Appendix 2: Dynamical programming
2.6 Boundary controllability of the 1D heat equation
2.6.3 Controllability and the problem of moments
In this section we introduce several notions of controllability.
LetT >0 and define, for any initial data u0∈L2(Ω), the set of reachable states
R(T;u0) ={u(T) :usolution of (2.156) withf ∈L2(0, T)}. (2.158) An element ofR(T, u0) is a state of (2.156) reachable in timeT by starting fromu0 with the aid of a controlf.
As in the previous section, several notions of controllability may be defined.
Definition 2.6.1 System (2.156) is approximately controllable in time T if, for every initial data u0 ∈L2(Ω), the set of reachable states R(T;u0) is dense in L2(Ω).
Definition 2.6.2 System (2.156)isexactly controllable in time T if, for every initial data u0 ∈ L2(Ω), the set of reachable states R(T;u0) coincides with L2(Ω).
Definition 2.6.3 System (2.156)isnull controllable in time T if, for ev- ery initial data u0 ∈ L2(Ω), the set of reachable states R(T;u0) contains the element0.
Remark 2.6.1 Note that the regularity of solutions stated above does not guarantee that u(T) belongs toL2(Ω). In view of this it could seem that the definitions above do not make sense. Note however that, due to the regular- izing effect of the heat equation, if the control f vanishes in an arbitrarily small neighborhood of t =T then u(T) is inC∞ and in particular in L2(Ω).
According to this, the above definitions make sense by introducing this minor restrictions on the controls under consideration.
Remark 2.6.2 Let us make the following remarks, which are very close to those we did in the context of interior control:
• The linearity of the system under consideration implies that R(T, u0) = R(T,0) +S(T)u0 and, consequently, without loss of generality one may assume thatu0= 0.
• Due to the regularizing effect the solutions of (2.156) are inC∞ far away from the boundary at time t =T. Hence, the elements of R(T, u0) are C∞ functions in [0,1). Then, exact controllability may not hold.
• It is easy to see that if null controllability holds, then any initial data may be led to any final state of the form S(T)v0 withv0∈L2(Ω).
Indeed, let u0, v0∈L2(Ω) and remark thatR(T;u0−v0) =R(T;u0)− S(T)v0.Since 0∈R(T;u0−v0), it follows thatS(T)v0∈R(T;u0).
• Null controllability implies approximate controllability. Indeed we have that S(T)[L2(Ω)]⊂R(T;u0) andS(T)[L2(Ω)] is dense inL2(Ω).
• Note that u1 ∈ R(T, u0) if and only if there exists a sequence (fε)ε>0
of controls such that ||u(T)−u1||L2(Ω) ≤ ε and (fε)ε>0 is bounded in L2(0, T). Indeed, in this case, any weak limit inL2(0, T) of the sequence (fε)ε>0gives an exact control which makes that u(T) =u1.
Remark 2.6.3 As we shall see, null controllability of the heat equation holds in an arbitrarily small time. This is due to the infinity speed of propagation.
It is important to underline, however, that, despite of the infinite speed of propagation, the null controllability of the heat equation does not hold in an infinite domain. We refer to [164] for a further discussion of this issue.
The techniques we shall develop in this section do not apply in unbounded domains. Although, as shown in [164], using the similarity variables, one can find a spectral decomposition of solutions of the heat equation on the whole or half line, the spectrum is too dense and biorthogonal families do not exist.
In this section the null-controllability problem will be considered. Let us first give the following characterization of the null-controllability property of (2.156).
Lemma 2.6.1 Equation(2.156)is null-controllable in time T >0 if and only if, for any u0 ∈ L2(0,1) there exists f ∈ L2(0, T) such that the following relation holds
Z T 0
f(t)ϕx(t,1)dt= Z 1
0
u0(x)ϕ(0, x)dx, (2.159) for any ϕT ∈ L2(0,1), where ϕ(t, x) is the solution of the backward adjoint problem
ϕt+ϕxx= 0 x∈(0,1), t∈(0, T) ϕ(t,0) =ϕ(t,1) = 0 t∈(0, T)
ϕ(T, x) =ϕT(x) x∈(0,1).
(2.160)
Proof. Let f ∈L2(0, T) be arbitrary anduthe solution of (2.156). If ϕT ∈ L2(0,1) andϕis the solution of (2.160) then, by multiplying (2.156) by ϕand by integrating by parts we obtain that
0 = Z T
0
Z 1 0
(ut−uxx)ϕdxdt= Z 1
0
uϕdx
T
0
+ Z T
0
(−uxϕ+uϕx)dt
1
0
+
+ Z T
0
Z 1 0
u(−ϕt−ϕxx)dxdt= Z 1
0
uϕdx
T
0
+ Z T
0
f(t)ϕx(t,1)dt.
Consequently Z T
0
f(t)ϕx(t,1)dt= Z 1
0
u0(x)ϕ(0, x)dx− Z 1
0
u(T, x)ϕT(x)dx. (2.161) Now, if (2.159) is verified, it follows that R1
0 u(T, x)ϕT(x)dx = 0, for all ϕ1∈L2(0,1) andu(T) = 0.
Hence, the solution is controllable to zero andf is a control for (2.156).
Reciprocally, if f is a control for (2.156), we have that u(T) = 0. From (2.161) it follows that (2.159) holds and the proof finishes.
From the previous Lemma we deduce the following result:
Proposition 2.6.1 Equation(2.156)is null-controllable in timeT >0 if and only if for any u0∈L2(0,1), with Fourier expansion
u0(x) =X
n≥1
ansin(πnx), there exists a functionw∈L2(0, T)such that,
Z T 0
w(t)e−n2π2tdt= (−1)n an
2nπe−n2π2T, n= 1,2, .... (2.162) Remark 2.6.4 Problem (2.162) is usually refered to as problem of mo- ments.
Proof. From the previous Lemma we know thatf ∈L2(0, T) is a control for (2.156) if and only if it satisfies (2.159). But, since (sin(nπx))n≥1 forms an orthogonal basis in L2(0,1), (2.159) is verified if and only if it is verified by ϕ1n= sin(nπx),n= 1,2, ....
If ϕ1n = sin(nπx) then the corresponding solution of (2.160) is ϕ(t, x) = e−n2π2(T−t)sin(nπx) and from (2.159) we obtain that
Z T 0
f(t)(−1)nnπe−n2π2(T−t)=an
2 e−n2π2T. The proof ends by takingw(t) =f(T−t).
The control property has been reduced to the problem of moments (2.162).
The latter will be solved by using biorthogonal techniques. The main ideas are due to R.D. Russell and H.O. Fattorini (see, for instance, [78] and [79]).
The eigenvalues of the heat equation are λn = n2π2, n ≥ 1. Let Λ = e−λnt
n≥1 be the family of the corresponding real exponential functions.
Definition 2.6.4 (θm)m≥1 is abiorthogonal sequence to Λ in L2(0, T) if and only if
Z T 0
e−λntθm(t)dt=δnm, ∀n, m= 1,2, ....
If there exists a biorthogonal sequence (θm)m≥1, the problem of moments (2.162) may be solved immediately by setting
w(t) = X
m≥1
(−1)m am
2mπe−m2π2Tθm(t). (2.163)
As soon as the series converges in L2(0, T), this provides the solution to (2.162).
We have the following controllability result:
Theorem 2.6.2 Given T > 0, suppose that there exists a biorthogonal se- quence (θm)m≥1 toΛ in L2(0, T)such that
||θm||L2(0,T)≤M eωm, ∀m≥1 (2.164) whereM andω are two positive constants.
Then(2.156)is null-controllable in time T.
Proof. From Proposition 2.6.1 it follows that it is sufficient to show that for anyu0∈L2(0,1) with Fourier expansion
u0=X
n≥1
ansin(nπx),
there exists a function w∈L2(0, T) which verifies (2.162).
Consider
w(t) = X
m≥1
(−1)m am
2mπe−m2π2Tθm(t). (2.165) Note that the series which defines wis convergent inL2(0, T). Indeed, X
m≥1
(−1)m am
2mπe−m2π2Tθm
L2(0,T)
= X
m≥1
|am|
2mπe−m2π2T||θm||L2(0,T)≤
≤M X
m≥1
|am|
2mπe−m2π2T+ωm<∞
where we have used the estimates (2.164) of the norm of the biorthogonal sequence (θm).
On the other hand, (2.165) implies that w satisfies (2.162) and the proof finishes.
Theorem 2.6.2 shows that, the null-controllability problem (2.156) is solved if we prove the existence of a biorthogonal sequence (θm)m≥1 to Λ inL2(0, T) which verifies (2.164). The following sections are devoted to accomplish this task.