J. Math. Anal. Appl.

410 (2014) 455–468

Contents lists available at ScienceDirect

Journal of Mathematical Analysis and

Applications
www.elsevier.com/locate/jmaa

Optimal control of a singular PDE modeling transient MEMS

with control or state constraints
Christian Clason a , Barbara Kaltenbacher b,∗
a
Institute for Mathematics and Scientific Computing, Karl-Franzens-Universität Graz, Heinrichstr. 36, 8010 Graz, Austria
b
Institute of Mathematics, Alpen-Adria-Universität Klagenfurt, Universitätsstr. 65–67, 9020 Klagenfurt, Austria

a r t i c l e i n f o a b s t r a c t

Article history: A particular feature of certain microelectromechanical systems (MEMS) is the appearance
Received 6 September 2012 of a so-called “pull-in” instability, corresponding to a singularity in the underlying PDE
Available online 4 September 2013 model. We here consider a transient MEMS model and its optimal control via the dielectric
Submitted by P. Yao
properties of the membrane and/or the applied voltage. In contrast to the static case, the
Keywords:
control problem suffers from low dimensionality of the control compared to the state and
Optimal control of PDEs hence requires different techniques for establishing first order optimality conditions. For
Control constraints this purpose, we here use a relaxation approach combined with a localization technique.
State constraints © 2013 Elsevier Inc. All rights reserved.
Singular PDEs
Microelectromechanical systems (MEMS)

1. Introduction

Consider the initial boundary value problem

⎧
⎪
⎪ 2 b(t )a(x)
⎨ ytt + c yt + dy + ρ  y − η y + (1 + y )2 = 0 in Q = (0, T ) × Ω,
⎪
(1.1)
⎪
⎪ y = ∂ν y = 0 on (0, T ) × ∂Ω,
⎪
⎩ 0 1
y=y , yt = y in {0} × Ω,

for Ω ⊆ Rn , n ∈ {1, 2, 3} (typically n = 2), which models the deflection of the membrane of a microelectromechanical system
(MEMS), where y is the mechanical displacement, a is the reciprocal of the dielectric coefficient, and b is a dimensionless
number proportional to the applied voltage [6]. The constants c , d  0, ρ , η > 0 are material parameters, with the term c yt
modeling possible damping and the term dy potentially taking into account the reset force of a spring in the system. The
boundary conditions used here correspond to a clamped setting; for a number of different possible boundary conditions we
refer to [6]. Fig. 1 shows the schematic of the type of MEMS we are considering here.
The case y (t , x) = −1 corresponds to the so-called “pull-in” instability, in which the applied voltage leads to a sufficiently
large deflection of the membrane for it to touch the ground plate, possibly damaging the device. This undesirable situation
manifests itself in the equation as a potential singularity.
In practical applications, either the dielectric properties, the applied voltage, or both are available as design variables.
Consequently, we will consider optimization problems of the form

* Corresponding author.
E-mail addresses: christian.clason@uni-graz.at (C. Clason), barbara.kaltenbacher@uni-klu.ac.at (B. Kaltenbacher).

0022-247X/$ – see front matter © 2013 Elsevier Inc. All rights reserved.
http://dx.doi.org/10.1016/j.jmaa.2013.08.058
456 C. Clason, B. Kaltenbacher / J. Math. Anal. Appl. 410 (2014) 455–468

Fig. 1. Schematic of a MEMS.

⎧
⎪
⎪ 1 α
⎪
⎪ min  y − yd 2L 2 ( Q ) + u 2U =: J (u , y )
⎪
⎨ y ∈Y, u ∈U 2 2
βu
⎪
⎪ s.t. ytt + c yt + dy + ρ 2 y − η y + = 0,
⎪
⎪ ( 1 + y )2
⎪
⎩
y |∂Ω = ∂ν y |∂Ω = 0, y (0) = y 0 , yt (0) = y 1 ,

with Y denoting the state space and the control u ∈ U being defined by one of the following three cases:

(i) control by dielectric properties: U = L 2 (Ω) and β = b ∈ L 2 (0, T ) fixed,

(ii) control by applied voltage: U = L 2 (0, T ) and β = a ∈ L 2 (Ω) fixed,
(iii) control by both: U is a subspace of L 2 (Ω) × L 2 (0, T ) (to be defined below) and β ≡ 1.

For simplicity of exposition and since it is also of high practical relevance, we restrict ourselves to a tracking type cost
function for a prescribed target displacement yd . The existence results below (Theorems 4.1 and 4.6) remain valid for any
cost functional J (u , y ) that is bounded from below, U-coercive with respect to u (this condition may be omitted in the
control constrained case), and weakly lower semi-continuous on U and Y.
A straightforward approach to prevent instabilities such as the “pull-in” instability at y = −1 is to impose control con-
straints

u U  M u
with M u sufficiently small to indirectly – via the PDE – guarantee that the state never reaches the critical value y = −1.
However, the singularity can also be prevented by imposing pointwise state constraints

− y (t , x)  M y < 1.
As already demonstrated in [7] for the simpler static MEMS model, only the latter approach is able to attain states corre-
sponding to large deflections, which is relevant in applications for achieving a sufficiently large stroke of the device. In the
transient situation with state constraints, due to the reduced dimensionality of the control, the approach from [2] used in
[7] is not applicable directly any more. We therefore apply the relaxation approach from [3] together with a localization
technique as in [5]. Specifically, we introduce a new independent variable in place of the possibly singular nonlinearity
and penalize the deviation from the original minimizers. Taking the limit with respect to the penalty parameter in the
corresponding optimality conditions yields the optimality system for the original problem.
This work is organized as follows. In Section 2, we provide some necessary results on the (linearized) state and adjoint
equations. Section 3 briefly discusses existence of and optimality conditions for each of the three types of controls above in
the control constrained case. The corresponding results for the state constrained case, which form the main contribution of
this work, are given in Section 4.
In the following, C 0 (Ω) denotes the completion of the space of all continuous functions with compact support in the
simply connected domain Ω ⊆ Rn with respect to the norm ·C (Ω) , and M(Ω) denotes the space of regular Borel measures
(which can be identified with the topological dual of C 0 (Ω)). Likewise, for a Banach space X with dual X  , M(0, T ; X  )
denotes the dual of C (0, T ; X ).

2. State equation

We start with a well-posedness result for a linear problem related to (1.1):

⎧ 2
⎨ ytt + c yt + dy + ρ  y − η y + w = 0 in Q ,
⎪
y = ∂ν y = 0 on (0, T ) × ∂Ω, (2.1)
⎪
⎩ 0 1
y=y , yt = y in {0} × Ω.
 C. Clason, B. Kaltenbacher / J. Math. Anal. Appl. 410 (2014) 455–468 457

For this purpose we require the following regularity and compatibility conditions on the initial data:
   
y 0 , y 1 ∈ H 2 (Ω) × L 2 (Ω), y 0 ∂Ω = ∂ν y 0 ∂Ω = 0. (2.2)
We also introduce for further reference the spaces
   
Y̆ := C 0, T ; H 02 (Ω) ∩ C 1 0, T ; L 2 (Ω) ,
     
Ỹ := C 0, T ; H 02 (Ω) ∩ C 1 0, T ; L 2 (Ω) ∩ H 2 0, T ; H −2 (Ω) ,
     
Y̌ := C 0, T ; L 2 (Ω) ∩ C 1 0, T ; H −2 (Ω) ∩ C 2 0, T ; H −4 (Ω) .

Lemma 2.1. For any w ∈ M(0, T ; L 2 (Ω)), and y 0 , y 1 satisfying (2.2), there exists a unique solution y ∈ Y̆ to (2.1). Furthermore, there
exists a constant C > 0 independent of w , y 0 , y 1 such that
 
 y Y̆ + c  yt L 2 (0,T ; L 2 (Ω))  C y 0 H 2 (Ω) + y 1 L 2 (Ω) +  w M(0,T ; L 2 (Ω)) .

If additionally w ∈ L 2 (0, T ; H −2 (Ω)), then y ∈ Ỹ and there exists a C > 0 such that for all w ∈ M(0, T ; L 2 (Ω)) ∩ L 2 (0, T ; H −2 (Ω)),
there holds
 
 y Ỹ + c  yt L 2 (0,T ; L 2 (Ω))  C y 0 H 2 (Ω) + y 1 L 2 (Ω) +  w M(0,T ; L 2 (Ω))∩L 2 (0,T ; H −2 (Ω)) . (2.3)

Proof. Existence of a solution can be found, e.g., in [6], and the energy estimate (2.3) can be obtained by means of the
multiplier yt and integration by parts:

t
1
t
 yt 2L 2 (Ω) + d y 2L 2 (Ω) + ρ  y 2L 2 (Ω) + η∇ y 2L 2 (Ω) 0
+c  yt 2L 2 (Ω)
2
0
t 
=− w yt dx dt   w M(0, T ; L 2 (Ω)) sup  yt  L 2 (Ω) . (2.4)
t ∈(0, T )
0 Ω

The estimate of the second-order time derivative needed to obtain y ∈ Y̆ for w ∈ L 2 (0, T ; H −2 (Ω)) can be easily verified
using
 
ytt = −c yt − dy − ρ 2 y + η y − w ∈ L 2 0, T ; H −2 (Ω) . 2

Since in the state constrained case the adjoint equation will have a measure-valued right hand side, the following result
will be useful for assessing regularity of the adjoint solution. It is obtained by applying the solution operator (−)−1 of
the Laplace equation with homogeneous Dirichlet boundary conditions to both sides of the equation. The regularity and
compatibility conditions on the initial data in this case reduce to
  
y 0 , y 1 ∈ L 2 (Ω) × H −2 (Ω), (−)−1 ∂ν y 0 ∂Ω = 0. (2.5)

Corollary 2.2. For any w ∈ M(0, T ; M(Ω)) and y 0 , y 1 satisfying (2.5), there exists a unique solution y ∈ C (0, T ; L 2 (Ω)) ∩
C 1 (0, T ; H −2 (Ω)) to (2.1). Furthermore, there exists a constant C > 0 independent of w , y 0 , y 1 such that
 
sup (−)−1 yt (t ) L 2 (Ω) + sup y (t ) L 2 (Ω)  C y 0 L 2 (Ω) + (−)−1 y 1 L 2 (Ω) +  w M(0, T ;M(Ω)) .
t ∈(0, T ) t ∈(0, T )

Proof. We apply (−)−1 (which can be continuously extended to an operator from H −2 (Ω) to L 2 (Ω) by duality) to both
sides of (2.1). For any g ∈ C ∞ (0, T ; M(Ω)) and q = n2n
+2 we can estimate

(−)−1 g (t )  C 1 (−)−1 g (t ) W 1,q (Ω)  C 2 g (t ) M ,
L 2 (Ω)

where we have used Sobolev’s embedding W 1,q (Ω)

→ L 2 (Ω) in the first inequality and [9, Theorem 7.7] with n2n +2 < n−1
n

∞
(for n ∈ {1, 2, 3}) in the second inequality. Using this and the density of C (0, T ; M(Ω)) in M(0, T ; M(Ω)), we get for any
w ∈ M(0, T ; M(Ω)) that

(−)−1 w  C 2  w M(0,T ;M(Ω)) .
M(0, T ; L 2 (Ω))

Hence we can apply Lemma 2.1 to conclude the assertion. 2

458 C. Clason, B. Kaltenbacher / J. Math. Anal. Appl. 410 (2014) 455–468

Similarly we obtain

Corollary 2.3. For any w ∈ L 2 (0, T ; H −2 (Ω)) and y 0 , y 1 satisfying (2.5), there exists a unique solution y ∈ Y̌ to (2.1). Furthermore,
there exists a constant C > 0 independent of w , y 0 , y 1 such that
 
 y Y̌  C y 0 L 2 (Ω) + (−)−1 y 1 L 2 (Ω) +  w L 2 (0,T ; H −2 (Ω)) . (2.6)

For the derivation of the optimality system in the control constrained case, we will rely on a reduced approach using
the control-to-state map. For this purpose we have to show well-posedness of the PDE for all admissible controls. Here and
below we will make use of the continuous lifting from L 2 (Ω) to C 0 (Ω) via the inverse Laplacian with Dirichlet conditions,
i.e., the existence of a constant C L > 0 such that for all v ∈ L 2 (Ω),

(−)−1 v  C L  v L 2 (Ω) . (2.7)
C (Ω)

In order to deal with a linear state space later on (see (4.1) below) we set – without loss of generality – y 0 = y 1 = 0 in
the following.

Theorem 2.4. There exists a constant M > 0 such that for all f ∈ M(0, T ; L 2 (Ω)) with

 f M(0,T ; L 2 (Ω))  M ,

there exists a unique solution y ∈ Y̆ to

⎧ f
⎪ 2
⎨ ytt + c yt + dy + ρ  y − η y + (1 + y )2 = 0 in Q = (0, T ) × Ω,
⎪
(2.8)
⎪
⎪ y = ∂ν y = 0 on (0, T ) × ∂Ω,
⎩
y = 0, yt = 0 in {0} × Ω.
Moreover, there exists an m > 0 independent of f such that

 y C (0,T ;C 0 (Ω))  m < 1. (2.9)

Proof. Proceeding similarly to the static case in [7] (see also the well-posedness proof in [6]), we use Banach’s fixed point
theorem. Let

W = v ∈ Y̆: sup  v C (Ω)  m, v (0) = y 0 , v t (0) = y 1
t ∈(0, T )

for some m ∈ (0, 1) to be chosen below, and consider the fixed point operator T : W → W, v → T v := y, where y solves
⎧
⎪
⎪ 2 f
⎨ ytt + c yt + dy + ρ  y − η y = − (1 + v )2
⎪ in Q ,

⎪ y = ∂ν y = 0 on (0, T ) × ∂Ω,
⎪
⎪
⎩
y = y0, yt = y 1 in {0} × Ω.
Well-definedness of T follows from Lemma 2.1.
To show that T is a self-mapping, consider an arbitrary v ∈ W. Then we have by (2.4) and Young’s inequality (canceling
the  yt C (0, T ; L 2 (Ω)) terms) that

CL
sup ( T v )(t ) C (Ω)  √  f M(0,T ; L 2 (Ω)) ,
t ∈(0, T ) ρ (1 − m)2
where C L is the constant in (2.7), and the right hand side is not larger than m if
√
ρ m(1 − m)2
 f M(0,T ; L 2 (Ω))  . (2.10)
CL
Contractivity of T follows from the fact that for any v , w ∈ W, by the identity

1 1 (1 + w )2 − (1 + v )2 (1 + w + 1 + v )( w − v )
− = =
(1 + v )2 (1 + w )2 (1 + v )2 (1 + w )2 (1 + v )2 (1 + w )2
 
1 1
= + ( w − v ),
(1 + v )2 (1 + w ) (1 + v )(1 + w )2
 C. Clason, B. Kaltenbacher / J. Math. Anal. Appl. 410 (2014) 455–468 459

there holds
 
CL 1 1
sup ( T v − T w )(t ) C (Ω)  √  f M(0, T ; L 2 (Ω)) sup + ( v − w )(t )

t ∈(0, T ) ρ t ∈(0, T ) (1 + v ) (1 + w ) (1 + w ) (1 + v )
2 2
C (Ω)
2C L
√  f M(0,T ; L 2 (Ω)) sup ( v − w )(t ) C (Ω) .
ρ (1 − m)3 t ∈(0, T )

Hence contractivity holds if

√
ρ (1 − m)3
 f M(0,T ; L 2 (Ω)) < . (2.11)
2C L
The maximum over m ∈ (0, 1) of the minimum of the two right hand sides in (2.10) and (2.11) can be found by equili-
√ √
ρ m(1−m)2 ρ (1−m)3
brating CL
= 2C L
, which yields the optimal bound

1
m= .
3
Therefore, a solution to (1.1) exists if
√
4 ρ
 f M(0,T ; L 2 (Ω))  M := . 2
27C L

Hence, for
M( L 2 )  
D S ⊆ BM (0) = f ∈ M(0, T ; L 2 (Ω)):  f M(0,T ; L 2 (Ω))  M
with M > 0 sufficiently small, the operator

S : D S → Y̆,
f → y solving (2.8),
from which we will derive the control-to-state mapping S later on, is well defined. For proving smoothness of this mapping,
we will consider the linearization of (2.8).
⎧
⎪
⎪ 2f
⎪ 2
⎨ ytt + c yt + dy + ρ  y − η y − (1 + y )3 y = w in Q ,
(2.12)
⎪
⎪ y = ∂ν y = 0 on (0, T ) × ∂Ω,
⎪
⎩
y = 0, yt = 0 in {0} × Ω,

for given f ∈ D S , y ∈ Y̆, w ∈ M(0, T ; L 2 (Ω)).

M( L 2 )
Lemma 2.5. There exists a constant M > 0 such that for all f ∈ D S ⊆ B M (0), y = S( f ), and all w ∈ M(0, T ; L 2 (Ω)), the
linearized state equation (2.12) has a unique solution y ∈ Y̆, which depends continuously on w.

2f
Proof. By Theorem 2.4 the coefficient e = − (1+ y )3 is contained in M(0, T ; L 2 (Ω)). Since e might take negative values –
leading to negative contributions on the left hand side of an energy estimate obtained as in (2.4), – it is clear that we
have to impose smallness of e. This can be obtained again using Theorem 2.4 from smallness of f . We can thus apply a
fixed point argument as in the proof of Theorem 2.4, this time with the (linear) fixed point operator T lin : Wlin → Wlin ,
v → T lin v := y, where y solves
⎧
⎪
⎪ 2f
⎪ 2
⎨ ytt + c yt + dy + ρ  y − η y = (1 + y )3 v + w in Q ,
⎪
⎪ y = ∂ν y = 0 on (0, T ) × ∂Ω,
⎪
⎩
y = 0, yt = 0 in {0} × Ω.
By Lemma 2.1, the operator T is well defined, contractive, and a self-mapping on

Wlin = v ∈ Y̆: sup  v C (Ω)  R , v (0) = 0, v t (0) = 0
t ∈(0, T )

with R sufficiently large so that

460 C. Clason, B. Kaltenbacher / J. Math. Anal. Appl. 410 (2014) 455–468

CL 2f
R√ +  w M(0,T ; L 2 (Ω))  R .
ρ (1 + y ) M(0,T ;L 2 (Ω))
3

From this we deduce the assertion. 2

To show weak continuity of S as needed for the existence of a minimizer later on, we will use compactness of the
embedding of Y̆ in C (0, T ; C 0 (Ω)). This is established in the following lemma, which is based on the Dubinskii theorem
quoted here for the sake of convenience.

Proposition 2.6. (See [10, Theorem 4.1].) Assume that E 0 , E , E 1 are reflexive Banach spaces, with continuous embeddings E 0
→
E
→ E 1 , and compact embedding E 0
→ E, and let 1 < q, q1 < ∞. Moreover, assume that the set M is bounded in L q (0, T ; E 0 ) and
consists of functions u (t ) equicontinuous in C (0, T ; E 1 ). Then M is relatively compact in L q1 (0, T ; E ) and in C (0, T ; E 1 ).

Lemma 2.7. For all σ ∈ (0, 1 − n4 ) and ε ∈ (0, 2(1 − n

4
− σ )) we have
   
Y̆
→ W σ ,∞ 0, T ; W ε,∞ (Ω)
→ C 0, T ; C 0 (Ω) ,

where both embeddings are continuous and the latter is compact.

Proof. Continuity of the first embedding follows by continuity of the embeddings L 2 (Ω)
→ W −s,r (Ω) and H 2 (Ω)
→
W t ,r (Ω) for s = n2 , t = 2 − n2 and any r ∈ [1, ∞), and by interpolation

      
W 1,r 0, T ; W −s,r (Ω) , L r 0, T ; W t ,r (Ω) θ,r = W (1−θ ),r 0, T ; W −(1−θ )s+θ t ,r (Ω) .

(Here the interpolation spaces are defined either by the J- or by the K-method of real interpolation; see, e.g., [1, Chapter 7].)
With r  n8 and θ = 1 − σ − 2r , the latter space is continuously embedded in W σ ,∞ (0, T ; W ε,∞ (Ω)) for σ ∈ (0, 1 − n4 ),
ε ∈ (0, 2(1 − n4 − σ ) − n+r 4 ), where we may let r tend to infinity. Compactness of the second embedding in (4.1) can be con-
ε
cluded from Proposition 2.6 applied to E 0 = W ε,2p (Ω) and E 1 = W 2 , p (Ω) with p > max{1, 2n
ε } so that E 1
→ C 0 (Ω). 2

Finally, we address differentiability of the control-to-state mapping, which will be needed for deriving first order opti-
mality conditions (see Theorem 3.1 below).

Theorem 2.8. There exists a constant M > 0 such that S : D S → Y̆ is globally Lipschitz continuous, weakly continuous, and Fréchet
differentiable with respect to the L 2 ( Q ) topology in preimage space. The derivative S ( f )( f̃ − f ) is given by the solution y ∈ Y̆ to
⎧
⎪
⎪ 2f 1
⎪ 2
⎨ ytt + c yt + dy + ρ  y − η y − y=− ( f̃ − f ) in Q ,
(1 + y )3 (1 + y )2
(2.13)
⎪
⎪ y = ∂ν y = 0 on (0, T ) × ∂Ω,
⎪
⎩
y = 0, yt = 0 in {0} × Ω,

where y = S( f ).

Proof. Lipschitz continuity follows by applying a similar argument to the one in Lemma 2.5 to the equation
⎧
⎪
⎪ f̃ (2 + ỹ + y ) 1
⎪ 2
⎨ ztt + czt + dz + ρ  z − η z − z=− ( f̃ − f ) in Q ,
(1 + y )2 (1 + ỹ )2 (1 + y )2
⎪
⎪ z = ∂ν z = 0 on (0, T ) × ∂Ω,
⎪
⎩
z = 0, zt = 0 in {0} × Ω,

which is satisfied by the difference z = ỹ − y between two solutions ỹ = S( f̃ ) and y = S( f ).

Weak continuity of S can be obtained by a subsequence–subsequence argument using the fact that Y̆ is compactly
embedded in C (0, T ; C 0 (Ω)) (see Lemma 2.7), which allows taking limits in the nonlinear part of the PDE defining S (see
the proof of Theorem 2.2 in [7]).
Fréchet differentiability follows from the fact that y ∈ Y̆ is well defined by (2.13) due to Lemma 2.1, and z := S( f̃ ) −
S( f ) − y = ỹ − y − y solves
 C. Clason, B. Kaltenbacher / J. Math. Anal. Appl. 410 (2014) 455–468 461

⎧
⎪
⎪ 2f
⎪
⎪ ztt + czt + dz + ρ 2 z − η z − z
⎪
⎪ (1 + y )3
⎪
⎪
⎨ 2 f̃ (3 + y − ỹ )
= ( f̃ − f )( ỹ − y ) − ( ỹ − y )2 in Q ,
⎪
⎪ (1 + y ) 3 (1 + y )3 (1 + ỹ )2
⎪
⎪
⎪
⎪ z = ∂ν z = 0 on (0, T ) × ∂Ω,
⎪
⎪
⎩
z = 0, zt = 0 in {0} × Ω.

The M(0, T ; L 2 (Ω)) norm of the right hand side can be estimated by a multiple of  f̃ − f 2M(0, T ; L 2 (Ω)) due to the Lipschitz
continuity shown above, so that Lemma 2.5 yields the desired quadratic estimate of the first order Taylor remainder. 2

3. Control constrained optimal control

For comparison, we first consider the control constrained case. We distinguish between separate control by either the
dielectric properties (case (i) above) or the applied voltage (case (ii) above) and combined control by both (case (iii) above).

3.1. Control by dielectric properties or voltage

Since cases (i) and (ii) only differ in the domain on which the control is defined, we can treat both using the same
approach. We define the unconstrained control space as U := L 2 ( X ) with X = Ω in case (i) and X = (0, T ) in case (ii) and
consider the control constrained optimal control problem
⎧ 1 α
⎪
⎪ min J (u , y ) =  y − yd 2L 2 ( Q ) + u 2L 2 ( X )
⎪
⎪
⎪
⎪ y ∈Y̆,u ∈U 2 2
⎪
⎨ βu
s.t. ytt + c yt + dy + ρ 2 y − η y + = 0 in Q , (Pcc )
⎪
⎪ (1 + y )2
⎪
⎪
⎪
⎪ y = ∂ν y = 0 on (0, T ) × ∂Ω, y = yt = 0 in {0} × Ω,
⎪
⎩
u L 2 (Ω)  M u .
The bound M u is chosen small enough so that  y C (0, T ;C 0 (Ω)) < 1, which is possible due to (2.9). The corresponding con-
strained control space is defined as
 
U M = u ∈ U:  u  L 2 ( X )  M u .
Since the state equation is well-posed for every u ∈ U M (see Theorem 2.4), we can use the control-to-state mapping S :
U M → Y̆, to obtain the reduced problem
   )
min J u , S (u ) . (Pcc
u ∈U M

Note that by Theorem 2.8, S is weakly continuous and Fréchet differentiable.

Since U M is nonempty and weakly sequentially compact, S is weakly continuous for all u ∈ U M by Theorem 2.8, and
J is weakly Y̆ × U lower semi-continuous and bounded from below, we obtain the existence of a minimizer u ∗ ∈ U M by
standard arguments; cf., e.g., [9].
Similarly, we obtain first order necessary optimality conditions. For a local minimizer u ∗ of (Pcc
 ) and y ∗ := S (u ∗ ) ∈ Y̆,
∗
we can introduce the adjoint state p ∈ Y̆ solving
⎧
⎪
⎪ 2β u ∗
⎪ ptt∗ − cpt∗ + dp ∗ + ρ 2 p ∗ − η p ∗ −
⎨ p ∗ = −( y − yd ) in Q ,
(1 + y ∗ )3
∗ ∗ (3.1)
⎪ p = ∂ν p = 0
⎪ on (0, T ) × ∂Ω,
⎪
⎩ ∗ ∗
p = 0, pt = 0 in { T } × Ω,

due to Lemma 2.5 with t replaced by T − t.

Furthermore, a Slater condition is trivially satisfied for the inequality constraint (take u = 0). From, e.g., [4, Proposi-
tion 3.2] using differentiability of S (cf. Theorem 2.8), we thus deduce the existence of a corresponding Lagrange multiplier
λ∗ ∈ R and hence the following optimality system.

Theorem 3.1. Let u ∗ ∈ U M be a local minimizer of (Pcc

 ), y ∗ := S (u ∗ ) ∈ Y̆, and p ∗ ∈ Y̆ satisfy (3.1). Then there exists λ∗ ∈ R, λ∗  0

satisfying
462 C. Clason, B. Kaltenbacher / J. Math. Anal. Appl. 410 (2014) 455–468

⎧ 
⎪ ∗ ∗ β
⎪
⎨ (α + 2 λ ) u = p ∗ ds,
(1 + y ∗ )2
Z
⎪
⎩ λ∗ 
⎪
u ∗ 22

− M u2 = 0,
L (Ω)

with Z = (0, T ) if X = Ω and Z = Ω if X = (0, T ).

3.2. Control by dielectric properties and voltage

In the case (iii), we take as control variable

 
u (t , x) = v (x) 1 + ϕ (t ) for all (t , x) ∈ Q .
where ϕ is normalized to have zero mean, see (3.3) below. (Note that a straightforward representation u (t , x) = v (x)ϕ (t ),
with v and ϕ elements of linear spaces, would be non-unique: For any such pair ( v , ϕ ) representing u and any constant
c = 0, the pair (cv , c −1 ϕ ) would also represent u.) Correspondingly, we define the unconstrained control space as
 
U = v · (1 + ϕ ): v ∈ L 2 (Ω), ϕ ∈ L 2♦ (0, T ) , (3.2)
where
 T 
2 2
L ♦ (0, T ) = ψ ∈ L (0, T ): ψ dt = 0 (3.3)
0

and 1 ∈ L ∞ ( Q ) is the constant function with value one. Furthermore, we set

u 2U :=  v 2L 2 (Ω) + ϕ 2L 2 (0,T ) . (3.4)

The following lemma allows us to identify U with a closed affine subspace of L 2 (Ω) × L 2 (0, T ), which implies that u U
is well defined.

Lemma 3.2. For every u ∈ U, the representation u = v · (1 + ϕ ) with v ∈ L 2 (Ω) and ϕ ∈ L 2♦ (0, T ) is unique. Moreover, we have
 
v · (1 + ϕ ) 22 =  v 2L 2 (Ω) T + ϕ 2L 2 (0,T ) . (3.5)
L (Q )

Proof. For proving uniqueness of the representation, assume that for v 1 , v 2 ∈ L 2 (Ω), ϕ1 , ϕ2 ∈ L 2 (0, T ) we have
   
v 1 (x) · 1 + ϕ1 (t ) = v 2 (x) · 1 + ϕ2 (t )
for almost all x ∈ Ω and t ∈ (0, T ). This implies that there exists a λ ∈ R, λ = 0 such that
1
v1 = v2 and 1 + ϕ1 = λ(1 + ϕ2 ),
λ
the latter implying

(λ − 1)1 + λϕ2 − ϕ1 = 0.
T
Taking the L 2 (0, T ) inner product with 1 and using the normalization condition 0
ϕi dt = 0 for i = 1, 2, we easily see that
this implies λ − 1 = 0 and hence v 1 = v 2 , ϕ1 = ϕ2 . 2

We now consider the optimization problem

⎧ 
⎪ 1 α
⎪
⎪ min J (u , y ) =  y − yd 2L 2 (Ω) + u 2U
⎪
⎪ y ∈Y̆,u ∈U 2 2
⎪
⎪
⎪
⎨
0
u (t , x)
ytt + c yt + dy + ρ 2 y − η y + = 0 in Q , (Pcc
2
)
⎪ s.t.
⎪ (1 + y )2
⎪
⎪
⎪
⎪ y = ∂ν y = 0 on (0, T ) × ∂Ω, y = yt = 0 in {0} × Ω,
⎪
⎪
⎩
u U  M u .
The corresponding constrained control space is defined by
 
U M = u ∈ U: u U  M u .
 C. Clason, B. Kaltenbacher / J. Math. Anal. Appl. 410 (2014) 455–468 463

In the following we will identify U with L 2 (Ω) × (1 + L 2♦ (0, T )), which is a closed (linear) subspace of L 2 (Ω) × L 2 (0, T ),
and use the topology of L 2 (Ω) × L 2 (0, T ) on U, i.e.,

u U = ( v , ϕ ) L 2 (Ω)×L 2 (0,T ) ,

cf. (3.4). Any ball

 
BrU (u 0 ) = u = v · (1 + ϕ ) ∈ U: u − u 0 U  r

with u 0 = v 0 · (1 + ϕ0 ) ∈ U is then convex, (weakly) closed and weakly compact in U in the sense that

 
( v , ϕ ) ∈ L 2 (Ω) × L 2♦ (0, T ): ( v − v 0 , ϕ − ϕ0 ) L 2 (Ω)×L 2 (0,T )  r

has these properties. Moreover, we can make use of Theorems 2.4 and 2.8, since by

u 1 − u 2 L 2 ( Q ) = v 1 (1 + ϕ1 ) − v 2 (1 + ϕ2 ) L 2 ( Q )
  v 1 L 2 (Ω) ϕ1 − ϕ2 L 2 (0,T ) + ϕ2 L 2 (0,T )  v 1 − v 2 L 2 (Ω) ,

smallness of u 1 − u 2 U implies smallness of u 1 − u 2  L 2 ( Q ) .

Using these facts we can proceed as in Section 3.1 to obtain existence of a minimizer and first order optimality condi-
tions.

Theorem 3.3. The reduced problem

 
min J u , S (u )
u ∈U M

has a global minimizer u ∗ = v ∗ · (1 + ϕ ∗ ) ∈ U M . Furthermore, let y ∗ := S (u ∗ ) ∈ Y̆ for any local minimizer u ∗ . Then there exist p ∗ ∈ Y̆
satisfying (3.1) and λ∗ ∈ R with λ∗  0 such that
⎧
⎪
⎪ T
⎪
⎪
 
∗ ∗ 1 + ϕ∗ ∗
⎪
⎪ α + 2λ v = p dt ,
⎪
⎪ (1 + y ∗ )2
⎪
⎨ 0

  v∗
⎪
⎪ α + 2 λ∗ ϕ ∗ = p ∗ dx,
⎪
⎪ (1 + y ∗ )2
⎪
⎪
⎪
⎪ Ω
⎩ λ∗ 
⎪ 2 
v ∗ 22 + ϕ ∗ L 2 (0,T ) − M u2 = 0.
L (Ω)

4. State constrained optimal control

Here we use the pointwise state constraints

− y (t , x)  M y < 1 for all (t , x) ∈ Q

to prevent the singularity of (1.1) at y (t , x) = −1. The unconstrained state space is defined as

 
Y = y ∈ Ỹ: y (0) = 0, yt (0) = 0 (4.1)

where

   
Ỹ
→ W σ ,∞ 0, T ; W 0ε,∞ (Ω)
→ C 0, T ; C 0 (Ω) ,

for σ ∈ (0, 1 − n4 ) and ε ∈ (0, 2(1 − n4 − σ )), see Lemma 2.7 and note that Ỹ ⊆ Y̆.
For M y < 1, we define the constrained state space as

 
Y M := y ∈ Y: − y (t , x)  M y for all (t , x) ∈ Q .
464 C. Clason, B. Kaltenbacher / J. Math. Anal. Appl. 410 (2014) 455–468

4.1. Control by dielectric properties or voltage

We again start with cases (i) and (ii), i.e., U = L 2 ( X ) with X = Ω in case (i) and X = (0, T ) in case (ii), and consider the
state constrained optimal control problem
⎧ 
⎪ 1 α
⎪
⎪ min J ( u , y ) =  y − yd 2L 2 (Ω) + u 2L 2 (Ω)
⎪
⎪ y ∈Y,u ∈U 2 2
⎪
⎪
⎪
⎨ 0
βu
s.t. ytt + c yt + dy + ρ 2 y − η y + = 0 in Q , (Psc )
⎪
⎪ (1 + y )2
⎪
⎪
⎪
⎪ y = ∂ν y = 0 on (0, T ) × ∂Ω, y = yt = 0 in {0} × Ω,
⎪
⎪
⎩
− y  M y in Q .
We also define
 
G : U × Y M → L 2 0, T ; H −2 (Ω)

by the weak form of the PDE with boundary and initial conditions on y in (Psc ).

Theorem 4.1. There exists a minimizer (u ∗ , y ∗ ) ∈ U × Y M of (Psc ).

Proof. The set of feasible pairs (u , y ) satisfying the equality and inequality constraints is nonempty (take (u , y ) = (0, 0)).
By the boundedness of J from below and the coercivity of the functional in u, we obtain the existence of a minimizing
sequence whose u component is bounded in U. The equality constraint G (u , y ) = 0 together with Lemma 2.1 implies that
the y component of the minimizing sequence is uniformly bounded in Y. Hence, there exists a subsequence, denoted by
{(un , yn )}n∈N , that converges weakly in the reflexive space U × Ys ⊃ U × Y to (u ∗ , y ∗ ) ∈ U × Ys , where
     
Ys := L s 0, T ; H 02 (Ω) ∩ W 1,s 0, T ; L 2 (Ω) ∩ W 2,s 0, T ; H −2 (Ω)

with arbitrarily large s ∈ (1, ∞). Due to the compact embedding of Y in C (0, T ; C 0 (Ω)) and the continuity of Ψ : r → (1+1r )2 ,
we have that a subsequence of (1+ y )2 converges strongly to (1+ y ∗ )2 , with β fixed, and that y ∗ satisfies the inequality
β β
n
constraint. Thus, we can pass to the limit in (the weak formulation of) G (un , yn ) = 0 to obtain G (u ∗ , y ∗ ) = 0, and hence by
Lemma 2.1, y ∗ ∈ Y. 2

Let again (u ∗ , y ∗ ) denote a local minimizer of (Psc ), and let r denote the radius of the neighborhood in U × Y in which
(u ∗ , y ∗ ) is optimal.
Since the state equation is not well-posed for arbitrary u ∈ U (i.e., it admits a solution only for sufficiently small u; see
Theorem 2.4), we cannot use a control-to-state mapping for deriving optimality conditions. Furthermore, considering the
optimization directly, i.e., without control-to-state mapping, poses difficulties since a regular point condition according to,
e.g., [2] would require existence of (u 0 , y 0 ) ∈ U × Y M such that ( y ∗ + y 0 ) ∈ int Y M and
    
G y u ∗ , y ∗ y 0 + G u u ∗ , y ∗ u 0 − u ∗ = 0,

where G y denotes the Fréchet derivative of G with respect to y. However, in the cases considered here, u 0 is too low-
dimensional to guarantee existence of such a y 0 in case y ∗ ∈
/ int Y M . Therefore we follow the relaxation approach from [3]
combined with a localization technique as in [5]: We introduce the independent variable w in place of the nonlinearity
βu
(1+ y )2
and consider the family of control problems

⎧
⎪ min J ε (u , w , y )
⎪
⎪ u ∈U, y ∈Y, w ∈ L 2 ( Q )
⎪
⎨
s.t. ytt + c yt + dy + ρ 2 y − η y + w = 0 in Q , (Psc , ε )
⎪
⎪ y = ∂ν y = 0 on (0, T ) × ∂Ω, y = yt = 0 in {0} × Ω,
⎪
⎪
⎩
− y  M y in Q ,
where
2 2
1 βu 1 2 1 ∗
J ε (u , w , y ) = J (u , y ) + w− + u − u ∗ L 2 (Ω) + w − βu ,
2ε (1 + y )2 L 2 ( Q ) 2δ 2 (1 + y ∗ )2 L 2 ( Q )
 C. Clason, B. Kaltenbacher / J. Math. Anal. Appl. 410 (2014) 455–468 465

with C as in (2.3) and

r2
δ< . (4.2)
2 max{1, C }2 J (u ∗ , y ∗ )

Here, (u ∗ , y ∗ ) is a fixed local minimizer of (Psc ). The penalty term involving 1

2ε
will be used to enforce convergence of
β u∗
minimizers (u ∗ε , w ∗ε , y ∗ε ) of (Psc , ε ) to (u ∗ , (1+ y ∗ )2 , y ∗ ) as ε → 0, while the penalty term with the (fixed) factor 21δ will
ensure that (u ∗ε , y ∗ε ) always remains sufficiently close to the local minimizer (u ∗ , y ∗ ) in order to exploit its optimality
property (i.e., the latter penalty “localizes” the relaxation).
We now define
 
G lin : L 2 ( Q ) × Y M → L 2 0, T ; H −2 (Ω)

by the weak form of the PDE with boundary and initial conditions on y in (Psc , ε ). Similarly to Theorem 4.1 one can then
show existence of a minimizer (u ∗ε , w ∗ε , y ∗ε ).

Theorem 4.2. There exists a minimizer (u ∗ε , w ∗ε , y ∗ε ) ∈ U × L 2 ( Q ) × Y M of (Psc , ε ).

Optimality conditions for the minimizers follow from a regular point condition. In the following, let
   
P := C 0, T ; L 2 (Ω) ∩ C 1 0, T ; H −2 (Ω)
and set again Z = (0, T ) in case X = Ω and Z = Ω in case X = (0, T ).

Lemma 4.3. Let (u ∗ε , w ∗ε , y ∗ε ) ∈ U × L 2 ( Q ) × Y M be a local minimizer of (Psc , ε ). Then there exist μ∗ε ∈ M(0, T ; M(Ω)) and p ∗ε ∈ P
satisfying
⎧  
⎪ ∗ ∗ ∗ 2 ∗ ∗
 ∗  ∗ 1 2β u ∗ε ∗ β u ∗ε
⎪
⎪ p ε tt − c p ε t + dp ε + ρ  p ε − η p ε = − y ε − yd − με + wε − ,
⎪
⎪ ε (1 + y ∗ε )3 (1 + y ∗ε )2
⎪
⎪
⎪
⎪   
⎪
⎪ ∗ 1 ∗ ∗
 1 β ∗ β u ∗ε
⎪
⎨ αuε + uε − u = w ε − ds,
δ ε (1 + y ∗ε )2 (1 + y ∗ε )2
Z (4.3)
⎪
⎪ 1
⎪
⎪ β u ∗ε
 
β u∗

⎪
⎪ w ∗ε − = − ∗
− + p ∗ε ,
⎪
⎪ ∗ )2
w ε ∗ )2
⎪
⎪ ε ( 1 + y ε (1 + y
⎪
⎪
⎩ ∗ 
με , y − y ∗ε M,C  0 for all y ∈ YM .

Proof. We wish to apply [2, Theorem 2.1]. Due to linearity of G lin , the regular point condition
    
G lin, y w ∗ε , y ∗ε y 0 + G lin, w w ∗ε , y ∗ε w 0 − w ∗ε = 0

with ( w 0 , y 0 ) ∈ L 2 ( Q ) × Y M such that ( y ∗ε + y 0 ) ∈ int Y M can be satisfied by setting

y 0 = −C̃ y ∗ε ∈ Y M , w 0 = (1 − C̃ ) w ∗ε ∈ L 2 ( Q ),

with
My
C̃ = .
2 sup(t ,x)∈ Q y ∗ε (t , x)

It remains to show surjectivity of the linearized equality constraint. An inspection of Case 3 in the proof of Theorem 2.2
and of Lemma 2.3 a) in [2] shows that by linearity of G lin, y , it suffices to assume surjectivity of G lin, y : Ŷ → V̂, where Ŷ is
dense in a superset of Y. Setting
   
Ŷ := y ∈ Y: G lin, y w ∗ε , y ∗ε ∈ L 2 ( Q ) ,

we have by definition that G lin, y ( w ∗ε , y ∗ε ) : Ŷ → V̂ := L 2 ( Q ) is surjective. Furthermore, by density of L 2 ( Q ) in L 2 (0, T ;

H −2 (Ω)) and the stability estimate (2.6) of Corollary 2.3, Ŷ is dense in Y̌ ⊇ Y. We thus obtain existence of an adjoint state
in V̂∗ = L 2 ( Q ). Together with Corollary 2.2, this yields the desired first order optimality conditions. 2

Similarly to [3, Lemma 3], one can show convergence of the regularized solutions.
466 C. Clason, B. Kaltenbacher / J. Math. Anal. Appl. 410 (2014) 455–468

Lemma 4.4. Let (u ∗ , y ∗ ) be a local minimizer of (Psc ), and let {(u ∗ε , w ∗ε , y ∗ε )}ε>0 be a family of global minimizers of (Psc , ε ). Then as
ε → 0,
β u∗
u ∗ε → u ∗ in U, w ∗ε → in L 2 ( Q ), y ∗ε → y ∗ in Y.
(1 + y ∗ )2

Proof. By minimality of (u ∗ε , w ∗ε , y ∗ε ) we have

 
  β u∗  
J ε u ∗ε , w ∗ε , y ∗ε  J ε u ∗ , , y∗ = J u∗ , y∗ . (4.4)
(1 + y ∗ )2
Hence {(u ∗ε , w ∗ε )}ε>0 is bounded in U × L 2 ( Q ); and thus by G lin ( w ∗ε , y ∗ε ) = 0 and Lemma 2.1, {(u ∗ε , w ∗ε , y ∗ε )}ε>0 is bounded
in U × L 2 ( Q ) × Y. Moreover, due to the penalty term 21δ u − u ∗ 2U and our choice of δ according to (4.2) with (2.3), we have
(u ∗ε , y ∗ε ) ∈ BrU (u ∗ ) × BrY ( y ∗ ) for all ε > 0. By compactness of the embedding Y
→ C (0, T ; C 0 (Ω)), we can therefore extract
from any subsequence of {(u ∗ε , w ∗ε , y ∗ε )}ε>0 a further subsequence (denoted by the same symbols) such that as ε → 0,
 
u ∗ε  û in U, w ∗ε  ŵ in L 2 ( Q ), y ∗ε  ŷ in Y, y ∗ε → ŷ in C 0, T ; C 0 (Ω) (4.5)

for some (û , ŵ , ŷ ) ∈ BrU (u ∗ ) × L 2 ( Q ) × (BrY ( y ∗ ) ∩ Y M ). Here we have used weak closedness of BrU (u ∗ ) and BrY ( y ∗ ) ∩ Y M .
Taking now the weak limit in the PDE, we see that G lin ( ŵ , ŷ ) = 0. The uniform boundedness (4.4) and the penalty term
with factor ε1 in J ε imply that
2
∗ ∗
w − β uε → 0,
ε (1 + y ∗ )2 2
ε L (Q )

hence we deduce from (4.5) and the weak lower semicontinuity of the norm that

β û
ŵ − = 0. (4.6)
(1 + ŷ )2
Together with G lin ( ŵ , ŷ ) = 0, this implies that G (û , ŷ ) = 0. Thus, (û , ŷ ) is admissible for (Psc ) and contained in BrU (u ∗ ) ×
(BrY ( y ∗ ) ∩ Y M ). We can now use minimality of (u ∗ , y ∗ ) in BrU (u ∗ ) × BrY ( y ∗ ) with respect to (Psc ), relation (4.4), and weak
lower semicontinuity of J and the norms to conclude
   
J (û , ŷ )  J u ∗ , y ∗  lim sup J ε u ∗ε , w ∗ε , y ∗ε
ε →0
 2 
  1 2 1 ∗ β u∗
 lim sup J u ∗ε , y ∗ε + u ∗ε − u ∗ U + w −
ε →0 2δ 2 ε (1 + y ∗ )2 L 2 ( Q )
 2 
 ∗ ∗ 1 ∗ 1 ∗ β u∗
 lim inf J u ε , y ε + lim inf ∗ 2
uε − u U + w ε −
ε →0 ε →0 2δ 2 (1 + y ∗ )2 L 2 ( Q )

1 2 1 β u∗
2
 J (û , ŷ ) + û − u ∗ U + ŵ −
. (4.7)
2δ 2 (1 + y ∗ )2 L 2 ( Q )
β u∗
This implies that û = u ∗ and – by (4.6) – that (1+ ŷ )2 = (1+ y ∗ )2 . The uniqueness claim in Lemma 2.1 then yields ŷ = y ∗ .
β û

Strong convergence of the u and w components along a subsequence immediately follows from (4.7), and hence the
stability estimate of Lemma 2.1 yields strong convergence of the y part. Using a subsequence–subsequence argument, we
arrive at the assertion. 2

As in [3, Theorem 2], we can thus take the limit as ε → 0 in (4.3) to derive necessary optimality conditions for (Psc ).

Theorem 4.5. Let (u ∗ , y ∗ ) ∈ U × Y M be a local minimizer of (Psc ). Then there exist μ∗ ∈ M(0, T ; M(Ω)) and p ∗ ∈ P satisfying
⎧
⎪ ∗ 2β u ∗  
⎪
⎪ ptt − cpt∗ + dp ∗ + ρ 2 p ∗ − η p ∗ − p ∗ = −γ y ∗ − yd − μ∗ ,
⎪
⎪ (1 + y ∗ )3
⎪
⎨ 
∗ β
γ αu = p ∗ ds, (4.8)
⎪
⎪ ( 1 + y ∗ )2
⎪
⎪
⎪ Z
⎩ μ∗ , y − y ∗ 
⎪
 0 for all y ∈ Y M ,
M,C

where γ ∈ {0, 1} and γ +  p ∗ P > 0.

 C. Clason, B. Kaltenbacher / J. Math. Anal. Appl. 410 (2014) 455–468 467

β u∗
Proof. Eliminating the term 1ε ( w ∗ε − (1+ y ∗ )2 ) by means of the last equation in (4.3), we obtain
⎧    
⎪
⎪ ∗ ∗ ∗ 2 ∗ ∗
 ∗  ∗ 2β u ∗ε ∗ β u∗ ∗
⎪
⎪ p − c p + dp + ρ  p − η  p = − y − y d − μ + − w − + p ε ,
⎪
⎪
ε tt εt ε ε ε ε ε
(1 + y ∗ε )3 ε
(1 + y ∗ )2
⎪
⎨     
∗ 1 ∗ ∗
 β ∗ β u∗ ∗ (4.9)
⎪ α u ε + u − u = − w − + p ε ds,
⎪
⎪ δ ε (1 + y ∗ε )2 ε
(1 + y ∗ )2
⎪
⎪
⎪ Z
⎩ μ∗ , y − y ∗ 
⎪
 0 for all y ∈ Y .
ε ε M,C M
∗ ∗
If { p ε , με }ε>0 has a bounded subsequence in P × M(0, T ; M(Ω)), we can take (weak) limits in (4.9) to obtain (4.8) with
γ = 1.
Otherwise,
 
{rε }ε>0 := p ∗ε P + μ∗ε M(0,T ;M(Ω)) ε>0 → ∞
p∗ μ∗ε
as ε → 0, and we can divide { p ∗ε , μ∗ε }ε>0 by rε to get a bounded sequence { p ∗ε , μ∗ε }ε>0 := { r ε , rε
}ε>0 . Inserting this into
ε
(4.9) yields
⎧    
⎪
⎪ ∗ ∗ ∗ 2 ∗ ∗ 1 ∗  ∗ 2β u ∗ε 1 ∗ β u∗ ∗
⎪ p
⎪ tt − c p + dp + ρ  p − η  p = − y − y d − μ + − w − + p ,
⎪
⎪
ε ε t ε ε ε
rε ε ε
(1 + y ∗ε )3 rε ε
(1 + y ∗ )2 ε
⎪ 
⎨      
1 ∗ 1 ∗ ∗
 β 1 ∗ β u∗
⎪ α u ε + u ε − u = − w ε − + p ∗ε ds,
⎪
⎪ rε δ (1 + y ∗ε )2 rε (1 + y ∗ )2
⎪
⎪
⎪ Z
⎩ μ∗ , y − y ∗ 
⎪
 0 for all y ∈ Y .
ε ε M,C M

Taking (weak) limits, we then obtain (4.8) with γ = 0. To show that in this case p ∗ = 0, we assume that p ∗ = 0 and note
that the first line of (4.8) then yields με  μ∗ = 0. By [3, Lemma 1], we therefore have με M(0, T ;M(Ω)) → 0, which
implies
 p ∗ε P  p ∗ε P
 p ε P = = ∗ → 1,
 p ∗ε P ∗
+ με M(0,T ;M(Ω))  p ε P + με∗ M(0,T ;M(Ω))
in contradiction with p ε → p ∗ = 0. 2

4.2. Control by dielectric properties and voltage

Finally, we consider case (iii), i.e., we use the control space U defined by (3.2) and u U defined by (3.4). The corre-
sponding state constrained control problem is
⎧ 
⎪
⎪ 1 α
⎪
⎪ min J ( u , y ) =  y − yd 2L 2 (Ω) + u 2U
⎪
⎪ y ∈Y, u ∈U 2 2
⎪
⎪
⎨ 0
u (t , x)
s.t. ytt + c yt + dy + ρ 2 y − η y + = 0 in Q , (P2sc )
⎪
⎪ (1 + y )2
⎪
⎪
⎪
⎪ y = ∂ν y = 0 on (0, T ) × ∂Ω, y = yt = 0 in {0} × Ω,
⎪
⎪
⎩
− y  M y in Q .
Proceeding as in Section 4.1, we can obtain existence of a minimizer and first order optimality conditions.

Theorem 4.6. There exists a local minimizer (u ∗ , y ∗ ) ∈ U × Y M of (P2sc ) with u ∗ = v ∗ · (1 + ϕ ∗ ). Furthermore, there exist μ∗ ∈
M(0, T ; M(Ω)) and p ∗ ∈ P satisfying
⎧ 2β u ∗  
⎪
⎪ p ∗
− cp ∗
+ dp ∗
+ ρ  2 ∗
p − η  p ∗
− p ∗ = −γ y ∗ − yd − μ∗ ,
⎪
⎪ tt t
+ ∗ )3
⎪
⎪
(1 y
⎪
⎪ T
⎪
⎪ 1 + ϕ∗ ∗
⎪
⎪ ∗
⎨ γ αv = p dt ,
(1 + y ∗ )2 (4.10)
⎪
⎪ 
0
⎪
⎪ v∗
⎪
⎪ γ αϕ ∗
= p ∗ dx,
⎪
⎪ (1 + y ∗ )2
⎪
⎪
⎪
⎪
⎩ ∗ 
Ω
μ , y − y ∗ M,C  0 for all y ∈ YM ,
where γ ∈ {0, 1} and γ +  p ∗ P > 0.
468 C. Clason, B. Kaltenbacher / J. Math. Anal. Appl. 410 (2014) 455–468

The proof of this theorem closely follows the proofs in Section 4.1, additionally making use of the results on U derived
in Section 3.2 (especially the weak closedness and weak compactness of balls BrU (u ∗ ) with respect to  · U ). Note that first
order optimality conditions in qualified form, i.e., with γ = 1 in (4.10), would require using  ·  L 2 ( Q ) as a penalty term.
However, the term  · U is essential for proving existence of a minimizer, since  ·  L 2 ( Q ) would not yield a U-coercive
cost functional, cf. (3.5). (For the control constrained case, U-coercivity was not required, but the low dimensionality of the
control still prevents optimality conditions in qualified form.)

5. Conclusions

Using techniques from [3] and [5], we have extended our results on control and state constrained optimal control of static
MEMS [7] to the time-dependent case, considering as practically relevant controls the dielectric properties of the membrane
and/or the applied voltage. The approach followed in this work is able to deal with the relatively low dimensionality of the
control and thus promises to be useful for boundary control of singular PDEs (such as the Westervelt equation) arising from
high intensity ultrasound applications [8]. Future work in the context of MEMS applications is concerned with the design
and control of devices containing magnetostrictive layers, which leads to optimal control problems for PDEs with hysteresis
properties.

Acknowledgments

The authors wish to thank Ira Neitzel, Christian Meyer, and Daniel Wachsmuth for pointing them to the localization
technique from [5].
The work of the first-named author was supported by the Austrian Science Fund (FWF) under grant SFB F32 (SFB “Math-
ematical Optimization and Applications in Biomedical Sciences”). The second-named author would like to thank the Austrian
Science Fund (FWF) for financial support within the grant P24970 (“Mathematics of Nonlinear Acoustics: Analysis, Numerics,
and Optimization”).
Finally, the authors thank one of the referees for helpful comments leading to an improved presentation.

References

[1] R. Adams, J. Fournier, Sobolev Spaces, Elsevier, Oxford, 2003.

[2] J.-J. Alibert, J.-P. Raymond, A Lagrange multiplier theorem for control problems with state constraints, Numer. Funct. Anal. Optim. 19 (1998) 697–704.
[3] J.F. Bonnans, E. Casas, Optimal control of semilinear multistate systems with state constraints, SIAM J. Control Optim. 27 (1989) 446–455.
[4] J.F. Bonnans, A. Shapiro, Perturbation Analysis of Optimization Problems, Springer-Verlag, New York, 2000.
[5] E. Casas, F. Tröltzsch, Error estimates for the finite-element approximation of a semilinear elliptic control problem, Control Cybernet. 31 (2002)
695–712.
[6] D. Cassani, B. Kaltenbacher, A. Lorenzi, Direct and inverse problems related to MEMS devices, Inverse Problems 25 (2009).
[7] C. Clason, B. Kaltenbacher, On the use of state constraints in optimal control of singular PDEs, Systems Control Lett. 62 (2013) 48–54.
[8] C. Clason, B. Kaltenbacher, S. Veljovic, Boundary optimal control of the Westervelt and the Kuznetsov equation, J. Math. Anal. Appl. 356 (2009) 738–751.
[9] F. Tröltzsch, Optimal Control of Partial Differential Equations. Theory, Methods and Applications, Grad. Stud. Math., vol. 112, American Mathematical
Society, Providence, RI, 2010, translated from the 2005 German original by Jürgen Sprekels.
[10] M. Vishik, A. Fursikov, Mathematical Problems in Statistical Hydromechanics, Kluwer, Boston, 1988.