You are on page 1of 23

Finite-time control for the bilinear heat equation Recommended by Prof.

T Parisini

Journal Pre-proof

Finite-time control for the bilinear heat equation

M. Ouzahra

PII: S0947-3580(19)30114-1
DOI: https://doi.org/10.1016/j.ejcon.2020.06.010
Reference: EJCON 459

To appear in: European Journal of Control

Received date: 14 March 2019


Revised date: 21 February 2020
Accepted date: 19 June 2020

Please cite this article as: M. Ouzahra, Finite-time control for the bilinear heat equation, European
Journal of Control (2020), doi: https://doi.org/10.1016/j.ejcon.2020.06.010

This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition
of a cover page and metadata, and formatting for readability, but it is not yet the definitive version of
record. This version will undergo additional copyediting, typesetting and review before it is published
in its final form, but we are providing this version to give early visibility of the article. Please note that,
during the production process, errors may be discovered which could affect the content, and all legal
disclaimers that apply to the journal pertain.

© 2020 Published by Elsevier Ltd on behalf of European Control Association.


Finite-time control for the bilinear
heat equation
M. Ouzahra
M2TS Laboratory, University of Sidi Mohamed Ben Abdellah
P.O. Box 5206, Bensouda, Fès, Morocco
m.ouzahra@yahoo.fr

Abstract

In this paper, we deal with the multidimensional heat equation governed by a bilinear control. We
first investigate the finite-time stability and null controllability properties. Then, we study the global and
partial approximate controllability. These results are further applied to establish the exact controllability
of the system at hand. Nonnegative controllability results are also discussed. The approaches are based
on linear semigroup theory and the controllability of the heat equation with additive controls. An other
essential tool that we use here is the classical maximum principle.

Keywords: Bilinear control, parabolic equation, finite-time stabilization, approximate control-


lability, exact controllability.

I. Introduction
Let Ω be an open bounded domain of Rn , n ≥ 1 with smooth boundary ∂Ω and let us consider
the following Dirichlet problem:

 yt = ∆y + v( x, t)y, in Q = Ω × R∗+
y = 0, on Σ = ∂Ω × R∗+ (1)

y (0) = y0 , in Ω

where, for each time t the state y(t) is given by the function y(·, t) ∈ L2 (Ω) and the multiplicative
control function v(·, t) will be chosen in a suitable space, which can be viewed as controlling
the reaction rate of the processus (1). In terms of applications, equation (1) provides practical
description of various real problems (see [32], p. 1). In the present paper, we first study the
question of finite-time stability and null controllability for the equation (1). A recent study on
the two notions has been conducted by Coron and Nguyen [17] in the context of the 1D −heat
equation with variable coefficients. The main idea builds on the design of time-varying feedbacks
via the backstepping approach which leads to a constructive steering control. Moreover, it has
been shown that finite time stability can be achieved in arbitrary time by means of periodic time-
varying feedback laws. In [43], the question of finite-time stability for linear equation evolving
in a Hilbert space was considered. The approach is based on the implicit Lyapunov function
method in connection with the notion of homogeneity. As an application, the 1D −heat equation
was considered with Dirichlet boundary conditions, where it was shown that for some ρ, γ > 0
the feedback control v( x, t) = −ρky(t)k−γ guarantees the finite time stability of (1) on Ω = (0, 1)
for a settling time which depends on the initial condition. Moreover, achieving the stability
property in a prescribed time (i.e. the settling time is independent of the initial state) may present

1
advantages in many practical situations (see e.g. [18, 26, 45, 48]). Such a question has attracted
many researchers. In [20], the authors studied the problem of fixed-time stabilization of the
1D −reaction-diffusion with constant-parameters by means of continuous boundary time-varying
feedbacks. Based on the backstepping approach, it has been shown that the proposed time-varying
control transforms the initial boundary controlled equation to another equivalent target system
like (1) with a time-dependent parameter v( x, t) = −c(t) ≤ 0. The time-depending function c(t) is
constructed through Lyapunov techniques to prove the fixed-time stability for the target system in
L2 (0, 1). In [47], the prescribed finite time stability for the 1D − heat equation was investigated in
the sense of the H 1 −norm. That is, using the backstepping method, the authors selected a target
equation whose state as well as its spatial derivative converge to zero within the prescribed time.
In the second part of the paper, we discuss controllability properties of system (1), i.e. we
try to find a control v such that the associated state y possesses a desired behavior at a finite
time T > 0. Here the steering control is not necessary a feedback law. It is well understood that
the duality approach, used to study the controllability of linear systems, do not apply to the
bilinear case due to the nonlinear dependance of the state with respect to control. However, we
will see that in the context of equation (1), it is possible to exploit the controllability properties of
linear systems. The standard problem of controllability consists, for a given time T > 0, initial
and terminal data y0 and yd belonging to specified sets (spaces, cones, e.g.), in finding a suitable
control v such that the solution of the system, starting from the initial state y0 and provided with
this control function reaches (in appropriate sense) the state yd at time T. In the case yd = 0, one
recovers the notion of null controllability. Another notion of controllability is to determine the set
of states that are reachable at any time, rather than a specified time T. In other words, instead
of studying the set A T (y0 ) of all attainable states from y0 at time T (problem of attainability at
time T) one can rather consider the union R(y0 ) = ∪ T >0 A T (y0 ) (problem of reachability). On
the other hand, due to the maximum principle, system (1) is not controllable in any reasonable
state space X (or even X − (0)). Alternatively, it seems natural to investigate the possibility of
reaching suitable classes of target states for (1). In [4], the authors have studied the approximate
controllability of the rod and wave equations. The question of exact controllability of the rod
and wave equations has been studied in [9, 10]. In [27], the controllability problem has been
considered in the context of simultaneous control of the rod and Schrödinger equations. The
author in [7, 8] proved the local controllability of the bilinear Schrödinger equation around an
appropriate reference trajectory by a real-valued multiplicative control. In [39], approximate and
exact controllability of a class of equilibrium states have been studied for the multidimensional
wave equation in high energy state spaces. Optimal control approaches have been investigated
for various type of PDEs equations in [11, 12, 35, 36] to make the system’s state close to a desired
profile. Controllability for semilinear heat equations with additive controls have been treated in
various works (see [19, 21, 22, 25, 34, 49] and the references therein). In [29, 30, 31, 32], the authors
studied the multiplicative controllability of various equations of parabolic and hyperbolic types.
Based on these works, the exact controllability of equations like (1) with inhomogeneous Dirichlet
conditions has been treated in [37, 40]. In the context of system (1), the nonnegative approximate
controllability has been studied for the one-dimensional version of (1), which concerns initial and
target states of the same sign, i.e. y0 ≥ 0 and yd ≥ 0, a.e. in Ω [28]. Moreover, the multidimensional
case has been treated in [29]. In [14], the approximate controllability properties have been derived
in the one-dimensional case for a couple (y0 , yd ) of initial and target states which are allowed to
change their sign, provided this change of sign holds in the same manner for y0 and yd , at finitely
many zero. These results have been extended in [15] to the case of the presence of a Lipschitz
nonlinear term. Recently, the approximate controllability of the n−dimensional system (1) has
been studied in [41] under the sign condition : y0 ( x )yd ( x ) ≥ 0, for almost every (a.e.) x ∈ Ω

2
with the same points of vanishing. Moreover, under additional regularity assumptions on the
target state to be achieved, the exact controllability has been treated in the one-dimensional case.
We observe that, in the aforementioned papers, the control acts all over the evolution domain.
Moreover, the multiplicative controllability of the reaction-diffusion (1) has been established in
a time which depends on the initial and target state. Here, we will provide various situation in
which the time of steering can be taken independent of the initial and target states. Our results
here also include the case of locally supported control. Our approach is based on linear semigroup
properties for approximate controllability, and for exact controllability we use a combination of
approximate controllability of (1) with the null controllability of linear reaction diffusion equation.
The paper is organized as follows: In the second section, we prove that the system (1) is
finite-time stabilizable and null-controllable using various type of controllers. Then, we study the
possibility of reaching trajectories of the free system. In Section 3, we provide a class of target states
that are approximately accessible from y0 at a suitable time T > 0, which depends on the choice
of (y0 , yd ) and the desirable precision of steering e > 0. An application to partial approximate
controllability is also provided. Special attention will be given to nonnegative controllability with
locally distributed control and preassigned control time. In Section 4, we will apply the global
and partial approximate controllability to identify a set of target states, which can be globally
or partially exactly achieved through the equation (1) at a preassigned time T > 0. Section 4 is
devoted to some illustrative examples, and we end the paper with a conclusion in Section 5.

II. Finite time stability and controllability to trajectories


In this section, we will first study the finite time stability which consists on finding a feedback
control law for which the solution of (1) vanishes at some settling time. Then we will consider the
question of approximate null controllability. The latter will be used to establish the controllability
to trajectories for system (1).
Let us consider the state space L2 (Ω) equipped with its natural norm denoted by k · k, and let us
introduce the unbounded operator A = ∆ with domain D( A) = H01 (Ω) ∩ H 2 (Ω) which generates
a contraction semigroup S(t) in L2 (Ω).

i. Finite time stability and null controllability


i.1 Finite time stability

In the next result we establish a finite-time stability of (1) using a feedback control when the
settling time depends on the initial state. However, if the stabilizing feedback control is allowed to
depend on time, then we show that one can consider a prescribed time.

Theorem II.1 Let y0 ∈ L2 (Ω). Then for any ρ > 0 and 0 < µ < 1, we consider the following feedback
law
v(·, t) = v(t) = −ρky(·, t)k−µ 1{t≥0: y(t)6=0} · (2)

Then there is a time T = T (µ, ρ, ky0 k) such that the respective solution y(t) to (1) satisfies the estimate:

ky(t)k = O(( T − t)1/µ ), as t → T − ,


1
and y(t) = 0, ∀t ≥ T. Furthermore, the settling time T∗ is such that T∗ ≤ ρµ k y0 k .
µ

3
Proof 1 Let T > 0 be the time of stability which is to be determined, and let us set v(t) = −ρky(t))k−µ , µ ∈
(0, 1). Then observing that for 0 < µ < 1, the function ϕ : y 7→ α1 kykα with α := 2 − µ > 1
is convex and lower semi-continuous, we deduce that the nonlinear operator ∇ ϕ : y 7→ kyk−µ y is
maximal monotone. Hence, the problem (1) possesses a unique strong local solution y(t) such that
y(t) ∈ H01 (Ω) ∩ H 2 (Ω), ∀t ∈ (0, T ] (see [6], p. 257). Moreover, for almost every t ∈ (0, T ) we have
d
ky(t)k2 ≤ −2ρky(t)k2−µ . (3)
dt
Let us set w(t) = ky( ttT 2 2
+1 )k , t ≥ 0 so that limt→+∞ w ( t ) = k y ( T )k .
Then the estimate (3) implies
2Tρ µ
ẇ(t) ≤ − w1− 2 (t), a.e. t > 0,
( t + 1)2
which by integrating gives
2 2Tρ 2
w(t)µ/2 ≤ + w(0)µ/2 − 2Tρ, ∀t ≥ 0· (4)
µ t+1 µ
As a necessary condition for the inequality (4) to hold for all t ≥ 0, the following inequality-type condition
1
must be fulfilled: T ≤ ρµ w(0)µ/2 .
1
Then taking T = ρµ k y0 k , we get
µ
2
(ρTµ) µ
w(t) ≤ 2 , ∀ t ≥ 0·
( t + 1) µ
Thus for all t ∈ [0, T ), we have
 µ1
ky(t)k ≤ ρµ( T − t) ,
which implies y( T ) = 0. Moreover, since y(t) is a global solution of the bilinear equation (1), we deduce
that y(t) = 0, ∀t ≥ T.
1
Finally, by definition of the settling time T∗ we have T∗ ≤ ρµ k y0 k µ .

Remark 1 Feedback like (2) has been considered for 1D − version of (1) in [43]. Moreover, the case µ = 1
has been also considered in ([6], pp. 250-252) in the context of a multivalued heat equation.

In the next result we investigate the stability in a prescribed time under time-varying feedback
controls.

Theorem II.2 For any T > 0, there exists a continuous time-varying gain control k(·) ≥ 0 such for every
γ > 0, the following time-varying feedback control

v(·, t) = v(t) = −k(t)1[0,T ) ky(·, t)kγ , (5)

results in a unique mild solution y which is such that ky(t)k = O(( T − t)1/γ ), as t → T − ·

Proof 2 First let us note that for any continuous function k(·), the control (5) gives rise to a unique local
mild solution y(t) for (1), which is defined in a maximal interval of existence [0, tmax ) (see [42], p. 185).
Moreover, for any fixed τ ∈ (0, tmax ) we consider a sequence ( Fk ) ⊂ C([0, τ ]) that uniformly converges to
F : t 7→ −k(t)ky(t)kγ y(t) on [0, τ ]. Let (yk0 ) ⊂ D ( A) which converges to y(0) = y0 in H and let yk
denote the classical solution of the problem

ẏk = ∆yk (t) + Fk (t), yk (0) = yk0 ·

4
We have
 Z t 
y k ( t ) − y ( t ) = S ( t ) y k (0) − y (0) + S(t − s) Fk (s) − F (s) ds, ∀t ∈ [0, τ ],
0

from which we deduce that yk (t) → y(t) in H as k → +∞.


Moreover, we have
d
ky (t)k2 ≤ 2h Fk (t), yk (t)i, t ∈ (0, τ ).
dt k
Then by integrating we get
Z t
kyk (t)k2 − kyk (s)k2 ≤ 2 h Fk (r ), yk (r )idr, 0 ≤ s ≤ t ≤ τ,
s

which gives by letting k → +∞


Z t
ky(t)k2 − ky(s)k2 ≤ −2 k(r )ky(r )kγ+2 dr, 0 ≤ s ≤ t ≤ τ. (6)
s

In particular ky(t)k decreases (in time) and so the following estimate holds:

ky(t)k ≤ ky0 k, ∀ ∈ [0, τ ].


Then y(t) is bounded in [0, tmax ) and hence it is a global solution, i.e. tmax = +∞ (see [42], p. 185).
Let us now fixe a T > 0 and let us set z(t) = y( ttT 2
+1 ) and w ( t ) = k z ( t )k , t ≥ 0. Then (6) implies
Z t
T Tr γ
w ( t ) − w ( s ) ≤ −2 k( )w1+ 2 (r )dr, ∀t ≥ s ≥ 0. (7)
s (r + 1)2 r + 1
Now in order to estimate the integral term via the decreasing of w(t), we have to choose k(·) such that
1 Tr
2
k( ) ≥ β > 0, ∀r ≥ 0·
(r + 1) r+1
λ
Let us take k(t) = 1
( T −t)2 [0,T )
with λ > 0, so that one has

T Tr λ
k( ) = , ∀r ≥ 0.
(r + 1)2 r + 1 T
Then, letting wk = w(k) and taking s = k and t = k + 1 in (7), it comes
λ γ +1
wk+1 − wk ≤ −2 wk2+1 , ∀k ≥ 0,
T
which implies (see e.g. [1])
M
∃ M > 0, wk ≤ 2 , ∀k ≥ 1.
( k + 1) γ
Then since w(t) decreases, we deduce that
 
1
w(t) = O 2 , as t → +∞·
( t + 1) γ
Thus  
1/γ
ky(t)k = O ( T − t) , as t → T − ,

and hence y( T ) = 0.

5
Remark 2 1. Note that both the controls (2) and (5) are unbounded w.r.t time and are dominated by
T − t. However, the full input v( x, t)y( x, t) is bounded.
2. If we take γ = 0 in (5), we obtain a control which has the same effect on the system (1) as the
feedback control v(t) = − (T −λ t)2 1[0,T ) sign (|y(t)|), λ > 0 and leads to the estimate ky(t)k ≤
−σt
Me T−t ky0 k, ∀t ∈ [0, T ) (for some M, σ > 0).

i.2 Approximate null controllability


In this part, we study the null-controllability problem, which consists in steering (1) from its initial
state to the zero-state. For this end, we consider the following linear system, starting from y0 :

 ϕt = ∆ϕ + u0 ( x, t), in Q
ϕ = 0, on Σ (8)

ϕ(·, 0) = y0 , in Ω
where y0 is the initial state of the system (1) and u0 ( x, t) is an additive control. It is apparent that
in order to deduce the controllability properties of system (1) from the linear system (8), one can
u ( x,t)
formally use the multiplicative control: v( x, t) = ϕ0( x,t) . This relation will be a key point to get
our results. In the sequel, we will use the notations y, y(t), y(·, t) to design any ( x, t)−dependent
function y( x, t). Here, we will construct an appropriate control u0 (·, t) that steers the linear system
(8) to 0, and then try to find a control v for system (1) such that the corresponding solution y is
such that: vy = u0 in Q T := Ω × (0, T ), where T > 0 is the control time.
Theorem II.3 Let y0 ∈ L2 (Ω). Then for all time T > 0 and e > 0, there exists a time-dependent control
v(·) ∈ L∞ (0, T ) and a unique corresponding mild solution y to (1) such that ky( T )k < e.
Proof 3 Let us consider the system (8), starting from the initial state y0 of the bilinear system (1), and let
us consider the control u0 (., t) = −T1 S(t)y0 . Then, the system (8) possesses a unique mild solution given by
the following variation of constant formula:
Z t
ϕ(·, t) = S(t)y0 + S(t − s)u0 (·, s)ds, t ∈ [0, T ]. (9)
0

Then we can observe that


1
u0 ( x, t) = ϕ( x, t), a.e., ( x, t) ∈ Ω × (0, T ).
t−T
For ρ ∈ (0, T ), we consider the control defined by:
 1
 t− T , if t ∈ [0, T − ρ)
v( x, t) = vρ (t) = (10)

0, if t ∈ [ T − ρ, T ]
Since vρ ∈ L∞ (0, T ) for all 0 < ρ < T, we deduce that the system (1) has a unique solution y ∈
C([0, T ]; L2 (Ω)), which is given by the following integral equation:
Z t
y ( t ) = S ( t ) y0 + S(t − s)vρ (s)y(s)ds, t ∈ [0, T ]· (11)
0

Observing that vρ (t) ϕ(·, t) = u0 (·, t) in Q T −ρ , we deduce that ϕ satisfies (11) in Q T −ρ , and hence y = ϕ
in Q T −ρ . Then, from the definition of vρ it comes: y( T ) = S(ρ)y( T − ρ) = S(ρ) ϕ( T − ρ). Moreover,
using the fact that S(t) is a C0 −semigroup of contractions, we deduce that S(ρ) ϕ( T − ρ) → ϕ( T ) = 0, as
ρ → 0+ . This gives the claimed approximate steering at T.

6
Remark 3 1. Note that the control (10) may be replaced by the following time-varying feedback:
1[0,T −ρ)
vρ (t) = sign(|y(t)|),
t−T
as they have the same effect on the bilinear system (1).

2. The control v0 corresponding to ρ = 0 in (10), which lies in L1loc ([0, T )) ∩ Lr ([0, T ]) for 0 < r < 1,
results in the exact null controllability of (1). Indeed, for all t ∈ [0, T ) we have v0 (t) = t−1T , so that
v0 (t) ϕ( x, t) = u0 ( x, t) in [0, T ) × Ω. Moreover, for t = T we have ϕ( x, T ) = 0, a.e. x ∈ Ω which
implies that ϕ = 0 and so u0 = 0 in Q T . Thus, we have v0 (t) ϕ( x, t) = u0 ( x, t) in [0, T ] × Ω, and
hence ϕ is the unique solution of (1) corresponding to v0 . Then v0 (t) may also be written in the form
v0 (t) = t−1T sign(|y(t)|). Moreover, the backward uniqueness property prevents us to consider the
problem of exact null controllability of (1) under a square integrable bilinear control v( x, t) (see [5]).

3. The null-controllability of (1) holds for any time T > 0, which seems to be natural since parabolic
equations have an infinite speed of propagation.

4. In the context of a linear system, the null controllability is equivalent to exact controllability to
trajectories of the uncontrolled system (see [23], pp. 1396-1400). However, this is not the case for
the bilinear system (1). Indeed, for y0 ≥ 0 a.e. in Ω, we have (according to the maximum principle)
that y(t) ≥ 0, t ∈ [0, T ], so that (1) can not be steered to yd := S( T )y1 for y1 < 0. In the next
subsection, we will provide a set of trajectories for which this implication holds.

ii. Exact controllability to trajectories


In this part, we apply the null controllability result to study the controllability properties to
trajectories. We will look for a control u0 in the form u0 (., t) = k(t)S(t)w1 , where k ∈ L1 (0, T ) and
w1 ∈ L2 (Ω) are to be determined so that the corresponding solution to (8) satisfies ϕ( T ) = yd
(where yd is the target state for system (1)). Note that with this kind of control, we have
ϕ( T ) = S( T )(y0 + Tw1 ) so with this approach, only trajectories of the free system (i.e. v = 0 in
(1)) can be exactly reached.
As a matter of notation, for any map f : Ω → R, the inequality f ≥ 0 means that for, a.e. x ∈ Ω,
we have: f ( x ) ≥ 0. We also write f ≥ g or g ≤ f when f − g ≥ 0. Moreover for a set E, the symbol
1 E denotes the characteristic function of E. We also define for all f , g ∈ L2 (Ω), the following
correspondences: [ f × g]S : ( x, t) 7→ (S(t) f )( x )(S(t) g)( x ), for all t > 0 and for a.e. x ∈ Ω and
( f , g) ∈ L2 (Ω) × L2 (Ω) 7→ [ f × g]S ∈ L1 (Ω)·
For any T > 0, we consider the following sets:

Λ T (y0 ) = {y1 ∈ L2 (Ω); [y0 × y1 ]S ≥ 0 and ∃α, β > 0; βS(t)y0 ≤ S(t)y1 ≤ αS(t)y0 , in Q T },

Λ(y0 ) = ∪ T >0 Λ T (y0 ), ET (y0 ) = {S( T )ξ : ξ ∈ Λ T (y0 )} and E(y0 ) = ∪ T >0 ET (y0 ).
Note that R+ y0 ⊂ Λ T (y0 ). Moreover, if y0 ≥ 0 a.e. in Ω, then, it follows from the maximum
principle that for all T > 0, we have

{S( T )y1 / y1 ∈ L2 (Ω) and βy0 ≤ y1 ≤ αy0 , a.e. in Ω for some α, β > 0} ⊂ ET (y0 ).

Furthermore, we can easily show that Λ(y0 ) (and so is E(y0 )) is a convex cone.
The next result concerns the attainability of elements of E(y0 ).
Theorem II.4 Let y0 ∈ L2 (Ω). Then for all yd ∈ E(y0 ), there exist a time T = T (y0 , yd ) and a control
v ∈ L∞ ( Q T ) such that the system (1) possesses a unique mild solution y such that y( T ) = yd .

7
Proof 4 Let T > 0 and y1 ∈ L2 (Ω) be such that yd = S( T )y1 with [y1 × y0 ]S ≥ 0 and βS(t)y0 ≤
S(t)y1 ≤ αS(t)y0 , in Q T (α, β > 0). Let us reconsider the linear system (8) with the control u0 (., t) =
k (t)S(t)w1 , where k ∈ L1 (0, T ) is a scalar value function and w1 ∈ L2 (Ω) to be determined such that
ϕ( T ) = yd . Then, the system (8) admits a unique mild solution ϕ, and we have:
Z T
ϕ ( T ) = S ( T ) y0 + ( k (s)ds)S( T )w1 .
0
RT 1
Thus taking 0 k(s)ds = 1 and w1 = y1 − y0 , we get ϕ( T ) = yd . In the sequel, we take k(t) = T, ∀t ∈
[0, T ]. Then, with the control u0 (·, t) = T1 S(t)(y1 − y0 ), we have
t t
ϕ ( t ) = (1 − )S(t)y0 + S(t)y1 , ∀t ∈ [0, T ]. (12)
T T
Taking into account that [y0 × y1 ]S ≥ 0 in Q T , it comes from (12) that

for all ( x, t) ∈ Q T ; ϕ( x, t) = 0 ⇒ u0 ( x, t) = 0 (13)

Let Γ = {( x, t) ∈ Q T ; ϕ( x, t) 6= 0}. Then, since [y0 , y1 ]S ≥ 0, we have

u0 ( x, t) | S ( t ) y1 | + | S ( t ) y0 |
| 1Γ | ≤
ϕ( x, t) ( T − t)|S(t)y0 | + t|S(t)y1 |
Let us determine an upper bound of the right-hand side of this inequality. Using the two assumptions
[y0 , y1 ]S ≥ 0 and βS(t)y0 ≤ S(t)y1 ≤ αS(t)y0 , , we can suppose that S(t)y0 > 0 and S(t)y1 > 0. In
this case, we obtain:
u ( x, t) α+1
| 0 1Γ | ≤
ϕ( x, t) T − t + βt
u0 ( x, t)
which implies that 1 Γ ∈ L ∞ ( Q T ).
ϕ( x, t)
Let us consider the control defined by:

u0 ( x, t)
v( x, t) = 1Γ , ∀( x, t) ∈ Q T , (14)
ϕ( x, t)

and let y be the corresponding mild solution of (1), which is given by:
Z t
y ( t ) = S ( t ) y0 + S(t − s)v( x, s)y(s)ds, ∀t ∈ [0, T ]. (15)
0

Remarking that (13) leads to

u0 ( x, t)
1Γ ϕ( x, t) = u0 ( x, t), ∀( x, t) ∈ Q T , (16)
ϕ( x, t)

we can see that ϕ is a mild solution of (1) corresponding to control (14). Thus, by uniqueness we deduce
that y = ϕ in [0, T ], which gives y( T ) = ϕ( T ). This achieves the proof.

Remark 4 1. Note that since [y0 × y1 ]S ≥ 0 inQ T , it follows that Γ = Γy0 ∪ Γy1 (where for all
ξ ∈ L2 (Ω), we set Γξ := {( x, t) ∈ Q T ; S(t)ξ ( x ) 6= 0}). Moreover, it comes from the maximum
principle that for any y0 ∈ L2 (Ω) such that y0 ≥ 0 and y0 6= 0, the zero set Q T \ Γy0 of S(·)y0 is
of zero measure. Moreover, in space dimension n = 1 this set is discrete (see [2]). Note also that by
unique continuation (see [46]) the set Q T \ Γy0 has, in general, empty interior.

8
2. In the case y1 = λy0 , λ > 0, we have v( x, t) = T +(λλ−−1 1)t 1Γy0 ∈ L∞ ( Q T ). If in addition y0 ≥ 0 and
y0 6= 0, then (according to the maximum principle) the control is time-dependent only.

As a direct consequence, we have the following result which provides a situation in which the
assumptions of Theorem II.4 are easily checkable.

Corollary 1 Let y0 , y1 ∈ L2 (Ω) be such that βy0 ≤ y1 ≤ αy0 , a.e. in Ω (for some α, β > 0). Then, for
all time T > 0, there exists a control v ∈ L∞ ( Q T )) such that for the respective solution to (1), we have
y ( T ) = S ( T ) y1 .

III. Approximate controllability


In this section, we will derive sufficient conditions for partial (resp. global) approximate con-
trollability of system (1) with locally (resp. globally) distributed control. Let us introduce the
space:
p
L+ (Ω) := { f ∈ L p (Ω); f ≥ 0, a.e. on Ω}, 1 ≤ p ≤ +∞.

i. Approximate controllability under global distributed control


In this part, we apply the result of Theorem II.4 to study the reachable set of system (1).

Theorem III.1 Let y0 ∈ L2 (Ω) and let yd ∈ Λ(y0 ). Then for any e > 0, there are a time T =
T (y0 , yd , e) > 0 and a control v ∈ L∞ ( Q T ) such that for the respective solution to (1), the following
estimate holds:
ky( T ) − yd k < e. (17)

Proof 5 First, let us observe that the case yd = 0 follows from Theorem II.3. In the sequel, we assume that
yd 6= 0. Let T0 > 0 be such that [yd × y0 ]S ≥ 0 and βS(t)y0 ≤ S(t)yd ≤ αS(t)y0 in Q T0 , (α, β > 0). It
follows from the proof of Theorem II.4 that for all T ∈ (0, T0 ), there exists a control v ∈ L∞ ( Q T ) such that
y( T ) = S( T )yd . Since S( T )yd → yd , as T → 0+ in L2 (Ω), the result follows.
We can further estimate the time of approximate controllability. Indeed, let us fix e > 0. Then for
yd ∈ D ( A), we have
Z T
y( T ) − yd = S( T )yd − yd = S(t) Ayd dt.
0

Since S(t) is a semi-group of contractions, we deduce that

ky( T ) − yd k ≤ T kyd kD( A) .


e
Thus for any e > 0, the estimate (17) holds for all 0 < T < kyd kD( A)
.
Now, for yd ∈ L2 (Ω), there exists ỹd ∈ D ( A) such that d d
ky − ỹ k < e/4. Then, using again the fact that
S(t) is a contraction semigroup, we derive

ky( T ) − yd k ≤ 2kyd − ỹd k + kS( T )ỹd − ỹd k.


e
Thus (17) holds for 0 < T < ·
2kỹd k D( A)

As an application of Theorem III.1, we have the following partial approximate controllability


for system (1).

9
Corollary 2 Let ω be a nonempty open set of Ω. Let y0 ∈ L2+ (Ω) and let yd ∈ L2 (ω ) be such that
βy0 ≤ yd ≤ αy0 , a.e. ω, for some α, β > 0. Then for any e > 0, there exist a time T = T (y0 , yd , e) > 0
and a control v ∈ L∞ ( Q T ) such that for the respective solution to (1), we have

ky( T )|ω − yd k < e.



yd , in ω
Proof 6 Let us consider the following extended target state ŷd =
0, otherwise
Combining the results of Theorem III.1 and Corollary 1, we deduce that for all e > 0, there exist a time
T = T (y0 , yd , e) > 0 and a control v ∈ L∞ ( Q T ) such that the respective solution y to (1) satisfies the
estimate :
ky( T ) − ŷd k < e.
Then, since ŷd |ω = yd , it follows that ky( T )|ω − yd k < e.

ii. Partial approximate controllability under locally distributed control


In this part, we discuss the nonnegative partial approximate controllability for system (1), which
consists in achieving yd only on a part of Ω, for a preassigned control time T by using a locally
supported input. In [24], the nonnegative partial (local) approximate controllability was studied
for the one dimensional version of (1) by making use of Fourier series expansions. In the next
result we provide a multidimensional version of the result of [24], by using an approach based on
semigroup properties and maximum principle.

Theorem III.2 Assume that y0 ∈ L2+ (Ω) − (0). Then for any T > 0, the set {y( T )|ω ; y is a solution of
(1) with v = 1ω v ∈ L∞ 2
+ ( Q T )} is dense in S ( T )y0 |ω + L+ ( ω ).

Proof 7 Let yd ∈ S( T )y0 |ω + L2+ (ω ), which we extend to Ω by 0. Let us further denote by yd this
extension. Inspired by the work [24], we consider the following locally supported control:
1 d
u0 ( x, t) = (y − S( T )y0 )1ω ×(T −h,T ) , (0 < h < T ).
h
Thus the corresponding solution to (8) is given by :


 S ( t ) y0 , t ∈ [0, T − h]

ϕ(t) = Z t


 S(t)y0 + 1h S(t − s)(yd − S( T )y0 )1ω ds, t ∈ [ T − h, T ]
T −h

Then, we have Z h
1
ϕ ( T ) = S ( T ) y0 + S(s)(yd − S( T )y0 )1ω ds, (18)
h 0
and using semigroup properties of S(t), we get

ϕ( T ) → S( T )y0 + (yd − S( T )y0 )1ω , as h → 0+ .

It follows that
1ω ϕ( T ) → 1ω yd , as h → 0+ .
In other hand, since yd ∈ S( T )y0 |ω + L2+ (ω ), we have u0 ( x, t) ≥ 0 in ω × (0, T ). This, together with the
fact that y0 ≥ 0 in Ω implies (by the maximum principle) that for all ( x, t) ∈ ω × [ T − h, T ], we have:

ϕ( x, t) = 0 ⇒ u0 ( x, t) = 0

10
Moreover, it follows from the strong maximum principle (see for instance [44], p. 23) that S(t)y0 > 0 in
Ω × (0, T ]. Thus we have:
u0
ϕ 1 = u0 1ω , in Q T
ϕ {( x,t)∈ω ×(0,T ); ϕ( x,t)6=0}

Then, following the same method as in the proof of Theorem II.4, we deduce that ϕ is a solution of (1)
corresponding to control
u0
v= 1 , (Γ = {( x, t) ∈ Q T ; ϕ( x, t) 6= 0}).
ϕ ω ×(0,T )∩Γ

We have shown that ϕ is a solution of (1) and is such that: ϕ( T )|ω ∈ S( T )y0 |ω + L2+ (ω ) and ϕ( T )|ω →
yd |ω , as h → 0+ .
It remains to show that the control v can be chosen in L∞ + ( Q T ). Let e > 0 be fixed, let us set f =
(yd − S( T )y0 )1ω , and let f˜ be a continuous and nonnegative function in Ω with compact support K, such
that k f − f˜k < e. Let us introduce the following additive control:

1 ˜
ũ0 ( x, t) = f1 ,
h ω ×[T −h,T ]
and let ϕ̃(t) be the corresponding solution of the system (8).
From the strong maximum principle, we have ρ := minK ×[T −h,T ] S(t)y0 > 0. Then ϕ̃(t) ≥ S(t)y0 ≥ ρ,
a.e. in K × [ T − h, T ], and so we can define the following multiplicative control:
ũ0
ṽ = 1
ϕ̃ K ×[T −h,T ]

which lies in L∞
+ ( Q T ). Then we have

k ϕ̃( T )|ω − yd |ω k ≤ k ϕ̃( T )|ω − ϕ( T )|ω k + k ϕ( T )|ω − yd |ω k

≤ k f˜ − f k + k ϕ( T )|ω − yd |ω k < 2e
Moreover, from the definition of the control ṽ, we have ṽ ϕ̃ = ũ0 . It follows that ϕ̃ is a solution of (1). This
completes the proof

We have the following consequence regarding target states in L2+ (ω ).

Corollary 3 Let ω be a nonempty open subset of Ω. Let y0 ∈ L2+ (Ω) − (0) and yd ∈ L2+ (ω ). Then, for all
T > 0 and for all e > 0, there exists a control v = 1ω v ∈ L∞ ( Q T ) for which we have ky( T )|ω − yd k < e.

Proof 8 Let e > 0 and let T1 ∈ (0, T ) be fixed. For all σ > 0, the unbounded operator A − σI generates
in L2 (Ω) the semigroup T (t) = e−σt S(t). Then, taking the constant control v(t) = −σ in (0, T1 ) with
2k y k
σ > T1 ln( e 0 ); the solution of (1) satisfies the estimate ky( T1 )k < e/2.
1
Since S(t) is a contraction semigroup, we also have kS( T )y( T1 )k < e/2.
Let us now consider the two systems

 yt = ∆y + 1ω v( x, t)y, in Ω × ( T1 , T )
y = 0, on ∂Ω × ( T1 , T ) (19)

y( T1 ) = y( T1− ), in Ω
and

11

 ϕt = ∆ϕ + u0 ( x, t), in Ω × ( T1 , T )
ϕ = 0, on ∂Ω × ( T1 , T ) (20)

ϕ( T1 ) = y( T1− ), in Ω

Let us extend yd to Ω by 0 and denote by yd this extension.


It follows from the proof of Theorem III.2 that, with the additive control:

1 d 
u0 ( x, t) = y + S( T )y( T1 ) − S( T − T1 )y( T1 ) 1ω ×(T −h,T ) ,
h

the state ϕ of (20) is such that

1ω ϕ( T ) → 1ω yd + 1ω S( T )y( T1 ), as h → 0+ .

u0
Furthermore, ϕ is a solution of (19) corresponding to control v = ϕ 1ω ×( T −h,T ) with h ∈ (0, T − T1 ).
Let h ∈ (0, T − T1 ) be such that

k ϕ( T )|ω − (yd + S( T )y( T1 ))|ω k < e/2.

Then, we have

k1ω ϕ( T ) − 1ω yd k ≤ k1ω ϕ( T ) − 1ω (yd + S( T )y( T1 ))k + kS( T )y( T1 )k

We conclude that the corresponding solution to (1) with the control:



 −σ,
 t ∈ (0, T1 )
v( x, t) = (21)

 1 u0 ( x,t)
ω ×( T −h,T ) ϕ( x,t) , t ∈ ( T1 , T )

guarantees the following desired estimate:

k1ω ϕ( T ) − 1ω yd k ≤ e·

As a direct consequence, we have the following global nonnegative controllability result :

Corollary 4 Assume that y0 ∈ L2+ (Ω) − (0). Then, for any T > 0, the set {y( T ); y is solution of (1)
with v ∈ L∞ ( Q T )} is dense in L2+ (Ω).

Remark 5 In [24], it has been shown that in the context of system (1) with proper control support O the
controllability result of Theorem III.2 does not hold for the whole state, unless the control support O is
allowed to move with time (i.e. O = O(t)).

IV. Exact controllability

In this section, we reconsider the system (1) and we will use the results of the previous section to
investigate the exact controllability for (1).

12
i. Global exact controllability
Let us recall the following null controllability result for linear system, which will be needed to
achieve the exact controllability later.

Lemma IV.1 [22] Consider the system



 ψt = ∆ψ + b( x, t)ψ + 1ω u2 ( x, t), in Ω × (t0 , T )
ψ = 0, on ∂Ω × (t0 , T ) (22)
 2
ψ(·, t0 ) = g ∈ L (Ω), in Ω

where 0 ≤ t0 < T and ω is a nonempty open subset of Ω. If b ∈ L∞ (Ω × (t0 , T )), then there is a control
u2 ∈ L∞ (Ω × (t0 , T )) such that the corresponding solution to (22) vanishes at T. Furthermore, we have:

ku2 k L∞ (Ω×(t0 ,T )) ≤ C k gk L2 (Ω) , (23)

where C is a positive constant depending on T − t0 .

In the sequel, the same letter C will denote a generic positive constant (which is independent
of T) that can change from one line to another one. Also, for a fixed target state yd we set
Λ := { x ∈ Ω; yd ( x ) 6= 0}.
Our exact controllability result is stated as follows.

Theorem IV.2 Let y0 ∈ L2+ (Ω) − (0) and yd ∈ L2+ (Ω) satisfy the following assumptions:
∆yd
(i) yd ∈ H01 (Ω) ∩ H 2 (Ω) and yd Λ
1∈ L∞ (Ω), and (ii) yd ≥ µ > 0, a.e. in some nonempty open subset
O of Ω. Then for any T∗ > 0, there exists a control v ∈ L∞ ( Q T∗ ) for which the solution of system (1)
satisfies : y( T∗ ) = yd , a.e. in Ω.

Proof 9 Let e > 0 be fixed. According to Corollaries 3 & 4, we deduce that for any T∗ > 0 and for all
0 < T < T∗ , there exist h, σ > 0 for which the following control:

 −σ,
 t ∈ (0, T2 ), (i.e. T1 = T2 )
v( x, t) =

 u0 ( x,t) , t ∈ ( T , T )
ϕ( x,t) 2

is such that v ∈ L∞ ( Q T ) and the respective mild solution z(t) satisfies the estimate:

ky( T ) − yd k L2 (Ω) < e. (24)

1 d
Recall that ϕ in the definition of v is the solution of (20) corresponding to control u0 ( x, t) = h (y +
S( T )y( T1 ) − S( T − T1 )y( T1 ))1(T −h,T ) with T1 = T2 .
Let us consider the following system

 zt = ∆z + v( x, t)(z + yd ) + ∆yd , in Ω × ( T, T∗ )
z = 0, on ∂Ω × ( T, T∗ ) (25)

z( T ) = y( T − ) − yd , in Ω

Then, the problem of steering the system (1) from y0 to yd at time T∗ turns out to steering the system (25)
from z( T ) to 0 at T∗ .
The main idea in the remainder of the proof consists in using the null controllability of the heat equation

13
under additive control and small initial state.
We will determine a time T 0 := T + η, (for η > 0 small enough and independent of T) such that
kz( T 0 )k L∞ (Ω) ≤ e.
For this end, we consider the following time-independent control on ( T, T 0 ) :

∆yd
v = b := − 1Λ , a.e. x ∈ Ω.
yd
With this control, the system (25) is reduced to the following homogeneous one :

 zt = ∆z + b( x )z, in Ω × ( T, T 0 )
z = 0, on ∂Ω × ( T, T 0 ) (26)
 − d
z( T ) = y( T ) − y , in Ω
whose solution is given by
Z t
z(t) = S(t − T )z( T ) + S(t − s)b( x )z(s)ds, ∀t ∈ ( T, T 0 ]. (27)
T

Now for λ > kbk L∞ (Ω) , the solution of the following system:
 
 ζ t = ∆ζ + b( x ) − λ ζ, in Ω × ( T, T 0 )
ζ = 0, on ∂Ω × ( T, T 0 ) (28)

ζ ( T ) = z ( T ), in Ω
satisfies the following estimate (see [6], p. 259)
−n
kζ (t)k L∞ (Ω) ≤ C (t − T ) 2 kz( T )k L1 (Ω) , t ∈ ( T, T 0 ), (29)
for some C > 0.
Observing that the solution of (28) is such that
ζ (t) = e−λ(t−T ) z(t), t ∈ ( T, T 0 ),
where z is the solution of (26), we deduce from (29) that
−n
kz(t)k L∞ (Ω) ≤ C (t − T ) 2 eλ(t−T ) kz( T )k L1 (Ω) , t ∈ ( T, T 0 )
and so
kz( T 0 )k L∞ (Ω) ≤ C kz( T )k L1 (Ω) , C = C (η ) > 0.
This, together with (24) implies
kz( T 0 )k L∞ (Ω) ≤ Ce, C = C (η ) > 0. (30)
Let us now proceed to the final phase of determining the time and control of the exact steering.
In the sequel, we consider the control
v( x, t) = b( x ) + v1 ( x, t), ( x, t) ∈ Ω × ( T 0 , T∗ ),
where v1 is a new control to be determined such that z( T∗ ) = 0.
With such a control v; the system (25) becomes

 zt = ∆z + b( x )z + v1 ( x, t)(z + yd ), in Ω × ( T 0 , T∗ )
z = 0, on ∂Ω × ( T 0 , T∗ ) (31)
 0 0−
z ( T ) = z ( T ), in Ω

14
Let us introduce the following linear auxiliary system :

 ψt = ∆ψ + b( x )ψ + 1O u2 ( x, t), in Ω × ( T 0 , T∗ )
ψ = 0, on ∂Ω × ( T 0 , T∗ ) (32)
 0 0−
ψ ( T ) = z ( T ), in Ω

From Lemma IV.1, we deduce that for any T∗ = T 0 + η, (where η > 0 is an arbitrary positive constant and
small enough), there exists a control u2 ∈ L∞ (Ω × ( T 0 , T∗ )) such that the corresponding solution to (32) is
such that ψ( T∗ ) = 0, a.e. in Ω, and we have

ku2 k L∞ (Ω×(T 0 ,T∗ )) < C kψ( T 0 )k L2 (Ω) , for some C = C (η ) > 0. (33)

As above, we derive from ([6], p. 259) that for λ > kbk∞ , the solution of

 ξ t = ∆ξ + (b( x ) − λ)ξ + 1O w( x, t), in Ω × ( T 0 , T∗ )
ξ = 0, on ∂Ω × ( T 0 , T∗ ) (34)
 0 0−
ξ ( T ) = z ( T ), in Ω

is such that:
 Z t 
0 −n 0
kξ (t)k L∞ (Ω) ≤ C (t − T ) 2 kz( T )k L1 (Ω) + kw(·, s)k L∞ (Ω) ds , t ∈ ( T 0 , T∗ ),
T0
0
for some C > 0, where w(t) = eλ(T −t) u2 (t) ∈ L∞ (Ω × ( T 0 , T∗ )).
Then it comes from the variation of constants formula that
Z t Z t
0 0
eλ(t− T ) ξ ( t ) = S ( t − T 0 ) z ( T 0 ) + S(t − s)1O u2 (s)ds + S(t − s)b( x )eλ(T −s) ψ(s)ds,
T0 T0

from which we can see that


0
ψ ( t ) = e λ ( t − T ) ξ ( t ),
and hence for some C > 0, we have
 Z t 
−n
kψ(t)k L∞ (Ω) ≤ C (t − T 0 ) 2 kz( T 0 )k L1 (Ω) + ku2 (·, s)k L∞ (Ω) ds , t ∈ ( T 0 , T∗ )·
T0

From this and (30) & (33), we deduce that

kψ(t)k L∞ (Ω) < C∗ e, ∀t ∈ [ T 0 , T∗ ],


where C∗ > 0 is a constant which depends on η but not on T.
Since all the constants appeared in this proof can be chosen only η −dependent (i.e. do not depend on T),
and since the small nonnegative parameter η can be arbitrary chosen, we can choose e small enough (e.g.
0 < e < C∗ ) so that |ψ( x, t) + yd | ≥ µ̄, a.e. on O × ( T 0 , T∗ ) for some µ̄ > 0. This enables us to define a
µ

control v1 ( x, t) on Ω × ( T 0 , T∗ ) by the following relation

v1 ( x, t)(ψ( x, t) + yd ) = 1O u2 ( x, t), a.e. ( x, t) ∈ Ω × ( T 0 , T∗ ),

Moreover, since u2 ∈ L∞ (Ω × ( T 0 , T∗ )), we have v1 ∈ L∞ (Ω × ( T 0 , T∗ )).


Then, controlling the system (31) with this control v1 ( x, t) leads to the following system

u2 ( x,t) d 0
 zt = ∆z + b( x )z + ψ( x,t)+yd 1O (z + y ), in Ω × ( T , T∗ )

z = 0, on ∂Ω × ( T 0 , T∗ ) (35)


z( T 0 ) = z( T 0− ), in Ω

15
which admits ψ as a solution. Hence, by uniqueness, we have z = ψ, a.e in Ω × ( T 0 , T∗ ).
In conclusion, for any T > 0, the control defined by :


 −σ, (0, T2 )

 u ( x,t)

 ϕ0( x,t) ( T2 , T )
v( x, t) = d

 − ∆yyd Λ
1 , ( T, T + η )



 − ∆y d u2 ( x,t)
yd Λ
1 + ψ( x,t 1 ,
)+yd O
( T + η, T + 2η )

lies in L∞ (Ω × (0, T∗ )) and the respective solution of the initial system (1) is such that y( T∗ ) = yd . This
achieves the proof.

Remark 6 1. The time T∗ of exact controllability can be chosen uniform with respect to initial and
target states.
2. For any fixed initial state y0 ∈ L2 (Ω), the set (identified in Theorem IV.2) of target state which are
exactly accessible from y0 is a cone convex.
3. The exact controllability of the one-dimensional version of (1) has been considered in [41]. The result
of Theorem IV.2 improves the one of [41] in term of assumptions on y0 and yd (which are relaxed)
and also in term of the control time, which is here, uniform with respect to initial and target states.
4. The exact controllability of equations like (1) with inhomogeneous Dirichlet conditions has been
considered in [37, 40]. However, the assumptions imposed there do not allow the homogeneous
Dirichlet conditions.

ii. Partial exact controllability


Let us show the following exact controllability result.

Theorem IV.3 Let ω be an open subset of Ω. Let p ∈ R be such that p ≥ 2 and p > n. Let y0 ∈
p
L+ (Ω) − (0) and yd ∈ W 2,∞ (ω ). Assume that for some open subset O of Ω such that ω ⊂ O with
O ⊂ Ω, the target state yd possesses an extension ŷd ∈ H01 (Ω) ∩ W 2,∞ (Ω) such that (i) ŷd ≥ 0 a.e. in Ω
and ŷd ≥ µ > 0, a.e. in O and (ii) ∆ŷd = 0 in Ω \ O.
Then for any T∗ > 0, there exists a control v ∈ L∞ (Ω × (0, T∗ )), for which the solution of system (1) is
such that : y( T∗ ) = yd , a.e. in ω.

Proof 10 We will follows the strategy used in the proof of Theorem IV.2. Let e > 0 be fixed and let
T∗ > 0 to be determined. Since p ≥ 2, one can use the results of Corollaries 3 & 4 to deduce that for all
0 < T < T∗ , there are h, σ > 0 such that the following control

 −σ,
 t ∈ (0, T1 ), (with T1 := T2 )
v( x, t) =

 u0 ( x,t) , t ∈ ( T , T )
ϕ( x,t) 1

where ϕ is the solution of (20) corresponding to control u0 ( x, t) = 1h (ŷd + S( T )y( T1 ) − S( T − T1 )y( T1 ))1(T −h,T ) ,
lies in L∞ ( Q T ) and provides the following estimates

ky( T1 )k L2 (Ω) < e

and
ky( T ) − ŷd k L2 (Ω) < e

16
Under the control u0 defined above, the respective solution ϕ to (20) satisfies the following integral relation
Z h  
1 d
ϕ( T ) = S( T − T1 )y( T1 ) + S(s) ŷ + S( T )y( T1 ) − S( T − T1 )y( T1 ) ds
h 0

Moreover, we know that for some C0 > 0 we have (see [16], p. 44)
n
kS(t)yk L∞ (Ω) ≤ C0 t− 4 kyk L2 (Ω) , ∀t > 0, ∀y ∈ L2 (Ω).

It follows that
n
kS( T )y( T1 )k L∞ (Ω) < C0 T − 4 e. (36)
It comes from the regularity effect of the heat semigroup that RS(t)y0 ∈ D ( A) for all t > 0 and all
r
y0 ∈ L2 (Ω) . Then, having in mind that A is closed and that 0 S(t)y0 dt ∈ D ( A), for all r > 0 and
all y0 ∈ L2 (Ω) (see [16, 42]), we can see from the above expression of ϕ( T ) that ϕ( T ) ∈ D ( A) and
∆ϕ( T ) → ∆ ŷd + S( T )y( T1 ) , as h → 0+ . On the other hand, by the Sobolev embedding, we deduce that
ϕ( T ) → ŷd + S( T )y( T1 ), as h → 0+ in W 1,p (Ω) and hence (recall that p > n) ϕ( T ) → ŷd + S( T )y( T1 ),
as h → 0+ in L∞ (Ω) (see [13], pp. 168-169). This together with (36) implies for h > 0 small enough

k ϕ( T ) − ŷd k L∞ (Ω) < C1 e, (C1 > 0)·

Moreover, from the definition of the control v(t) in [ T1 , T ], we deduce that y = ϕ in [ T1 , T ] for h > 0 small
enough.
Let us now consider the following system

 zt = ∆z + v( x, t)(z + ŷd ) + ∆ŷd , in Ω × ( T, T∗ )
z = 0, on ∂Ω × ( T, T∗ ) (37)
 − d
z( T ) = y( T ) − ŷ , in Ω

Then, the problem of steering the system (1) in ω from y0 to yd at time T∗ turns out to steering the linear
system (37) in ω from z( T ) to 0 at T∗ .
For h > 0 small enough we have ϕ = y( T ) = z( T ) + ŷd , so we deduce that

kz( T )k L∞ (Ω) ≤ C1 e. (38)

Let us introduce the following linear system:



 φt = ∆φ + 1O u2 ( x, t), in Ω × ( T, T∗ )
φ = 0, on ∂Ω × ( T, T∗ ) (39)

φ ( T ) = z ( T − ), in Ω

Then, for all T∗ = T + η, (where η > 0 is an arbitrary positive constant and small enough), there exists a
control u2 ∈ L∞ (O × ( T, T∗ )) such that the corresponding solution to (39) is such that : φ( T∗ ) = 0, a.e.
in Ω, and we have

ku2 k L∞ (O×(T,T∗ )) < C2 kφ( T )k L2 (Ω) , for some C2 = C2 (η ) > 0. (40)

For all t ∈ [ T, T∗ ], we have


Z t
φ(t) = S(t − T )z( T ) + S(t − s)1O u2 (s)ds. (41)
T

17
Using the same arguments as in the proof of Theorem IV.2, one can shows that:

kφ(t)k L∞ (Ω) < C∗ e, ∀t ∈ [ T, T∗ ],

where C∗ is some constant which depends on η but not on T.


Hence, we can choose e small enough such that |φ( x, t) + ŷd | ≥ µ̄, a.e. on O × ( T, T∗ ), for some µ̄ > 0.
This enables us to define a multiplicative control v( x, t) on Ω × ( T, T∗ ) through the following relation:

v( x, t)(φ( x, t) + ŷd ) + ∆ŷd = 1O u2 ( x, t), a.e. ( x, t) ∈ Ω × ( T, T∗ ), (e > 0)

Thus letting
u2 ( x, t) − ∆ŷd
v( x, t) = 1O , a.e. ( x, t) ∈ Ω × ( T, T∗ ),
φ( x, t) + ŷd
we can see, since ŷd ∈ W 2,∞ (O) and u2 ∈ L∞ (Ω × ( T, T∗ )) that v ∈ L∞ (0, T∗ ; L2 (Ω)).
Then controlling the system (37) with such a control v( x, t) leads to the following system:

1 u ( x,t)−∆ŷd
 zt = ∆z + O φ2( x,t)+ŷd (z + ŷd ) + ∆ŷd , in Ω × ( T, T∗ )

z = 0, in ∂Ω × ( T, T∗ ) (42)


z ( T ) = z ( T − ), in Ω

which admits φ as a solution.


In conclusion for any T > 0, the control defined by:


 −σ, (0, T2 )
 u0 ( x,t)
v( x, t) = ϕ( x,t)
( T2 , T )

 d
 u2 ( x,t)−∆ŷ 1O , ( T, T + η )
φ( x,t)+ŷd

lies in L∞ (Ω × (0, T∗ )), where ϕ is the solution of (20) corresponding to u0 ( x, t) = 1h (ŷd + S( T )y( T2 ) −
S( T2 )y( T2 ))1(T −h,T ) , for h > 0 small enough.
Consequently, we have y( T∗ ) = ŷd , and so y( T∗ )|ω = yd .

Remark 7 1. In the case n = 1 one can get the partial controllability in the usual state space L2 (Ω).
2. The problem of partial controllability occurs when one is unable to achieve the controllability in the
whole domain. Furthermore, it may happen that the system at hand is not controllable in Ω but it can
be controllable in a sub-domain ω which may be very close to Ω.

Examples 1 In this part, we give some illustrating examples.

Situation 1: Finite-time stability. √


Let Ω = (0, 1) and let us take the initial state y0 = 2 sin(πx ), x ∈ (0, 1) be the unitary eigenfunction
of the Laplacian associated to the first eigenvalue λ0 = −π 2 . Then, observing that y0 ≥ 0 on (0, 1), one
can look for the solution in the form y(t) = λ(t)y0 , with λ(t) ≥ 0.
By replacing in the equation we can see that y satisfies (1) iff λ satisfies the following Bernoulli type
equation:
λ̇(t) − λ0 λ(t) + ρλ1−µ (t) = 0, λ(0) = 1
whose solution is given for all t ∈ [0, T∗ ) (T∗ being the settling time, so that λ does not vanish in [0, T∗ ))
by:
ρ  µ1
λ(t) = e−µλ0 t c +
λ0

18
where c = 1 −
ρ −1 ln(1 − λ0
λ0 . Hence the settling time is given by T∗ = λ0 µ ρ ).
k y0 k µ
Let us check the estimate T∗ ≤ µρ founded in Theorem II.1.
Observing that
k y0 k µ π2 π2
T∗ ≤ ⇔ ln(1 + )≤ ,
µρ ρ ρ
k y0 k µ
we deduce the claimed estimate for T∗ . In fact, here we have T∗ < µρ .

Situation 2: Global exact steering.


Let Ω = (0, 1), y0 = x and yd = sin(πx ). First, let us note that the results established in [37, 40] do not
apply here, as the boundary conditions in (1) are homogeneous. Moreover, according to the existing results
from the literature [14, 15, 29], one can only get the approximate steering at a time depending on y0 and
yd . Now, observing that the assumptions of Theorem IV.2 are satisfied with O = ( a, b), 0 < a < b < 1,
we deduce that for any a priori given T > 0, there is a control v ∈ L∞ ( Q T ) that steers (1) from y0 to yd at T.

Situation 3: Partial steering.


Let Ω = (0, 3) and 
 x, x ∈ [0, 3/2]
y0 =

− x + 3, x ∈ [3/2, 3]
Let us set  2 2
yd = 1 − ( x − 1)( x − 2) e( x−1) ( x−2) , x ∈ Ω.

Here the assumptions of Theorem IV.2 are not fulfilled for the full state yd .
Now let ω = O = (1, 2). Then observing that (yd )00 (1) = (yd )00 (2) = 0, we can build the extention yd |ω
by letting,
 d 0
 (y )+ (1)( x − 1) + yd (1), x ∈ [0, 1]
d
ŷ = y d ( x ), x ∈ [1, 2]
 d 0
(y )− (2)( x − 2) + yd (2), x ∈ [2, 3]
where (yd )0− (resp. (yd )0+ ) refers to left-hand (resp. right-hand) derivatives of yd .
Thus 
 x, x ∈ [0, 1]
d  ( x −1)2 ( x −2)2
ŷ = 1 − ( x − 1)( x − 2) e , x ∈ [1, 2]

− x + 3, x ∈ [2, 3]
Hence according to Theorem IV.3, the system (1) can be exactly steered from y0 to yd in ω at a small time.

V. Conclusion
In this work, various kind of controllability are discussed in the context of a multidimensional
reaction-diffusion equation governed by bilinear control. First, the finite-time stability is inves-
tigated using feedback law in the case of a settling time depending on the inial condition. Also
the case of a prescribed settling time is discussed by making use of time-varying feedback laws.
Then a multiplicative null-controllability result is established via linear controllability problem,
and extended to exact controllability of states of the free system. Then, several results on partial
and global approximate controllability are given. These results are applied to study the partial
and global exact controllability. Formulae for steering controls are given as well. Applications to

19
nonnegative controllability are also provided. Furthermore, in the case of nonnegative approximate
and exact controllability the control time can be chosen arbitrary small and uniform with respect
to initial and desirable states. The techniques used for finite-time stability of the equation (1) is
promising to tackle the same question for general bilinear equations.

References
[1] K. Ammari and M. Tucsnak, Stabilization of second order evolution equations by a class of
unbounded feedbacks, ESAIM Control Optim. Calc. Var., 6 (2001), 361–386.

[2] S. Angenent, The zero set of a solution of a parabolic equation. J. Reine Angew. Math, 390
(1988), 79-96.

[3] Ball, J. M. (1978). On the asymptotic behavior of generalized processes, with applications to
nonlinear evolution equations. Journal of differential equations, 27(2), 224-265.

[4] J.M. Ball, J.E. Marsden, and M. Slemrod, Controllability for distributed bilinear systems.
SIAM J. Control and Optim., 20 (1982), 575-597.

[5] C. Bardos & L. Tartar, (1973). Sur l’unicité rétrograde des équations paraboliques et quelques
questions voisines. Archive for Rational Mechanics and Analysis, 50(1), 10-25.

[6] V. Barbu, Analysis and control of nonlinear infinite dimensional system, Academic Press,
Boston, 1993.

[7] K. Beauchard, Local controllability of a 1-D Schrödinger equation, J. Math. Pures Appl., 84
(2005), 851-956.

[8] K. Beauchard and J.-M. Coron, Controllability of a quantum particle in a moving potential
well, J. Funct. Anal., 232 (2006), 328-389.

[9] K. Beauchard, Local controllability of a 1-dimensional beam equation, SIAM J. Control.


Optim., 47 (2008), 1219-1273.

[10] K. Beauchard, Local controllability and non-controllability for a 1D wave equation with
bilinear control. J. Differential Equations., 250 (2011), 2064-2098.

[11] M. E. Bradley, S. Lenhart, Bilinear optimal control of a Kirchhoff plate. Systems & Control
Letters., 22 (1994) 27-38.

[12] 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.

[13] H. Brezis. Analyse fonctionnelle. Théorie et applications. Masson, Paris, 1983.

[14] P. Cannarsa and A. Khapalov, Multiplicative controllability for reaction-diffusion equation


with target states admitting finitely many changes of sign, Discrete and Continuous Dynamical
Systems Series B, 14 (2010), 1293-1311.

[15] P. Cannarsa, G. Floridia, & A. Y. Khapalov, Multiplicative controllability for semilinear


reaction-diffusion equations with finitely many changes of sign. Journal de Mathématiques
Pures et Appliquées, 108(4) (2017), 425-458.

20
[16] T. Cazenave and A. Haraux, An introduction to semilinear evolution equations, Oxford
Lecture Series in Mathematics and its Applications 13, Oxford University Press, Oxford, 1998

[17] Coron, J. M., & Nguyen, H. M. (2017). Null controllability and finite time stabilization for
the heat equations with variable coefficients in space in one dimension via backstepping
approach. Archive for Rational Mechanics and Analysis, 225(3), 993-1023.

[18] CortéS, J. (2006). Finite-time convergent gradient flows with applications to network consensus.
Automatica, 42(11), 1993-2000.

[19] A. Doubova, E. Fernandez-cara, E. Zuazua, On the controllability of parabolic systems with


a nonlinear term involving the state and the gradient. SIAM. J. Control Optim. 41, 798-819,
(2003).

[20] Espitia, N., Polyakov, A., Efimov, D., & Perruquetti, W. Boundary time-varying feedbacks for
fixed-time stabilization of constant-parameter reaction-diffusion systems. Automatica, 103,
398-407, (2019).

[21] C. Fabre, J.P. Puel, E. Zuazua, Approximate controllability of the semilinear heat equation.
Proc. Roy. Soc. Edinburgh 125A, 31-61, (1995).

[22] E. Fernàndez-Cara, E. Zuazua, Null and approximate controllability for weakly blowing-up
semilinear heat equations, Ann. Inst. H. Poincaré Anal. Non Linéaire. 17, (2000), 583-616.

[23] E. Fernández-Cara and S. Guerrero. Global Carleman inequalities for parabolic systems and
applications to controllability. SIAM J. Control Optim. 45 (2006), 1395-1446.

[24] L. A. Fernández and A. Y. Khapalov. Controllability properties for the one-dimensional Heat
equation under multiplicative or nonnegative additive controls with local mobile support.
ESAIM: Control Optim. Calc. Var., Volume 18, Number 4, (2012), 1207-1224.

[25] A.V. Fursikov, O.Yu. Imanuvilov, Controllability of Evolution Equations, in: Lecture Notes
Series, vol. 34, Seoul National University, Korea, (1996).

[26] Holloway, J. C., & Krstic, M. (2019). Prescribed-time observers for linear systems in observer
canonical form. IEEE Transactions on Automatic Control, to appear.

[27] K. Kime, Simultaneous control of a rod equation and a simple Schrödinger equation, Systems
& Control Letters., 24 (1995), 301-306.

[28] A. Y. Khapalov, Global non-negative controllability of the semilinear parabolic equation


governed by bilinear control, ESAIM: Control Optim. Calc. Var., 7 (2002), 269-283.

[29] A. Y. Khapalov, Controllability of the semilinear parabolic equation governed by a multiplica-


tive control in the reaction term: A qualitative approach, SIAM J. Control. Optim., 41 (2003),
1886-1900.

[30] A. Y. Khapalov, Controllability properties of a vibrating string with variable axial load,
Discrete Cont. Dyn. Syst., 11 (2004), 311-324.

[31] A. Y. Khapalov, Reachability of nonnegative equilibrium states for the semilinear vibrating
string by varying its axial load and the gain of damping, ESAIM: Control Optim. Calc. Var.,
12 (2006), 231-252.

21
[32] A. Y. Khapalov, Controllability of partial differential equations governed by multiplicative
controls, A Lecture Notes in Mathematics 1995, Springer-Verlag, Berlin, 2010.
[33] O. A. Ladyzhenskaya, V. A. Solonnikov, and N. N. Uraltseva, Linear and Quasilinear Equa-
tions of Parabolic Type, Am. Math. Soc, Providence, R. I., 1968.
[34] G. Lebeau and L. Robbiano, Contrôle exacte de l’équation de la chaleur, Comm. PDE 20
(1995), 335-356.
[35] S. Lenhart, Optimal control of convective-diffusive fluid problem, Math. Models Methods
Appl. Sci., 5 (1995), 225-237.
[36] M. Liang, Bilinear optimal control for a wave equation, Mathematical Models and Methods
in Applied Sciences, 9 (1999), 45-68.
[37] P. Lin, Zhongcheng Zhou and Hang Gao, Exact controllability of the parabolic system with
bilinear control. Applied Mathematics Letters, 19 (2006), 568-575.
[38] Mitrinovic, D. S., Pecaric, J., & Fink, A. M. (1991). Inequalities involving functions and their
integrals and derivatives (Vol. 53). Springer Science & Business Media.
[39] M. Ouzahra, Controllability of the wave equation with bilinear controls, European Journal of
Control, 20 (2014), 57-63.
[40] M. Ouzahra, A. Tsouli and A. Boutoulout. Exact controllability of the heat equation with
bilinear control. Mathematical methods in the applied sciences, 38(18) (2015), 5074-5084.
[41] M. Ouzahra, Approximate and exact controllability of a reaction-diffusion equation governed
by bilinear control, European Journal of Control, 32 (2016), 32-38.
[42] Pazy, A. Semi-groups of linear operators and applications to partial differential equations,
Springer Verlag, New York, 1983.
[43] Polyakov, A., Coron, J. M., & Rosier, L. (2018). On homogeneous finite-time control for linear
evolution equation in hilbert space. IEEE Transactions on Automatic Control, 63(9), 3143-3150.
[44] A.I. Prilepko, D.G. Orlovsky and I.A. Vasin, Methods for solving inverse problems in mathe-
matical physics. Marcel Dekker Inc., New York (2000).
[45] Zarchan, P. (2012). Tactical and strategic missile guidance. American Institute of Aeronautics
and Astronautics, Inc.
[46] Saut, J. C., & Scheurer, B. Unique continuation for some evolution equations. Journal of
differential equations, 66(1) (1987), 118-139.
[47] Steeves, D., Krstic, M., & Vazquez, R. (2019, June). Prescribed-time H 1 −stabilization of
reaction-diffusion equations by means of output feedback. In 2019 18th European Control
Conference (ECC) (pp. 1932-1937). IEEE.
[48] Song, Y., Wang, Y., Holloway, J., & Krstic, M. (2017). Time-varying feedback for regulation of
normal-form nonlinear systems in prescribed finite time. Automatica, 83, 243-251.
[49] E. Zuazua, Approximate controllability for semilinear heat equations with globally Lipschitz
nonlinearities. Control Cybern. 28, (1999), 665-683.

Received xxxx 20xx; revised xxxx 20xx.

22

You might also like