See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/241705285

Optimal Control of Shear-Thinning Fluids

Article in SIAM Journal on Control and Optimization · January 2012

DOI: 10.1137/110842612

CITATIONS READS

26 68

1 author:

Nadir Arada

43 PUBLICATIONS 757 CITATIONS

SEE PROFILE

All content following this page was uploaded by Nadir Arada on 30 March 2015.

The user has requested enhancement of the downloaded file.

 1

Optimal control of shear-thinning fluids

Nadir Arada1

Abstract
The aim of this paper is to establish necessary optimality conditions for optimal control
problems governed by steady, incompressible Navier-Stokes equations with shear-dependent
viscosity. The main difficulty is related with the differentiability of the control-to-state map-
ping and is overcome by introducing a family of smooth approximate control problems, and
by passing to the limit in the corresponding optimality conditions.

Key words. Optimal control, steady Navier-Stokes equations, shear-thinning flows, neces-
sary optimality conditions.

AMS Subject Classification. 49K20, 76D55, 76A05.

1 Introduction
This paper deals with optimal control problems associated with a viscous, incompressible fluid
described by the following partial differential equations that generalize the Navier-Stokes system


 −∇ · (τ (Dy)) + y · ∇y + ∇π = u in Ω,

∇·y =0 in Ω, (1.1)


y=0 on Γ,


where y is the velocity field, π is the pressure, τ is the extra stress tensor, Dy = 21 ∇y + (∇y)T

is the symmetric part of the velocity gradient ∇y, u is the given body force and Ω ⊂ IRn (n = 2
or n = 3) is a bounded domain with boundary Γ. We assume that τ : IRn×n n×n
sym −→ IRsym has a
potential, i.e. there exists a function Φ ∈ C 2 (IR+ +
n , IRn ) with Φ(0) = 0 such that

∂Φ(|η|2 )
τij (η) = ∂ηij = 2Φ0 (|η|2 ) ηij for all η ∈ IRn×n
sym , τ (0) = 0.

Here IRn×n


sym consists of all symetric (n × n)-matrices. Moreover, we assume that the following
assumptions hold
A1 - There exists a positive constant γ such that for all i, j, k, ` = 1, · · · , n

∂τk` (η)
α−2

∂ηij ≤ γ 1 + |η|2 2
for all η ∈ IRn×n
sym .

A2 - There exists a positive constant ν such that

X ∂τ
k` (η)

α−2
τ 0 (η) : ζ : ζ = ∂ηij ζk` ζij ≥ ν 1 + |η|2 2
|ζ|2 for all η, ζ ∈ IRn×n
sym .
ijk`
1 Departamento de Matemática, Faculdade de Ciências e Tecnologia da Universidade Nova de Lisboa. E-mail:
[email protected].
 2

These assumptions are usually used in the literature and cover a wide range of non-Newtonian
fluids. Typical prototypes of extra tensors used in applications are

α−2 α−2
τ (η) = 2ν 1 + |η|2 2
η or τ (η) = 2ν (1 + |η|) η.

We recall that a fluid is called shear-thickening if α > 2 and shear-thinning if α < 2. For the
special case τ (η) = 2νη (α = 2), we recover the Navier-Stokes equation with viscosity coeficient
ν > 0.
The paper is concerned with the following optimal control problem
Z Z
2 2
Minimize J(u, y) = 12 |y − yd | dx + λ2 |u| dx
(Pα ) Ω Ω
Subject to (u, y) ∈ Uad × W01,α (Ω) satisfies (1.1) for some π ∈ Lα (Ω)

where yd is some desired velocity field, λ is a positive constant, the set of admissible controls
3n
Uad is a nonempty convex closed subset of L2 (Ω) and n+2 ≤ α ≤ 2. Although the analysis
of several results can be more general, in order to simplify the redaction, we will assume that
Uad ⊂ {v ∈ L2 (Ω) | kvk2 ≤ U } for some U > 0.
The considered class of fluids is described by partial differential equations of the quasi-linear
type. It was first proposed by Ladyzhenskaya in [17], [18] and [19] as a modification of the
Navier-Stokes system (the viscosity depending on the shear-rate), and was similarly suggested
by Lions in [20]. Existence of weak solutions was proved by both authors using compactness
arguments and the theory of monotone operators. Since these pioneering results, much has been
done and we emphasize the works by Nečas et al. who proved existence of weak and measure-
2n
valued solutions under the less restrictive assumption α > n+2 (see for example [25] and [11]).
In the absence of flow convection, optimal control problems governed by generalized Stokes
systems can be studied following the ideas developed in [4] and [6] for problems governed by
quasilinear elliptic equations. Similar underlying difficulties, consequence of the nonlinearity
of the extra-stress tensor, are related with the differentiability of the control-to-state mapping.
The corresponding analysis cannot be achieved in Sobolev spaces and the natural setting for the
linearized equation and the adjoint state equation involves weighted Sobolev spaces. The lack of
regularity of the state variable in the case of shear-thinnig fluids creates an additional difficulty
that can be overcome by considering a family of approximate problems falling into the case α = 2.
Differentiability of the approximate control-to-state mapping can then be established, allowing
to derive the approroximate optimality conditions, and the optimality conditions by passing to
the limit.
The case of problems governed by generalized Navier-Stokes equations is more delicate since
another difficulty arises in connection with the convective term and the uniqueness of the state
variable, guaranteed under some constraint on the data. It is similarly encountered when studying
problems governed by the Navier-Stokes equations for which the necessary optimality conditions
can be established by restraining all the admissible controls to satisfy this constraint (see for
example [10] and [26]).
The difficulties related with the nonlinearity of the extra stress tensor and the convective term
can be more easily handled (especially in the case of shear-thinning flows) if the gradient of the
velocity is bounded. The corresponding viscosity, although non constant, is also bounded and
the system can be studied as in the case of Navier-Stokes equations. These regularity results are
few, difficult to obtain in general and do not seem to be available for the three-dimensional case.
For the two-dimensional steady case, the boundedness of the gradient was proved by Kaplický et
al. in [22] enabling Slawig to derive the corresponding optimality conditions in [27]. Similarly,
 3

Wachsmuth and Roubı́ček used the regularity results established in [21] to derive the optimality
conditions for a two-dimensional unsteady system describing the flow of shear-tickening fluids
(see [28]). Related to this aspect, we also mention Gunzburger and Trenchea who used the reg-
ularity results obtained in [14] to derive the optimality conditions for a problem governed by a
three-dimensional modified Navier-Stokes system coupled with Maxwell equations (see [15]).
There are few works dealing with these problems when no higher regularity results are available.
We mention the recent paper by De los Reyes [8], who considered a problem governed by the
Bingham nonlinear mixed variational inequality. Besides difficulties induced by the nonlinearity
of the viscoplastic and the convective terms, the non-regularity of the model has to be man-
aged. By exploiting the specific structure of the non-differentiable term, a family of regularized
problems is introduced, the corresponding optimality systems are derived and the optimality
conditions for the original problem are obtained by passing to the limit. We also mention our
work dealing with steady shear-thickening fluids where the restriction on the set of admissible
controls has been relaxed and the optimality conditions obtained under a precise condition on
the optimal control (see [2] and [3]).
In the case of shear-thinning fluids, the problems are more difficult to handle. The techniques
developped in [2] cannot be directly applied because of the combined effect of the convective
term and the nonlinear stress tensor. Moreover, unlike the case of problems governed by gen-
eralized Stokes systems and unless we restrict all the admissible controls, the differentiability of
the approximate control-to-state mapping is not guaranteed, the approximate control problem
does not fall in the case α = 2 and further analysis is needed. Let us finally mention that in [2],
[3] and in the present work, the considered potential is C 2 . Nevertheless, the problems are still
challenging especially in the case of shear-thinning flows. A further interesting aspect would be
the adaptation of the techniques developed in [8] to the case of a less regular potential. (In this
respect, see also the paper by Casas and Fernandez [5].)
In the present paper, we establish explicite estimates, carefully analyse the related equations
and derive optimality conditions without restraining the set of admissible controls. The only
constraint concerns the optimal control. The plan is as follows. Assumptions, notation and
some preliminary results are given in Section 2. Section 3 is devoted to existence and uniqueness
results for the state equation and to the derivation of corresponding estimates. Section 4 deals
with existence of an optimal control while necessary optimality conditions are given in Section
5. In Section 6, we introduce a family of approximate control problems, study the properties of
the corresponding control-to-state mapping and establish the approximate optimality conditions.
By passing to the limit, we prove the optimality conditions for the control problem in Section 7.

2 Notation and preliminary results

Throughout the paper Ω ⊂ IRn (n=2 or n=3) is a bounded domain with a boundary Γ of class
C 2 . Since many of the quantities occuring in the paper are vector-valued functions, we will use
the same notation of norms for scalar, vector and matrix-valued functions for the sake of brevity.

2.1 Function spaces and classical inequalities

Let us define some useful function spaces. The space of infinitely differentiable functions with
compact support in Ω will be denoted by D(Ω). The standard Sobolev spaces are denoted by
W k,α (Ω) (k ∈ IN and 1 < α < ∞), and their norms by k · kk,α . We set W 0,α (Ω) ≡ Lα (Ω) and
k · kLα ≡ k · kα . In order to eliminate the pressure in the weak formulation of the state equation,
 4

we will work in divergence-free spaces. Consider

V = {ϕ ∈ D(Ω) | ∇ · ϕ = 0} ,

and denote by Vα the closure of V in the Lα -norm of gradients, i.e.

n o
Vα = ϕ ∈ W01,α (Ω) | ∇ · ϕ = 0 .

Given y ∈ W01,α (Ω), we can associate two weighted Sobolev spaces Vαy and Hαy , where Vαy is the
set of functions z ∈ V2 such that the norm k · k defined by
α−2
kzk = (1 + |Dy|2 ) 4 Dz
2

Hαy
is finite, and is the completion of V in Vαy .
It may be verified that Vαy and Hαy are Hilbert
spaces and that Hα ⊂ Vα . Moreover, we have V2 ⊂ Hαy ⊂ Vα if α ≤ 2, with continuous
y y

injections. Weighted Sobolev Spaces of this type have been studied by Coffman et al. [7],
Murthy and Stampacchia [24].
Let us now collect some useful auxiliary results. We begin by three classical inequalities.
Lemma 2.1 (Poincaré’s inequality.) Let y be in W01,α (Ω) with 1 < α ≤ 2. Then the following
estimate holds
1

 n−1√ |Ω| n
n
if α = 2
kykα ≤ CP,α k∇ykα with CP,α = 1
 α(n−1) √ |Ω| n if α < 2.
2(n−α) n

Proof. See for example [12], Chapter 2. 

Lemma 2.2 (Sobolev’s inequality.) Let y be in W01,α (Ω) with 1 < α < 2. Then the following
estimate holds
α(n−1)
kyk nα ≤ 2(n−α) √ k∇yk .
n α
n−α

Proof. See for example [12], Chapter 2. 

Lemma 2.3 (Korn’s inequality.) Let y be in W01,α (Ω)
with 1 < α ≤ 2. Then there exists a
positive constant CK,α ≤ 1 only depending on α and Ω such that

CK,α k∇yk1,α ≤ kDykα .

Moreover, CK,2 = √1 .
2

Proof. See for example [25] and [16]. 

As a consequence, we have the following useful result.
Lemma 2.4 Let w be in L2 (Ω) and y be in W01,α (Ω) with n+2
2n
< α ≤ 2. Then the following
estimates hold
|(w, y)| ≤ Cα kwk2 kDykα
with  √
2(n−1) 1
 √
n
|Ω| n if α = 2
Cα = (n+2)α−2n
α(n−1) 1
 √
2(n−α) n CK,α
|Ω| 2αn if α < 2
and where CK,α is the constant of Korn.
 5

Proof. For α = 2, due to the Hölder, the Poincaré and the Korn inequalities, we have
√
1 2(n−1) 1
n−1
|(w, y)| ≤ kwk2 kyk2 ≤ √ |Ω| n
n
kwk2 k∇yk2 = √
n
|Ω| n kwk2 kDyk2

2n
wich gives the first estimate. Similarly, if α < 2 (and since α > n+2 ) we have L2 (Ω) ,→ W01,α (Ω)
and by using the Hölder, the Sobolev and the Korn inequalities, we deduce that
(n+2)α−2n
|(w, y)| ≤ kwk nα kyk nα ≤ |Ω| 2αn kwk2 kyk nα
(n+1)α−n n−α n−α

α(n−1) (n+2)α−2n
≤ √ |Ω|
2(n−α) n
2αn kwk2 k∇ykα ≤ Cα kwk2 kDykα

and the second estimate is proven. 

Finally, we point out some notable facts related with the trilinear form b defined by

b(y1 , y2 , y3 ) = (y1 · ∇y2 , y3 ) .

Lemma 2.5 Let w, y and z be in W01,α (Ω) with n+2

3n
≤ α ≤ 2. Then the following estimate
holds
|b (w, y, z)| ≤ κ1 kDwkα kDykα kDzkα
with  3
1
2 2 (n−1)


n |Ω| n(n−1) if α = 2
κ1 = 2 (n+2)α−3n
1 (α(n−1))

 3
CK,α 4n(n−α)2 |Ω| nα if α < 2

and where CK,α is the constant of Korn.

Proof. For α = 2, due to Lemma 1.1, Chapter VIII in [12], we have
1
n−1
|b(w, y, z)| ≤ n |Ω|
n(n−1) k∇wk2 k∇yk2 k∇zk2

and the conclusion follows by using the Korn inequality. If α < 2, Hölder’s and Sobolev’s
2α nα 3n
inequalities together with classical embedding results show that if α−1 ≤ n−α (and thus α ≥ n+2 )
then (n+2)α−3n
|b (w, y, z)| ≤ kwk 2α k∇ykα kzk 2α ≤ |Ω| nα kwk nα k∇ykα kzk nα
α−1 α−1 n−α n−α

(α(n−1))2 (n+2)α−3n
≤ 4n(n−α)2 |Ω|
nα k∇wkα k∇ykα k∇zkα .
The conclusion follows by using the Korn inequality. 
Lemma 2.6 Let w be in Vα and let y and z be in W01,α (Ω) with 3n
n+2 ≤ α ≤ 2. Then

b (w, y, z) = −b (w, z, y) and b (w, y, y) = 0.

2.2 The stress tensor

Let us recall that assumptions A1 -A2 imply the following standard continuity and monotonicity
properties for τ (see [25], Chapter 5)
n2 γ

α−2
|τ (η)| ≤ α−1 1 + |η|2 2
|η|, (2.1)

α−2 2
(τ (η) − τ (ζ)) : (η − ζ) ≥ ν 1 + |η 2 | + |ζ|2 2
|η − ζ| . (2.2)
The next auxiliary results deal with properties related with the tensor τ .
 6

α
Lemma 2.7 Let 1 < α < 2 and let f ∈ L 2−α (Ω), g ∈ L1 (Ω) and h ∈ Lα (Ω) be non negative
functions satisfying
h(x)2 ≤ f (x)g(x) for a.e. x ∈ Ω.
Then,
2
khkα ≤ kf k α kgk1 .
2−α

Proof. Taking into account the condition satisfied by f , g, h, integrating and using the Hölder
inequality, we obtain
Z Z
α
α
2 2 α α α α α α
khkα = h(x) dx ≤ f (x) 2 g(x) 2 dx ≤ f 2 2 g 2 2 = kf k 2 α kgk12
2−α α 2−α
Ω Ω

and the proof is complete. 

Lemma 2.8 Let y and z be in W01,α (Ω) with 1 < α ≤ 2. Then
νkD(y−z)k2α
(τ (Dy) − τ (Dz) , D(y − z)) ≥ 2−α .
(|Ω|+kDykα α
α +kDzkα )
α

Proof. If α = 2, then the result is a direct consequence of the monotonicity condition (2.2).
Assume then that α < 2. Since y and z belong to W01,α (Ω), by setting

2−α
f = 1 + |Dy|2 + |Dz|2 2
, g= 1
ν (τ (Dy) − τ (Dy)) : D(y − z), h = |D(y − z)|

and taking into account the monotonicity condition (2.2), we can see that the assumptions of
Lemma 2.7 are fulfilled. Therefore
2 2−α
1
kD(y − z)kα ≤ (1 + |Dy|2 + |Dz|2 ) 2
α ν (τ (Dy) − τ (Dz)) : D(y − z) 1
2−α
2−α
1
≤ ν (|Ω| + kDykα α
α + kDzkα )
α
(τ (Dy) − τ (Dz) , D(y − z))

and the result is proven. 

Lemma 2.9 Let 1 < α ≤ 2 and let z and y be in H01 (Ω). Then
2
α−2 kDzk2α
(1 + |Dy|2 ) 4 Dz ≥ 2−α .
2 (|Ω|+kDykα
α)
α

Proof. The case α = 2 is obvious. In case α < 2, since y and z belong to H01 (Ω), by setting

2−α
α−2
f = 1 + |Dy|2 2
, g = |Dz|2 1 + |Dy|2 2
, h = |Dz|,

we can see that the assumptions of Lemma 2.7 are satisfied and then
2 2−α α−2
kDzkα ≤ (1 + |Dy|2 ) 2
α
|Dz|2 (1 + |Dy|2 ) 2

2−α 1
2−α α−2
≤ (|Ω| + kDykα
α)
α
|Dz|2 (1 + |Dy|2 ) 2
1

which gives the result. 

 7

3 State equation
This section is devoted to existence and uniqueness results for the state equation and to derivation
of explicite estimates useful for the subsequent analysis.
First mathematical investigations of (1.1) under conditions (2.1)-(2.2), were performed by J. L.
3n
Lions who proved existence of a weak solution for α ≥ n+2 (see [20] for more details). The
restriction on the exponent α ensures that the convective term belongs to L1 when considering
test functions in Vα (cf. Lemma 2.5).
Multiplying equation (1.1) by test functions ϕ ∈ Vα and integrating, we obtain the following
weak formulation.
Definition 3.1 Let u ∈ L2 (Ω). A function y ∈ Vα is a weak solution of (1.1) if

(τ (Dy) , Dϕ) + b (y, y, ϕ) = (u, ϕ) for all ϕ ∈ Vα .

Let us recall that, having a solution satisfying the formulation given in Definition 3.1, it is
standard to construct the corresponding pressure π ∈ Lα0 (Ω) such that

(τ (Dy) , Dϕ) + b (y, y, ϕ) − (π, ∇ · y, ϕ) = (u, ϕ) for all ϕ ∈ W01,α (Ω).

We will involve the pressure only in the formulations of the theorems and lemmas but not in the
proofs, since it can always be reconstructed uniquely.
We begin by stating an existence result for the state equation and related useful estimates.
3n
Theorem 3.2 Assume that A1 -A2 are fulfilled with n+2 ≤ α ≤ 2. Then for u ∈ L2 (Ω), equation
(1.1) admits at least a weak solution yu ∈ Vα . Moreover, the following estimates hold
  2−α
kDyu kα ≤ κ2 1 + kuk
α−1 kuk
2 2
ν ν , (3.1)
 2−α α
 α−1
α kuk
kDyu kα ≤ 2 2 Cα ν 2 + |Ω|, (3.2)

 α  2−α
α
|Ω|
where κ2 = Cα α−1 + Cαα−1 with Cα defined in Lemma 2.4.
3n
Proof. As already observed, existence of a weak solution for problem (1.1) with α ≥ n+2 is well
known. To establish the estimates, we split the proof into two steps.
Step 1. Let us set ϕ = yu in the weak formulation of (1.1) and use Lemma 2.8, Lemma 2.6 and
Lemma 2.4 to obtain
νkDyu k2α
2−α ≤ (τ (Dyu ) , Dyu ) = (u, yu ) ≤ Cα kuk2 kDyu kα
(|Ω|+kDyu kαα ) α

If α = 2, the estimate is direct. If α < 2, we have

2−α
Cα kuk2 α
kDyu kα ≤ ν (|Ω| + kDyu kα ) α

and thus  α
 2−α
α Cα kuk2 α
kDyu kα2−α ≤ ν (|Ω| + kDyu kα ) . (3.3)
On the other hand, the Young inequality yields
  α α
  α
Cα kuk2 2−α α Cα kuk2 (2−α)(α−1)
ν kDyu kα ≤ (2 − α) kDy u k 2−α
α + (α − 1) ν . (3.4)
 8

Combining (3.3) and (3.4), we deduce that

α
  α   α
Cα kuk2 2−α Cα kuk2 (2−α)(α−1)
(α − 1) kDyu kα2−α ≤ ν |Ω| + (α − 1) ν

and consequently
2−α
  α  α
 α−1
|Ω| Cα kuk2 Cα kuk2
kDyu kα ≤ α−1 + ν ν
2−α
 α  2−α   α  α
 α−1
|Ω| α−1 α kuk2 kuk2
≤ Cα α−1 + Cα 1+ ν ν
 α  2−α   2−α
|Ω| α kuk2 α−1 kuk2
≤ Cα α−1 + Cαα−1 1+ ν ν

which gives (3.1).

Step 2. Let us now prove (3.2). Similar arguments together with (2.2) show that
Z Z
α
kDyu kα = |Dyu (x)|α dx + |Dyu (x)|α dx
{x||Dyu (x)|≥1} {x||Dyu (x)|<1}
Z
|Dyu (x)|2
≤ |Dyu (x)|2−α dx + |Ω|
{x||Dyu (x)|≥1}
Z
2−α |Dyu (x)|2
≤2 2 2−α dx + |Ω|
2
{x||Dyu (x)|≥1} (1+|Dyu (x)| ) 2
2−α
Z
2 2
≤ ν τ (Dyu (x)) : Dyu (x) dx + |Ω|
{x||Dyu (x)|≥1}
2−α
2−α
2 2 u


≤ ν (τ (Dyu ), Dyu ) + |Ω| = 2 2
ν , yu + |Ω|
2−α kuk2
≤2 2 Cα ν kDyu kα + |Ω|. (3.5)
The Young inequality yields
 α
 α−1
2−α kuk2 2−α kuk2 α
α−1 1
2 2 Cα ν kDyu kα ≤ α 2 2 Cα ν + α kDyu kα (3.6)

and the claimed result follows by combining (3.5) and (3.6). .

The next result deals with uniqueness of weak solutions.
3n
Theorem 3.3 Assume that A1 -A2 are fulfilled with n+2 ≤ α ≤ 2 and that u ∈ L2 (Ω) satisfies
  2(2−α)
kuk2 α−1 kuk2
κ̄ 1 + ν ν2 < 1, (3.7)

where κ̄ = κ1 κ2 κ3 with κ1 defined in Lemma 2.5, κ2 defined in Theorem 3.2 and κ3 =

2−α
 (4−α)α α  α
α−1
3|Ω| + 2 2(α−1) Cα . Then, equation (1.1) admits a unique weak solution yu ∈ Vα .
Proof. Assume that yu and χu are two weak solutions of (1.1) corresponding to u. Setting
ϕ = yu − χu in the corresponding weak formulation and taking into account Lemma 2.8 and
Lemma 2.6, we obtain
νkD(yu −χu )k2α
2−α ≤ (τ (Dyu ) − τ (Dχu ), D (yu − χu ))
(|Ω|+kDyu kαα +kDχu kαα ) α

= b (χu , χu , yu − χu ) − b (yu , yu , yu − χu )
= −b (yu − χu , χu , yu − χu ) . (3.8)
 9

Lemma 2.5 and estimate (3.1) then yield

2
|b (yu − χu , χu , yu − χu )| ≤ κ1 kD (yu − χu )kα kDχu kα
  2−α
≤ κ1 κ2 1 + kuk
α−1 kuk 2
ν kD (yu − χu )kα .
2 2
ν (3.9)

On the other hand, by taking into account estimate (3.2), we have

2−α
 α  α
2−α  2−α
 α−1
α α kuk
(|Ω| + kDyu kα + kDχu kα ) α
≤ 3|Ω| + 2 2 2 Cα ν 2
2−α
 α α  α
(4−α)α
  α−1
α−1 kuk2
= 3|Ω| + 2 × 2 2(α−1) Cα 2ν

  2−α
kuk2 α−1
≤ κ3 1 + ν . (3.10)

By combining (3.8), (3.9) and (3.10), we deduce that

2−α
!
  α−1
ν kuk2 kuk2 2
  2−α − κ1 κ2 1 + ν ν kD (yu − χu )kα ≤ 0
kuk
κ3 1+ ν 2 α−1

and thus yu = χu if condition (3.7) is satisfied. 

Remark 3.4 Notice that in the case of the Navier-Stokes equations (α = 2), condition (3.7)
reduces to √
kuk2 n3
ν2 < 1 . (3.11)
4(n−1)2 |Ω| n−1

Remark 3.5 Condition (3.7) is fulfilled if the term kukν

2
is ”small enough”, and can be inter-
preted either as a constraint on the size of kuk2 (small body force u) or as a restriction on the
viscosity parameter ν (large viscosity parameter ν).

4 Existence of an optimal control

3n
Theorem 4.1 Assume that A1-A2 are fulfilled with n+2 ≤ α ≤ 2. Then problem (Pα ) admits
at least a solution.

Proof. The proof is split into three steps.

Step 1. Considering a minimizing sequence (uk , yk )k ⊂ Uad × Vα , let us establish related
estimates and preliminary convergence results. Since (uk )k is uniformly bounded in the closed
convex set Uad , by taking into account (3.2) we obtain
 2−α α
 α−1  2−α α
 α−1
α ku k
kDyk kα ≤ 2 2 Cα νk 2 + |Ω| ≤ 2 2 Cα Uν + |Ω| (4.1)

and the sequence (yk )k is then bounded in Vα . On the other hand, the continuity condition (2.1)
implies that for η 6= 0

n2 γ

α−2 n2 γ n2 γ
|τ (η)| ≤ α−1 1 + |η|2 2
|η| ≤ α−1 |η|
α−2
|η| = α−1 |η|
α−1
.
 10

The previous inequality is also valid for |η| = 0 and implies

α
n2 γ α
kτ (Dyk )k α−1
α ≤ α−1 kDyk kα
α−1

α
which together with (4.1) show that sequence (τ (Dyk ))k is uniformly bounded α
in L α−1 (Ω).
There then exist a subsequence, still indexed by k, u ∈ Uad , y ∈ Vα and τe ∈ L α−1 (Ω) such that
(uk )k weakly converges to u in L2 (Ω), (yk )k weakly converges to y in Vα and (τ (Dyk ))k weakly
α
2n
converges to τe in L α−1 (Ω). Moreover, since α > n+1 , by using compactness results on Sobolev
α
spaces, we deduce that (yk )k strongly converges to y in L α−1 (Ω).
Step 2. Let us now prove that (u, y) is an admissible pair for (Pα ). Taking into account the
convergence results obtained in Step 1, we deduce that for every ϕ ∈ V, we have

|b (yk , yk , ϕ) − b (y, y, ϕ)| ≤ |b (yk − y, yk , ϕ)| + |b (y, yk − y, ϕ)|

= |b (yk − y, yk , ϕ)| + |b (y, ϕ, yk − y)|
(4.2)
≤ (k∇yk kα kϕk∞ + kykα k∇ϕk∞ ) kyk − yk α
α−1

−→ 0 when k → +∞.

Moreover, by passing to the limit in the weak formulation corresponding to yk , we obtain

(e
τ , Dϕ) + b (y, y, ϕ) = (u, ϕ) for all ϕ ∈ V
2α
3n
and by using the fact that V is dense in Vα and that y ∈ L α−1 (Ω) if α ≥ n+2 , it follows that

(e
τ , Dϕ) + b (y, y, ϕ) = (u, ϕ) for all ϕ ∈ Vα . (4.3)

In particular, by taking into account Lemma 2.6, we have

(e
τ , Dy) = (e
τ , Dy) + b (y, y, y) = (u, y) . (4.4)

On the other hand, the monotonicity assumption (2.2) implies

(τ (Dyk ) − τ (Dϕ) , Dyk − Dϕ) ≥ 0 for all ϕ ∈ Vα . (4.5)

Since (τ (Dyk ) , Dyk ) = (uk , yk ), by substituing in (4.5), we obtain

(uk , yk ) − (τ (Dyk ) , Dϕ) − (τ (Dϕ) , Dyk − Dϕ) ≥ 0 for all ϕ ∈ Vα

and by passing to the limit, we get

(u, y) − (e
τ , Dϕ) − (τ (Dϕ) , Dy − Dϕ) ≥ 0 for all ϕ ∈ Vα .

This inequality together with (4.4) then yields

τ − τ (Dϕ) , Dy − Dϕ) ≥ 0
(e for all ϕ ∈ Vα

and by setting ϕ = y − tψ with t > 0, we obtain

τ − τ (Dy − tDψ) , Dψ) ≥ 0

(e for all ψ ∈ Vα .

Letting t tend to zero and using the continuity of τ , we deduce that

τ − τ (Dy) , Dψ) ≥ 0
(e for all ψ ∈ Vα
 11

and thus
(e
τ , Dψ) = (τ (Dy) , Dψ) for all ψ ∈ Vα . (4.6)
Combining (4.3) and (4.6), we deduce that

(τ (Dy) , Dϕ) + b (y, y, ϕ) = (u, ϕ) for all ϕ ∈ Vα

showing that
yk −→ y weakly in W01,α (Ω)
and that (u, y) satisfies (1.1).
Step 3. Finally, from the convexity and continuity of J, it follows the lower semicontinuity of
J in the weak topology and

J(u, y) ≤ lim inf J(uk , yk ) = inf(Pα ),

k

showing that (u, y) is a solution for (Pα ). 

5 Statement of the necessary optimality conditions

In order to obtain the necessary optimality conditions for (Pα ) stated in Theorem 5.1 below, a
family of problems (Pαε )ε whose solutions converge towards a solution of (Pα ) is introduced and
the corresponding optimality conditions are derived in Section 6. We pass to the limit in these
conditions in Section 7.
Let us now formulate our main result.
3n
Theorem 5.1 Assume that A1 -A2 are fullfilled with n+2 ≤ α ≤ 2. Let ū be a solution of (Pα )
satisfying condition (3.7) and let ȳ ∈ Vα be the associated state. There then exists p̄ ∈ Vα such
that the following conditions hold


 −∇ · (τ (Dȳ)) + ȳ · ∇ȳ + ∇π̄ = ū in Ω,

∇ · ȳ = 0 in Ω,


ȳ = 0 on Γ,


0 T
 −∇ · (τ (Dȳ) : Dp̄) − ȳ · ∇p̄ + (∇ȳ) p̄ + ∇e


π = ȳ − yd in Ω,
∇ · p̄ = 0 in Ω, (5.1)


p̄ = 0 on Γ,


(p̄ + λū, v − ū) ≥ 0 for all v ∈ Uad . (5.2)

Moreover, p̄ satisfies
(τ 0 (Dȳ) : Dp̄, Dp̄) + (p̄ · ∇ȳ, p̄) ≤ (ȳ − yd , p̄) . (5.3)
Notice that the optimality conditions for (Pα ) are obtained under a constraint on the optimal
control, the same that guarantees uniqueness of the corresponding state. This result seems
interesting in the sense that we do not need to impose any other constraint on the admissible
set of controls. Notice also that for α < 2, condition (5.3) implies that p̄ belongs to Vαȳ and not
necessarily to Hαȳ . Therefore, the adjoint equation is to be understood in the distributional sense

(τ 0 (Dȳ) : Dp̄, Dϕ) + (∇ȳ)T p̄ − ȳ · ∇p̄, ϕ = (ȳ − yd , ϕ)

for all ϕ ∈ V.
 12

Let us finish this section by considering the case of the Navier-Stokes equations. It corresponds
to α = 2 and Vαȳ ≡ Hαȳ ≡ V2 . The first order optimality conditions we obtain in this case are less
restrictive than the ones obtained in [9], [10], [26] where all the admissible controls are subject
to a condition that ensures the uniqueness of the corresponding states. Condition (3.7) reduces
to (3.11) and guarantees uniqueness of both optimal state and optimal adjoint state. It implies
that the set Uad of admissible controls satisfies the property (C), introduced by Gunzburger et
al. [15], at (ū, ȳ). Our result can then be seen as a qualified version of the optimality conditions
already established by Abergel and Casas in [1] for a slightly different functional.
Corollary 5.2 Assume that the extra-stress tensor has the form τ (η) = 2νη. Let (ū, ȳ) be
a solution of (Pα ) with ū satisfying (3.11). There then exists a unique p̄ ∈ V2 such that the
following conditions hold


 −ν ∆ȳ + (ȳ · ∇) ȳ + ∇π̄ = ū in Ω,

∇ · ȳ = 0 in Ω,


ȳ = 0 on Γ,




 −ν ∆p̄ − (ȳ · ∇) p̄ + (∇ȳ)T p̄ + ∇e
π = ȳ − yd in Ω,

∇ · p̄ = 0 in Ω,


p̄ = 0 on Γ,


(p̄ + λū, v − ū) ≥ 0 for all v ∈ Uad .

6 Approximate optimal control problem

When deriving the first order optimality conditions, we have to manage several combined diffi-
culties related with the local Lipschitz continuity (and thus with the Gâteaux differentiability)
in adequate functional spaces of the control-to-state mapping u 7−→ yu .
To clarify the ideas, let us first assume that we are dealing with generalized Stokes systems (no
convective term). In the case α ≥ 2, we can follow the ideas developed in [4] and [6] to study
optimal control problems governed by quasi-linear elliptic equations and prove that the sequence
(zρ )ρ>0 defined by
y −yu
zρ = u+ρ(v−u)
ρ , u, v ∈ Uad , ρ ∈]0, 1[,
converges weakly in the weighted Sobolev space Hαyu and strongly in V2 . To prove that the limit
belongs to Hαyu , it is essential that (zρ )ρ>0 be uniformly bounded in Vα ⊂ Hαyu . If α < 2, we
can argue similarly and show that (zρ )ρ>0 is uniformly bounded in Vα . Nevertheless, in this case
Hαyu ⊂ Vα and we can only prove that there exist subsequences converging to elements wich are
solutions of a linearized system in the distributional sense and belonging to Vαyu . Differentiability
of the control-to-state mapping is equivalent to the equality of all these limit points and this
problem of uniqueness leads us to the problem of density of the function space V in Vαyu . To
overcome this difficulty, we can consider a family of approximate problems (Pαε ) governed by the
following regularized equation

−ε ∆y − ∇ · (τ (Dy)) + ∇π = u in Ω, ∇·y =0 in Ω, y=0 on Γ

yε ε

−yu
and falling into the case α = 2. Adequate estimates for zρε = u+ρ(v−u) ρ ρ>0
can be estab-
lished in V2 (here yuε denotes the solution of the previous equation corresponding to u), allowing
 13

to derive the approximate optimality conditions and to obtain optimality conditions for (Pα ) by
passing to the limit.
The case of problems governed by Navier-Stokes equations and generalized Navier-Stokes equa-
tions is more delicate since a direct adaptation of these arguments, managing the convective
term, may restrain not only u but also v (and by extension, all the admissible controls) to satisfy
condition (3.11) (see for example [10], [26] and [27]). This difficulty is overcome in the case
of shear-thickening and Navier-Stokes fluids by observing that, in order to establish uniform
estimates of (zρ )ρ>0 (consequence of the local Lipchitz continuity of the state with respect to
the control) in Vα , the terms we need to restrain are related to the convective term and only
depend on u. This fact is particularly important and enables us, when deriving the necessary
optimality conditions, to impose a constraint only on the optimal control (see [3]). In the case
of shear-thinning flows, the problems are even more difficult to handle because of the combined
effect of the convective term and the nonlinear stress tensor. Unlike the case of generalized Stokes
systems, obtaining uniform estimates for (zρε )ρ>0 in V2 without restraining both u and v is not
an easy issue. In this section, by carrying out a careful analysis, we prove that uniform estimates
for (zρε )ρ>0 can be established under a condition involving ρ and the regularization parameter
ε and by imposing restriction (3.7) only on u. The approximate optimality conditions are then
derived.

6.1 Setting and approximate optimality conditions

For ε > 0 and u in L2 (Ω), consider the following problem


 −ε ∆y − ∇ · (τ (Dy)) + y · ∇y + ∇π = u in Ω,

∇·y =0 in Ω, (6.1)


y=0 on Γ.


Let (ū, ȳ) be a fixed solution of (Pα ) and assume that ū satisfies condition (3.7). Introduce the
cost functional Z
I(u, y) = J(u, y) + 12 |u − ū|2 dx
Ω
and the control problem
(
ε
minimize I(u, y ε )
(Pα )
subject to (u, y ε ) ∈ Uad × H01 (Ω) satisfies (6.1) for some πε ∈ L2 (Ω).
The main result of this section deals with the necessary optimality conditions for the approximate
problem (Pαε ).
3n
Theorem 6.1 Assume that A1 -A2 are fulfilled with n+2 ≤ α ≤ 2. For each ε > 0, there exists
at least one solution (ū , ȳ ) of (Pα ). Moreover, if ū satisfies (3.7), then there exists p̄ε ∈ V2
ε ε ε ε

such that
ε ε ε ε ε ε

 −ε∆ȳ − ∇ · (τ (Dȳ )) + ȳ · ∇ȳ + ∇π = ū in Ω,

∇ · ȳ ε = 0 in Ω,

 ε
ȳ = 0 on Γ,
ε 0 ε ε ε T ε ε ε ε ε

 −ε∆p̄ − ∇ · (τ (Dȳ ) : Dp̄ ) + (∇ȳ ) p̄ − ȳ · ∇p̄ + ∇π̃ = ȳ − yd in Ω,

∇ · p̄ε = 0 in Ω, (6.2)

 ε
p̄ = 0 on Γ,
(p̄ε + (λ + 1)ūε − ū, v − ūε ) ≥ 0 for all v ∈ Uad . (6.3)
 14

6.2 Approximate state equation

In the following proposition we state an existence and uniqueness result for the approximate
state equation (6.1) and related a priori estimates.
3n
Proposition 6.2 Assume that A1 -A2 are fulfilled with n+2 ≤ α ≤ 2. Then for u ∈ L2 (Ω),
ε
equation (6.1) admits at least a weak solution yu ∈ V2 and the following estimates hold
  2−α
kDyuε kα ≤ κ2 1 + kuk
α−1 kuk
2 2
ν ν , (6.4)
 2−α α
 α−1
α kuk
kDyuε kα ≤ 2 2 Cα ν 2 + |Ω|, (6.5)
2ε kDyuε k2 ≤ C2 kuk2 , (6.6)
 1
 α−1 
2 2−α kuk 1
2ε kDyuε k2 ≤ Cα kuk2 2 2 Cα ν 2 + |Ω| α . (6.7)

Moreover, if u satisfies condition (3.7), then the solution is unique.

Proof. The weak formulation associated with problem (6.1) reads as
(2εDy + τ (Dy), Dϕ) + b (y, y, ϕ) = (u, ϕ) for all ϕ ∈ V2 .
Since 2εI + τ satisfies assumptions A1 -A2 with exponent 2, classical arguments show that the
approximate system (6.1) admits at least a weak solution yuε in V2 . Setting ϕ = yuε in the weak
formulation of (6.1) yields
2
2ε kDyuε k2 + (τ (Dyuε ) , Dyuε ) + b (yuε , yuε , yuε ) = (u, yuε ) .
Estimates (6.4) and (6.5) can then be obtained with arguments similar to those used in the proof
of Theorem 3.2. On the other hand, by taking into account the monotonicity condition (2.2),
Lemma 2.6 and Lemma 2.4, we obtain
2 2
2ε kDyuε k2 ≤ 2ε kDyuε k2 + (τ (Dyuε ) , Dyuε ) = (u, yuε ) ≤ C2 kuk2 kDyuε k2
which gives estimate (6.6). Similarly, by taking into account Lemma 2.4 and estimate (6.5) we
have
2
2ε kDyuε k2 ≤ (u, yuε ) ≤ Cα kuk2 kDyuε kα
and estimate (6.7) is proven. The uniqueness result can be obtained with arguments similar to
those used in the proof of Theorem 3.3. 
Remark 6.3 Notice that the existence result as well as the estimates stated in Proposition 6.2
are valid for α ∈]1, 2]. To prove the uniqueness result, we need to restrain the values of α in
2α
order to guarantee the embedding of W01,α (Ω) into L α−1 (Ω) (cf. Lemma 2.5).

6.3 Linearized equation

We next investigate the following linearized equation


 −ε∆z − ∇ · (τ 0 (Dyuε ) : Dz) + z · ∇yuε + yuε · ∇z + ∇π = w in Ω,

∇·z =0 in Ω, (6.8)


z=0 on Γ,


where u ∈ L2 (Ω), yuε ∈ V2 a corresponding solution of (6.1) and w ∈ L2 (Ω).

 15

3n
Proposition 6.4 Assume that A1 -A2 are fulfilled with n+2 ≤ α ≤ 2. Let u ∈ L2 (Ω) satisfying
(3.7) and let yuε ∈ V2 be the corresponding solution of (6.1). For w ∈ L2 (Ω), problem (6.8) admits
ε
a unique solution zuw in V2 . Moreover, the following estimates hold
 
ε
kDzuw kα ≤ Cα L kuk ν
2
kwk2 ,
ε
2ε kDzuw k2 ≤ C2 kwk2
 
k2 ≤ Cα2 L kuk
ε 2 2
2ε kDzuw ν
2
kwk2 ,

2−α
κ3 (1+t) α−1
where L(t) = 2(2−α) with κ̄ = κ1 κ2 κ3 .
ν−κ̄t(1+t) α−1

Proof. Let us first recall that a function z is a weak solution of (6.8) if

2ε (Dz, Dϕ) + (τ 0 (Dyuε ) : Dz, Dϕ) + b (z, yuε , ϕ) + b (yuε , z, ϕ) = (w, ϕ) for all ϕ ∈ V2 .

Consider then the bilinear form defined by

B(z1 , z2 ) = 2ε (Dz1 , Dz2 ) + (τ 0 (Dyuε ) : Dz1 , Dz2 ) + b (z1 , yuε , z2 ) + b (yuε , z1 , z2 ) .

Taking into account Lemma 2.6, we get

2
B(z, z) = 2ε kDzk2 + (τ 0 (Dyuε ) : Dz, Dz) + b (z, yuε , z) + b (yuε , z, z)
2
= 2ε kDzk2 + (τ 0 (Dyuε ) : Dz, Dz) + b (z, yuε , z)

for every z ∈ V2 . On the other hand, by using assumption A2 , Lemma 2.9, estimate (6.5) and
arguing as in (3.10), we deduce that
Z

α −1
(τ 0 (Dyuε ) : Dz, Dz) ≥ ν 1 + |Dyuε |2 2 |Dz|2 dx
Ω
νkDzk2α νkDzk2α
≥ 2−α ≥   2−α .
ε kα kuk
( |Ω|+kDyu α ) α
κ3 1+ ν 2 α−1

Moreover, due to Lemma 2.5 and estimate (6.4), we have

  2−α
2 kuk2 α−1 kuk2 2
|b (z, yuε , z)| ≤ κ1 kDyuε kα kDzkα ≤ κ1 κ2 1 + ν ν kDzkα .

Therefore, we obtain
!
  2−α
2 kuk2 α−1 kuk2
B(z, z) ≥ 2ε kDzk2 + 
ν
 2−α − κ1 κ2 1 + ν ν kDzk2α (6.9)
kuk
κ3 1+ ν 2 α−1

which shows that B is coercive on V2 since u satisfies (3.7). Let us now prove that B is continuous.
Similarly, due to A1 , (6.4) and the fact that u satisfies condition (3.7), we have
Z

α−2
1
γ |(τ 0
(Dyu
ε
) : Dz1 , Dz 2 )| ≤ 1 + |Dyuε |2 α |Dz1 ||Dz2 | dx
Ω
Z
≤ |Dz1 ||Dz2 | dx ≤ kDz1 k2 kDz2 k2
Ω
 16

and
|b (z1 , yuε , z2 ) + b (yuε , z1 , z2 )| ≤ 2κ1 kDyuε kα kDz1 kα kDz2 kα
2−α
≤ 2κ1 |Ω| kDyuε kα kDz1 k2 kDz2 k2
α

2−α
  2−α
≤ 2κ1 κ2 |Ω| α 1 + kuk
α−1 kuk

ν kDz1 k2 kDz2 k2
2 2
ν
2−α
2|Ω|
ν α
≤   2−α kDz1 k2 kDz2 k2
kuk
κ3 1+ ν 2 α−1

≤ 2ν kDz1 k2 kDz2 k2

for every z1 , z2 ∈ V2 . Therefore,

B (z1 , z2 ) ≤ (2ε + γ + 2ν) kDz1 k2 kDz2 k2 .

The bilinear form B is then continuous and coercive on V2 . Applying the Lax-Milgram theorem,
ε
we deduce that problem (6.8) admits a unique solution zuw in V2 . Taking into account (6.9), we
obtain
!
  2−α
ν kuk2 α−1 kuk2 ε
  2−α − κ1 κ2 1 + ν ν kDzuw k2α ≤ B (zuwε ε
, zuw ε
) = (w, zuw )
kuk2 α−1
κ3 1+ ν

ε
≤ Cα kwk2 kDzuw kα .

This gives the first estimate which combined with (6.9) imply
 ε
 C2 kwk2 kDzuw k2
ε 2 ε ε
2εkDzuw k2 ≤ B (zuw , zuw ) ≤ 
kuk2

2
 Cα kwk kDzuw ε 2
2 k α ≤ Cα L ν kwk2

and we derive the second and third estimate. 

6.4 Analysis of the control-to-state mapping

In order to study the local Lipschitz continuity of the approximate state with respect to the
control, we first establish some useful estimates.
3n
Lemma 6.5 Assume that A1 -A2 are fulfilled with n+2 ≤ α ≤ 2. Let u1 , u2 be in L2 (Ω) and let
yuε 1 and yuε 2 be two corresponding solutions of (6.1). Then the following estimate holds
2 2
2ε D yuε 1 − yuε 2 ν
D yuε 1 − yuε 2



2
+   2−α α
ku1 k2 ku2 k2 α−1
κ3 1+ 2ν + 2ν

  2−α
ku2 k2 α−1 ku2 k2 2
D yuε 1 − yuε 2 + C2 ku1 − u2 k2 D yuε 1 − yuε 2



≤ κ1 κ2 1 + ν ν α 2
.

Proof. The arguments are very similar to those used in the proof of Theorem 3.3. For the
confort of the reader, we will give the principal ideas. Setting ϕ = yuε 1 − yuε 2 in the corresponding
weak formulation and taking into account Lemma 2.8 and Lemma 2.6, we obtain
2
2 ν kD(yu ε
−yuε
)k
2ε D yuε 1 − yuε 2

1 2
2
+ 
α
 2−α
α α
|Ω|+kDyuε
1
k + kDy u2 k
ε α
α α
 17

2
≤ 2ε D yuε 1 − yuε 2 + τ (Dyuε 1 ) − τ (Dyuε 2 ), D yuε 1 − yuε 2




2

= −b yuε 1 − yuε 2 , yuε 2 , yuε 1 − yuε 2 + u1 − u2 , yuε 1 − yuε 2 .

(6.10)
Lemma 2.5 and estimate (3.1) then yield

2
b yuε 1 − yuε 2 , yuε 2 , yuε 1 − yuε 2 ≤ κ1 D yuε 1 − yuε 2 α Dyuε 2 α

  2−α
2
≤ κ1 κ2 1 + kuν2 k2
α−1 ku k
D yuε 1 − yuε 2
2 2


ν α
. (6.11)

On the other hand, by taking into account estimate (3.2), we obtain

2−α
2−α
  2−α α
 α−1  2−α α  α
 α−1
α α
ku k ku k
|Ω| + Dyuε 1 α
+ Dyuε 2 α
α
≤ 3|Ω| + 2 2 Cα ν1 2 + 2 2 Cα ν2 2
  2−α
ku1 k2 ku2 k2 α−1
≤ κ3 1 + 2ν + 2ν . (6.12)

The conclusion follows by combining (6.10), (6.11), (6.12) and by taking into account Lemma
2.4. 
A direct consequence of the previous lemma is that the control-to-state mapping G : u 7−→ yuε
ε

is locally Lipschitz continuous in V2 if the following condition holds

  2−α   2−α
ku1 k2 ku2 k2 α−1 ku2 k2 α−1 ku2 k2
κ̄ 1 + 2ν + 2ν 1+ ν ν2 <1

with κ̄ = κ1 κ2 κ3 . Unlike the Navier-Stokes case or the shear-thickening case, where the corre-
sponding conditions only involve u2 (see [3]), the previous sufficient constraint is quite restrictive
since it requires both controls u1 and u2 to be sufficiently ”small”. It holds if both controls
u1 and u2 satisfies (3.7) and guarantees the Lipchitz continuity only for restricted admissible
controls.
To overcome this difficulty, we impose condition (3.7) only on u2 and refine the analysis of the
result obtained in Lemma 6.5.
3n
Lemma 6.6 Assume that A1 -A2 are fulfilled with n+2 ≤ α ≤ 2. Let u1 , u2 be in L2 (Ω) and
ε ε
let yu1 and yu2 be two corresponding solutions of (6.1). If u2 satisfies condition (3.7), then the
following estimate holds
 2−α
 α−1
max(ku1 k2 −ku2 k2 ,0)
2ε D yuε 1 − yuε 2 D yuε 1 − yuε 2



2
≤ νsα 2ν 2
+ C2 ku1 − u2 k2 (6.13)

with sα = 0 if α = 2 and sα = 1 if α < 2.

Proof. Let us first notice that if ku1 k2 ≤ ku2 k2 , then
  2−α  2−α
 α−1
ku1 k2 ku2 k2 α−1 ku2 k2
1+ 2ν + 2ν ≤ 1+ ν .

Due to Lemma 6.5, it follows that

!
  2−α
2 ku2 k2 α−1 ku2 k2 2
yuε 1 yuε 2 ν
D yuε 1 − yuε 2



2ε D − 2
+   2−α − κ1 κ2 1 + ν ν α
ku k
κ3 1+ ν2 2 α−1
 18

≤ C2 ku1 − u2 k2 D yuε 1 − yuε 2

2
which by taking into account the fact that u2 satisfies (3.7) gives (6.13). Similarly, observing
that 2−α
α−1 ∈ [0, 1] and that for x > y ≥ 0 we have
2−α 2−α 2−α
(1 + x) α−1 − (1 + y) α−1 ≤ sα (x − y) α−1
we deduce from Lemma 6.5 that if ku1 k2 > ku2 k2 then
!
  2−α
2 ku2 k2 α−1 ku2 k2 2
yuε 1 yuε 2 ν
D yuε 1 − yuε 2



2ε D − 2
+   2−α − κ1 κ2 1 + ν ν α
ku k
κ3 1+ ν2 2 α−1

2 2
ν kD (yu
ε
−yuε
)kα ν kD (yu
ε ε
−yu )kα

≤ 
1 2
 2−α − 
1 2
 2−α + C2 ku1 − u2 k2 D yuε 1 − yuε 2 2
ku2 k2 α−1 ku1 k2 ku2 k2 α−1
κ3 1+ ν κ3 1+ 2ν + 2ν

  2−α
νsα ku1 k2 −ku2 k2 α−1
2

≤ κ3 2ν D yuε 1 − yuε 2 α
+ C2 ku1 − u2 k2 D yuε 1 − yuε 2 2
2−α  2−α
 α−1
νsα |Ω| α ku1 k2 −ku2 k2
2

≤ κ3 2ν D yuε 1 − yuε 2 2
+ C2 ku1 − u2 k2 D yuε 1 − yuε 2 2
 2−α
 α−1
ku1 k2 −ku2 k2
2

≤ νsα 2ν D yuε 1 − yuε 2 2
+ C2 ku1 − u2 k2 D yuε 1 − yuε 2 2

which gives the claimed result. 

Remark 6.7 From Lemma 6.6, we deduce that the Lipschitz continuity of Gε in V2 holds if u1
and u2 satisfy
  2−α
νsα max(ku1 k2ν
2 −ku2 k2 ,0)
α−1
< 2ε
with u2 satisfying condition (3.7). The previous inequality is obviously valid if α = 2 or if
ku1 k2 ≤ ku1 k2 , but also if the difference ku1 k2 − ku2 k2 can be ”controlled”.
Let u, v be in Uad with u satisfying condition (3.7) and let ρ in ]0, 1[. Set uρ = u + ρ(v − u), and
let yuε ρ be a solution of (6.1) corresponding to uρ and yuε be the solution of (6.1) corresponding
to u. In the remaining part of this section, and in order to simplify the notation, we set y ε and
y ε −y ε
yρε instead of yuε and yuε ρ and zρε = ρ ρ .
Lemma 6.8 For every ρ such that
  α−1
2−α
ν ε
ρ< U νsα , (6.14)

the following estimate holds

ε Dzρε 2
≤ C2 ku − vk2 .
Proof. First recall that since u e v belong to Uad , then kuk2 ≤ U and kvk2 ≤ U . Due to Lemma
6.6, we have
  2−α
max(kuρ k2 −kuk2 ,0) α−1
2ε Dzρε 2 ≤ νsα 2ν Dzρε 2 + C2 kv − uk2
  2−α
kuρ −uk2 α−1
≤ νsα 2ν Dzρε 2
+ C2 kv − uk2
  2−α 2−α
kv−uk2 α−1
≤ νsα 2ν ρ α−1 Dzρε 2
+ C2 kv − uk2

U

2−α 2−α
≤ νsα ν
α−1
ρ α−1 Dzρε 2
+ C2 kv − uk2 .
 19

Choosing ρ such that

U

2−α
α−1
2−α
νsα ν ρ α−1 < ε
we obtain the claimed result. 
Remark 6.9 Notice that in the case α = 2, condition (6.14) holds for every ρ ∈]0, 1[. Moreover,
a direct consequence of Lemma 6.8 is that (yρε )ρ>0 strongly converges to y ε in H01 (Ω).


Lemma 6.10 If zρεk k weakly converges to z ε in V2 for some sequence (ρk )k converging to zero,
then
lim ρ1k b yρεk , yρεk , ϕ − b (y ε , y ε , ϕ) = b (z ε , y ε , ϕ) + b (y ε , z ε , ϕ)



k→+∞

lim ρ1k τ (Dyρεk ) − τ (Dy ε ), Dϕ = (τ 0 (Dy ε ) : Dz ε , Dϕ)

k→+∞
for all ϕ ∈ V.
Proof. Since (ρk )k converges to zero, we can assume without loss of generality that ρk satisfies
condition (6.14). The rest of the proof is split into two steps.
Step 1. Notice that
1 ε ε

ε ε

ε ε ε ε
ρk b yρk , yρk , ϕ − b (y , y , ϕ) − (b (z , y , ϕ) + b (y , z , ϕ))



= b zρεk , yρεk , ϕ + b y ε , zρεk , ϕ − (b (z ε , y ε , ϕ) + b (y ε , z ε , ϕ))



≤ b zρεk , yρεk , ϕ − b (z ε , y ε , ϕ) + b y ε , zρεk , ϕ − b (y ε , z ε , ϕ)




≤ b zρεk − z ε , yρεk , ϕ + b z ε , yρεk − y ε , ϕ + b y ε , zρεk − z ε , ϕ



≤ zρεk − z ε 4 ∇yρεk 2 kϕk4 + kz ε k4 ∇ yρεk − y ε 2 kϕk4 + b y ε , zρεk − z ε , ϕ


The result is then a consequence of the strong convergence of yρεk k to y ε in H01 (Ω), the weak


convergence of zρεk k to z ε in H01 (Ω) and its strong convergence in L4 (Ω).
Step 2. To prove the second result, fix ϕ ∈ V and use the mean value theorem to get
Z Z 1
1 ε ε
τ 0 (Dy ε (x) + θD(yρεk − y ε )(x)) : Dzρεk (x) : Dϕ(x) dθ dx


ρk τ (Dyρk ) − τ (Dy ), Dϕ =
Ω 0
= τ 0 σρϕ,ε : Dzρεk , Dϕ = Dϕ : τ 0 σρϕ,ε , Dzρεk ,





k k
(6.15)


where σρϕ,ε
k
(x) = θρϕ,ε
k
(x) Dyρεk (x) − Dy ε (x) + Dy ε (x) with 0 < θρϕ,ε k
(x) < 1 being a number
(depending on ϕ(x)) arising when applying the mean values theorem to the integral in the
interval [0, 1]. Convergence of (σρεk )k to Dy ε in L2 (Ω) and continuity of τ 0 imply that
Dϕ : τ 0 (σρεk ) −→ Dϕ : τ 0 (Dy ε ) a.e. in Ω.
On the other hand, due to A1 , for all x ∈ Ω and i, j, m, ` = 1, · · · , n

α−2
τ 0 (σρεk )(x) ijm` ≤ γ 1 + |σρεk (x)|2 2 ≤ γ

and thus X
Dϕ(x) : τ 0 (σρεk )(x) m` ≤ γ


|Dij ϕ(x)| ≤ nγ|Dϕ|.
i,j

Due to the dominated convergence theorem, we deduce that Dϕ : τ 0 (σρεk ) k strongly converges

to Dϕ : τ 0 (Dy ε ) in L2 (Ω), which together with the weak convergence of (Dzρεk )k to Dz ε in L2 (Ω)
prove the claimed result. 
 20

Proposition 6.11 If zρεk k weakly converges to z ε in V2 for some sequence (ρk )k converging to
zero, then z ε is
the unique solution of the linearized problem (6.8) corresponding to (y ε , v − u).
Moreover, zρεk k converges strongly to z ε in V2 .
Proof. Since (ρk )k converges to zero, we can assume without loss of generality that ρk satisfies
condition (6.14). The first assertion of the proposition is a direct consequence of Lemma 6.10
together with the density of V in V2 . To prove the strong convergence, let us set
M ε (x) = 2εI + τ 0 (Dy ε (x)), Mρε (x) = 2εI + τ 0 (σρε (x)),
where I denotes the identity matrix n × n and where σρε is defined as in (6.15) with ϕ substituted
by yρε − y ε . Due to A2 , the matrices
M ε (x)+(M ε (x))T Mρε (x)+(Mρε (x))T
M ε,S (x) = 2 , Mρε,S (x) = 2 ,
are symmetric and positive definite. Applying the Cholesky method to M (x) and Mρε,S (x), ε,S

we deduce the existence of lower triangular matrices Lε (x) and Lερ (x) such that
M ε,S (x) = Lε (x)(Lε (x))T and Mρε,S (x) = Lερ (x)(Lερ (x))T .
Substituing in the weak formulation of (6.1), we obtain
2ε D(yρε − y ε ), Dϕ + τ (Dyρε ) − τ (Dy ε ), Dϕ = b (y ε , y ε , ϕ) − b yρε , yρε , ϕ + ρ(v − u, ϕ)

for all ϕ ∈ V2 . Therefore, taking into account (6.15), Lemma 2.6, Lemma 2.5, Lemma 2.4,
Lemma 6.8 and estimate (6.6), we have
2
(Lερk )T Dzρεk 2 = Mρεk : Dzρεk , Dzρεk

= −b y ε , zρεk , zρεk − b zρεk , yρεk , zρεk + v − u, zρεk

= −b zρεk , yρεk , zρεk + v − u, zρεk

2
≤ κ1 Dyρεk 2
Dzρεk 2
+ C2 kv − uk2 Dzρεk 2
C2 2 C22 2
≤ ε2 κ2ε 1 C2 U κ 1 C2



2 kuρk k2 + 1 kv − uk2 ≤ ε 2ε2 + 1 kv − uk2 (6.16)
and the sequence ((Lερk )T Dzρεk )k is then bounded in L2 (Ω). On the other hand, due to A1 we
have √
2
Lερk (x)
= Mρεk (x) ≤ 2 nε + C(γ, n) for all x ∈ Ω.
Taking into account the convergence of Dyρk k to Dy into L (Ω) and the continuity of τ 0 , we
ε

ε 2

deduce that Mρεk (x) converges to M ε (x) and, consequently, Lερk (x) converges to Lε (x) for a.e.
x ∈ Ω. The dominated convergence theorem then implies
Lερk −→ Lε strongly in L2 (Ω) (6.17)
which together with the weak convergence of (zρεk )k
to z in V2 , guarantees that ε
(Lερk )T Dzρεk
ε T ε 2
weakly converges to (L ) Dz in L (Ω). Moreover, taking into account (6.16), we deduce that
2 2
(Lε )T Dz ε 2
≤ lim inf (Lερk )T Dzρεk 2
k
2
≤ lim sup (Lερk )T Dzρεk = lim sup Mρεk : Dzρεk , Dzρεk


2
k k

zρεk , yρεk , zρεk + v − u, zρεk

= lim sup −b
k

= −b (z ε , y ε , z ε ) + (v − u, z ε ) = −b (y ε , z ε , z ε ) − b (z ε , y ε , z ε ) + (v − u, z ε )
2
= (M ε : Dz ε , Dz ε ) = (Lε )T Dz ε 2
.
 21

Weak convergence together with norm convergence implies strong convergence of ((Lερk )T Dzρεk )k
to (Lε )T Dz ε in L2 (Ω). There then exist a subsequence, still indexed by ρk , and a function
hε ∈ L2 (Ω) such that
(Lερk (x))T Dzρεk (x) ≤ hε (x) for a.e. x ∈ Ω and k > k0 ,

(Lερk (x))T Dzρεk (x) −→ (Lε (x))T Dz ε (x) for a.e. x ∈ Ω.

Therefore, taking into account A2 , we obtain
2
2ε Dzρεk (x) ≤ Mρεk (x) : Dzρεk (x) : Dzρεk (x) = (Dzρεk (x))T Mρεk (x)Dzρεk (x)
2
= (Lερk )T (x)Dzρεk (x) ≤ (hε (x))2 for a.e. x ∈ Ω and k > k0 .

−1
−1
Since (6.17) implies that (Lερk (x))T converges to (Lε (x))T for a.e. x ∈ Ω, we deduce
that
−1

Dzρεk (x) = (Lερk (x))T (Lερk (x))T Dzρεk (x)

−1

−→ (Lε (x))T (Lε (x))T Dz ε (x) = Dz ε (x)
for a.e. x ∈ Ω. The conclusion follows by applying the dominated convergence theorem. 

6.5 Proof of the approximate optimality conditions

Let us now prove Theorem 6.1. Existence of an optimal solution (ūε , ȳ ε ) for problem (Pαε ) can
be established arguing as in the proof of Theorem 4.1. Assume in the rest of the proof that ūε
satisfies condition (3.7). For ρ ∈]0, 1[ satisfying (6.14) and v ∈ Uad , let uερ = ūε + ρ(v − ūε ),
y ε −ȳ ε

yρε = yuε ερ and zρε = ρ ρ . Due to Lemma 6.8, we deduce that zρε ρ is bounded in V2 . There



then exist a subsequence zρεk k and z ε ∈ V2 such that zρεk k weakly converges to z ε in V2 . Due
to Proposition 6.11, z ε is the unique solution of


 −ε∆z − ∇ · (τ 0 (Dȳ ε ) : Dz) + z · ∇ȳ ε + ȳ ε · ∇z + ∇π = v − ūε in Ω,

∇·z =0 in Ω,


z=0 on Γ,



and zρεk k strongly converges to z ε in V2 . Therefore
I(uερ ,yρε )−I(ūε ,ȳ ε )
lim k k
ρk = (z ε , ȳ ε − yd ) + ((λ + 1)ūε − ū, v − ūε ) . (6.18)
k→+∞

On the other hand, since (uερk , yρεk ) is admissible for (Pαε ) and (ūε , ȳ ε ) is an optimal solution, we
deduce that
I(uερ ,yρε )−I(ūε ,ȳ ε )
lim k k
ρk ≥0 for all v ∈ Uad . (6.19)
k→+∞
ε
Let p̄ ∈ V2 be the unique solution of (6.2) (existence and uniqueness of a solution can be obtained
with arguments similar to those used in the proof of Proposition 6.4). Setting φ = z ε and taking
into account the weak formulation of problem (6.8), we obtain
(ȳ ε − yd , z ε ) = (2εDp̄ε , Dz ε ) + (τ 0 (Dȳ ε ) : Dz ε , Dp̄ε ) + (∇ȳ ε )T p̄ε − ȳ ε · ∇p̄ε , z ε

= (2εDz ε , Dp̄ε ) + (τ 0 (Dȳ ε ) : Dz ε , Dp̄ε ) + (z ε · ∇ȳ ε , p̄ε ) + (ȳ ε · ∇z ε , p̄ε )

= (v − ūε , p̄ε ) . (6.20)
The result follows by combining (6.18), (6.19) and (6.20). 
 22

7 Proof of the necessary optimality conditions

We begin this section by establishing a useful convergence result.
3n


Proposition 7.1 Assume that A1 -A2 are fulfilled with n+2 ≤ α ≤ 2. Let vε , yvεε be an
admissible pair for (Pαε ). There exists a subsequence (εk )k converging to zero and (u, y), an
admissible pair for (Pα ), such that

vεk −→ u weakly in L2 (Ω) and yvεεk −→ y strongly in W01,α (Ω).

k

Proof. To simplify the redaction, let us set yε = yvεε . Since (vε )ε is uniformly bounded in
L2 (Ω), estimate (6.5) and arguments similar to those used in αthe proof of Theorem 4.1 show
that (yε )ε and (τ (Dyε ))ε are uniformly bounded in Vα and L α−1 (Ω), respectively. αThere then
exists a subsequence (εk )k converging to zero and u ∈ L2 (Ω), y ∈ Vα and τe ∈ L α−1 (Ω) such
that (vεk )k weakly converges to u in L2 (Ω), (yεk )k weakly converges to y in Vα and (τ (Dyεk ))k
α √

weakly converges to τe in L α−1 (Ω). On the other hand, estimate (6.7) implies that εk Dyεk k
is bounded in L2 (Ω) and that for ϕ ∈ V2 , we have
√ √
|εk (Dyεk , Dϕ)| ≤ εk k εk Dyεk k2 kDϕk2 −→ 0 when εk → 0. (7.1)

Taking into account (7.1) and passing to the limit in the weak formulation corresponding to yεk ,
we deduce that
(e
τ , Dϕ) + b (y, y, ϕ) = (u, ϕ) for all ϕ ∈ V
and thus for all ϕ ∈ Vα . On the other hand, (2.2) gives

(τ (Dyεk ) − τ (Dϕ) , Dyεk − Dϕ) ≥ 0 for all ϕ ∈ Vα ,

and since
2
(τ (Dyεk ) , Dyεk ) = (vεk , yεk ) − 2εk kDyεk k2 ,
we obtain
2
(vεk , yεk ) − (τ (Dyεk ) , Dϕ) − (τ (Dϕ) , Dyεk − Dϕ) ≥ 2εk kDyεk k2 ≥ 0

for all ϕ ∈ Vα . By passing to the limit in the previous inequality, and arguing as in the proof of
Theorem 4.1, we can prove that

(e
τ , Dϕ) = (τ (Dy), Dϕ) for all ϕ ∈ Vα

and thus y is a solution of (1.1) corresponding to u. To prove the strong convergence of (yεk )k
to y in W01,α (Ω), notice that estimate (6.5) together with (6.12) yield
2
ν kD(yεk −y)k
(τ (Dyεk ) − τ (Dy) , D (yεk − y)) ≥ α
2−α
(|Ω|+kDyεk kαα +kDykαα ) α

2 2
ν kD(yεk −y)k ν kD(yεk −y)k
≥ 
α
 2−α ≥ 2−α
α
.
κ3 1+
kuk2
+
kvε k2
k
α−1 κ3 (1+ U
ν )
α−1
2ν 2ν

Therefore, by taking into account the previous convergence results, we deduce that
ν 2
2−α lim sup kD(yεk − y)kα
(
κ3 1+ U
ν ) α−1 k
 23

≤ lim sup (τ (Dyεk ) − τ (Dy) , D (yεk − y)) ≤ lim sup (τ (Dyεk ) , D (yεk − y))
k k

≤ lim sup ((vεk , yεk ) − (τ (Dyεk ) , Dy)) ≤ (u, y) − (e

τ , Dy) = 0
k

and the claimed result is proven. 

Next, we prove that the solutions of problems (Pαε ) form an approximating family for (Pα ).
3n
Proposition 7.2 Assume that A1 -A2 are fulfilled with n+2 ≤ α ≤ 2. Let (ūε , ȳ ε ) be a solution
ε
of (Pα ). There exists a subsequence (εk )k converging to zero such that

lim kūεk − ūk2 = 0, lim kȳ εk − ȳk1,α = 0, lim I(ūεk , ȳ εk ) = J(ū, ȳ).
k→+∞ k→+∞ k→+∞

Proof. Setting vε = ū for all ε > 0 and vε = ūε and applying Proposition 7.1, we deduce that
there exists a subsequence (εk )k such that (yūεk )k converges in W01,α (Ω) to ȳ (the unique solution
of (1.1) corresponding to ū), (ūεk )k weakly converges in L2 (Ω) to some u and (ȳ εk )k converges
in W01,α (Ω) to y (a solution of (1.1) corresponding to u). Using the lower semicontinuity of I
and the admissibility of (ū, yūεk ) for (Pαεk ), we obtain
2 2 2
1
2 ky − yd k2 + λ
2 kuk2 + 1
2 ku − ūk2 ≤ lim inf I(ūεk , ȳ εk )
k

≤ lim sup I(ūεk , ȳ εk )

k
2 2
≤ lim I(ū, yūεk ) = 1
2 kȳ − yd k2 + λ
2 kūk2
k

and consequently
1 2
J(u, y) + 2 ku − ūk2 ≤ J(ū, ȳ).
Since (ū, ȳ) is solution of (Pα ), we have J(ū, ȳ) ≤ J(u, y) and thus u = ū. Recalling that ū
satisfies condition (3.7), we deduce that y = ȳ and thus

lim I(ūεk , ȳ εk ) = J(ū, ȳ).

k→+∞

Finally, observing that

 
2 2
1
2 lim sup kūεk − ūk2 = lim sup I(ūεk , ȳ εk ) − 21 kȳ εk − yd k22 − λ
2 kūεk k2
k k
≤ J(ū, ȳ) − 12 kȳ − yd k22 − λ
2 lim inf kūεk k22
k
2 2
= λ
2 kūk2 − λ
2 lim inf kūεk k2 ≤0
k

we conclude that (ūεk )k converges to ū strongly in L2 (Ω). 

Proof of Theorem 5.1. Let (ūεk , ȳ εk ) be the solution of (Pαεk ) given in Proposition 7.2. Since ū
satisfies condition (3.7), we deduce that there exists k1 ∈ IN such that ūεk also satisfies condition
(3.7) for every k > k1 . The rest of the proof is split into two steps.
Step 1. Let us first prove (5.1) and (5.2). Due Theorem 6.1, there exists p̄εk ∈ V2 such that


 −εk ∆p − ∇ · (τ 0 (Dȳ εk ) : Dp) + (∇ȳ εk )T p − ȳ εk · ∇p + ∇π̃ εk = ȳ εk − yd ,

∇ · p = 0, (7.2)


p|Γ = 0,

 24

(p̄εk + (λ + 1)ūεk − ū, v − ūεk ) ≥ 0 for all v ∈ Uad . (7.3)

With arguments similar to those used in the proof of Proposition 6.4, we obtain the following
estimates  εk 
kDp̄εk kα ≤ Cα L kū ν k2 kȳ εk − yd k2 ≤ Cα L Uν kȳ εk − yd k2 ,


(7.4)
 εk 
2εk kDp̄εk k2 ≤ Cα2 L kū ν k2 kȳ εk − yd k2 ≤ Cα2 L Uν kȳ εk − yd k2 .
2 2
2

√
Therefore, (p̄εk )k is bounded in Vα and ( εk Dp̄εk )k is bounded in L2 (Ω). There then exist a
subsequence, still indexed by k, and p̄ such that (p̄εk )k weakly converges to p̄ in Vα and, by
using compactness results on Sobolev spaces, (p̄εk )k strongly converges to p̄ in Lq (Ω) for every
nα
q < n−α . Moreover, for ϕ ∈ V, we have
√ √
|εk (Dp̄εk , Dϕ)| ≤ εk k εk Dp̄εk k2 kDϕk2 −→ 0 when εk → 0. (7.5)

Similarly, we may prove that

(∇ȳ εk )T p̄εk − ȳ εk · ∇p̄εk , ϕ = (∇ȳ)T p̄ − ȳ · ∇p̄, ϕ .

lim (7.6)
k→+∞

Taking into account the convergence of (Dȳ εk )k to Dȳ in Lα (Ω) and the continuity of τ 0 , with
arguments similar to those used in the proof of Lemma 6.10, we deduce that

lim kτ 0 (Dȳ εk ) : Dϕ − τ 0 (Dȳ) : Dϕk α = 0.

k→+∞ α−1

This result together with the convergence of (Dp̄ε )k to Dp̄ in the weak topology of Lα (Ω) imply

lim (τ 0 (Dȳ εk ) : Dϕ, Dp̄εk ) = (τ 0 (Dȳ) : Dϕ, Dp̄) . (7.7)

k→+∞

Taking into account (7.5), (7.6) and (7.7), and passing to the limit in (7.2) and (7.3), we obtain

(τ 0 (Dȳ) : Dϕ, Dp̄) + (∇ȳ)T p̄ − ȳ · ∇p̄, ϕ = (ȳ − yd , ϕ) for all ϕ ∈ V,

and
(p̄ + λū, v − ū) ≥ 0 for all v ∈ Uad
which gives the claim result.
Step 2. Let us now prove (5.3). Set

M (x) = τ 0 (Dȳ(x)), M ε = τ 0 (Dȳ ε (x)),

and
M (x)+(M (x))T M ε (x)+(M ε (x))T
M S (x) = 2 , M ε,S (x) = 2 .
S ε,S
Due to A2 , the matrices M (x) and M (x) are symmetric and positive definite. Applying the
Cholesky method, we deduce the existence of lower triangular matrices L(x) and Lε (x) such that

M S (x) = L(x)(L(x))T and M ε,S (x) = Lε (x)(Lε (x))T .

 25

Therefore, setting ϕ = pεk in the weak formulation of (7.2) and due to lemma 2.5, Lemma 2.4
and estimates (6.5) and (7.4) , we obtain
2
(Lεk )T Dp̄εk 2
= (M εk : Dp̄εk , Dp̄εk )
2

= −2εk kDp̄εk k2 − (∇ȳ εk )T p̄εk − ȳ εk · ∇p̄εk , p̄εk + (ȳ εk − yd , p̄εk )
2
= −2εk kDp̄εk k2 − (p̄εk · ∇ȳ εk , p̄εk ) + (ȳ εk − yd , p̄εk )
≤ − (p̄εk · ∇ȳ εk , p̄εk ) + (ȳ εk − yd , p̄εk )
2
≤ κ1 kDp̄εk kα kDȳ εk kα + Cα kȳ εk − yd k2 kDp̄εk kα
  2−α εk

1 

≤ Cα2 L Uν + κ1 2 2 Cα kū ν k2 α−1 + |Ω| α L2 Uν kȳ εk − yd k2

1 2

  2−α

1 1

 2
≤ Cα2 L Uν + κ1 2 2 Cα Uν α−1 + |Ω| α L2 Uν kȳ εk − yd k2

and the sequence ((Lεk )T Dp̄εk )k is bounded in L2 (Ω). On the other hand, due to A1 we have
2
|Lεk (x)| = |M εk (x)| ≤ C(γ, n) for all x ∈ Ω.

Taking into account the convergence of (Dȳ εk )k to Dȳ in Lα (Ω) and the continuity of τ 0 , we
deduce that (M εk (x))k converges to M (x) and thus (Lεk (x))k converges to L(x) for a.e. x ∈ Ω.
The dominated convergence theorem then implies the convergence of (Lεk )k to L in Lq (Ω) for
α
every q > 1 and in particular for q = α−1 . Since (Dp̄εk )k weakly converges to Dp̄ in Lα (Ω), we
deduce that
(Lεk )T Dpεk −→ LT Dp weakly in L2 (Ω).
Therefore,
2 2
LT Dp̄ 2
≤ lim inf (Lεk )T Dp̄εk 2
k
2
≤ lim sup (Lεk )T Dp̄εk 2
= lim sup (M εk : Dp̄εk , Dp̄εk )
k k
εk
≤ lim sup − (p̄ · ∇ȳ , p̄ ) + (ȳ εk − yd , p̄εk )
εk εk
k

= − (p̄ · ∇ȳ, p̄) + (ȳ − yd , p̄) ,

and the claimed result is proven. 

References

[1] F. Abergel and E. Casas, Some optimal control problems of multistate equations appearing
in fluid mechanics, RAIRO Modél. Math. Anal. Numér. 27 (1993), 223-247.
[2] N. Arada, Distributed control for multistate modified Navier-Stokes equations, accepted for
publication in ESAIM:COCV.
[3] N. Arada, Optimal control of shear-thickening flows, Technical Report 3, Departamento de
Matemática, FCT-UNL, Portugal (2012).
[4] E. Casas, L.A. Fernández, Boundary control of quasilinear elliptic equations, Rapport de
Recherche 782, INRIA, 1988.
 26

[5] E. Casas, L.A. Fernández, Optimal control of quasilinear elliptic equations with nondiffer-
entiable coefficients at the origin, Rev. Mat. Univ. Complut. Madrid 4 (1991), 227-250.
[6] E. Casas, L.A. Fernández, Distributed control of systems governed by a general class of
quasilinear elliptic equations, J. Differential Equations 35 (1993), 20-47.
[7] C. V. Coffman, V. Duffin, V. J. Mizel, Positivity of weak solutions of non-uniformly elliptic
equations, Ann. Mat. Pura Appli. 104 (1975), 209-238.
[8] J. C. De los Reyes, Optimization of mixed variational inequalities arising in flow of vis-
coplastic materials, Comput. Optim. Appl. DOI:10.1007/s10589-011-9435-x
[9] J. C. De los Reyes, F. Tröltzsch, Optimal control of the stationary Navier-Stokes equations
with mixed control-state constraints, SIAM Journal on Control and Optimization, 46 (2007),
604-629.
[10] J. C. De los Reyes, R. Griesse, State-constrained optimal control of the three-dimensional
stationary Navier-Stokes equations, J. Math. Anal.Appl. 343 (2008), 257-272.
[11] J. Frehse, J. Málek, M. Steinhauer, An existence result for fluids with shear dependent
viscosity- Steady flows, Nonlin. Anal. 30 (1997), 3041-3049.
[12] G. P. Galdi, An introduction to the mathematical theory of the NavierStokes equations,
Vol. I and II, Springer Tracts in Natural Philosophy 38, 39, 2nd edition, Springer-Verlag,
New York, 1998.
[13] M. D. Gunzburger, L. Hou, T. P. Svobodny, Boundary velocity control of incompressible
flow with an application to viscous drag reduction, SIAM J. Control Optim. 30 (1992),
167-181.
[14] M. D. Gunzburger, O. A. Ladyzhenskaya, J. Peterson, On the global unique solvability
of initial-boundary value problems for the coupled modified Navier-Stokes and Maxwell
equations, J. Math. Fluid. Mech. 6 (2004), 462-482.
[15] M. D. Gunzburger, C. Trenchea, Analysis of an optimal control problem for the three-
dimensional coupled modified Navier-Stokes and Maxwell equations, J. Math. Anal. Appl.
333 (2007), 295-310.
[16] C. O. Horgan, Korn’s inequalities and their applications in continuum mechanics, SIAM
Rev. 37 (1995), 491-511.
[17] O. A. Ladyzhenskaya, New equations for the description of motion of viscous incompressible
fluids and solvability in the large of boundary value problems for them, Proc. Stek. Inst.
Math. 102 (1967), 95-118.
[18] O. A. Ladyzhenskaya, On some modifications of the Navier-Stokes equations for large gra-
dients of Velocity, Zap. Naučn. Sem. Leningrad. Otdel. Mat. Inst. Steklov (LOMI) 7 (1968),
126-154.
[19] O. A. Ladyzhenskaya, The mathematical theory of viscous incompressible flow, Gordon and
Beach, New York, 1969.
[20] J.-L. Lions, Quelques méthodes de résolution des problèmes aux limites non linéaires, Dunod,
Gauthier-Villars, Paris, 1969.
[21] P. Kaplický, Regularity of flows of a non-Newtonian fluid subject to Dirichlet boundary
conditions, Z. Anal. Anwendungen 24 (2005), 467-486.
 27

[22] P. Kaplický, J. Málek, J. Stará, C 1,α -solutions to a class of nonlinear fluids in two
dimensions-stationary Dirichlet problem, Zap. Nauchn. Sem. POMI 259 (1999), 89-121.
[23] K. Kunisch, X. Marduel, Optimal control of non-isothermal viscoelastic fluid flow. J. Non-
Newtonian Fluid Mech. 88 (2000), 261-301.
[24] M. K. V. Murthy, G. Stampacchia, Boundary value problems for some degenerate elliptic
operators, Ann. Mat. Pura Appli. 80 (1968) 1-122.
[25] J. Nečas, J. Málek, J. Rokyta, M. Ružička, Weak and measure-valued solutions to evolu-
tionary partial differential equations, Applied Mathematics and Mathematical Computation,
Vol. 13, Chapmann and Hall, London, 1996.
[26] T. Roubı́ček and F. Tröltzsch, Lipschitz stability of optimal controls for the steady-state
Navier-Stokes equations, Control Cybernet. 32 (2003), 683-705.
[27] T. Slawig, Distributed control for a class of non-Newtonian fluids, J. Differential Equations
219 (2005), 116-143.
[28] D. Wachsmuth, T. Roubı́ček, Optimal control of incompressible non-Newtonian fluids, Z.
Anal. Anwend 29 (2010), 351-376.

View publication stats