OPTIMIZING THE LINEAR QUADRATIC MINIMUM–TIME PROBLEM FOR DISCRETE DISTRIBUTED SYSTEMS
M OSTAFA RACHIK ∗ , A HMED ABDELHAK ∗
∗ Faculté des Sciences Ben M’Sik, Département de Mathématiques Université Hassan II Mohammadia, Casablanca, Morocco
e-mail: abdelllak@hotmail.com
With reference to the work of Verriest and Lewis (1991) on continuous finite-dimensional systems, the linear quadratic minimum-time problem is considered for discrete distributed systems and discrete distributed time delay systems. We treat the problem in two variants, with fixed and free end points. We consider a cost functional J which includes time, energy and precision terms, and then we investigate the optimal pair (N, u) which minimizes J .
Keywords: discrete distributed systems, time delay systems, minimum time, optimal control
1. Introduction
The linear quadratic minimum-time problem was con- sidered before (Athans and Falb, 1996; Schwartz and Gourdeau, 1989), but is was not fully exploited. Verri- est and Lewis (1991) treat the case of continuous finite- dimensional systems. Discrete systems in the finite- dimensional case were considered later (El Alami et al., 1998). In the present paper, we investigate discrete-time distributed systems. In the first part of this work, we con- sider systems described by
x(i + 1) = Ax(i) + Bu(i), 0 ≤ i ≤ N − 1, x(0) = x 0 ,
(1)
where N is taken to be free, x(i) ∈ X is the state vari- able and u(i) ∈ U is the input variable. X and U are Hilbert spaces, the operators A and B are bounded (A ∈ L(X) and B ∈ L(U, X)).
We consider a cost functional J (N, u) which in- cludes time and energy, that is to say,
J (N, u) = ϕ(N ) +
N −1
X
i=0
hu(i), Ru(i)i, (2)
where u = (u(0), . . . , u(N − 1)) ∈ U N , ϕ: N → R + is assumed to be positive and increasing, i.e.
ϕ(N ) ≥ 0, ∀ N ∈ N,
N ≤ M ⇒ ϕ(N ) ≤ ϕ(M ), ∀ N, M ∈ N, and
lim
N →+∞ ϕ(N ) = +∞. (3)
R ∈ L(U ) is a self-adjoint positive definite operator.
Then we investigate the optimal pair (N ∗ , u ∗ ) ∈ N ∗ × U N
∗which minimizes the cost functional J (N, u) un- der constraints
(N, u) ∈ {(M, v) ∈ N × U N : x v (M ) = x d }, where N ∗ is taken to be as small as possible, x d is a given desired final state, x v (·) is the trajectory of system (1) corresponding to the control v, and N ∗ is the set of all non-zero integers.
We establish that the optimal solution (N ∗ , u ∗ ) ex- ists, is unique and is obtained by solving a finite sequence of algebraic equations and by minimizing a time func- tional over a finite sub-interval of N. An example is given to illustrate the results. The case where the final end point x(N ) is free, is also considered. In this case, the func- tional cost includes time, energy and precision terms, i.e.
J (N, u) = ϕ(N )+
N −1
X
i=0
hu(i), Ru(i)i+hx(i), M x(i)i
+ hx(N ), Gx(N )i, (4)
where M, G ∈ L(X) are self-adjoint positive operators.
Since J contains both the final time N and quadratic components of x(i) and u(i), we shall call J a linear quadratic minimum-time performance index. In the sec- ond part of this paper, we treat the case of discrete dis- tributed time delay systems. To settle the problem, we de- fine a new state variable which satisfies a discrete system without delays.
In what follows, we denote by h·, ·i and h·, ·i U the
inner products defined respectively on X and U . We also
denote by N ∗ and R ∗ the set of non-zero integers and the set of non-zero reals, respectively.
2. The Case of a Fixed End Point
Consider the linear discrete-time system given by
x(i + 1) = Ax(i) + Bu(i), 0 ≤ i ≤ N − 1, x(0) = x 0 ,
(5)
where N is free, x(i) ∈ X is the state variable and u(i) ∈ U is the input variable. X and U are Hilbert spaces, A ∈ L(X) and B ∈ L(U, X). Let ϕ be a posi- tive increasing function such that
N →+∞ lim ϕ(N ) = +∞. (6)
The problem can be stated as follows: Given the per- formance index
J (N, u) = ϕ(N ) +
N −1
X
i=0
hu(i), Ru(i)i (7)
and a desired final state x d ∈ X, we investigate the opti- mal pair (N ∗ , u ∗ ) ∈ N ∗ × U N
∗where N ∗ is as small as possible and
J (N ∗ , u ∗ ) = min
(N,u)∈V
J (N, u), (8)
with V = {(N, u) ∈ N ∗ × U N : x(N ) = x d }.
Definition 1. An integer N is said to be admissible if there exists a control sequence u ∈ U N such that x(N ) = x d .
To determine the optimal sequence (N ∗ , u ∗ ), we proceed as follows: For each admissible inte- ger N , we determine an optimal control u N = (u N (0), . . . , u N (N − 1)) which minimizes the cost J (N, u) over all controls u = (u(0), . . . , u(N −1)) such that x(N ) = x d . The optimal time N ∗ is the smallest integer which minimizes J (N, u N ) over all admissible integers N .
Let N ∈ N be a fixed integer. From (5) it follows that for every control u = (u(0), . . . , u(N − 1)) ∈ U N , we have
x(N ) = A N x 0 + H N u, (9) where H N is the operator defined by
U N → X, H N :
(u(0), . . . , u(N − 1)) ,→
N −1
X
j=0
A N −1−j Bu(j). (10)
Consider the inner product on U N given by
hu, vi R =
N −1
X
i=0
hu(i), Rv(i)i U , (11)
u = (u(0), . . . , u(N − 1)), v = (v(0), . . . , v(N − 1)), and let H N ∗ be the adjoint operator of H N defined with respect to the inner products h·, ·i and h·, ·i R , i.e.
hH N u, xi = hu, H N ∗ xi R , ∀ u ∈ U N , ∀ x ∈ X. (12) Define the functional k · k F
Nby
kxk F
N= kH N ∗ xk R , ∀ x ∈ X, (13) where k · k R is the norm corresponding to the inner prod- uct h·, ·i R . Then the functional k · k F
Ndescribes a semi- norm on X and a norm on F 0 , where F 0 is the subspace of X defined by
F 0 = Im H N = (Ker H N ∗ ) ⊥ . (14) Indeed, if x ∈ F 0 and kxk F
N= 0, then we deduce that x ∈ (Ker H N ∗ ) ∩ (Ker H N ∗ ) ⊥ , which implies that x = 0.
We denote by h·, ·i N the inner product on F 0 given by hx, yi N = hH N ∗ x, H N ∗ yi R , ∀ x, y ∈ F 0 . (15) Now, we introduce the operator
F 0 → F 0 , Λ N :
x ,→ H N H N ∗ x. (16) For every x ∈ F 0 , we have
kΛ N xk F
N= kH N ∗ Λ N xk R = kH N ∗ H N H N ∗ xk R
≤ kH N ∗ H N k kxk F
N.
Hence Λ N is a bounded operator on F 0 endowed with the norm k · k F
N.
Let F N be the completion of F 0 with respect to the norm k · k F
N. Since we have
|hΛ N x, yi| = |hx, yi N | ≤ kxk F
Nkyk F
N, ∀ x, y ∈ F 0 , (17) it is classical that Λ N has a unique extension denoted also by Λ N and defined from F N to its dual F N
0(Li- ons, 1988). Indeed, for any x ∈ F 0 we define the map ψ x by
F 0 → R, ψ x :
y ,→ hΛ N x, yi. (18)
The map ψ x is linear and continous with respect to the
norm k · k F
N. Since F 0 is dense in F N , ψ x can be ex-
tended to a bounded linear operator denoted by ψ x which
belongs to the space F N
0. Now consider the map π de- fined by
Λ N (F 0 ) → F N
0, π :
Λ N x ,→ ψ x . (19) We verify that the map π is well defined on Λ N (F 0 ).
Moreover, π is linear and injectif. This allows us to iden- tify the space Λ N (F 0 ) with a subspace of F N
0. Using the operator π, we rewrite the operator Λ N as follows:
F 0 → F N
0, Λ N :
x ,→ ψ x .
(20)
We show that Λ N is linear and continous with respect to the norm k · k F
N, which implies that Λ N has a linear and bounded extension also denoted by Λ N and defined from F N to its dual F N
0. Moreover, this extension is an isomorphism from F N to F N
0. To show this, we prove that
hΛ N x, xi F
0N
,F
N= kxk 2 F
N
, ∀ x ∈ F N , (21) where we denote by hφ, xi F
0N
,F
Nthe range of x ∈ F N
by the operator φ ∈ F N
0. From (21) it follows that Λ N
is injectif. Consequently, Λ N is an isomorphism from F N to Λ N (F N ). This implies that Λ N (F N ) is a closed subspace of F N
0, and hence Λ N (F N ) = Λ N (F N ). On the other hand, if A ⊂ F N
0, we denote by A ◦ the subspace of F N given by
A ◦ = {x ∈ F N /hφ, xi F
0N
,F
N= 0, ∀ φ ∈ A}. (22) If B ⊂ F N , we denote by B ◦ the subspace of F N
0given by
B ◦ = {φ ∈ F N
0/hφ, xi F
0N
,F
N= 0, ∀ x ∈ B}. (23) Let x ∈ (Λ N (F N )) ◦ . Then from (22) it follows that
hΛ N y, xi F
0N
,F
N= 0, ∀ y ∈ F N . (24) This implies
hΛ N x, xi F
0N
,F
N= 0 = kxk 2 F
N.
Hence x = 0. Consequently, (Λ N (F N )) ◦ = {0}. Thus Λ N (F N ) = Λ N (F N ) = ((Λ N (F N )) ◦ ) ◦ = {0} ◦ = F N
0,
(25) which implies that Λ N is an isomorphism from F N
to F N
0.
Remark 1. Suppose that x ∈ ImH N . Then there exists u ∈ U N such that x = H N u. Consider the function ϕ x
defined by
F 0 → R, ϕ x :
y ,→ hx, yi. (26)
We have
|ϕ x (y)| = |hH N u, yi|
= |hu, H N ∗ yi R | ≤ kuk R kyk F
N, ∀ y ∈ F 0 . Hence ϕ x is a bounded operator on F 0 endowed with the norm F N . Using the Hahn-Banach theorem, we de- duce that ϕ x ∈ F N
0. Consequently, we may assume that ImH N ⊂ F N
0since the map i given by
ImH N → F N
0, ϕ x :
x ,→ ϕ x
(27)
is injectif.
Now, we can formulate the following proposition which characterizes the admissible integers.
Proposition 1. An integer N is admissible if and only if x d − A N x 0 ∈ F N
0.
Proof. If x d − A N x 0 ∈ F N
0, then there exists a unique f ∈ F N such that Λ N f = x d − A N x 0 . Consider the control u = H N ∗ f . Then
x(N ) = A N x 0 + H N u = A N x 0 + Λ N f = x d (28) and hence N is admissible.
Conversely, if N is admissible, then there exists a control u such that x(N ) = x d , which implies x d − A N x 0 = H N u. Hence x d − A N x 0 ∈ ImH N ⊂ F N
0(see Remark 1).
Proposition 2. If N is an admissible integer, then the control u N being a solution to the optimization problem
J (N, u N ) = min
u∈U
NJ (N, u)
subject to x(N ) = x d is given by u N = H N ∗ f , where f ∈ F N is the unique solution of the algebraic equation
Λ N f = x d − A N x 0 . Moreover, the corresponding cost is
J (N, u N ) = ϕ(N ) + kf k 2 F
N.
Proof. Let N be an admissible integer. From Proposi- tion 1 it follows that there exists a unique f ∈ F N such that Λ N f = x d − A N x 0 . Define u = H N ∗ f ∈ U N . Then
x(N ) = A N x 0 + H N u = A N x 0 + Λ N f = x d . On the other hand, for each control v ∈ U N such that x v (N ) = x d , we have
x(N ) = x v (N ) = x d ,
where x v (·) denotes the trajectory of system (5) corre- sponding to the control v. Hence
H N u = H N v, which implies
hH N (u − v), f i = 0 or
hu − v, H N ∗ f i R = 0.
Since u = H N ∗ f , we deduce that hu, ui R = hv, ui R ≤ kvk R kuk R . Thus kuk R ≤ kvk R , ∀ v ∈ U N .
Remark 2.
(a) By convention, if N is not admissible, we set J (N, u N ) = +∞.
(b) In order to obtain the minimizing control u N , we have to solve the algebraic equation Λ N f = x d − A N x 0 . However, we do not in general have an ex- plicit expression for the operator Λ −1 N , so we propose the Galerkin method to approximate f (the bilinear form F N × F N → R : (x, y) ,→ hΛ N x, yi is coer- cive).
Finally, the optimal sequence (N ∗ , u ∗ ) is given by the following proposition.
Proposition 3. Let A be the set of all admissible integers.
If A is bounded, then N ∗ is the smallest integer that minimizes J (N, u N ) over A. Otherwise, consider N 0 ∈ A and M ∈ A such that ϕ(M ) > J(N 0 , u N
0). Then N ∗ is the smallest integer that minimizes J (N, u N ) over the interval [1, M ].
Proof. If A is bounded, the result is obvious. Suppose that A is not bounded and consider N 0 , M ∈ A such that ϕ(M ) > J (N 0 , u N
0). It follows that N 0 ∈ [1, M ].
Indeed, if it is not, then ϕ(N 0 ) ≥ ϕ(M ), which implies J (N 0 , u N
0) ≥ ϕ(N 0 ) ≥ ϕ(M ) > J (N 0 , u N
0), a contradiction. Thus N 0 ∈ [1, M ].
On the other hand, for each N ∈ N such that N >
M , we have ϕ(N ) ≥ ϕ(M ). Consequently, J (N, u N ) ≥ ϕ(N ) ≥ ϕ(M ) > J (N 0 , u N
0).
Example 1. Consider the discrete-time system described by
x(i + 1) = Ax(i) + Bu(i), i = 0, . . . , N − 1, x(0) = 0,
(29)
where N is free, Ω =]0, 1[, x(i) ∈ L 2 (Ω) is the state variable, u i ∈ R is the input variable and
A = S(δ) ∈ L L 2 (Ω), (30) S(t) t≥0 being the strongly continuous semigroup gener- ated by the Laplacian operator ∆, i.e.
S(δ)x =
∞
X
i=1
e −i
2π
2δ hx, e i ie i , ∀ x ∈ L 2 (Ω), (31)
where h·, ·i is the usual inner product on L 2 (Ω), δ > 0 and e i (s) = √
2 sin (iπs), ( (e i ) i is a basis of L 2 (Ω) ).
The operator B is defined by
B = Z δ
0
S(σ)D dσ, (32)
where
R → L 2 (Ω) D :
u ,→ ue 1 (·).
Remark 3. The difference equation (29) can be inter- preted as the sampling version ot the following continuous diffusion system:
∂x
∂t − ∆x = g(s)u(t), s ∈ Ω, t ∈ [0, T ], x(0, ·) = x 0 (·) in Ω,
x(t, s) = 0 in ∂Ω×]0, T [,
(33)
where g = e 1 .
The linear quadratic minimum-time problem consists in determining the optimal pair (N ∗ , u ∗ ) which mini- mizes the cost functional
J (N, u) = N 2 +
N −1
X
i=0
Ru 2 (i) (34)
while driving the system from x 0 = 0 to x d = αe 1 ,
where α ∈ R ∗ is given.
Lemma 1. The space F 0 defined by F 0 = ImH N is given by
F 0 = E(e 1 ),
where E(e 1 ) is the subspace of L 2 (Ω) spanned by the vector e 1 .
Proof. For every N ≥ 1 and every u ∈ R N , we have H N u =
N −1
X
i=0
A N −1−i Bu(i)
=
N −1
X
i=0
S((N − 1 − i)δ)u(i) Z δ
0
S(σ)e 1 dσ
=
N −1
X
i=0
u(i) Z δ
0
S((N − 1 − i)δ + σ)e 1 dσ
=
N −1
X
i=0
u(i) Z δ
0
∞
X
j=1
e −j
2π
2((N −1−i)δ+σ) he 1 , e j ie j dσ
=
N −1
X
i=0
u(i) Z δ
0
e −π
2((N −1−i)δ+σ) dσe 1
= (c
N −1
X
i=0
u(i)e −π
2((N −1−i)δ) )e 1 ,
where c is the constant given by c = R δ
0 e −π
2σ dσ.
Hence ImH N ⊂ E(e 1 ). Conversely, if x ∈ E(e 1 ), there exists β ∈ R such that x = βe 1 . Choose u = (u(0), . . . , u(N −1)) such that u(0) = · · · = u(N −2) = 0 and u(N − 1) = β/c. Then H N u = x and
ImH N = E(e 1 ). (35)
Consequently,
F 0 = ImH N = E(e 1 ), F N = E(e 1 ).
Now, for every integer N ≥ 1 we have x d − A N x 0 ∈ ImH N , since x 0 = 0 and x d ∈ ImH N . Hence from Remark 1 and Proposition 1 it follows that every in- teger N ≥ 1 is admissible. In order to solve the equation Λ N f = x d , we first determine the adjoint operators B ∗ and H N ∗ . By simple calculations we establish that for ev- ery x ∈ L 2 (Ω), we have
B ∗ x = che 1 , xi, (36)
H N ∗ x = (H N ∗ x) 0 , . . . , (H N ∗ x) N −1 , (H N ∗ x) i = R −1 B ∗ A N −1−i x
= c
R e −π
2(N −1−i)δ hx, e 1 i. (37)
Let f ∈ F N (= E(e 1 )) be such that Λ N f = x d in F N
0. Then
hΛ N f, xi = hx d , xi, ∀ x ∈ F 0 . (38) Since f = a N e 1 for some a N ∈ R, (38) implies
a N hΛ N e 1 , βe 1 i = hx d , βe 1 i, ∀ β ∈ R or, equivalently,
a N hH N ∗ e 1 , H N ∗ e 1 i R = hx d , e 1 i = α.
Thus
a N = α
kH N ∗ e 1 k 2 = α ke 1 k 2 F
N
. (39)
Consequently, the optimal cost corresponding to u N is J (N, u N ) = N 2 + kf k 2 F
N
= N 2 + a 2 N ke 1 k 2 F
N
= N 2 + α 2 ke 1 k 2 F
N
. (40)
Using (37), we establish
ke 1 k 2 F
N= c 2 (e −2π
2(N −1)δ − e 2π
2δ ) R(1 − e 2π
2δ ) .
For numerical simulation we take α = 10, δ = 0.1, R = 1, N 0 = 7. Then we apply Proposition 3 to deduce that the minimum time N ∗ exists in the interval [1, 147]
and is equal to 4. The optimal control is u ∗ = H N
∗f , where f = (2132.4)e 1 . The evolution of J (N, u N ) with respect to N is given in Fig. 1.
Fig. 1. The evolution of J (N, u
N) with respect to N .
3. The Case of a Free End Point
In this case, we consider a cost functional J (N, u) which includes time, energy and precision terms, i.e.
J (N, u) = ϕ(N ) +
N −1
X
i=0
hu(i), Ru(i)i + hx(i), M x(i)i
+ hx(N ), Gx(N )i, (41)
where M ∈ L(X), G ∈ L(X) are self-adjoint positive operators and R ∈ L(U ) is a self-adjoint positive definite operator.
Then we investigate the optimal sequence (N ∗ , u ∗ ) where N ∗ is as small as possible and
J (N ∗ , u ∗ ) = min
(N,u)∈N×U
NJ (N, u). (42) To show that this problem has a unique solution (N ∗ , u ∗ ), we proceed in two steps: In the first one, for any fixed integer N , we determine the optimal control u N which minimizes the cost J (N, u) over all controls u ∈ U N . In the second step, we minimize J (N, u N ) over all integers N . By convention, we set
J (0, u) = ϕ(0) + hx 0 , Gx 0 i, ∀ N ∈ N, ∀u ∈ U N . (43) For a fixed N ∈ N ∗ , if we denote by u N ∈ U N the optimal control which satisfies
J (N, u N ) = min
u∈U
NJ (N, u), (44) then u N is unique and given by the following proposition:
Proposition 4. Let N ∈ N ∗ and K i : X → X, i = 0, . . . , N − 1 be a family of operators given by
K i+1 = A ∗ K i (I + BR −1 B ∗ K i ) −1 A + M, i = 0, . . . , N − 1, K 0 = G.
Given an initial condition x 0 ∈ X, the optimal control u N is given in feedback form by
u N (i) = −R −1 B ∗ K N −1−i (I + BR −1 B ∗ K N −1−i ) −1
× Ax(i), i = 0, . . . , N − 1.
The corresponding cost is
J (N, u N ) = hK N x 0 , x 0 i.
Proof. For the proof, see (Zabczyk, 1974).
Finally, the optimal pair (N ∗ , u ∗ ) being a solution of (42) is determined by the following result:
Proposition 5. Consider (N 0 , M ) ∈ N 2 such that ϕ(M ) > J (N 0 , u N
0). Then the minimum time N ∗ is the smallest integer that minimizes J (N, u N ) over the inter- val [0, M ]. Moreover, we have u ∗ = u N
∗.
Proof. The proof is similar to the one of Proposi- tion 3.
4. Discrete Time Delay Systems
Consider the discrete time delay system described by
x(i + 1) =
m
X
j=0
A j x(i − j)
+
q
X
j=0
B j u(i − j), i = 0, . . . , N − 1, x(0) = x 0 ,
x(r) = α r , −m ≤ r ≤ −1, u(r) = µ r , −q ≤ r ≤ −1,
(45)
where x(i) ∈ X, u(i) ∈ U, X and U are Hilbert spaces, A j ∈ L(X), j = 0, . . . , m and B j ∈ L(U, X), j = 0, . . . , q. Furthermore, (α r ) r and (µ r ) r are fixed initial conditions. Here m ≥ 0 and q ≥ 1 are given integers.
Given the cost functional J (N, u) = ϕ(N ) +
N −1
X
i=0
hu(i), Ru(i)i (46) and a desired final state x d , we investigate the optimal pair (N ∗ , u ∗ ) which steers the system from the initial state (x 0 , (α r ) −m≤r≤−1 ) to x d with a minimal cost. We recall that ϕ: N → R + is a positive increasing map satisfying (6) and R ∈ L(U ) is a self adjoint positive definite operator. Similarly to the case of discrete sys- tems without delays, the determination of the optimal pair (N ∗ , u ∗ ) follows from solving the following optimization problems:
Find u N ∈ U N such that J (N, u N ) = min
u∈U
NJ (N, u), (47) and x(N ) = x d , where N is an admissible integer.
The determination of N ∗ is then performed by mini- mizing J (N, u N ) over an appropriate bounded subset of N.
First, we establish some results which are useful for the sequel. Define a new state variable e(i) ∈ X m+1 × U q by
e(i) = x(i), x(i − 1), . . . , x(i − m), u(i − 1), . . . , u(i − q) T
. (48)
Then e(·) satisfies the difference equation
( e(i + 1) = Φe(i) + ¯ Bu(i), i = 0, . . . , N − 1, e(0) = e 0 ,
(49) where
Φ =
A
0A
1. . . A
mB
1. . . B
qI 0 0 0 0
. . . . . . .. .
0 I 0 0 0
0 0 0 . . . 0 .. . I . . . .. .
. . . 0 . . . I 0
, ¯ B =
B
00 .. . 0 I 0 .. . 0
, (50)
and e 0 = (x 0 , α −1 , . . . , α −m , µ −1 , . . . , µ −q ) T .
Let P ∈ L(X m+1 × U q , X) be the projection oper- ator defined by
X m+1 × U q → X, P :
(y 1 , . . . , y m+1 , v 1 , . . . , v q ) ,→ y 1 .
(51)
Then from (49) it follows that
x(N ) = P e(N ) = P Φ N e 0 + P ¯ H N u, (52) where ¯ H N is the operator
U N → X, H ¯ N :
(u(0), . . . , u(N − 1)) ,→
N −1
X
i=0
Φ N −1−i Bu(i). ¯ (53)
Let K N = P ¯ H N and G 0 = ImK N . Then consider the semi-norm k · k G
Ndefined on X by
kxk G
N= kK N ∗ xk R , ∀ x ∈ X, (54) where K N ∗ is the adjoint operator of K N defined with respect to the inner products h·, ·i and h·, ·i R . Since G 0 = ImK N = (KerK N ∗ ) ⊥ , we deduce that the func- tional k · k G
Nis a norm on G 0 . Denote by G N the completion of G 0 under the norm k · k G
Nand consider the operator L N given by
G 0 → G 0 , L N :
x ,→ K N K N ∗ x. (55) Clearly, L N defines a bounded operator on G 0 en- dowed with the norm k · k G
N. By standard results (Lions, 1988), the operator L N may be extended to an isomor- phism denoted also by L N and defined from G N to its dual G
0N .
Proposition 6. An integer N ≥ 1 is admissible if and only if x d − P Φ N e 0 ∈ G
0N .
Proof. If N is admissible, then there exists a con- trol sequence u ∈ U N such that x(N ) = x d , which implies P e(N ) = P Φ N e 0 + K N u = x d , or x d − P Φ N e 0 = K N u. Since ImK N ⊂ G
0N , we deduce that x d − P Φ N e 0 ∈ G
0N . Conversely, suppose that x d − P Φ N e 0 ∈ G
0N . Then there exists y ∈ G N such that L N y = x d − P Φ N e 0 . Hence x d = P Φ N e 0 + K N K N ∗ y or x d = x(N ), where u = K N ∗ y. Thus N is admissi- ble.
Proposition 7. For each admissible integer N , the con- trol u N exists, is unique and given by u N = K N ∗ g, where g ∈ G N is the unique solution of the algebraic equation
L N g = x d − P Φ N e 0 .
Moreover, the optimal cost is J (N, u N ) = ϕ(N ) + kgk 2 G
N