)¯ y
0. (5) Using (4), we obtain
¯
z(t) = (K ¯ y
0)(t), (6) where K : H
2(Ω) × H
1(Ω) −→ O = L
2(0, T ; R
q) given by Ky = CS(.)y.
Let w
mbe the basis of eigenfunctions of the operator
A and λ
mbe the associated eigenvalues with multiplici-
ties r
m. Then the semigroup (S(t))
t≥0generated by the
operator A is given by S(t)(y
1, y
2)
=
⎛
⎜ ⎜
⎜ ⎜
⎜ ⎜
⎜ ⎜
⎜ ⎜
⎜ ⎜
⎜ ⎜
⎜ ⎜
⎝
∞ m=1rm
j=1
y
1, w
mjcos
−λ
mt
+ √ 1
−λ
my
2, w
mjsin
−λ
mt
× w
mj(·)
∞ m=1rm
j=1
−
−λ
my
1, w
mjsin
−λ
mt +
y
2, w
mjcos
−λ
mt
w
mj(·)
⎞
⎟ ⎟
⎟ ⎟
⎟ ⎟
⎟ ⎟
⎟ ⎟
⎟ ⎟
⎟ ⎟
⎟ ⎟
⎠ .
Consider a regular boundary portion Γ of ∂Ω with a positive measure and let ω an open part of Ω with a regu- lar boundary ∂ω such that Γ ⊂ ∂Ω ∩ ∂ω. Then one can consider restriction operators defined by
χ
ω: H
2(Ω) × H
1(Ω) −→ H
2(ω) × H
1(ω) (y
1, y
2) −→ χ
ω(y
1, y
2)
= (y
1, y
2) |
ω,
χ
Γ: H
3/2(∂Ω) × H
1/2(∂Ω) −→ H
3/2(Γ) × H
1/2(Γ) (y
1, y
2) −→ χ
Γ(y
1, y
2)
= (y
1, y
2) |
Γ, and the trace operator
γ
0: H
2(Ω) × H
1(Ω) −→ H
3/2(∂Ω) × H
1/2(∂Ω), where γ
0∗, χ
∗Γand χ
∗ωdenote respectively the adjoints of γ
0, χ
Γand χ
ω. We denote by K
∗the adjoint of K and obtain what follows.
Definition 1.
1. The system (1) augmented with (2) is said to be exac- tly (resp. weakly) observable in ω if Imχ
ωK
∗= H
2(ω) × H
1(ω) (resp. Imχ
ωK
∗= H
2(ω) × H
1(ω)).
2. The system (1) augmented with (2) is said to be exac- tly (resp. weakly) observable on Γ if Imχ
Γγ
0K
∗= H
3/2(Γ) × H
1/2(Γ) (resp Imχ
Γγ
0K
∗= H
3/2(Γ) × H
1/2(Γ)).
Remark 1. If the system is exactly (resp. approximately) observable in ω, then it is exactly (resp. approximately) observable in every subset ω
1⊂ ω.
Problem. Given the system (1) augmented with the out- put (2), is it possible to reconstruct the initial state of the system (1) on Γ?
From the above definitions we have the following.
Proposition 1. If the system (1) augmented with (2) is exactly (resp. weakly) observable in ω, then it is exactly (resp. weakly) observable on Γ.
Proof. (Part 1) Let us show that if the system (1) is exactly observable in ω then it is exactly observable on Γ. For this purpose, it is sufficient to show that
H
3/2(Γ) × H
1/2(Γ) ⊂ Imχ
Γγ
0K
∗.
Let (y
1, y
2) ∈ H
3/2(Γ) × H
1/2(Γ) and let (˜ y
1, ˜ y
2) be its extension to H
3/2(∂Ω) × H
1/2(∂Ω).
Applying the trace theorem, there exists a continuous harmonic operator
R : H
3/2(∂Ω) × H
1/2(∂Ω) −→ H
2(Ω) × H
1(Ω) such that
γ
0R(˜y
1, ˜ y
2) = (˜ y
1, ˜ y
2), which yields
χ
ωR(˜y
1, ˜ y
2) ∈ H
2(ω) × H
1(ω).
Since the system (1) is exactly observable in ω, there exists z
1∈ O such that
χ
ωK
∗z
1= χ
ωR(˜y
1, ˜ y
2) and then
γ
0(χ
∗ωχ
ωK
∗z
1) = γ
0(χ
∗ωχ
ωR(˜y
1, ˜ y
2)).
Thus
χ
Γ(γ
0(χ
∗ωχ
ωK
∗z
1)) = χ
Γ(γ
0(χ
∗ωχ
ωR(˜y
1, ˜ y
2))
= (y
1, y
2).
Using the fact that
χ
Γ( γ
0(χ
∗ωχ
ωK
∗z
1)) = χ
Γ(γ
0(K
∗z
1)), we have
χ
Γ( γ
0(K
∗z
1)) = (y
1, y
2)
which means that the system (1)–(2) is exactly observable on Γ.
(Part 2) We must now show that
∀ε > 0, ∀(y
1, y
2) ∈ H
3/2(Γ) × H
1/2(Γ), ∃z
1∈ O, χ
Γ( γ
0(K
∗z
1)) − (y
1, y
2) < ε.
Let (y
1, y
2) ∈ H
3/2(Γ) × H
1/2(Γ) and (˜ y
1, ˜ y
2) be its extension to H
3/2(∂Ω) × H
1/2(∂Ω). By the trace theorem, there exists R(˜y
1, ˜ y
2) ∈ H
2(Ω) × H
1(Ω) such that
γ
0(R(˜ y
1, ˜ y
2)) = (˜ y
1, ˜ y
2).
Since
χ
ω(R(˜ y
1, ˜ y
2)) ∈ H
2(ω) × H
1(ω)
230
and the system (1)–(2) is weakly observable in ω, we have
∀ε > 0, ∃z
1∈ O,
χ
ω(K
∗z
1)) − χ
ω(R(˜ y
1, ˜ y
2)) < ε. (7)
By the continuity of the trace map γ we have γ
0(χ
∗ωχ
ω(K
∗z
1) − χ
∗ωχ
ω(R(˜ y
1, ˜ y
2)))
≤ χ
∗ωχ
ω(K
∗z
1) − χ
∗ωχ
ω(R(˜ y
1, ˜ y
2))
H2(Ω)×H1(Ω)= χ
ω(K
∗z
1) − χ
ω(R(˜ y
1, ˜ y
2))
H2(ω)×H1(ω)and
χ
Γγ
0(χ
∗ωχ
ω(K
∗z
1) − χ
∗ωχ
ωR(˜y
1, ˜ y
2))
≤ γ
0(χ
∗ωχ
ω(K
∗z
1) − χ
∗ωχ
ω(R(˜ y
1, ˜ y
2))), but
χ
Γγ
0(χ
∗ωχ
ΓK
∗z
1− χ
∗ωχ
ωR(˜y
1, ˜ y
2))
= χ
Γγ
0(K
∗z
1− R(˜y
1, ˜ y
2))
= χ
Γγ
0K
∗z
1− (y
1, y
2), which gives
χ
Γγ
0K
∗z
1− (y
1, y
2) ε.
Then the system (1) is weakly observable on Γ.
Proposition 2. We have the equivalence between the fol- lowing statements:
1. The system (1)–(2) is exactly observable on Γ.
2. There exists c > 0 such that
||y||
H32(Γ)×H12(Γ)≤ c||H
∗y ||
O,
∀y = (y
1, y
2) ∈ H
32(Γ) × H
12(Γ), (8)
where H = χ
Γγ
0K
∗.
Proof. The proof uses the following general results (see (Avdonin and Ivanov, 1978)). Let E, F, and G, be reflexi- ve Banach spaces and f ∈ L(E, G), g ∈ L(F, G). Then we have the equivalence between the statements below:
1. Im(f ) ⊂ Im(g).
2. There exists c > 0 such that ||f
∗y ||
E∗≤ ||g
∗y ||
F∗. We set E = G = H
32(Γ) × H
12(Γ), F = O, f = Id
H32(Γ)×H12(Γ)and g = H, and obtain the inequality (8).
3. Characterization of Γ-strategic sensors
In this section, we shall characterize sensors which ensure approximate regional boundary observability in a portion Γ of the boundary ∂Ω. Let us reconsider the system (1) with measurements given by
z(t) = (z
1(t), . . . , z
q(t)), t ∈]0, T [, (9) where
z
i(t) =
⎧ ⎪
⎪ ⎪
⎨
⎪ ⎪
⎪ ⎩
y(b
i, t), b
i∈ ¯Ω in the pointwise case,
Di
y(x, t)f (x)dx, D
i⊂ ¯Ω in the zonal case.
Definition 2. A sequence of sensors (D
i, f
i)
1≤i≤qis said to be Γ-strategic if the observed system is weakly obse- rvable on Γ, which is equivalent to Ker(Kγ
0∗χ
∗Γ) = {0}.
We define the restriction operators χ
1Γ: H
12(∂Ω) −→ H
12(Γ) and
χ
2Γ
: H
32(∂Ω) −→ H
32(Γ),
while χ
1∗Γand χ
2∗Γdenote respectively the adjoints of χ
1Γand χ
2Γ. Let the trace function be given by
γ
01: H
1(Ω) −→ H
12(∂Ω) and
γ
02: H
2(Ω) −→ H
32(∂Ω),
where γ
01∗and γ
02∗denote respectively the adjoints of γ
01and γ
02.
Assume that (χ
1Γγ
10
w
mj)
1≤j≤rm;1≤mform a com- plete set in H
12(Γ) and that (χ
2Γγ
20
w
mj)
1≤j≤rm;1≤mform a complete set in H
32(Γ). Suppose that r = sup r
m< ∞.
Then we have the following result.
Proposition 3. If the observation time T is large enough, the sequence of sensors (D
i, f
i)
1≤i≤qis Γ-strategic if and only if
(i) q ≥ r,
(ii) rank G
m= r
m, ∀m ≥ 1, where
(G
m)
i,j=
w
mj, f
iL2(Ωi)
for the zonal case,
w
mj(b
i) for the pointwise case,
for 1 ≤ i ≤ q and 1 ≤ j ≤ r
m.
Proof. (Sufficiency) Let us show that if rank G
m= r
m, ∀m ≥ 1, then the system (1)–(9) is weakly observable on Γ.
We suppose that Ker(Kγ
0∗χ
∗Γ
) = {0}. Then the- re exists z
∗= (z
∗1, z
∗2) ∈ H
32(Γ) × H
12(Γ) such that (z
1∗, z
2∗) = 0 and Kγ
0∗χ
∗Γz
∗= 0. Therefore
Kγ
0∗χ
∗Γz
∗=
∞ m=1rm
j=1
γ
02∗χ
2∗Γz
1∗, w
mjcos
−λ
mt
+ √ 1
−λ
mγ
01∗χ
1∗Γz
2∗, w
mjsin
−λ
mt
× w
mj, f
i= 0, ∀i = 1, q, where ψ
m1j= χ
1Γγ
10
w
mjand ψ
2mj= χ
2Γγ
20
w
mj. For T large enough, the set
{cos(
−λ
m·), sin(
−λ
m·)}
m≥1forms a complete set in L
2(0, T ), which gives
⎧ ⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪ ⎨
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎩
rm
j=1
γ
2∗0
χ
2∗Γz
1∗, w
mjw
mj, f
i= 0,
∀m ≥ 1, ∀i = 1, . . . , q.
rm
j=1
γ
01∗χ
1∗Γz
2∗, w
mjw
mj, f
i= 0,
∀m ≥ 1, ∀i = 1, . . . , q, i.e.,
⎧ ⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪ ⎨
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎩
rm
j=1
z
∗1, χ
2Γ
γ
20
w
mjw
mj, f
i= 0,
∀m ≥ 1, ∀i = 1, . . . , q.
rm
j=1
z
∗2, χ
1Γ
γ
10
w
mjw
mj, f
i= 0,
∀m ≥ 1, ∀i = 1, . . . , q.
(10)
Since z
∗2∈ H
12(Γ), we obtain
z
2∗=
∞ m=1rm
j=1
z
2∗, χ
1Γγ
01w
mjH12(Γ)
χ
1Γγ
01w
mj. If z
∗2= 0, there exists m
0≥ 1 and 1 ≤ j
0≤ r
m0such that
z
∗2, χ
1Γ
γ
10
w
m0j0
H12(Γ)
= 0.
Consider z
2m0
= z
2∗, χ
1Γ
γ
10
w
m01
H12(Γ)
, . . . ,
z
2∗, χ
1Γ
γ
10
w
m0rm0
H12(Γ)
T.
Using (10), we obtain G
m0
z
m20= 0,
which shows that rank G
m0= r
m0. Similar results can be obtained if we take z
1∗= 0.
(Necessity) Conversely, we show that if the system (1)–(9) is weakly observable on Γ, then
rank G
m= r
m, ∀m ≥ 1.
Suppose that there exists m
0≥ 1 such that rank G
m0=
r
m0, that is, there exists
z
m0= (z
m01, . . . , z
m0rm0)
T= 0 and
G
m0
z
m0= 0.
Let z
2∗∈ H
12(Γ) and z
1∗∈ H
32(Γ) satisfying
⎧ ⎪
⎪ ⎨
⎪ ⎪
⎩
z
2∗, χ
1Γγ
01w
m0j= z
m0j, ∀j = 1, . . . , r
m0,
z
2∗, χ
1Γγ
01w
mj= 0, ∀m = m
0and
∀j = 1, . . . , r
m, (11) and ⎧
⎪ ⎪
⎨
⎪ ⎪
⎩
z
1∗, χ
2Γγ
02w
m0j= z
m0j, ∀j = 1, . . . , r
m0,
z
1∗, χ
2Γγ
02w
mj= 0, ∀m = m
0and
∀j = 1, . . . , r
m. (12) Since G
m0z
m0= 0 and using (10)–(12), the system is not weakly observable on Γ which contradicts the above
assumption.
Remark 2. The choice of q = 1 can be sufficient to ensu- re the system observability. Indeed, we can show that by means of a weak perturbation of the boundary of the sys- tem domain Ω, the multiplicity of the eigenvalues may be reduced to one.
In the following, we shall give two approaches which enable the reconstruction of the initial conditions on the boundary part Γ of ∂Ω.
4. Boundary state reconstruction
4.1. Direct approach. Let us consider the decomposi- tion of the initial state ¯ y
0in the form
¯ y
0=
⎧ ⎪
⎪ ⎨
⎪ ⎪
⎩
¯
y
01on Γ,
¯
y
02on ∂Ω \ Γ,
¯
y
03in Ω,
with ¯ y
01= (y
10, y
11), ¯ y
20= (y
02, y
21), ¯ y
03= (y
30, y
13). Our
objective is the reconstruction of ¯ y
0, the initial state on Γ.
232
Let ε be the observation error between the output function ¯ z and the observation model given by
ε(¯ y
10, ¯ y
02, ¯ y
03) = ||¯ z − K ¯y
0||
2O,
where ε is to be meant as a function of the initial state and its components.
Consider the following optimization problem:
min ε(¯ y
01, ¯ y
20, ¯ y
30)
¯
y
01∈ H
32(Γ) × H
12(Γ).
(13)
Theorem 1. If the system (1) is observable in ¯ Ω, then the problem (13) admits a unique solution given by
¯
y
01= D
†Ψ, which coincides with the regional ini- tial state to be observed on the boundary part Γ, where D
†denotes the pseudoinverse of D, D and Ψ are given by
⎧ ⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪ ⎨
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎩
D = Bγ
∗0χ
∗Γ(χ
Γγ
0B
∗Bγ
0∗χ
∗Γ)
−1χ
Γγ
0B
∗B,
Ψ = (I − B(B
∗B)
−1+ Bγ
0∗χ
∗Γ(χ
Γγ
0B
∗Bγ
0∗χ
∗Γ)
−1×χ
Γγ
0B
∗ν, R = γ
0K
∗, M = (RR
∗)
−1,
ν = (I + R
∗M R − K(K
∗K)
−1K
∗), B = R
∗M RK.
(14) Proof. We have
RR
∗∈ L(H
32(∂Ω) × H
12(∂Ω), H
32(∂Ω) × H
12(∂Ω)) and
RR
∗z
∗, z
∗= ||Kγ
0∗z
∗||
2≤ c||z
∗||
2H32(∂Ω)×H12(∂Ω)
,
∀z
∗∈ H
32(∂Ω) × H
12(∂Ω), which shows that the operator RR
∗is invertible.
Solution of the problem (13) can be implemented using the followings steps:
Step 1: We minimize ε(¯ y
10, ¯ y
02, ¯ y
03) with respect to ¯ y
30. By developing ε and using the fact that ¯ y
0= ¯ y
10∗+ ¯ y
0∗2+ ¯ y
0∗3, where
¯ y
0∗1=
y ¯
10in Γ, 0 on Ω \ Γ, ¯
¯ y
0∗2=
y ¯
20in ∂Ω \ Γ, 0 on Ω ∪ Γ, ¯
¯ y
0∗3=
y ¯
30in Ω, 0 on ∂Ω, we obtain the following problem:
min Ψ
1(¯ y
30∗),
γ
0K
∗(K ¯ y
30∗) = 0, (15)
where
Ψ
1(¯ y
30∗) = K ¯ y
30∗, K ¯ y
30∗O
+ 2 K ¯ y
30∗, K(¯ y
0∗1+ ¯ y
0∗2) − ¯ z
O. Consider the Lagrangian operator defined by Λ(¯ y
0∗3, λ
1)
= K ¯ y
30∗, K ¯ y
0∗3O
+ 2 K ¯ y
0∗3, K(¯ y
0∗1+ ¯ y
0∗2) − ¯ z
O+ λ
1, γ
0K
∗K ¯ y
0∗3H32(∂Ω)×H12(∂Ω)
.
Then the condition
∂Λ(¯ y
30∗, λ
1)
∂ ¯ y
0∗3= 0 is equivalent to
2K
∗K ¯ y
30∗+ 2K
∗K(¯ y
0∗1+ ¯ y
0∗2)
− 2K
∗z + K ¯
∗Kγ
∗0
λ
1= 0 Since the system (1) is observable in Ω, (K
∗K)
−1exists (Amouroux et al., 1994), which allows us to write
¯
y
30∗= −(¯ y
10∗+ ¯ y
20∗) + (K
∗K)
−1K
∗z ¯ − 1 2 γ
0∗λ
1. The constraint γ
0K
∗K ¯ y
30∗= 0 from the condition ¯ y
0∗3= 0 on ∂Ω gives
−γ
0K
∗K(¯ y
10∗+ ¯ y
0∗2) + γ
0K
∗z ¯ − 1
2 γ
0∗(K
∗K)γ
∗0λ
1= 0, and then
λ
1= 2M R(¯ z − K(¯y
0∗1+ ¯ y
20∗)), where M and R are given by (14).
Consequently,
˜
y
0∗3= (γ
∗0M RK − I)(¯y
10∗+ ¯ y
0∗2) + ((K
∗K)
−1K
∗− γ
0∗M R)¯ z.
Then the minimum is given by the following equiva- lent problem:
ε(¯ y
10∗, ¯ y
0∗2, ˜ y
30∗) = ||ν ¯ z − B(¯y
0∗1+ ¯ y
0∗2)||
2O, (16) where B and ν are given by (14).
Step 2: We minimize ε(¯ y
0∗1, ¯ y
0∗2, ˜ y
30∗) with respect to ¯ y
20∗, as in the first step, by developing (16). We obtain the fol- lowing problem:
min Ψ
2(¯ y
0∗2),
χ
Γγ
0B
∗B(¯ y
20∗) = 0, (17) where
Ψ
2(¯ y
0∗2) = B ¯ y
20∗, B ¯ y
0∗2O
+ 2 B ¯ y
20∗, B ¯ y
0∗2− ν¯z
O.
The Lagrangian is given by
Λ
2(¯ y
0∗2, λ
2) = B ¯ y
20∗, B ¯ y
20∗O
+ 2 B ¯ y
20∗, B ¯ y
20∗− ν¯z
O+ λ
2, χ
Γγ
0B
∗B ¯ y
0∗2H32(Γ)×H12(Γ)
. Then the condition
∂Λ(¯ y
20∗, λ
2)
∂ ¯ y
20∗= 0 is equivalent to
2B
∗B ¯ y
20∗+ 2B
∗B ¯ y
0∗1− 2B
∗ν ¯ z + B
∗Bγ
0∗χ
∗Γλ
2= 0.
We obtain
¯
y
0∗2= −¯ y
0∗1+ (B
∗B)
−1B
∗ν ¯ z − 1
2 γ
0∗χ
∗Γλ
2. The constraint χ
Γγ
0B
∗B(¯ y
20∗) = 0, which comes from
¯
y
0∗2= 0 on Ω ∪ Γ, gives
−χ
Γγ
0B
∗B ¯ y
10∗+ χ
Γγ
0B
∗ν ¯ z − 1
2 χ
Γγ
0B
∗Bγ
0∗χ
∗Γλ
2= 0.
We obtain
λ
2=2(χ
Γγ
0B
∗Bγ
∗0
χ
∗Γ
)
−1χ
Γγ
0B
∗ν ¯ z
− 2(χ
Γγ
0B
∗Bγ
0∗χ
∗Γ)
−1χ
Γγ
0B
∗B ¯ y
0∗1. Thus
˜
y
20∗= − ¯ y
0∗1+
(B
∗B)
−1− γ
0∗χ
∗Γ(χ
Γγ
0B
∗Bγ
0∗χ
∗Γ)
−1χ
Γγ
0B
∗ν ¯ z + γ
0∗χ
∗Γ(χ
Γγ
0B
∗Bγ
0∗χ
∗Γ)
−1χ
Γγ
0B
∗B ¯ y
10∗. The minimum is then given by the equivalent problem
ε(¯ y
10∗, ˜ y
0∗2, ˜ y
30∗) = ||Ψ¯ z − D¯y
10∗||
2O, (18) where Ψ and D are given by (14).
Step 3: The solution of the problem (13) turns out to solve the problem ε(¯ y
0∗1, ˜ y
20∗, ˜ y
30∗) given by (18) with respect to
¯
y
0∗1. We have ε(¯ y
10∗, ˜ y
0∗2, ˜ y
0∗3)
= D¯ y
0∗1, D ¯ y
0∗1O
− 2 D¯y
10∗, Ψ¯ z
O+ Ψ¯ z, Ψ¯ z
O. The condition
ε(¯ y
0∗1, ˜ y
20∗, ˜ y
30∗)
∂ ¯ y
10∗= 0 gives
2D
∗D ¯ y
10∗− 2D
∗Ψ¯ z = 0.
Thus
˜
y
10∗= (D
∗D)
−1DΨ¯ z = D
†Ψ¯ z.
Since the operator (18) is strictly convex, then ˜ y
10∗is uni- que.
4.2. Hilbert uniqueness method approach. The sub- ject of this section is to update the Hilbert uniqueness me- thod developed by Lions (1988) to the case of regional observability of following hyperbolic system:
⎧ ⎪
⎪ ⎪
⎪ ⎪
⎨
⎪ ⎪
⎪ ⎪
⎪ ⎩
∂
2y(x, t)
∂t
2= Δy(x, t), Q,
y(x, 0) = y
0(x), ∂y
∂t (x, 0) = y
1(x), Ω,
∂y(ζ, t)
∂η = 0, Σ,
(19) where Δ is the Laplacian operator and we assume that (19) is augmented with the output function
z(t) = y(t), f
L2(D), (20) where f ∈ L
2(D).
We consider the following decomposition:
y
0=
y
10in ω, y
20in Ω \ ω, y
1=
y
11in ω, y
21in Ωω.
In the sequel, without loss of generality, we assume that the eigenfunctions (w
m)
m≥1of the operator Δ associated with the eigenvalues are simple.
Consider the set
G = {χ
∗ωχ
ω(ϕ
0, ϕ
1) | (ϕ
0, ϕ
1) ∈ D(A) × H
1(Ω) with ϕ
0= ϕ
1= 0 in Ω \ ω}.
For (ϕ
0, ϕ
1) ∈ G, the system
⎧ ⎪
⎪ ⎪
⎪ ⎪
⎨
⎪ ⎪
⎪ ⎪
⎪ ⎩
∂
2ϕ(x, t)
∂t − Δϕ(x, t) = 0, Q,
ϕ(x, 0) = ϕ
0(x), ∂ϕ(x, 0)
∂t = ϕ
1(x), Ω,
∂ϕ(ζ, t)
∂η = 0, Σ,
(21)
admits a unique solution
ϕ ∈ C(0, T ; H
2(Ω))∩C
1(0, T ; H
1(Ω))∩C
2(0, T ; L
2(Ω)) (see (Lions and Magenes, 1968)).
We define a semi-norm on G by
(ϕ
0, ϕ
1)
G=
T0
ϕ(t), f
2L2(D)dt
!
12, (22) and we consider the retrograde system
⎧ ⎪
⎪ ⎪
⎪ ⎪
⎨
⎪ ⎪
⎪ ⎪
⎪ ⎩
∂
2ψ(x, t)
∂t = Δψ(x, t)+ ϕ(t), f
L2(D)χ
Df (x), Q, ψ(x, T ) = 0, ∂ψ(x, T )
∂t = 0, Ω,
∂ψ(ζ, t)
∂η = 0 Σ,
(23)
234
which admits a unique solution
ψ ∈ C(0, T ; H
2(Ω))∩C
1(0, T ; H
1(Ω))∩C
2(0, T ; L
2(Ω)) (see (Lions and Magenes, 1968)).
Let the operator Λ be defined by Λ(ϕ
0, ϕ
1) = P(− ´ ψ(0), ψ(0)),
where P = χ
∗ωχ
ωand χ
∗ωis the adjoint operator of χ
ω. Consider the system
⎧ ⎪
⎪ ⎪
⎪ ⎪
⎨
⎪ ⎪
⎪ ⎪
⎪ ⎩
∂
2z(x, t) ¯
∂t = Δ¯ z(x, t) + z(t)χ
D(x)f (x), Q,
¯
z(x, T ) = 0, ∂ ¯ z(x, T )
∂t = 0, Ω,
∂ ¯ Z(ζ, t)
∂η = 0, Σ.
(24) We have ¯ z(0) = ¯ z
0, ¯ z
(0) = ¯ z
1, ψ(0) = ψ
0and ψ
(0) = ψ
1.
If ϕ
0, ϕ
1are convenably chosen on G, i.e., such that
¯
z
0= ψ
0and ¯ z
1= ψ
1in ω, then the observability pro- blem for the system (19) in the subregion ω amounts to solving the equation
Λ(ϕ
0, ϕ
1) = P(−¯ z
1, ¯ z
0). (25) Theorem 2. If the system (19) augmented with the output function (20) is weakly observable in ω, then the equation (25) admits a unique solution (ϕ
0, ϕ
1) ∈ G, which coinci- des with the initial conditions (y
01, y
11) in the subregion ω, and the initial conditions to be observed in the subregion Γ of ∂Ω are given by
(y
0Γ, y
Γ1) = χ
Γγ
0(ϕ
0, ϕ
1). (26) Proof. (Part 1) We show that if the sensor (D, f ) is ω- strategic, then the formula (22) defines a seminorm on G.
Indeed,
(ϕ
0, ϕ
1)
G
= 0 ⇔ ϕ(t), f
L2(D)= 0
⇔ CS(t)(ϕ
0, ϕ
1) = 0, where (S(t))
t≥0is the semigroup generated by 0 I
Δ 0
. Then we have
K(t)χ
∗ωχ
ω( ˜ ϕ
0, ˜ ϕ
1) = 0 with
(ϕ
0, ϕ
1) = χ
∗ωχ
ω( ˜ ϕ
0, ˜ ϕ
1).
Since the system (1) is weakly ω-observable, we have χ
ω( ˜ ϕ
0, ˜ ϕ
1) = 0
and then
( ˜ ϕ
0, ˜ ϕ
1) = 0.
Consequently, ϕ
0= ϕ
1= 0. Thus (22) is a norm.
(Part 2) Let ˆ G be the completion set of G with respect to the norm (22) equipped with the associated inner product ·, ·
Gˆand ˆ G
∗be its dual. We show that Λ is an isomor- phism from ˆ G onto ˆ G
∗.
Indeed,
Λ(ϕ
0, ϕ
1), (ϕ
0, ϕ
1)
=
(−ψ
1, ψ
0), (ϕ
0, ϕ
1) . On the other hand, multiplying (23) by ϕ (the solution to (21)) and integrating the result by part, we obtain
T0
ψ
(t), ϕ(t)
dt
=
ψ
(T ), ϕ(T )
−
ψ
(0), ϕ
0−
ψ(T ), ϕ
(T )
+
ψ(0), ϕ
1+
T0
ψ(t), ϕ
(t)
dt
=
T0
ψ(t), ϕ
(t)
dt +
(−ψ
(0), ψ(0)), (ϕ
0, ϕ
1)
(27)
Using Green formulae, we have
T0
Δψ(t), ϕ(t) dt
=
T0
ψ(t), Δϕ(t) dt +
∂ψ
∂η ϕ dσ
−
ψ ∂ϕ
∂η dσ
=
T0
ψ(t), Δϕ(t) dt.
(28)
Thus
T0
ψ
(t), ϕ(t)
− Δψ(t), ϕ(t) dt
=
(−ψ
(0), ψ(0)), (ϕ
0, ϕ
1)
. On the other hand, we have
T0
ϕ(t), f
L2(D)χ
Df, ϕ(t) dt
=
T0