Optimal approximation of internal

controls for a wave-type problem with

fractional Laplacian

Ionel Rovenţa

Department of Mathematics, University of Craiova, Romania

Dynamics, Equations and Applications (DEA 2019)

Krakow, September, 19, 2019

joint work with P. Lissy, CEREMADE, Université Paris-Dauphine

September 26, 2020

The control for the continuous wave equation

Given T ≥ 2 and (u0 , u1 ) ∈ L2 ((0, 1), C) × H −1 ((0, 1), C) there exists

a control function v ∈ C 0 ([0, T ], C) such that the solution of the wave
equation
 00

 u (t, x) − uxx (t, x) = 0 t ∈ (0, T ), x ∈ (0, 1),
u(t, 0) = 0 t ∈ (0, T ),

(1)
 u(t, 1) = v(t) t ∈ (0, T ),
0
 0 1
u(0, x) = u (x), u (0, x) = u (x) x ∈ (0, 1),


satisfies
u(T, x) = u0 (T, x) = 0 (x ∈ (0, 1)).
The discrete model of the wave equation

Let N ∈ N∗ and h = N1+1 . We consider the following semi-discrete

space approximation of the wave equation by the explicit
finite-differences method:
u (t)−2uj (t)+uj−1 (t)

u00 (t) − j+1 = 0 1 ≤ j ≤ N, t > 0,
 j h2


u0 (t) = 0 t ∈ (0, T ),
(2)
u (t) = v (t) t ∈ (0, T ),
 N +1 h


0 0 1
uj (0) = uj , uj (0) = uj 1 ≤ j ≤ N.

The discrete controllability problem

Given T ≥ 2, h > 0 and ((u0j , u1j ))1≤j≤N ∈ C2N , there exists a control
function vh ∈ C 0 ([0, T ]) such that the solution of the system of
equations (2) verifies

uj (T ) = u0j (T ) = 0 (j = 1, 2, ..., N ).
discrete model vs discreet model
discrete horses
the limit
The convergence property does not holds

The main problem

The sequence of discrete controls (vh )h>0 converges to a control v of
the continuous wave equation?

In general, there exist high-frequency spurious solutions generated by

the discretization process that make the discrete controls diverge
when the mesh-size h goes to zero. (Glowinski - Li - Lions (’90),
Infante - Zuazua (’99), Micu (’03), Zuazua (’05), etc...)
The first strategy???
The aim

Basically, this difficulty can be overcame by using an appropriate

filtering technique to eliminate the short wave length components of
the solutions/initial data of the discrete system, i.e. the large
frequencies (of order |n| = N ) of the discretized problem.
To filtering the initial data in an optimal range in order to
restore the uniform controllability property.

To obtain a relation between the range of filtration and the

minimal time of control.
The main result

Theorem (Lissy, Rovenţa, Math. Comp. 2019)

Let (u0 , u1 ) ∈ L2 ((0, 1), C) × H −1 ((0, 1), C) a filtered initial data such
that
X
(u0 , u1 )1≤j≤N = an Φn (x).
1≤|n|≤f (N )

4
Then, for any T > πΓ(f ) , there exists a control
1−sin( 2 )
0
vh ∈ C ([0, T ], C) bringing the solution of (2) to (0,0) such that the
sequence (vh )h>0 is bounded in C 0 ([0, T ], C).

Micu [Numer. Math. (’02)] proved that there exists regular initial
data (with f (N ) = N ) for which there exists no sequence of discrete
controls uniformly bounded in L2 (0, T ).
By adding a vanishing viscosity term the problem appearing in the
high frequencies has been solved (Micu, SICON (’08)).
The continuous model for the hinged beam

The boundary controlled transversal vibrations of a 1–D beam with

hinged boundary conditions are modelled by the following equation
 00

 u (t, x) + uxxxx (t, x) = 0 (t, x) ∈ (0, T ) × (0, 1)
u(t, 0) = u(t, 1) = u (t, 0) = 0 t ∈ (0, T )

xx


uxx (t, 1) = v(t) t ∈ (0, T ) (3)
0
u(0, x) = u (x) x ∈ (0, 1)



 0

u (0, x) = u1 (x) x ∈ (0, 1),
 
u
The vector represents the state and v is the control acting on
u0
the extremity x = 1 of the beam.
The same result for the beam

The convergence of the approximate boundary controls corresponding

to initial data in the finite energy space cannot be guaranteed( Leon
and Zuazua (ESAIM COCV, 2002)).

By adding a vanishing viscosity term the problem appearing in the

high frequencies has been solved (Bugariu, Micu, Roventa, Math.
Comp,. (2016)).
What about clamped beam equation?

The controllability problem associated to the beam clamped at both

extremities
Let T > 0. For every initial data (y0 , y1 ) ∈ L2 (0, 1) × H −2 (0, 1) there
exists a control v ∈ L2 (0, T ) such that the solution of
 00

 y (x, t) + yxxxx (x, t) = 0, (x, t) ∈ (0, 1) × (0, T )
y(0, t) = y(1, t) = 0, t ∈ (0, T )


 ∂x y(0, t) = 0, ∂ x y(1, t) = v(t), t ∈ (0, T )
y(x, 0) = y0 (x), ẏ(x, 0) = y1 (x), x ∈ (0, 1),

(4)
verifies
y(x, T ) = ẏ(x, T ) = 0 (x ∈ (0, 1)).
The main result

(Cindea, Micu, Roventa, SICON, 2017)

It is proved the convergence of the approximate boundary controls
corresponding to initial data in the finite energy space, if the initial
data is filtered in the range δN ).

since the discrete finite-difference operator is no more the square

of finite-differences discrete Laplacian, the eigenvalues cannot be
explicitly computed.
for the eigenvalues: algebraic computations combined with
Rouché’s theorem
for the eigenvectors: asymptotic estimates and a discrete
multiplier method.
Between the beam and the wave

the approximation of controls for the wave equation:

2 nπh
λn = sin
h 2

the approximation of controls for hinged beam equation:

4 nπh
λn = 2
sin2
h 2

What is between the BEAM and WAVE?

We consider the approximation of controls a problem with
fractional Laplacian:
 α
2 nπh
λn = sin , α ∈ (1, 2).
h 2
Internal controls for a problem with fractional
Laplacian using finite-difference method
We consider the 1-D Dirichlet-Laplace operator ∆D on
H 2 (0, 1) ∩ H01 (0, 1) with state space L2 (0, 1).
−∆D : D(∆D ) → L2 (0, 1) is a positive definite operator with compact
2 2
resolvent, the k-th eigenvalue is λ √k = k π , with associated
normalized eigenvector ek (x) := 2 sin (kπx) .

Let s > 0. Then, we introduce

∞
X 2
D((−∆D )s ) = {f ∈ L2 (Ω) λ2s
k < f, ek >L2 (0,1) < ∞}.
k=1
P∞
For f = k=1 < f, ek >L2 (0,1) ek ∈ D((−∆D )s ), we define the
spectral fractional Laplace operator
∞
X
(−∆D )s f (x) = λsk < f, ek >L2 (0,1) ek (x).
k=1
The continuous fractional control problem

Let T > 0, α ∈ (1, 2) and 0 < a < b < 1.

 00

 u (t, x) + (−∆D )α u(t, x) = χ(a,b) (x)v(t, x), t ∈ (0, T ), x ∈ (0, 1),
u(t, 0) = 0, t ∈ (0, T ),

 u(t, 1) = 0, t ∈ (0, T ),
u(0, x) = u0 (x), u0 (0, x) = u1 (x),

x ∈ (0, 1),

(5)
where v ∈ L2 ((0, T ) × (0, 1)) is a control localized in (a, b).

We study the existence of a control function v ∈ L2 ((0, T ) × (0, 1))

such that the solution of equation (5) verifies

u(T ) = u0 (T ) = 0.
The space of initial conditions:
( )
X X
α/2 2 0 1 k 2
H = D((−∆D ) )×L (0, 1) = (u , u ) = ak Φ (x) | |ak | < ∞ ,
k∈Z∗ k∈Z∗

where
1 T
Φk (x) = (sin(kπx), −i sgn(k)|k|α sin(kπx)) , k ∈ Z∗ .
sgn(k)|k|α

Note that (i sgn(k)|k|α )k∈Z∗ is the family of eigenvalues of the elliptic

operator associated to the continuous problem (5).
The discrete control problem

Let N ∈ N∗ and h = 1
N +1 . We introduce
     0  1
u1 (t) v1 (t) u1 u1
 u2 (t)   v2 (t)  0
 u02  1
 u12 
U (t) = 
 . . .  , Vh (t) =  . . .  , U =  . . .  and U =  . . .  .
      

uN (t) vN (t) u0N u1N

We consider the following semi-discrete space approximation:

 00
U (t) + Aα h U (t) = Bh (Vh (t)), t ∈ (0, T ),
(6)
U (0, ·) = U 0 , U 0 (0, ·) = U 1 .
Ah ∈ MN (R) is the discrete Laplacian matrix given by
 
2 −1 0 . . . . . . . . . 0
−1 2 −1 0 . . . . . . 0 
 
 0 −1 2 −1 0 . . . 0 
1  
Ah := 2  . . . . . . . . . . . . . . . . . . . . ..
h 
 . . . . . . . . . . . . . . . . . . . . . 

 0 ... 0 0 −1 2 −1
0 . . . . . . . . . 0 −1 2

The discrete control operator Bh ∈ MN (R) is given by

Bh (Vh (t)) := (χ(a,b) (jh)vj (t))T1≤j≤N .

Uniformly boundedness of the sequence of discrete
controls

Let f : N∗ → N∗ be a increasing filtration function.

Theorem
Let T > 0 and α ∈ (1, 2). Consider (ak )k∈Z∗ ∈ l2 (Z∗ , C). Assume that

f (N )
lim f (N ) = ∞ and Γ := lim sup < 1. (7)
N →∞ N →∞ N

For (u0j , u1j )1≤j≤N = 1≤|n|≤f (N ) an Φnh , there exists a uniformly

P

bounded sequence of controls (Vh )h>0 for the discrete control problem
(6) verifying that for any h > 0, the solution (u1 , . . . uN ) of (6)
verifies (u1 (T ), . . . , uN (T )) = (0, . . . , 0).
The uniform controllability may fail if we do not filtrate
the initial conditions

Theorem
Let T > 0 and α ∈ (1, 2). There exists a sequence (ak )k∈Z∗ ∈ l2 (Z∗ , C)
for which no sequence of controls (Vh )h>0 for the discrete control
problem (6) with initial condition (u0j , u1j )1≤j≤N = 1≤|n|≤N an Φnh
P
can be uniformly bounded in h.
The convergence result

Theorem
Let (u0 , u1 ) = n∈Z∗ an Φn ∈ H be the initial datum for the
P
continuous problem (5) and let us consider the initial datum of the
semi-discrete problem (6)
X
(u0j , u1j )1≤j≤N = an Φ̃n (h), (8)
1≤|n|≤f (N )

where Φ̃n (h) is the discretization of the eigenfunction Φn .

Let v ∈ L2 ((0, T ) × (0, 1)) be a weak limit of a subsequence of the
bounded controls (Vh )h>0 , with initial datum given in (8). Then v is a
control for the continuous problem (5).
A variational lemma for the continuous moment
problem
Lemma
Problem (5) is controllable if and only if, for any initial datum
(u0 , u1 ) ∈ H decomposed as in (18), there exists v ∈ L2 ((0, T ) × (0, 1))
such that
Z 1
hw1 , u0 iD((−∆D )α/2 )0 ,D((−∆D )α/2 ) − u1 (x)w0 (x) dx
0
Z b Z T
= v(t, x)w(t, x) dt dx, (9)
a 0

for all (w0 , w1 ) ∈ H0 = L2 (0, 1) × D((−∆D )α/2 )0 , where w is the the

solution of the adjoint problem
 00
w (t, x) + (−∆D )α w(t, x) = 0 t ∈ (0, T ), x ∈ (0, 1),
w(T, x) = w0 (x), w0 (T, x) = w1 (x) x ∈ (0, 1).
(10)
The continuous moment problem

Theorem
The initial datum (u0 , u1 ) ∈ H is controllable if and only if there
exists v ∈ L2 ((0, T ) × (0, 1)) such that
Z b Z T
v(t, x)e−iνn t Φn (x) dt dx = 2an (n ∈ Z∗ ), (11)
a 0

where νn = π α sgn(n)|n|α and (iνn )n∈Z∗ are the eigenvalues of the

elliptic operator associated to the continuous problem (10).
The discrete adjoint system

We consider Lα
h ∈ M2N ×2N (R) given by
 
0 −In
Lα
h :=
 , (12)
α
Ah 0
where In ∈ Mn (R) is the identity matrix of size n.

For Z(t) = (W (t), W 0 (t))T and Z T = (W 0 , W 1 )T , we consider the

discrete adjoint system
 0
Z (t) + Lα
h Z(t) = 0, t > 0, (13)
Z(T ) = Z T .
 The eigenvalues and eigenvectors for the discrete
adjoint system

The eigenvalues of the discretization matrix Lα h are given by the

family (i λn (h))1≤|n|≤N , where
  α
2 |n|πh
λn (h) = sgn(n) sin , 1 ≤ |n| ≤ N,
h 2
and the corresponding eigenvectors are

ϕnh
 
1
Φnh = , 1 ≤ |n| ≤ N,
λn (h) −i λn (h)ϕnh
where

(ϕnh )1≤|n|≤N = (sin(nπh), sin(2nπh), · · · , sin(N nπh)) ∈ CN .

The discrete moment problem

Theorem
The discrete P
problem (2) is controllable if and only if, for any initial
datum Z 0 = 1≤|k|≤N αk Φkh ∈ C2N there exists a function

N
Vh ∈ L2 (0, T ) such that

X Z T
h χ(a,b) (jh) sin(jπnh) e−iλn t vj (t) dt = 2αn , 1 ≤ |n| ≤ N.
1≤j≤N 0
The biorthogonals

A sequence (θm )1≤|m|≤N ⊂ L2 − T2 , T2 is biorthogonal to the family

of exponential functions eiλn t 1≤|n|≤N in L2 − T2 , T2 if

Z T
2
θm (t)eiλn t dt = δmn (1 ≤ |m|, |n| ≤ N ) .
− T2
An explicit formula for the discrete control

Once we have a biorthogonal sequence (θm )1≤|m|≤N to the family

(eiλn t )1≤|n|≤N in L2 − T2 , T2 we can construct a control as:

2an Bhn −iλn T

 
X T
Vh (t) = e 2 Θn t − ,
||Bhn ||2 2
1≤|n|≤f (N )

where Bhn is given by

Bhn = h(χ(a,b) (jh) sin(jπnh))1≤j≤N .

Construction of the biorthogonals

We construct and evaluate an explicit
biorthogonal sequence to the

family (eiλn t )1≤|n|≤N in L2 − T2 , T2 in the following way:
1 We construct an entire function Pm , with the property that
Pm (λn ) = δmn .
Construction of the biorthogonals

We construct and evaluate an explicit
biorthogonal sequence to the

family (eiλn t )1≤|n|≤N in L2 − T2 , T2 in the following way:
1 We construct an entire function Pm , with the property that
Pm (λn ) = δmn .
2 We obtain an estimate of the product Pm on the real axis.
Construction of the biorthogonals

We construct and evaluate an explicit
biorthogonal sequence to the

family (eiλn t )1≤|n|≤N in L2 − T2 , T2 in the following way:
1 We construct an entire function Pm , with the property that
Pm (λn ) = δmn .
2 We obtain an estimate of the product Pm on the real axis.
3 We construct a smart multiplier Mm with rapid decay on the real
axis such that Pm Mm ∈ L1 (R) ∩ L2 (R) and Mm (λn ) = δmn .
Construction of the biorthogonals

We construct and evaluate an explicit
biorthogonal sequence to the

family (eiλn t )1≤|n|≤N in L2 − T2 , T2 in the following way:
1 We construct an entire function Pm , with the property that
Pm (λn ) = δmn .
2 We obtain an estimate of the product Pm on the real axis.
3 We construct a smart multiplier Mm with rapid decay on the real
axis such that Pm Mm ∈ L1 (R) ∩ L2 (R) and Mm (λn ) = δmn .
4 The Fourier transform of the entire function
ψm (z) := Pm (z)Mm (z) gives the element θm of a biorthogonal
sequence to the family (eiλn t )1≤|n|≤N in L2 − T2 , T2 .


Open problems

the extension to the N-dimensional case

to consider non-uniform meshes
to obtain general controllability results depending on gap
conditions between the eigenvalues