Optimal Control of Stationary Doubly Diffusive Flows On Two and Three Dimensional Bounded Lipschitz Domains: Numerical Analysis

See discussions, stats, and author profiles for this publication at: https://www.researchgate.
net/publication/379037320
OPTIMAL CONTROL OF STATIONARY DOUBLY DIFFUSIVE FLOWS ON TWO AND

THREE DIMENSIONAL BOUNDED LIPSCHITZ DOMAINS: NUMERICAL ANALYSIS
Preprint · March 2024
CITATIONS READS
0 76
3 authors, including:
Arbaz Khan Manil T. Mohan

Indian Institute of Technology Roorkee Indian Institute of Technology Roorkee
44 PUBLICATIONS 0 CITATIONS 160 PUBLICATIONS 0 CITATIONS
SEE PROFILE SEE PROFILE
All content following this page was uploaded by Manil T. Mohan on 18 March 2024.
The user has requested enhancement of the downloaded file.

OPTIMAL CONTROL OF STATIONARY DOUBLY DIFFUSIVE FLOWS ON TWO
AND THREE DIMENSIONAL BOUNDED LIPSCHITZ DOMAINS: NUMERICAL
ANALYSIS ∗
JAI TUSHAR† , ARBAZ KHAN‡ , AND MANIL T. MOHAN‡
Abstract. In this work, we propose fully nonconforming, locally exactly divergence-free discretizations based on
lowest order Crouziex-Raviart finite element and piecewise constant spaces to study the optimal control of stationary
double diffusion model presented in [Bürger, Méndez, Ruiz-Baier, SINUM (2019), 57:1318-1343]. The well-posedness of
the discrete uncontrolled state and adjoint equations are discussed using discrete lifting and fixed point arguments, and
arXiv:2403.10282v1 [math.NA] 15 Mar 2024
convergence results are derived rigorously under minimal regularity. Building upon our recent work [Tushar, Khan, Mohan,
arXiv (2023)], we prove the local optimality of a reference control using second-order sufficient optimality condition for the
control problem, and use it along with an optimize-then-discretize approach to prove optimal order a priori error estimates
for the control, state and adjoint variables upto the regularity of the solution. The optimal control is computed using a
primal-dual active set strategy as a semi-smooth Newton method and computational tests validate the predicted error decay
rates and illustrate the proposed scheme’s applicability to optimal control of thermohaline circulation problems.
Key words. Doubly diffusive flows, cross diffusion, optimal control, second order sufficient optimality conditions,
Mixed finite element method, a priori error estimation, semi-smooth Newton method
AMS subject classifications. 65N30, 65N50, 74F99, 74A50, 76S05.
1. Introduction. We consider the following bilaterally constrained flow control problem:

1 2 1 2 λ 2
(1.1) min J(u, y, U ) = ∥u − ud ∥0,Ω + ∥y − y d ∥0,Ω + ∥U ∥0,Ω ,
U ∈U ad 2 2 2
subject to
K −1 u + (u · ∇)u − div(ν(T )∇u) + ∇p = F (y) + U in Ω,
(1.2) div u = 0 in Ω,
− div(D∇y) + (u · ∇)y = 0 in Ω,
y = y D , u = 0 on Γ,
where u denotes the fluid velocity, p stands for the pressure field and y := (T, S)⊤ , S denotes the
concentration of a certain species within this fluid, and T denotes the temperature. The domain Ω ⊂
Rd (d = 2, 3) is bounded with Lipschitz boundary Γ. The set of admissible controls U ad is non-empty,
closed and convex:
U ad := {U = (U1 , . . . , Ud ) ∈ Ls (Ω) : Uaj (x) ≤ Uj (x) ≤ Ubj (x) a.e. in Ω, j = 1, . . . , d},
here the exponent s will be made precise later on and Uaj (x), Ubj (x) ∈ Ls (Ω) are given, satisfying
Uaj (x) < Ubj (x) for all j = 1, ..., d and a.e. x ∈ Ω. The desired states (ud , y d ) ∈ L2 (Ω) × [L2 (Ω)]2 , λ > 0
is the regularization parameter, ν denotes the temperature-dependent viscosity function, K(x) > 0 is
the permeability matrix, F (y) is a given function modelling buoyancy, and D is a 2 × 2 constant matrix
of the thermal conductivity and solutal diffusivity coefficients (possibly with cross-diffusion terms). We
make the following assumptions on the governing equation:
• Boundary data regularity, y D = (T D , S D )⊤ ∈ [H 1/2 (Γ)]2 .
• Lipschitz continuity and uniform boundedness of the kinematic viscosity,
|ν(T1 ) − ν(T2 )| ≤ γν |T1 − T2 | and ν1 ≤ ν(T ) ≤ ν2 ,
where γν , ν1 , ν2 are positive constants and | · | denotes the Euclidean norm in Rd .
∗
Funding: This work was supported by the SERB-CRG India (Grant Number : CRG/2021/002569).
† School
of Mathematics, Monash University, Melbourne, Australia ([email protected]).
‡ Department of Mathematics, Indian Institute of Technology, Roorkee, India ([email protected],
[email protected]).
1
2 J. TUSHAR, A. KHAN, AND M. T. MOHAN
• Lipschitz continuity of the buoyancy term, that is, there exist positive constants γF , CF such
that for all y 1 , y 2 , y ∈ R2
|F (y 1 ) − F (y 2 )| ≤ γF |y 1 − y 2 |, and |F (y)| ≤ CF |y|.
• K is a d × d permeability matrix which is assumed to be symmetric and uniformly positive
definite, hence, its inverse satisfies v ⊤ K −1 (x)v ≥ α1 |v|2 ∀ v ∈ Rd and x ∈ Ω for a constant
α1 > 0. The constant matrix D is assumed to be positive definite (though not necessarily
symmetric), that is, s⊤ Ds ≥ α2 |s|2 ∀ s ∈ R2 , for a constant α2 > 0.
1.1. Related works. With respect to the well-posedness of the governing equation (1.2) under
suitable assumptions, we first restrict the discussion to classical Bousinessq-type equations. For a
detailed literature review on their solvability analysis, we direct the interested readers to a detailed
review in Section 1.2 of [1, 2] and [3]. A diversity of numerical methods have been proposed in the
literature to simulate these equations, we mention some recent contributions and refer the interested
readers to references therein; stabilized finite elements using projection based techniques [4], mixed
formulations [5] and fully mixed H(div)-conforming formulations [6]. The main difference between
Bousinessq equations and the double-diffusive equations (1.2) is the vector-valued nature of y that diffuse,
and the solvability analysis presented in [1] is based on the restriction that the diffusion matrix D is
positive definite to maintain the coercivity of − div (D∇y) and under the assumption that the fluid
velocity u ∈ W 1,∞ (Ω) and y ∈ [L∞ (Ω)]2 , which is valid under the regularity assumptions (see Chapter
II, Section 1.3, Proposition 1.1 in [7]) on the domain and the data F (y) ∈ H 1 (Ω). However, in the
context of control problems this framework will not hold, since the data U is in a much weaker space.
Recently in [2] the authors have presented a solvability analysis to overcome this issue. A new
framework is presented using Faedo-Galerkin approximation techniques and minimal regularity results
on two and three dimensional bounded Lipschitz domains. The literature regarding optimal control
of doubly diffusive flows is scarce, in this direction we mention the applicative example explored in
[8] in which the author derives the optimality system using a Lagrangian approach of an unconstrained
optimal control problem governed by non-stationary double diffusion model and present some simulations.
Therefore, we restrict ourselves to the recent literature on optimal control of Boussinesq type equations.
A Neumann and Dirichlet boundary optimal control problem of the stationary Bousinessq equations is
studied by the authors in [9] and [10], respectively. They derive the optimality system using Lagrangian
multiplier techniques and prove the existence of optimal solutions. The optimal control of stationary
thermally convected fluid flows from a theoretical and numerical perspective is addressed in [11]. The
authors consider a vorticity type cost functional and choose thermal convection as a control mechanism,
then derive the necessary optimality condition for the constrained minimization problem and develop a
numerical method to solve it on cavity and channel type flows.
1.2. Main contributions and outline. To the best of the authors’ knowledge, no results are
available in the literature concerning the numerical analysis of optimal control of doubly diffusive flows.
In Section 2, we describe the functional setting of the proposed problem and recapitulate significant
results from [2] up to the formulation of first-order necessary optimality conditions and derivation of an
equivalent optimality (KKT ) system corresponding to the control problem (1.1)-(1.2). Expanding upon
these findings, in Section 3, we extend the methodology established in [12] to establish the local optimality
of the control using a second-order sufficient optimality condition to scenarios when Taylor expansion of
temperature-dependent viscosity parameter and the nonlinear buoyancy term do not terminate after the
second-order term with zero remainder (see Theorem 3.3). Moreover, thanks to the precise estimates of the
nonlinear state equation and the corresponding linear equations, we specify how small the residuals should
be to satisfy the second-order sufficient optimality condition (see Theorem 3.6). Section 4 is devoted to the
numerical analysis of the governing (or state) equation. We propose a fully non-conforming finite element
discretization of the uncontrolled state equation (1.2) based on the lowest order Crouziex-Raviart finite
element [13] and piecewise constant spaces in Section 4.1, which produces exactly divergence-free velocity
approximations and are of particular importance in ensuring that solutions remain locally conservative
as well as energy stable [14]. The well-posedness of the discrete state equation is studied using discrete
OPTIMAL CONTROL OF DOUBLY DIFFUSIVE FLOWS 3
lifting [15] and fixed point arguments in Theorems 4.6 and 4.7. In Section 4.1.4, we derive pressure robust
a priori error estimates under minimal regularity (see Theorems 2.5 and 4.8) for the uncontrolled state
equation. The numerical analysis of the adjoint equation (2.12) using the fully-nonconforming scheme is
studied in Section 5. In Section 6, we propose a piecewise constant discretization of control which leads to
a fully-discrete optimality system (see (6.2)-(6.4)) following the optimize-then-discretize approach. Then
we derive optimal a priori error estimates for the control, state and adjoint variables in their natural norms
(see Theorem 6.2) up to the regularity of the solutions. In section 7, we solve the optimal control problem
using a primal-dual active set strategy as a semi-smooth Newton method [16] and present results of two
different numerical experiments, namely, an accuracy test for a manufactured solution that confirms that
the experimentally observed orders of convergence are consistent with Theorem 6.2 in flow, Stokes and
Darcy regimes, and an illustration of optimal control of Soret and Dufour effects in a two-dimensional
porous cavity setup that shows that the state trajectories reach the desired state trajectories.
2. Functional setting and continuous formulation. Let Ω ⊂ Rd (d = 2, 3) be a bounded
domain with Lipschitz boundary Γ. We use the following notations throughout this article: The usual pth
1
integrable Lebesgue spaces are denoted by Lp (Ω) with norm ∥f ∥Lp (Ω) = Ω |f (x)|p dx p , for f ∈ Lp (Ω),
R
p ∈ [1, ∞). For p = ∞, L∞ (Ω) is the space of all essentially bounded Lebesgue measurable functions on
Ω with the norm ∥f ∥∞,Ω = ess supx∈Ω (|f (x)|), for f ∈ L∞ (Ω). For p = 2, we denote the norm by ∥·∥0,Ω
and (·, ·) represents the usual inner
product Rin L2 (Ω). The space of square integrable functions with zero
mean is defined as L0 (Ω) := w ∈ L2 (Ω) : Ω w dx = 0 . Sobolev spaces are denoted by the standard
2
1
notation W k,p (Ω) with the norm ∥f ∥W k,p (Ω) = α p p
, if k ∈ [1, ∞), for f ∈ W k,p (Ω).
P R
|α|≤k Ω |D f (x)|
For p = ∞ we denote the norm by ∥f ∥W k,∞ (Ω) = max|α|≤k ess supx∈Ω |Dα f (x)| , for f ∈ W k,∞ (Ω).
Moreover, for p = 2, weuse the notation W k,2 (Ω) = H k (Ω) with the corresponding norm denoted
by ∥·∥k,Ω and H 10 (Ω) := v ∈ H 1 (Ω) : v|Γ = 0 , where the vector valued functions in dimension d are
denoted by bold face and v|Γ = 0 is understood in the sense of trace. The dual space of H 10 (Ω) is denoted
by H −1 (Ω) with the following norm: ∥u∥−1,Ω := sup0̸=v∈H 10 (Ω) ⟨u,v⟩∥v∥ , where ⟨·, ·⟩ denotes the duality
1,Ω
pairing between H 10 (Ω) and H −1 (Ω).We will also need H(div; Ω) := {v ∈ L2 (Ω) : divv ∈ L2 (Ω)}, a
2 2 2
vector-valued Hilbert space. endowed with the norm ∥w∥ div ,Ω := ∥w∥0,Ω + ∥div w∥0,Ω . The outward
normal on Γ is denoted by nΓ . The duality product between a space V and its dual V ′ is denoted by
⟨f , v⟩, where f ∈ V ′ and v ∈ V . In the duality product ⟨f , v⟩, if f is replaced by a function U ∈ Lr (Ω),
′
then ⟨f, v⟩ = (U , v)r,r′ such that 1/r + 1/r′ = 1, for all v ∈ V ∩ Lr (Ω). Throughout the article C will
denote a generic positive constant. Standard Sobolev embeddings indicate that for 1 ≤ r′ < ∞ if d = 2
and r′ ∈ [1, 6] if d = 3, there exists Crd′ > 0 such that ∥w∥Lr′ (Ω) ≤ Crd′ ∥w∥1,Ω , for all w ∈ H 1 (Ω).
′
Thus v ∈ H 10 (Ω) implies v ∈ Lr (Ω) provided the above mentioned Sobolev embedding is satisfied. As a
consequence in the sequel, we fix the range of r as :
6
(2.1) 1 < r < ∞ if d = 2, or ≤ r ≤ 6 if d = 3.
5
.
2.1. Existence and uniqueness of the governing equation. The variational formulation of the
governing or the state equation (1.2) is obtained by testing against suitable functions and integrating by
parts, and can be formulated as follows: For some U ∈ Lr (Ω), find (u, p, y) ∈ H 10 (Ω) × L20 (Ω) × [H 1 (Ω)]2
satisfying y = y D on Γ and
a(y; u, v) + c(u, u, v) + b(v, p) = d(y, v) + (U , v)r,r′ ∀ v ∈ H 10 (Ω),

(2.2) b(u, q) = 0 ∀ q ∈ L20 (Ω),
ay (y, s) + cy (u, y, s) = 0 ∀ s ∈ [H01 (Ω)]2 .
The involved forms in (2.2) are defined as follows:
a(y; u, v) := (K −1 u, v) + (ν(y)∇u, ∇v), c(w, u, v) := ((w · ∇)u, v), b(v, q) := −(q, divv),
d(s, v) := (F (s), v), ay (y, s) := (D∇y, ∇s), cy (v, y, s) := ((v · ∇)y, s),
where, ν(y) is understood as the kinematic viscosity depending only on the first component of y.
Lemma 2.1 (Boundedness (cf. [2], Section 2.1)). For all u, v, w ∈ H 1 (Ω), q ∈ L2 (Ω),
and y, s ∈ [H 1 (Ω)]2 , there holds
|a(y; u, v)| ≤ Ca ∥u∥1,Ω ∥v∥1,Ω , |ay (y, s)| ≤ Ĉa ∥∇y∥0,Ω ∥∇s∥0,Ω ,
√
|b(v, q)| ≤ d ∥q∥0,Ω ∥∇v∥0,Ω , |d(y, v)| ≤ CF ∥y∥0,Ω ∥v∥0,Ω , |(U , v)r,r′ | ≤ ∥U ∥Lr (Ω) ∥v∥Lr′ (Ω) ,
|c(w, u, v)| ≤ C6d C3d ∥w∥1,Ω ∥∇u∥0,Ω ∥v∥1,Ω , |cy (w, y, s)| ≤ C6d C3d ∥w∥1,Ω ∥∇y∥0,Ω ∥s∥1,Ω .
Next, Poincaré’s inequality implies that the bilinear forms a(·; ·, ·) (for a fixed temperature) and ay (·, ·)
are coercive, that is,
2 2
(2.3) a(·; v, v) ≥ αa ∥v∥1,Ω ∀ v ∈ H 10 (Ω), and ay (y, y) ≥ α2 ∥∇y∥0,Ω ∀ y ∈ [H 1 (Ω)]2 .
Let X denote the kernel of b(·, ·), namely,
X := v ∈ H 10 (Ω) : b(v, q) = 0 ∀ q ∈ L20 (Ω) = v ∈ H 10 (Ω) : div v = 0 in Ω .

Then for all w ∈ X and v ∈ H 10 (Ω), the following properties of the trilinear form hold:
(2.4) c(w, v, v) = 0, c(w, u, v) = −c(w, v, u), and c(u, v, w) = ((∇v)⊤ w, u).
Furthermore, the following equivalence result holds (weak sense):

Lemma 2.2 (Equivalence (cf. [2], Lemma 3.2)). If (u, p, y) ∈ H 10 (Ω) × L20 (Ω) × [H 1 (Ω)]2 solves
(2.2) for all U ∈ Lr (Ω) under (2.1), then (u, y) ∈ X × [H 1 (Ω)]2 satisfies y|Γ = y D and
(
a(y; u, v) + c(u, u, v) − d(y, v) − (U , v)r,r′ = 0 ∀ v ∈ X,
(2.5)
ay (y, s) + cy (u, y, s) = 0 ∀ s ∈ [H01 (Ω)]2 .
Conversely, if (u, y) ∈ X ×[H 1 (Ω)]2 is a solution of the reduced problem (2.5), then for some U ∈ Lr (Ω),
there exists a p ∈ L20 (Ω) such that (u, p, y) is a solution of (2.2).
A lifting argument is used to deal with the non-homogeneous Dirichlet data appearing in the advection-
diffusion equations. We write y as y = y 0 + y 1 , where y 0 ∈ [H01 (Ω)]2 and y 1 is such that
(2.6) y 1 ∈ [H 1 (Ω)]2 with y 1 |Γ = y D .
2
Definition 2.3 (Weak solution). The pair (u, y) ∈ X × H 1 (Ω) is called a weak solution to the
2 2
system (2.5), if for y D ∈ H 1/2 (Γ) , U ∈ Lr (Ω) and (v, s) ∈ X × H01 (Ω) , (u, y) satisfies (2.5) with

y|Γ = y D .
Theorem 2.4 (Existence (cf. [2], Theorem 3.5)). Let the assumptions in section 1 on the governing
equation hold. For every U ∈ Lr (Ω), r satisfying (2.1) and every y D ∈ [H 1/2 (Γ)]2 there exists a lifting
y 1 ∈ [H 1 (Ω)]2 of satisfying (2.6) such that the problem (2.5) has a weak solution (u, y = y 0 + y 1 ) ∈
X × [H 1 (Ω)]2 . Furthermore, there exists positive constants Cu and Cy such that
(2.7) ∥u∥1,Ω ≤ Cu y D 1/2,Γ + ∥U ∥Lr (Ω) := Mu , ∥y∥1,Ω ≤ Cy y D 1/2,Γ + ∥U ∥Lr (Ω) := My .

2
Theorem 2.5 (Regularity of state (cf. [2], Theorem 3.8)). Let U ∈ Lr (Ω) and y D ∈ H 1+δ (Γ) be

given, for δ ∈ (0, 21 ). Then for any domain Ω of the class O2 (R2 ) or O3 (R3 ) (see [2, Section 3.2] for more
details), the weak solution to (1.2) satisfies (u, p, y) ∈ [H 10 (Ω) ∩ H 3/2+δ (Ω)] × [L20 (Ω) ∩ H 1/2+δ (Ω)] ×
[H 3/2+δ (Ω)]2 and
(2.8) ∥u∥3/2+δ,Ω + ∥y∥3/2+δ,Ω ≤ M,
where, M depends on the data ∥U ∥Lr (Ω) and ∥y D ∥1+δ,Γ (cf. [2, (36)]).
Theorem 2.6 (Uniqueness (cf. [2], Theorem 3.9)). Let U ∈ Lr (Ω) (see (2.1) for the range of r)
2
and y D ∈ H 1+δ (Γ) be given for δ ∈ (0, 12 ). Let (u, y) ∈ [X ∩ H 3/2+δ (Ω)] × [H 1 (Ω) ∩ H 3/2+δ (Ω)]2 be a

weak solution of the reduced problem (2.5), and Ω be the domain as defined in Theorem 2.5. Then under
γν Cp2d Cgn M My

γF My
(2.9) αa > C6d C3d + Mu + ,
α̂a α̂a
the solution of (2.5) is unique.
We will also need the following stability results in our subsequent analysis:
Lemma 2.7 (Stability (cf. [2], Lemma 3.10)). If U ∈ Lr (Ω) (see (2.1) for the range of r) and
2 2
y D ∈ H 1+δ (Γ) are given, and (ui , y i ) ∈ X ∩ H 3/2+δ (Ω) × H 1 (Ω) ∩ H 3/2+δ (Ω) satisfy (2.5), then

the following stability result also holds under (2.9):

∥u1 − u2 ∥1,Ω + ∥y 1 − y 2 ∥1,Ω ≤ C ∥U 1 − U 2 ∥Lr (Ω) ,
(2.10)
∥u1 − u2 ∥3/2+δ,Ω + ∥y 1 − y 2 ∥3/2+δ,Ω ≤ C ∥U 1 − U 2 ∥Lr (Ω) .
2.2. First order necessary optimality condition. Throughout this subsection, we assume that
F (·) and ν(·) are twice continuously Fréchet differentiable with bounded derivatives of order up to two.
In order to work with the L2 -neighbourhood of the reference control, we fix the following exponents (see
for reference [2, Remark 4.2] and [12]) r′ = 4, r = 4/3 and s = 2. Notice that for all U ∈ Ls (Ω), it holds
by the interpolation inequality that
2
(2.11) ∥U ∥Lr (Ω) ≤ ∥U ∥L1 (Ω) ∥U ∥Ls (Ω) .
Observe that (2.11) imposes further restrictions on r in (2.1), specifically 1 < r ≤ 2.
Definition 2.8 (Local optimal control). A control Ū ∈ U ad is called locally optimal in L2 (Ω),
if there exists a positive constant ρ such that J(ū, ȳ, Ū ) ≤ J(u, y, U ) holds for all U ∈ U ad with
Ū − U 0,Ω ≤ ρ, where ū, ȳ and u, y denote the states corresponding to Ū and U , respectively.
Theorem 2.9 (First-order necessary optimality condition (cf. [2], Theorem 4.5)). Let Ū be a locally
optimal control for (1.1) with associated states ū = u(Ū ), ȳ = y(Ū ). Then there exists a unique solution
2 2
(φ̄, η̄) ∈ X × H01 (Ω) of the adjoint equation for all (v, s) ∈ X × H01 (Ω) ,
(
a(ȳ, φ̄, v) − c(ū, φ̄, v) + c(v, ū, φ̄) + cy (v, ȳ, η̄) = (ū − ud , v),
(2.12)
ay (η̄, s) − cy (ū, η̄, s) + (νT (T̄ )∇ū : ∇φ̄, s) − ((F y (ȳ))⊤ φ̄, s) = (ȳ − y d , s),
s1
where (νT (T̄ )∇ū : ∇φ̄, s) = νT (T̄ )∇ 0
ū:∇φ̄
, s2 . Furthermore, the following variational inequality is
satisfied:

(2.13) λŪ + φ̄, U − Ū ≥ 0 ∀ U ∈ U ad .
Remark 2.10. Moreover, using a projection on the admissible control set U ad ,
(2.14) PU ad : U −→ U ad , PU ad (Uj (x)) = max(Uaj (x), min(Ubj (x), Uj (x))),
the variational inequality (2.13) can be expressed component-wise as follows for a.e. x ∈ Ω:

1 1
(2.15) Ūj (x) = PU ad − φ̄j (x) = max Uaj (x), min Ubj (x), − φ̄j (x) ,
λ λ
where the subscript j = 1, . . . , d (cf. [17]).
The solution (φ̄, η̄) of the adjoint equation (2.12) is called the adjoint state associated with (ū, ȳ)
and is the weak solution of the following adjoint PDE:

−1 ⊤ ⊤
K φ̄ + (∇ū) φ̄ − (ū · ∇)φ̄ − div(ν(T̄ )∇φ̄) + ∇ξ + (∇ȳ) η̄ = ū − ud in Ω,


div φ̄ = 0 in Ω,


(2.16) ⊤


 − div (D∇η̄) − (ū · ∇)η̄ − (F y (ȳ)) φ̄ + νT (T̄ )∇ū : ∇φ̄ = ȳ − y d in Ω,

 φ̄ = 0, η̄ = 0 on Γ.
Here, ξ is interpreted as the adjoint pressure which can be recovered using “De Rham’s Theorem” (see
[2, Remark 4.3]). The adjoint state satisfies the following estimates (cf. [2, Lemma 4.6]).
Lemma 2.11. The adjoint state (φ̄, η̄), given by (2.12), satisfies the following a priori estimate:

(2.17) ∥φ̄∥1,Ω ≤ Cφ ∥ū − ud ∥0,Ω + ∥ȳ − y d ∥0,Ω := Mφ ,

(2.18) ∥η̄∥1,Ω ≤ Cη ∥ū − ud ∥0,Ω + ∥ȳ − y d ∥0,Ω := Mη ,
where, Cφ and Cη are positive constants provided

CνT Cp2d Cgn M My

CFy My
(2.19) αa > C6d C3d + Mu + ,
α̂a α̂a
holds. Here ∥F y (ȳ)∥∞,Ω := CFy and ∥νT̄ ∥∞,Ω := CνT .
Remark 2.12. In summary, we have derived the following continuous optimality system for the
unknowns (ū, p, ȳ, φ̄, ξ, η̄, Ū ) ∈ H 10 (Ω) × L20 (Ω) × [H 1 (Ω)]2 × H 10 (Ω) × L20 (Ω) × [H01 (Ω)]2 × U ad satisfying
ȳ|Γ = y D , which can be used to determine the optimal control:
 1
a(ȳ; ū, v) + c(ū, ū, v) + b(v, p) = d(ȳ, v) + (Ū , v) ∀ v ∈ H 0 (Ω),

(2.20) b(ū, q) = 0 ∀ q ∈ L20 (Ω),

ay (ȳ, s) + cy (ū, ȳ, s) = 0 ∀ s ∈ [H01 (Ω)]2 ,

a(ȳ, φ̄, v) − c(ū, φ̄, v) + c(v, ū, φ̄) + b(v, ξ) + cy (v, ȳ, η̄) = (ū − ud , v) ∀ v ∈ H 10 (Ω),



(2.21) b(φ̄, q) = 0 ∀ q ∈ L20 (Ω),

ay (η̄, s) − cy (ū, η̄, s) + (νT (T̄ )∇ū : ∇φ̄, s) − ((F y (ȳ))⊤ φ̄, s) = (ȳ − y d , s) ∀ s ∈ [H01 (Ω)]2 ,

and

(2.22) λŪ + φ̄, U − Ū ≥ 0 ∀ U ∈ U ad .
Every solution (ū, p, ȳ, Ū ) to the optimal control problem (1.1)-(1.2) must, together with (φ̄, ξ, η̄), satisfy
this optimality system.
We will also need the following result on the regularity of adjoint variables in the subsequent analysis
to prove local optimality of the reference control using second order sufficient optimality conditions.
Theorem 2.13 (Regularity of adjoint state). Let U ∈ Lr (Ω), y D ∈ [H 1+δ (Γ)]2 , (u, p, y) ∈
[H 10 (Ω)
∩H 3/2+δ 2
(Ω)] × [L0 (Ω) ∩ H 1/2+δ
(Ω)] × [H 3/2+δ
(Ω)]2 , for δ ∈ (0, 1/2) be given. Then for any
2 2 3 3
domain Ω of the class O (R ) or O (R ) (see [2, Section 3.2] for more details), the weak solution to
(2.16) satisfies
(φ̄, ξ, η̄) ∈ [H 10 (Ω) ∩ H 3/2+δ (Ω)] × [L20 (Ω) ∩ H 1/2+δ (Ω)] × [H01 (Ω) ∩ H 3/2+δ (Ω)]2 ,
and
∥φ̄∥3/2+δ,Ω + ∥η̄∥3/2+δ,Ω ≤ C(ū, ȳ, y D , Ū , ud , y d ) =: M̄ .
Proof. Idea of the proof: One can rewrite the system (2.16) in Ω as


 −∆φ̄ + ∇ξ˜ =: F


 div φ = 0,
(2.23)


 − div(D∇η̄) = ȳ − y d + (ū · ∇)η̄ + (F y (ȳ))⊤ φ̄ − νT (T̄ )∇ū : ∇φ̄ =: f ,

φ̄ = 0, η̄ = 0 on Γ.

where,

1 νT (T̄ )∇T
F= ū − ud − K −1 φ̄ − (∇ū)⊤ φ̄ + (ū · ∇)φ̄ − (∇ȳ)⊤ η̄ + ∇ν(T̄ ) · ∇φ̄ − ξ ,
ν(T̄ ) ν(T )
and ξ˜ := ξ
ν(T ) − m, with m being the mean value of ξ
ν(T ) so that ξ˜ ∈ L20 (Ω).
Using ν1 ≤ ν(T̄ ) ≤ ν2 , if we can show that F ∈ H −1/2+δ (Ω), f ∈ [H −1/2+δ (Ω)]2 , then invoking the
regularity for the Stokes problem (see [18, Theorem 5.5 and Theorem 9.20] (or [2, Lemma 3.6])) and the
regularity for elliptic equations (see [19] or [2, Lemma 3.7]) on bounded Lipschitz domains, we can prove
the desired regularity for the adjoint variables.
The main difficulty lies in estimating the terms ∇ν(T̄ ) · ∇φ̄ and νT (T̄ )∇ū : ∇φ̄ in three dimensions,
which have similar nature and can be dealt analogously. We observe that in order to estimate them,
atleast H 3/2−δ (Ω) regularity is needed for φ, thus, we first prove the regularity in a less regular space
(φ̄, ξ, η̄) ∈ [H 10 (Ω) ∩ H 3/2−δ (Ω)] × [L20 (Ω) ∩ H 1/2−δ (Ω)] × [H01 (Ω) ∩ H 3/2−δ (Ω)]2 by showing that F ∈
H −1/2−δ (Ω) and f ∈ [H −1/2−δ (Ω)]2 . To do this, we borrow tools and techniques developed in [2, Theorem
3.8] and break the proof into two parts, dealing with two and three dimensions separately.
Part I: Two dimensions. Using the fractional form of the Gagliardo-Nirenberg (GN) inequality (cf.
[2, (12)]), we can deduce the following
C
∥(1/ν(T ))(∇ū)⊤ φ̄∥−1/2+δ,Ω ≤ ∥D−1/2+δ ((∇ū)⊤ φ̄)∥0,Ω ≤ C∥D1/2−δ (D−1/2+δ ((∇ū)⊤ φ̄))∥L4/(3−2δ) (Ω)
ν1
(2.24) ≤ C ∥ū∥W 1,p1 (Ω) ∥φ̄∥Lp2 (Ω) ,
where p11 + p12 = 3−2δ

4 . Choosing p2 = 2 implies p1 = 4/(1 − 2δ), now again an application of fractional
GN inequality yields
∥(1/ν(T ))(∇ū)⊤ φ̄∥−1/2+δ,Ω ≤ C ∥ū∥3/2+δ,Ω ∥φ̄∥0,Ω .
Analogously, ∥(∇ȳ)⊤ η̄∥−1/2+δ,Ω ≤ C ∥ȳ∥3/2+δ,Ω ∥η̄∥0,Ω . Other terms in F are bounded similarly to Part
I of Theorem 3.8 in [2] giving
∥(1/ν(T ))K −1 φ̄∥−1/2+δ,Ω ≤ C∥K −1 ∥L∞ (Ω) ∥φ̄∥1,Ω , ∥(1/ν(T ))ū − ud ∥−1/2+δ,Ω ≤ C∥ū − ud ∥0,Ω ,
∥(1/ν(T ))(ū · ∇)φ̄∥−1/2+δ,Ω ≤ C∥ū∥1,Ω ∥φ̄∥1,Ω , ∥(1/ν(T ))∇ν(T̄ ) · ∇φ̄∥ ≤ C∥ū∥1,Ω ∥φ̄∥3/2+δ,Ω .
Now an application of (2.7) and Theorem 2.5 results to
∥φ̄∥3/2+δ,Ω ≤ ∥F ∥−1/2+δ,Ω ≤ C(ū, ȳ, y D , Ū , ud , y d ).
Similarly it is not difficult see that the following also holds: ∥η̄∥3/2+δ,Ω ≤ C(ū, ȳ, y D , Ū , ud , y d ).
1 1 2+δ
Part II: Three dimensions. Analogous to (2.24), we can get the following with p1 + p2 = 3 ,
C
∥(1/ν(T ))(∇ū)⊤ φ̄∥−1/2−δ,Ω ≤ ∥D1/2+δ (D−1/2+δ ((∇ū)⊤ φ̄))∥L3/(2+δ) (Ω) ≤ C∥ū∥W 1,p1 (Ω) ∥φ̄∥Lp2 (Ω) .
ν1
Applying fractional GN after choosing p2 = 6 implies p1 = 6/(3 + 2δ) gives
∥(1/ν(T ))(∇ū)⊤ φ̄∥−1/2−δ,Ω ≤ C∥ū∥1−δ,Ω ∥φ̄∥1,Ω ≤ C∥ū∥1,Ω ∥φ̄∥1,Ω .
Analogously, ∥(∇ȳ)⊤ η̄∥−1/2−δ,Ω ≤ C ∥ȳ∥1,Ω ∥η̄∥1,Ω . To estimate ∥(1/ν(T ))∇ν(T̄ ) · ∇φ̄∥−1/2−δ,Ω , we first
consider 1
2(k+1) ≤ δ < 12 , and show that φ̄ ∈ H 3/2−kδ (Ω), for k = 1, 2, . . . . Then by using this regularity,
we prove that φ̄ ∈ H 3/2−(k−1)δ (Ω). We repeat this process (k −1) times to finally obtain φ̄ ∈ H 3/2−δ (Ω).
The details of this iterative technique can be found in Step 2 [2]. The rest of the bounds can be tackled
as in Step 1 of Part II and Step 2 of Part I in the proof of Theorem 3.8 in [2]. Now an application of
(2.7) and (2.8) results to
∥φ̄∥3/2−δ,Ω ≤ ∥F ∥−1/2−δ,Ω ≤ C(ū, ȳ, y D , Ū , ud , y d ).
Similarly the following also holds: ∥η̄∥3/2−δ,Ω ≤ C(ū, ȳ, y D , Ū , ud , y d ). With the regularity in a less
regular space in hand, we prove the desired regularity in three dimensions by first proving it in δa ∈ (0, 41 )
and then δb ∈ [ 41 , 12 ). Similar to (2.24), we have
∥(1/ν(T ))(∇ū)⊤ φ̄∥−1/2+δa ,Ω ≤ C∥ū∥W 1,p1 (Ω) ∥φ̄∥Lp2 (Ω) ,

1 1 2−δa
where, p1 + p2 = 3 . Choosing p2 = 6, p1 = 6/(3 − 2δ) and applying fractional GN gives
∥(1/ν(T ))(∇ū)⊤ φ̄∥−1/2+δa ,Ω ≤ C∥ū∥3/2+δa ,Ω ∥φ̄∥1,Ω .
Now the term ∥(1/ν(T ))∇ν(T̄ ) · ∇u∥−1/2+δa ,Ω is estimated as in [2, (55)] leading to the following bound:
∥(1/ν(T ))∇ν(T̄ ) · ∇ū∥−1/2+δa ,Ω ≤ C ∥ū∥3/2−δa ,Ω T̄ 3/2+2δa ,Ω

, where 2δa ∈ (0, 1/2).
From the above bounds analogous to previous steps, one can conclude that φ̄ ∈ H 3/2+δa (Ω). Using this
we can establish the following bound analogous to [2, (60)]:
∥(1/ν(T ))∇ν(T̄ ) · ∇ū∥−1/2+δb ,Ω ≤ C ∥ū∥3/2+δa ,Ω T̄ 3/2+(δb −δa ),Ω

, where (δb − δa ) ∈ (0, 1/2),
which is finite, leading to the desired regularity.

3. Second order sufficient optimality conditions. The sufficient conditions we present here are
inspired by the framework developed in [12], in which the authors prove the coercivity of the second
derivative of the Lagrangian for a subspace of all possible directions by using strongly active constraints.
Throughout this section, we assume that F (·) and ν(·) are thrice continuously Fréchet differentiable with
bounded derivatives of order up to three. Now, let us introduce the Lagrangian corresponding to the cost
functional (1.1) and the reduced state equation (2.5),
L : X × [H 1 (Ω)]2 × L2 (Ω) × X × [H01 (Ω)]2 −→ R such that

L(u, y, U , φ, η) = J(u, y, U ) − [a(y; u, φ) + c(u, u, φ) − d(y, φ) − (U , φ)]
(3.1) − [ay (y, η) + cy (u, y, η)] .
In the next Lemma, we discuss the Fréchet-differentiability of L.

Lemma 3.1. The Lagrangian L is thrice Fréchet-differentiable with respect to w = (u, y, U ) from
2
X ∩ H 3/2+δ (Ω) × H 3/2+δ (Ω) × L2 (Ω) to R. The second-order derivative at w̄ = (ū, ȳ, Ū ) fulfils

together with the adjoint state Λ̄ = (φ̄, η̄)
Lww (w̄, Λ̄) [(ζ 1 , µ1 , h1 ), (ζ 2 , µ2 , h2 )] = Luu (w̄, Λ̄) [ζ 1 , ζ 2 ] + Lyy (w̄, Λ̄) [µ1 , µ2 ] + LU U (w̄, Λ̄) [h1 , h2 ] ,
and
|Luu (w̄, Λ̄) [ζ 1 , ζ 2 ] | ≤ CL1 ∥ζ 1 ∥1,Ω ∥ζ 2 ∥1,Ω ,

|Lyy (w̄, Λ̄) [µ1 , µ2 ] | ≤ CL2 ∥µ1 ∥1,Ω ∥µ2 ∥1,Ω ,

|Lyyy (w̄, Λ̄)[µ1 , µ2 , µ3 ]| ≤ CL3 ∥µ1 ∥1,Ω ∥µ2 ∥1,Ω ∥µ3 ∥1,Ω ,
2
for all (ζ i , µi , hi ) ∈ X × H 1 (Ω) × L2 (Ω) with positive constants CLi independent of w̄, ζ i , µi . Here

the last inequality in three dimensions holds for δ ∈ [ 41 , 12 ), which can be further relaxed to include the
whole range of δ ∈ (0, 12 ) under reasonable assumptions (see Remark 3.5).
Proof. The first order derivatives with respect to u, y and U in the direction (ζ, µ, h) ∈ X ×
2
H (Ω) × L2 (Ω) are,
1
Lu (w̄, Λ̄)[ζ] = (ζ, ū − ud ) − a(ȳ; ζ, φ̄) − c(ζ, ū, φ̄) − c(ū, ζ, φ̄) − cy (ζ, ȳ, η̄),
Ly (w̄, Λ̄)[µ] = (µ, ȳ − y d ) − ay (µ, η̄) − cy (ū, µ, η̄) − ((νT (T̄ ))µT ∇ū, ∇φ̄) + ((F y (ȳ))µ, φ̄),
LU (w̄, Λ̄)[h] = λ(h, Ū ) + (h, φ̄),
2
where µ := (µT , µS ) and (ζ, µ, h) ∈ X × H 1 (Ω) × L2 (Ω). The mappings ū 7→ Lu and Ū 7→ LU are

affine and the linear parts are bounded due to Lemma 2.1 and hence continuous. On the other hand, the
mapping ȳ 7→ Ly is nonlinear. Using the regularity of ū and energy estimates of ū and φ̄ (see Lemmas
2.4 and 2.11), we have
(3.2) |((F y (ȳ))µ, v)| ≤ ∥(F y (ȳ))∥∞,Ω ∥µ∥0,Ω ∥v∥0,Ω ≤ CFy ∥µ∥1,Ω ∥v∥1,Ω ,
(3.3) |((νT (T̄ )µT ∇ū, ∇v))| ≤ CνT Cgn Cp2d M ∥µ∥1,Ω ∥v∥1,Ω .
Therefore, the mapping is bounded due to the bounds in Lemma 2.1, (3.2) and (3.3). Thus the mappings
are Fréchet-differentiable and twice Fréchet-differentiability of L is implied. The second order derivatives
with respect to u, y and U are,
(3.4) Luu (w̄, Λ̄)[ζ 1 , ζ 2 ] = (ζ 1 , ζ 2 ) − c(ζ 1 , ζ 2 , φ̄) − c(ζ 2 , ζ 1 , φ̄),

(3.5) Lyy (w̄, Λ̄)[µ1 , µ2 ] = (µ1 , µ2 ) − (νT T (T̄ )µT1 µT2 ∇ū, ∇φ̄) + (F yy (ȳ)(µ1 ⊗ µ2 ), φ̄),
(3.6) LU U (w̄, Λ̄)[h1 , h2 ] = λ(h1 , h2 ),
where µi := (µTi , µSi ) and the following estimates follow:
|Luu (w̄, Λ̄)[ζ 1 , ζ 2 ]| ≤ ∥ζ 1 ∥0,Ω ∥ζ 2 ∥0,Ω + 2 C6d C3d ∥ζ 2 ∥1,Ω ∥ζ 1 ∥1,Ω ∥φ̄∥1,Ω ,
≤ max {1, 2 C6d C3d Mφ } ∥ζ 1 ∥1,Ω ∥ζ 2 ∥1,Ω ,
|(F yy (ȳ)(µ1 ⊗ µ2 ), φ̄)| ≤ ∥Fyy (ȳ)∥∞,Ω ∥µ1 ∥L4 (Ω) ∥µ2 ∥L4 (Ω) ∥φ̄∥0,Ω
(3.7) ≤ CFyy C42d Mφ ∥µ1 ∥1,Ω ∥µ2 ∥1,Ω ,
|(νT T (T̄ )µT1 µT2 ∇ū, ∇φ̄)| ≤ ∥νT T (T̄ )∥∞,Ω ∥∇ū∥Lr1 (Ω) ∥µT1 ∥Lr2 (Ω) ∥µT2 ∥Lr3 (Ω) ∥∇φ̄∥Lr4 (Ω)
(
CνT T Cgn Cr2d Cr3d ∥ū∥H 3/2+δ (Ω) ∥µT1 ∥1,Ω ∥µT2 ∥1,Ω ∥φ̄∥1,Ω in 2D,
(3.8) ≤
CνT T ∥ū∥3/2+δ,Ω ∥µT1 ∥1,Ω ∥µT2 ∥1,Ω ∥φ̄∥3/2+δ,Ω in 3D,
where the last inequality is tackled by choosing r1 = r2 = r3 = 6 and r4 = 2 in two dimensions and
applying the fractional GN inequality (see [2, (12)]) with r1 = r4 = 3 and r2 = r3 = 6 in three dimensions.
Similar to the previous case, one can see that the mappings ū 7→ Luu , ȳ 7→ Lyy and Ū 7→ LU U are
twice Fréchet-differentiable and thrice Fréchet-differentiability of L is implied. The Lagrangian has third
order derivatives only with respect to y,
Lyyy (w̄, Λ̄)[µ1 , µ2 , µ3 ] = (F yyy (ȳ)(µ1 ⊗ µ2 ⊗ µ3 ), φ̄) − (νT T T (T̄ )µT1 µT2 µT3 ∇ū, ∇φ̄),
and the boundedness follows from Lemma 2.11 and,
|(F yyy (ȳ)(µ1 ⊗ µ2 ⊗ µ3 ), φ̄)| ≤ ∥Fyyy (ȳ)∥∞,Ω ∥µ3 ∥L6 (Ω) ∥µ2 ∥L6 (Ω) ∥µ1 ∥L6 (Ω) ∥φ̄∥0,Ω
(3.9) ≤ CFyyy Mφ ∥µ1 ∥1,Ω ∥µ2 ∥1,Ω ∥µ3 ∥1,Ω ,

|(νT T T (T̄ )µT1 µT2 µT3 ∇ū, ∇φ̄)| ≤ ∥νT T T (T̄ )∥∞,Ω ∥∇ū∥Lr1 (Ω) ∥µT1 ∥Lr2 (Ω) ∥µT2 ∥Lr3 (Ω) ∥µT3 ∥Lr4 (Ω) ∥∇φ̄∥Lr5 (Ω)
(
CνT T T Cgn ∥ū∥3/2+δ,Ω ∥µT1 ∥1,Ω ∥µT2 ∥1,Ω ∥µT3 ∥1,Ω ∥φ̄∥1,Ω in 2D,
(3.10) ≤
CνT T T Cgn ∥ū∥3/2+δ,Ω ∥µT1 ∥1,Ω ∥µT2 ∥1,Ω ∥µT3 ∥1,Ω ∥φ̄∥3/2+δ,Ω in 3D,
where the last inequality holds true for ri = 8 for i = 1, . . . , 4 and r5 = 2 in two dimensions. In three
dimensions, it can be bounded similar to (3.8) by choosing r1 = r5 = 4 and r2 = r3 = r4 = 6, for
δ ∈ [ 41 , 12 ).
Let w̄ = (ū, ȳ, Ū ) be a fixed admissible optimal triplet that satisfies the first order necessary
optimality conditions. We define the set of strongly active constraints
Definition 3.2 (Strongly active sets). For a fixed ε > 0 and all i = 1, . . . , n,

Ωε,i := x ∈ Ω : |λŪi (x) + φ̄i (x)| > ε ,
where, vi (x) denotes the ith component of a vector function v = (v1 , . . . , vd ) at x ∈ Ω. For U ∈ Lp (Ω)
and 1 ≤ p < ∞ we define the Lp -norm with respect to the set of positivity by
n
! p1
X p
∥U ∥Lp,Ωε := ∥Ui ∥Lp (Ωε,i ) .
i=1
Then for all U ∈ U ad , the following holds (see [12, Corollary 3.14] for a proof):
n Z
X
(3.11) λŪi (x) + φ̄i (x) Ui (x) − Ūi (x) ≥ ε U − Ū L1,Ωε
.
i=1 Ωε,i
We abbreviate [v, v] = [v]2 and [v, v, v] = [v]3 . Assume that the optimal triplet w̄ and the associated
adjoint pair Λ̄ satisfy the following coercivity assumption or the second-order sufficient optimality condition:


 ∃ ε > 0 and σ > 0 such that

 2 2



 Lww (w̄, Λ̄) [(ζ, µ, h)] ≥ σ ∥h∥Lr (Ω)

 2
holds for all (ζ, µ, h) ∈ X × H01 (Ω) × L2 (Ω) with





h = U − Ū ∀ U ∈ U ad , hi = 0 on Ωε,i for i = 1, . . . , n


(SSC)
and (ζ, µ) ∈ X × H 1 (Ω)2 , where µ := µT , µS , solves the linearized equation


 0
T

a(ȳ; ζ, v) + c(ζ, ū, v) + c(ū, ζ, v) + ((νT (T̄ ))µ ∇ū, ∇v) − ((F y (ȳ))µ, v) = (h, v),







 ay (µ, s) + cy (ū, µ, s) + cy (ζ, ȳ, s) = 0,
 2
∀ (v, s) ∈ X × H01 (Ω) .


Recall from Section 2.2 that in order to work with the L2 -neighbourhood of the reference control, we fix
the following exponents: r′ = 4, r = 4/3 and s = 2. Now we prove that (SSC) and first-order necessary
conditions together are sufficient for local optimality of w̄. In particular, we are able to show quadratic
growth of the cost functional with respect to Lr -norm in a Ls -neighbourhood of the reference control.
Theorem 3.3. Let w̄ = (ū, ȳ, Ū ) be the admissible triplet for our optimal control problem which
satisfies the first-order necessary optimality condition with corresponding adjoint state Λ̄ given in Theorem
2.9, such that Lemma 3.1 holds. Assume further that (SSC) is satisfied at w̄. Then there exists ϑ > 0
and ρ > 0 such that
2
J(w) ≥ J(w̄) + ϑ U − Ū Lr (Ω) ,
holds for all admissible triplets w = (u, y, U ) with U − Ū Ls (Ω)

≤ ρ.
Proof. Suppose w̄ = (ū, ȳ, Ū ) satisfies the assumptions of the theorem. Let w = (u, y, U ) be another
admissible triplet. Then we have
(3.12) J(w̄) = L(w̄, Λ̄) and J(w) = L(w, Λ̄).
Taylor-expansion of the Lagrangian around w̄ yields
(3.13) L(w, Λ̄) = L(w̄, Λ̄) + Lu (w̄, Λ̄)(u − ū) + Ly (w̄, Λ̄)(y − ȳ) + LU (w̄, Λ̄)(U − Ū )
1 2
+ Lww (w̄, Λ̄) [w − w̄] + RL .
2
Due the nature of nonlinearities that present themselves in the diffusion coefficient and the forcing terms,
we need to estimate the following higher order remainder term RL (see [20, Theorem 7.9-1]),
1 1
Z
3
RL := (1 − Θ)Lwww (w̄ + Θ(w − w̄), Λ̄) [w − w̄] dΘ.
3 0
Furthermore, we can verify that the first-order necessary conditions are equivalently expressed by
2
Lu (w̄, Λ̄)ζ = 0 ∀ ζ ∈ X, Ly (w̄, Λ̄)µ = 0 ∀ µ ∈ H01 (Ω) ,

(3.14)
LU (w̄, Λ̄)(U − Ū ) ≥ 0 ∀ U ∈ U ad ,
and are satisfied at w̄ with the adjoint state Λ̄. Using the above representation, the second and third
terms vanish in (3.13) and the fourth term is non-negative. Infact, using (3.11) we have the following
estimate for the fourth term in (3.13) on the subspace Ωε,i ,
(3.15) LU (w̄, Λ̄)(U − Ū ) ≥ ε U − Ū L1,Ωε

.
Next, we investigate the second derivative of L. Invoking (SSC) on the subspace Ωε,i , let us construct a
new admissible control Ũ ∈ Ls (Ω) for i = 1, ..., n by
(
Ūi (x) on Ωε,i
Ũi (x) =
Ui (x) on Ω\Ωε,i .
Then U − Ū = (U − Ũ ) + (Ũ − Ū ), such that h := Ũ − Ū = 0 on Ωε,i which is the assumption of (SSC).

2
The difference (ζ, µ) = (u − ū, y − ȳ) solves for all (v, s) ∈ X × H01 (Ω) the following equations,
(
a(ȳ; ζ, v) + c(ζ, ū, v) + c(ū, ζ, v) + ((νT (T̄ ))µT ∇ū, ∇v) − ((F y (ȳ))µ, v) = (U − Ū , v)r,r∗ − ⟨Ru , v⟩,
ay (µ, s) + cy (ū, µ, s) + cy (ζ, ȳ, s) = ⟨Ry , s⟩,
where the respective reminder terms ⟨Ru , v⟩ and ⟨Ry , s⟩ consist of quadratic and mixed higher-order
remainder terms due to the inherent nonlinear nature of the coupled problem and are defined for Θ ∈
(0, 1), respectively as follows:
1 1
Z
⟨Ru , v⟩ := c(u − ū, u − ū, v) + (1 − Θ)(Fyy (y + Θ(y − ȳ))(y − ȳ)2 , v) dΘ
2! 0
1
− (νT T (T + Θ(T − T̄ ))(T − T̄ )2 ∇ū, ∇v) − (νT (T + Θ(T − T̄ ))(T − T̄ )∇(u − ū), ∇v),
2!
⟨Ry , s⟩ := cy (u − ū, y − ȳ, s).
In order to apply (SSC), we split ζ = u − ū and µ = y − ȳ as follows: ζ = uh + ur and µ = y h + y r .

2
Here, (uh , y h ) with y h := (Th , Sh ) and (ur , y r ) with y r := (Tr , Sr ) for all (v, s) ∈ X × H01 (Ω) solve,
(
a(ȳ; uh , v) + c(uh , ū, v) + c(ū, uh , v) + ((νT (T̄ ))Th ∇ū, ∇v) − ((F y (ȳ))y h , v) = (h, v)r,r∗ ,
(3.16)
ay (y h , s) + cy (ū, y h , s) + cy (uh , ȳ, s) = 0,
and

a(ȳ; ur , v) + c(ur , ū, v) + c(ū, ur , v)

(3.17) + ((νT (T̄ ))Tr ∇ū, ∇v) − ((F y (ȳ))y r , v) = (U − Ũ , v)r,r∗ − ⟨Ru , v⟩,

ay (y r , s) + cy (ū, y r , s) + cy (ur , ȳ, s) = ⟨Ry , s⟩,

respectively. Equations (3.16) are linear and the auxiliary triplet (uh , y h , h) belongs to the subspace
where (SSC) applies. Using [2, Lemma 4.1] (by taking ⟨fˆ, v⟩ = (h, v)r,r′ and ⟨f˜, s⟩ = 0), we obtain the
following estimates for the auxiliary states:

(3.18) ∥uh ∥1,Ω + ∥y h ∥1,Ω ≤ C ∥h∥Lr (Ω) ≤ C ∥Ũ − U ∥Lr (Ω) + ∥U − Ū ∥Lr (Ω) .
To bound the remainder terms,, we follow the same steps as in the proof of [2, Lemma 4.3] and apply
Lemma 2.7, to get
2
|c(u − ū, u − ū, v)| ≤ C6d C3d ∥u − ū∥1,Ω ∥ur ∥1,Ω ≤ C∥U − Ū ∥2Lr (Ω) ∥v∥1,Ω ,
T̄ + Θ(T − T̄ ) (T − T̄ )2 ∇ū, ∇v)| ≤ CνT T Cgn Cr2d Cr3d M ∥y − ȳ∥23/2+δ,Ω ∥v∥1,Ω

|(νT T
≤ C∥U − Ū ∥2Lr (Ω) ∥v∥1,Ω ,
2
|(Fyy (ȳ + Θ(y − ȳ)) y − ȳ)2 , v | ≤ CFyy C42d ∥y − ȳ∥1,Ω ∥v∥1,Ω ≤ C∥U − Ū ∥2Lr (Ω) ∥v∥1,Ω ,

|(νT (T + Θ(T − T̄ ))(T − T̄ )∇(u − ū), ∇v)| ≤ CνT C6d Cgn ∥y − ȳ∥1,Ω ∥u − ū∥3/2+δ,Ω ∥v∥1,Ω
≤ C∥U − Ū ∥2Lr (Ω) ∥v∥1,Ω .
Thus,
(3.19) |⟨Ru , v⟩| ≤ C∥U − Ū ∥2Lr (Ω) ∥v∥1,Ω .
Similarly, we can get the following estimates for Ry :
(3.20) |⟨Ry , s⟩| ≤ C6d C3d ∥u − ū∥1,Ω ∥y − ȳ∥1,Ω ∥s∥1,Ω ≤ C∥U − Ū ∥2Lr (Ω) ∥s∥1,Ω .
Following the steps analogous to the proof of [2, Lemma 4.1], we can get the following estimates for the
auxiliary states solving (3.17):

∥ur ∥1,Ω + ∥y r ∥1,Ω ≤ C ∥Ũ − U ∥Lr (Ω) + ∥Ru ∥X ∗ + ∥Ry ∥−1,Ω
≤ C ∥Ũ − U ∥Lr (Ω) + ∥U − Ū ∥2Lr (Ω) .

(3.21)
Now we investigate the second order derivative of the Lagrangian. Denoting by wh = (uh , y h , h), we can
easily derive the following identities:
2 2 2 2
Lww (w̄, Λ̄) w − w̄ = Luu (w̄, Λ̄) uh + ur + Lyy (w̄, Λ̄) y h + y r + LU U (w̄, Λ̄) U − Ũ + h
2 2
= Lww (w̄, Λ̄) wh + Luu (w̄, Λ̄) [ur ] + 2Luu (w̄, Λ̄) uh , ur

2 2
+ Lyy (w̄, Λ̄) [y r ] + 2 Lyy (w̄, Λ̄) [y h , y r ] + LU U (w̄, Λ̄) U − Ũ

(3.22) + 2LU U (w̄, Λ̄) U − Ũ , h .
The first term is tackled by using (SSC). The second order derivatives with respect to control satisfy
2
LU U (w̄, Λ̄) U − Ũ + 2LU U (w̄, Λ̄) U − Ũ , h = λ∥U − Ũ ∥20,Ω + 2λ(U − Ũ , h).

Using the definition of Ũ and h, we can deduce that they are orthogonal to each other, so their inner
product is zero. Therefore, the following relation holds:
2
LU U (w̄, Λ̄) U − Ũ + 2LU U (w̄, Λ̄) U − Ũ , h = λ∥U − Ū ∥20,Ωε ≥ 0.

(3.23)
For our convenience, let use the following notations: z̃ := ∥U − Ũ ∥Lr (Ω) and z̄ := ∥U − Ū ∥Lr (Ω) . Then
the second order derivatives with respect to y in (3.22) are treated using Lemma 3.1 and estimates (3.18),
(3.21), as follows:
2 2
|Lyy (w̄, Λ̄) [y r ] + 2Lyy (w̄, Λ̄) [y h , y r ] | ≤ C ∥y r ∥1,Ω + ∥y r ∥1,Ω ∥y h ∥1,Ω
≤ C z̄ 4 + z̄ 3 + z̄z̃ + z̃z̄ 2 + z̃ 2 .

(3.24)
The same estimates hold for the second order derivatives with respect to u in (3.22) in an analogous way,
2 2
|Luu (w̄, Λ̄) [ur ] + 2Luu (w̄, Λ̄) [uh , ur ] | ≤ C ∥ur ∥1,Ω + ∥ur ∥1,Ω ∥uh ∥1,Ω
≤ C z̄ 4 + z̄ 3 + z̄z̃ + z̃z̄ 2 + z̃ 2 .

(3.25)
Substituting (3.23)-(3.25) in (3.22) results to

2 2
(3.26) Lww (w̄, Λ̄) [w − w̄] ≥ σ ∥h∥Lr (Ω) − C(z̄ 4 + z̄ 3 + z̄z̃ + z̃z̄ 2 + z̃ 2 ).
Next we eliminate h in (3.26) such that only terms containing Λ̄ and z̃ appear. It can be easily seen that
2
∥U − Ū ∥2Lr (Ω) = ∥U − Ũ + h∥2Lr (Ω) ≤ 2 ∥U − Ũ ∥2Lr (Ω) + ∥h∥Lr (Ω) ,

which implies
2 1 2
(3.27) ∥h∥Lr (Ω) ≥ z̄ − z̃ 2 .
2
Invoking (3.24), (3.25) and (3.27) in (3.26), we get
1 2 σ
Lww (w̄, Λ̄) [w − w̄] ≥ z̄ 2 − C z̄ 4 + z̄ 3 + z̄z̃ + z̃z̄ 2 + z̃ 2 .

(3.28)
2 4
Similarly, we investigate the third order derivative which presents itself in the remainder term RL of
(3.13). Proceeding analogously, we get
3 3 3
Lyyy (w̄ + Θ(w − w̄), Λ̄) [y − ȳ] = Lyyy (w̄ + Θ(w − w̄), Λ̄) [y h ] + Lyyy (w̄ + Θ(w − w̄), Λ̄) [y r ]
+ 3 Lyyy (w̄ + Θ(w − w̄), Λ̄) [y h , y h , y r ]
(3.29) + 3 Lyyy (w̄ + Θ(w − w̄), Λ̄) [y h , y r , y r ] .
Using Lemma 3.1 along with (3.18) and (3.21) gives

3 3
|Lyyy (w̄ + Θ(w − w̄), Λ̄) [y h ] | ≤ CL3 ∥y h ∥1,Ω ≤ C(z̃ + z̄)3 ,
3 3
|Lyyy (w̄ + Θ(w − w̄), Λ̄) [y r ] | ≤ CL3 ∥y r ∥1,Ω ≤ C(z̃ + z̄ 2 )3 ,
2 2
|Lyyy (w̄ + Θ(w − w̄), Λ̄) [y h , y h , y r ] | ≤ CL3 ∥y h ∥1,Ω ∥y r ∥1,Ω ≤ C (z̃ + z̄) z̃ + z̄ 2 ,

2 2
|Lyyy (w̄ + Θ(w − w̄), Λ̄) [y h , y r , y r ] | ≤ CL3 ∥y h ∥1,Ω ∥y r ∥1,Ω ≤ C (z̃ + z̄) z̃ + z̄ 2

.
On further simplifying after invoking the above bounds in RL , we can get the following estimate,
RL ≥ −C z̄ 6 + z̄ 5 + z̄ 3 + z̃ 2 z̄ + z̃ 2 z̄ 2 + z̃z̄ 2 + z̃z̄ 4 + z̃ 3 .

(3.30)
Adding (3.28) and (3.30) and applying Young’s inequality to separate powers of z̄ and z̃, we get
1 2 σ
Lww (w̄, Λ̄) [w − w̄] + RL ≥ z̄ 2 − C z̄ 6 + z̄ 5 + z̄ 4 + z̄ 3 + z̄z̃ + z̃ 2 z̄ + z̃ 2 z̄ 2 + z̃z̄ 2
2 4
+ z̃z̄ 4 + z̃ 2 + z̃ 3

σ
≥ z̄ 2 − C z̄ 8 + z̄ 6 + z̄ 5 + z̄ 4 + z̄ 3 + z̃ 4 + z̃ 3 + z̃ 2

8
σ
≥ z̄ 2 − C z̄ 6 + z̄ 4 + z̄ 3 + z̄ 2 + z̄ − C z̃ 4 + z̃ 3 + z̃ 2 .

(3.31)
8
If U is sufficiently close to Ū , that is, z̄ ≤ Ns,r ∥U − Ū ∥Ls (Ω) ≤ Ns,r ρ1 , then the term
σ σ
− C z̄ 6 + z̄ 4 + z̄ 3 + z̄ 2 + z̄ ≥

.
8 16
Thus we arrive at
1 2 σ 2
z̄ − C z̃ 4 + z̃ 3 + z̃ 2 .

(3.32) Lww (w̄, Λ̄) [w − w̄] + RL ≥
2 16
Focusing on z̃, by definition of our constructed control Ũ , (U − Ũ )i = 0 on Ω\Ωε,i and (Ũ − Ū )i = 0 on

Ωε,i . Hence using (2.11), we conclude the following:
(3.33) z̃ 2 ≤ ∥U − Ũ ∥L1 (Ω) ∥U − Ũ ∥Ls (Ω) = U − Ū L1,Ωε

U − Ū Ls,Ωε
≤ ρ1 U − Ū L1,Ωε
.
Using the above bound, (3.32), (3.15) and (3.12) in (3.13), we obtain
σ 2
J(w) ≥ J(w̄) + ε U − Ū L1,Ωε
+ U − Ū Lr (Ω)
16
3/2 1/2
+ ρ21 U − Ū

− C ρ1 + ρ1 U − Ū L1,Ωε L1,Ωε
+ U − Ū L1,Ωε
2 2 σ 2
≥ J(w̄) + ε U − Ū L1,Ωε
− Cρ2 U − Ū L1,Ωε
+ U − Ū Lr (Ω)
16
2 σ 2
≥ J(w̄) + (ε − Cρ2 ) U − Ū L1,Ωε
+ U − Ū Lr (Ω)
.
16
σ
For ρ2 small enough such that ε − Cρ2 > 0, the claim is proven with ϑ = 16 and ρ = min(ρ1 , ρ2 ).
The following result is an immediate consequence of Theorem 3.3.
Corollary 3.4. Let the assumptions of Theorem 3.3 hold. Then Ū is a locally optimal control in
the sense of Ls (Ω).
Remark 3.5. In three dimensions, the restriction on δ can be relaxed by choosing (ζ, µ) ∈ X ∩
H 3/2+δ (Ω) × [H 3/2+δ (Ω)]2 . This is reasonable since we have already shown in Theorem 2.13 that the
weak solution to the adjoint equation has this extra regularity, from which we can deduce that the
linearized equations in (SSC) will also have this characteristic.
In Theorem, 3.3 we assumed that the reference control satisfies (SSC). In the following theorem, we
investigate conditions under which it can be ensured that (SSC) holds.
Theorem 3.6. Let w̄ = (ū, ȳ, Ū ) be an admissible triplet for our optimal control problem. Suppose
w̄ satisfies the first-order necessary optimality conditions with associated adjoint state Λ̄. Then (SSC) is
fulfilled if the parameter λ is sufficiently large or the residuals ∥ȳ − y d ∥0,Ω and ∥ū − ud ∥0,Ω are sufficiently
small such that
1
Mψ Mφ C6d C3d + CFyy C42d ) + CνT T M̂ M̄ Mψ

(3.34) λ>
C2,r
is satisfied.
Proof. The second order derivative of the Lagrangian is given by
2 2 2
(3.35) Lww (w̄, Λ̄) [(ζ, µ, h)] = ∥ζ∥0,Ω − 2 c(ζ, ζ, φ̄) + ∥µ∥0,Ω − (νT T (T̄ )(µT )2 ∇ū, ∇φ̄)
2
+ (F yy (ȳ)(µ ⊗ µ), φ̄) + λ ∥h∥0,Ω .
The positivity of (3.35) can be disturbed by the second, fourth and fifth terms. The states (ζ, µ) are
weak solutions of the linear equations (see [2, (95)]). We can estimate these terms, thanks to the precise
estimates of the nonlinear and linearized system of equations (see [2, Lemma 4.2] and Lemma 2.11) and
(3.7), (3.8), as follows:
2 2
|c(ζ, ζ, φ̄)| ≤ C6d C3d ∥∇φ̄∥0,Ω ∥ζ∥1,Ω ≤ C6d C3d Mψ Mφ ∥h∥Lr (Ω) ,
2 2
|(F yy (ȳ)(µ ⊗ µ), φ̄)| ≤ CFyy C42d ∥µ∥1,Ω ∥φ̄∥1,Ω ≤ CFyy C42d Mψ Mφ ∥h∥Lr (Ω) ,
2
|(νT T (T̄ )(µT )2 ∇ū, ∇φ̄)| ≤ CνT T M̂ M̄ Mψ ∥h∥Lr (Ω) ,
where, M̂ is defined according to (3.8) in two and three dimensions, respectively. Thus using the
embeddings of Lr -spaces with the embedding constant C2,r depending only on the domain Ω, we obtain
2
Lww (w̄, Λ̄) [(ζ, µ, h)]
2 2 2
≥ ∥ζ∥0,Ω + ∥µ∥0,Ω + λC2,r − Mψ Mφ C6d C3d + CFyy C42d + CνT T M̄ M̂ Mψ ∥h∥Lr (Ω) .

This implies that Lww is positive definite, if (3.34) holds, that is, either λ is large enough or the residuals
∥ȳ − y d ∥0,Ω and ∥ū − ud ∥0,Ω are small enough.
4. Numerical analysis of the state equation. In this section we propose a discretization of the
governing equation (2.2) using nonconforming finite elements and study the well-posedness of resulting
discrete system. To this aim, let Th denote a partition of the domain Ω into a finite collection of open
subdomains Ki , i = 1, ..., N E (triangles or quadrilaterals in (2D) and tetrahedra or hexahedra in (3D))
such that:
Ω̄ = ∪Ki ∈Th K̄i and Ki ∩ Kj = ϕ, i ̸= j.
The mesh parameter h associated with Th is defined as h = maxK∈Th hK , where hK denotes the diameter
of K. We assume that Th is shape regular in the sense that there exists a real number ρ̄ > 0, independent
of h and K such that as h → 0,
hK
≤ ρ̄, ∀ K ∈ Th ,
ρK
where ρK := sup {diam(B) : B is a ball contained in K} . Let E(Th ) be the set of all edges in (2D) or the
set of all faces in (3D), which is divided into EI (Th ) and EB (Th ), the set of all interior and boundary edges,
respectively. For any e ∈ EI (Th ), we label by K + and K − the elements adjacent to e. Furthermore, he
denotes the edge length or the maximum diameter of the facet in (2D) and (3D), respectively. Let v, w
represent smooth vector and scalar fields, respectively, defined over Th . Then, by (v + , v − ) and (w+ , w− )
we will denote the traces of v and w on e taken from inside K + and K − , respectively. Similarly, we
−
denote by n+ e and ne the outward unit normal vectors to e corresponding to K
+
and K − , respectively.
The averages and jumps of v and w across e are defined, respectively, as follows
(v + + v − ) (w+ + w− )
, [[v]] := v − · n− + +
, [[w]] := w− n− + +

{{v}} := e + v · ne , {
{w}
} := e + w ne .
2 2
The respective notations for boundary jumps and averages are, {{v}} = v, [[v]] = v · nΓ , {{w}} = w, [[w]] =
w nΓ . In addition, ∇h denotes the broken gradient operator, that is, the operator ∇ taken element wise.
For any natural number k ≥ 1, Pk (K) denotes the space of piecewise polynomials of degree less than
or equal to k. In this section, for some U ∈ Lr (Ω), we use the following notations uh := uh (U ), ph :=
ph (U ), y h := y h (U ).
4.1. Nonconforming method. Consider the following lowest order Crouziex-Raviart space on the
partition Th ,
n o
d
V cr
h := v h ∈ L2
(Ω) : v h | K ∈ [P1 (K)] , v h is continuous at midpoint ∀ e ∈ EI (T h ) ,
V hcr0 := {v h ∈ V cr
h : v h at midpoint is zero ∀ e ∈ EB (Th )} ,
cr0
2
and let Mcrh and Mh denote their counterparts for all sh ∈ L2 (Ω) . Moreover, we also define the
following discontinuous finite element space
Qkh := qh ∈ L20 (Ω) : qh |K ∈ Pk (K) ∀ K ∈ Th .

Corresponding to these finite dimensional spaces, we consider the following nonconforming finite element
method which is based on an upwind discretization to tackle the convective terms (cf. [21]). For some
U ∈ Lr (Ω), find (uh , ph , y h ) ∈ V hcr0 × Q0h × Mcr h such that y h |Γ = y h and
D
a (y h ; uh , v h ) + ch (uh , uh , v h ) + bh (v h , ph ) = dh (y h , v h ) + (U , v h )r,r′ ∀ v h ∈ V cr
 h
h ,
0


(4.1) bh (uh , qh ) = 0 ∀ qh ∈ Q0h ,

ahy (y h , sh ) + chy (uh , y h , sh ) = 0 ∀ sh ∈ Mcr

h ,
0
with
Z Z
ah (sh ; uh , v h ) := K −1 uh · v h + ν(sh )∇h uh : ∇h v h dx, ahy (y h , sh ) :=

D ∇h y h : ∇h sh dx,
ZΩ X Z
Ω
h
c (wh , uh , v h ) := (wh · ∇h ) uh · v h dx + ŵup
h (uh ) · v h ds,
Ω K∈Th ∂K\Γ
Z X Z
chy (wh , y h , sh ) := (wh · ∇h ) y h · sh dx + ŵup
h (y h ) · sh ds,
Ω K∈Th ∂K\Γ
where bh (·, ·), dh (·, ·) allows for discontinuous elements and are defined analogously to b(·, ·), d(·, ·),
respectively. The fluxes are defined as follows:
1 1
ŵup
h (uh ) := (wh · nK − |wh · nK |) (ueh − uh ) and ŵup
h (y h ) := (wh · nK − |wh · nK |) (y eh − y h ),
2 2
and ueh and y eh are the traces of u and y = (T, S)⊤ , respectively, taken from within the exterior of K
over the edge under consideration and is zero on the boundary. Without loss of generality, by Πnc , we
2 1 cr
denote the nonconforming interpolation operator from H 1 (Ω) to Mcr h (H (Ω) to V h ) such that
Z
1
(4.2) Πnc s(me ) = s ds ∀ e ∈ E(Th ),
|e| e
where me denotes the mid-point of the edge e (cf. [15, 22]). Let
n 2 o
E h := s ∈ L2 (Γ) : s|e ∈ P0 (e) ∀ e ∈ EB (Th ) ,

denote the space of Lagrange multipliers on Γ. Following ([15, Section 2.3]), we use an analogue of Πnc
∂ ∂
2 2
restricted to the boundary in order to define y D D
h := Πnc y , where Πnc : L (Γ) −→ E h such that
Z
1
Π∂nc s(me ) = s ds ∀ e ∈ EB (Th ),
|e| e
where, |e| denotes the Lebesgue measure of the edge e. Due to the fact that divV cr 0
h ⊂ Qh , we have the
0
h
following discrete kernel corresponding to b (·, ·) :
cr0 cr0
X cr h 0

h := v h ∈ V h : b (v h , qh ) = 0 ∀ qh ∈ Qh = {v h ∈ V h : divv h = 0} .
4.1.1. Discrete boundedness properties. The space V cr h

0
is a subset of the piecewise Sobolev
space
H 1 (Th ) := v ∈ L2 (Ω) : v|K ∈ H 1 (K) ∀ K ∈ Th .

1 2 2
Similarly, Mcrh is also a subset of H (Th ) which is defined analogously for all sh ∈ L2 (Ω) . Now we
define the following mesh and parameter dependent norms over V cr cr
h and Mh as follows:
0
2
X 2 2 2
X 2 2
σ ∥v∥0,K +ν2 ∥∇v∥0,K , ∀ v ∈ H 1 (Th ), ∥s∥1,T nc := σ̄ ∥∇s∥0,K , ∀ s ∈ H 1 (Th ) ,

∥v∥1,T nc :=
h h
K∈Th K∈Th
−1
where σ = K ∞,Ω
, σ̄ = ∥D∥∞,Ω and ν2 is as defined in section 1. In the following lemma, we state
and prove the continuity properties of the discrete bilinear and trilinear forms.
Lemma 4.1. There holds
|ah (y h ; uh , v h )| ≤ Canc ∥uh ∥1,T nc ∥v h ∥1,T nc ∀ uh , v h ∈ V cr

h ,
0
h h
cr0
|ahy (y h , sh )| ≤ Ĉanc ∥y h ∥1,T nc ∥sh ∥1,T nc ∀ y h ∈ Mcr
h , sh ∈ Mh ,
h h
|ch (wh , uh , v h )| ≤ Cenc ∥wh ∥1,T nc ∥uh ∥1,T nc ∥v h ∥1,T nc ∀ wh , uh , v h ∈ V cr

h
0
h h h
|chy (wh , y h , sh )| ≤ Cenc ∥wh ∥1,T nc ∥y h ∥[L4 (Ω)]2 ∥sh ∥1,T nc ∀ wh ∈ V cr cr cr0
h , y h ∈ Mh , sh ∈ Mh ,
0
h h
1
|b (v h , qh )| ≤ Cd ∥v h ∥1,T nc ∥qh ∥0,Ω ∀ v h ∈ H (Th ), qh ∈ L20 (Ω).
h
h
Proof. We begin by establishing the continuity properties of the discrete bilinear form ah . The
element-wise contributions can be bounded for all v h ∈ V cr
h by the mesh-dependent norms as follows:
0
X X
ah (y h ; uh , v h ) ≤ σ ∥uh ∥0,K ∥v h ∥0,K + ν2 ∥∇h uh ∥0,K ∥∇h v h ∥0,K ≤ Ĉanc ∥uh ∥1,T nc ∥v h ∥1,T nc .
h h
K∈Th K∈Th
Similarly, we can derive the desired bound for ahy . To bound the discrete trilinear forms ch , we utilize
the discrete Sobolev embedding for the nonconforming setting (cf. [23, 24]): For r ∈ [1, ∞) if d = 2 and
r ∈ [1, 6] if d = 3 there exists a constant Cenc > 0 such that
(4.3) ∥v∥Lr (Ω) ≤ Cenc ∥v∥1,T nc ∀ v ∈ H 1 (Th ).

h
Indeed, the elementwise contribution of the convective term can be bounded using the Hölder’s inequality
and (4.3) as follows:
X Z X
(wh · ∇uh ) · v h dx ≤ ∥wh ∥L4 (K) ∥∇uh ∥L2 (K) ∥v h ∥L4 (K)
K∈Th K K∈Th
n X o1/4 n X o1/4 n X o1/2
4 4 2
≤ ∥wh ∥L4 (K) ∥v h ∥L4 (K) ∥∇uh ∥L2 (K)
K∈Th K∈Th K∈Th
≤ ∥wh ∥L4 (Ω) ∥v h ∥L4 (Ω) ∥uh ∥1,T nc ≤ Cenc ∥wh ∥1,T nc ∥v h ∥1,T nc ∥uh ∥1,T nc .
h h h h
Using the inverse inequality,

−1
(4.4) ∥v∥0,∂K ≤ CT hK 2 ∥v∥0,K ∀ v ∈ Pk (K),
with CT being independent of the mesh parameter h, and following similar steps as above, we can show
the following bound on the discrete upwind term:
X Z
ŵup nc
h (uh ) · v h ds ≤ Ce ∥w h ∥1,T nc ∥uh ∥1,T nc ∥v h ∥1,T nc .
h h h
K∈Th ∂K\Γ
Combining the above two bounds leads to desired bound for ch . To derive the bound for chy , we integrate
it by parts
X Z X Z
h
cy (wh , y h , sh ) := − (wh · ∇h ) sh · y h dx + ŵup
h (sh ) · y h ds,
K∈Th K K∈Th ∂K\Γ
and apply the same procedure used to get the bounds for ch .
In addition, for all wh ∈ X cr h , an integration by parts of the convective term and the fact that the
discrete velocities obtained are exactly divergence free, yield the following positivity properties,
Z
1 X
(4.5) ch (wh , uh , uh ) = |wh · ne | |[[uh ]]|2 ds ≥ 0 ∀ uh ∈ V cr
h ,
0
2 e e∈E(Th )
1 X Z
(4.6) chy (wh , sh , sh ) = |wh · ne | |[[sh ]]|2 ds ≥ 0 ∀ sh ∈ Mcr
h .
0
2 e
e∈E(Th )
Moreover, the following coercivity properties of ah and ahy also hold:

2 2
(4.7) ah (·; uh , uh ) ≥ αanc ∥v h ∥1,T nc ∀ v h ∈ V cr
h
0
and ahy (sh , sh ) ≥ α2nc ∥sh ∥1,T nc ∀ sh ∈ Mcr
h ,
0
h h
and bh (·, ·) satisfies the following discrete inf-sup condition for all qh ∈ Q0h :
bh (v h , qh )
(4.8) sup ≥ β̃ ∥qh ∥0,Ω , for some β̃ > 0 which is independent of h.
v h ∈V 0 \{0} ∥v h ∥1,T nc
cr
h h
4.1.2. Existence of discrete solutions. The following discrete analogue [2, Lemma 3.2] holds:
Lemma 4.2. If (uh , ph , y h ) ∈ V cr 0 cr
h × Qh × Mh solves (4.1), then for some U ∈ L (Ω), (uh , y h ) ∈
0 r
X cr
h
cr D
× Mh satisfies y h |Γ = y h and
( h
a (y h ; uh , v h ) + ch (uh , uh , v h ) − dh (y h , v h ) − (U , v h )r,r′ = 0 ∀ v h ∈ X cr
h ,
(4.9)
ay (y h , sh ) + cy (uh , y h , sh ) = 0 ∀ sh ∈ Mcr
h h
h .
0
Conversely, if (uh , y h ) ∈ X cr cr
h × Mh is a solution of the reduced problem (4.9), then for some U ∈ U ad ,
0
there exists ph ∈ Qh such that (uh , ph , y h ) is a solution of (4.1).
A discrete boundary lifting of y h is performed by setting
(4.10) y h = y h,0 + y h,1 with y h,0 ∈ Mcr cr D

h such that y h,1 ∈ Mh and y h,1 |Γ = y h .
0
Since trace of a nonconforming function to the boundary does not necessarily belong to [H 1/2 (Γ)]2 , a
discrete lifting theorem was proved in [15] using a discrete [H 1/2 (Γ)]2 -norm combined with an enriching
operator (see [25, Appendix B]) to convert the nonconforming approximation to a conforming one. The
combined results stating compatibility of discrete [H 1/2 (Γ)]2 -norm with the usual [H 1/2 (Γ)]2 -norm and
its relation with the ∥·∥1,T nc norm is stated in the following lemma (see [15, Lemmas 3.1 and 3.4]):
h
1 2
Lemma 4.3. Let g ∈ H 2 (Γ) then for Π∂nc g ∈ E h , there exists sh ∈ Mcr h satisfying sh (me ) =
Π∂nc g(me ) and
∥sh ∥1,Thnc ≤ C∥Π∂nc g∥1/2,E h ≤ C ∥g∥1/2,Γ ,
where C denotes a generic positive constant independent of the mesh parameter h.
Lemma 4.4. Let (uh , y h ) be a solution of (4.9) with the discrete boundary lifting as defined in (4.10).
Then there exist positive constants Cunc , Cync independent of the mesh parameter h, such that
∥uh ∥1,T nc ≤ Cunc yD + ∥U ∥Lr (Ω) := Munc , ≤ Cync yD

1/2,Γ
y h,0 1,Thnc 1/2,Γ
+ ∥U ∥Lr (Ω) , and
h
∥y h ∥1,T nc ≤ Cync yD + ∥U ∥Lr (Ω) := Mync .

h 1/2,Γ
Proof. Choose (v h , sh ) = (uh , y h,0 ) in (4.9) and use (4.5) and (4.6) to rewrite it as follows:
ah (y h ; uh , uh ) = dh (y h , uh ) + (U , uh )r,r′ ∀ uh ∈ X cr
h ,
ahy (y h,0 , y h,0 ) = −ahy (y h,1 , y h,0 ) − chy (uh , y h,1 , y h,0 ) ∀ y h,0 ∈ Mcr
h .
0
Due to the coercivity of the bilinear forms in (4.7), continuity properties stated in Lemma 4.1, we have
X
αanc ∥uh ∥1,T nc ≤ CF Cenc y h,0 0,K + y h,1 0,K + Cenc ∥U ∥Lr (Ω)

(4.11)
h
K∈Th
≤ CF Cenc + Cenc ∥U ∥Lr (Ω) ,

y h,0 1,Thnc
+ y h,1 1,Thnc
(4.12) α2nc y h,0 1,Thnc

≤ Cenc ∥uh ∥1,T nc y h,1 [L4 (Ω)]2
+ Ĉanc y h,1 1,Thnc
.
h
Put the bounds of (4.12) in (4.11) and apply [2, Lemma 3.3] which still holds for the discrete case with
CF (Cenc )2
(4.13) y h,1 < a < 1, to get
αanc α2nc [L4 (Ω)]2
CF Cenc Canc Cenc

(4.14) ∥uh ∥1,T nc ≤ max 1 + nc , y h,1 1,Thnc
+ ∥U ∥Lr (Ω) .
h (1 − a)αanc α2 (1 − a)αanc
Now an application of Lemma 4.3 on (4.14) and substituting the resulting bound in (4.12) leads to the
desired bound for y h,0 .
We make use of the following proposition which is a consequence of the Brouwer fixed point theorem to
show the existence of discrete solutions [7]:
Proposition 4.5. Let H be a finite dimensional Hilbert space with norm | · |H and inner product
(·, ·)H . Let R : H −→ H be a continuous map such that there exists a positive constant ρ satisfying
(R(v), v)H > 0 ∀ v ∈ H with |v|H = ρ.
Then there exists a w ∈ H such that R(w) = 0, |w|H ≤ ρ.

Theorem 4.6. Let y h,1 be the discrete lifting of y h as defined in Section 4.1.2. Then there exists a
discrete solution (uh , y h ) ∈ X cr cr
h × Mh to (4.9) satisfying the bounds of Lemma 4.4.
cr0 cr0
Proof. Define the maps Qu and Qy on X cr cr
h and Mh for all v h ∈ X h and sh ∈ Mh , respectively,
as follows:
(Qu (uh ), v h ) = ah (y h ; uh , v h ) + ch (uh , uh , v h ) − dh (y h , v h ) − (U , v h )r,r′ ,

(Qy (y h,0 ), sh ) = ah (y h,0 , sh ) + chy (uh , y h,0 , sh ) − ah (y h,1 , sh ) − chy (uh , y h,1 , sh ).
The Continuity of Qu and Qy readily follows by using the boundedness properties of Lemma 4.1. Let
Qu,y := Qu × Qy with (uh , y h,0 ) = ∥uh ∥X cr + y h,0 Mcr0 . Then the continuity of Qu,y follows from
h h
the continuity of maps Qu and Qy and the definition of Qu,y . In order to make use of Proposition 4.5,
which would imply that Qu,y has a fixed point, it remains to show the coercivity of Qu,y . Since uh ∈ X cr h
and thanks to (4.7) and (4.5), the form ah (·; ·, ·) + ch (·, ·, ·) is coercive, and on using the boundedness
properties and Lemma 4.4, we get
2
(Qu (uh ), uh ) ≥ αanc ∥uh ∥1,T nc − CF Cenc ∥y h ∥1,T nc ∥uh ∥1,T nc − Cenc ∥U ∥Lr (Ω) ∥uh ∥1,T nc
h h h h
2
≥ αanc ∥uh ∥1,T nc − CF Cenc Mync Munc − C̃ nc Munc ∥U ∥Lr (Ω) .
h
2
It follows that (Qu (uh ), uh ) > 0 for ∥uh ∥1,T nc = κnc
1 such that
h

1 nc nc
κnc
1 > Ce M u C F M nc
y + ∥U ∥ r
L (Ω) .
αanc
Similarly, invoking (4.7) and (4.6), the form ahy (·; ·, ·) + chy (·, ·, ·) is coercive, yielding
2
(Qy (y h,0 ), y h,0 ) ≥ α2nc y h,0 − Ĉanc y h,1 1,T nc y h,0 1,T nc − Cenc ∥uh ∥1,T nc y h,1
1,Thnc 1,Thnc
y h,0 1,Thnc
h h h

nc 2 nc nc
≥ α2 y h,0 1,T nc − Ĉa Cy y h,1 1,T nc + ∥U ∥Lr (Ω) y h,1 1,T nc
h h h

nc nc nc
− C e Mu C y y h,1 1,T nc + ∥U ∥Lr (Ω) y h,1 1,T nc .
h h
2
After an application of Lemma 4.3, we conclude that (Qy (y h,0 ), y h,0 ) > 0 for y h,0 1,T nc = κnc
2 such
h
that
1 nc nc D
κnc
2 > Ĉ a Cy y + ∥U ∥Lr (Ω) (1 + C e
nc
M nc
u ) y D
.
α2nc 1/2,Γ 1/2,Γ
From the coercivity of Qy and Qu and the definition of Qu,y , we deduce that (Qu,y (uh , y h,0 ), (uh , y h,0 )) >
0 for (uh , y h,0 ) = κnc nc
1 + κ2 > 0.
4.1.3. Uniqueness of discrete solutions.

Theorem 4.7. Let (uh , y h ) ∈ X cr
h ∩W
1,3
(Th ) × Mcr
h be a solution of the reduced problem (4.9), and
assume that
(4.15) ∥∇h uh ∥L3 (Ω) ≤ M̃ ,
where M̃ > 0 is chosen such that

Cenc Mync
(4.16) αanc > nc γν M̃ + CF + Cenc Munc
α2
holds. Then such a solution is unique under (4.16).

Furthermore, let (uh (U i ), y h (U i )) be solutions of (4.9) corresponding to U i ∈ Lr (Ω) for i = 1, 2,
then the following stability result holds:
(4.17) ∥uh (U 1 ) − uh (U 2 )∥1,T nc + ∥y h (U 1 ) − y h (U 2 )∥1,T nc ≤ C ∥U 1 − U 2 ∥Lr (Ω) .

h h
Proof. Let (uah , y ah ) and (ubh , y bh ) solve (4.9) and let (αh , β h ) := (uah −ubh , y ah −y bh ). Then subtracting
cr0
the corresponding variational formulations for all v h ∈ X cr h , sh ∈ Mh , we have
ah (y ah ; uah , v h ) − ah (y bh ; ubh , v h ) + ch (uah , uah , v h ) − ch (ubh , ubh , v h ) = dh (y ah , v h ) − d(y bh , v h ),

(4.18)
ahy (y ah , sh ) − ahy (y bh , sh ) + chy (uah , y ah , sh ) − chy (ubh , y bh , sh ) = 0.
We can rewrite the terms in (4.18) as
ah (y ah ; uah , v h ) − ah (y bh ; ubh , v h ) = ah (y ah ; αh , v h ) + ah (y ah ; ubh , v h ) − ah (y bh ; ubh , v h ),

ch (uah , uah , v h ) − ch (ubh , ubh , v h ) = ch (uah , αh , v h ) − ch (uah , ubh , v h ) + ch (ubh , ubh , v h ),
chy (uah , y ah , sh ) − chy (ubh , y bh , sh ) = chy (uah , β h , sh ) − chy (uah , y bh , sh ) + chy (ubh , y bh , sh ).
Utilizing Hölder’s inequality, stability properties discussed in Section 4.1.1, Lemma 4.4, discrete Sobolev
embedding (4.3) and the assumption (4.15), we can establish the following bounds:
X
|ah (y ah ; ubh , v h ) − ah (y bh ; ubh , v h )| ≤ ν(y ah ) − ν(y bh ) L6 (K) ∥∇h uah ∥L3 (K) ∥∇h v h ∥L2 (K)
K∈Th
≤ γν y ah − y bh L6 (Ω)
∥∇h uah ∥L3 (Ω) ∥v h ∥1,T nc
h
(4.19) ≤ γν M̃ ∥β h ∥1,T nc ∥v h ∥1,T nc ,

h h
|ch (ubh , ubh , v h ) − ch (uah , ubh , v h )| ≤ Cenc ∥αh ∥1,T nc ubh 1,Thnc
∥v h ∥1,T nc
h h
(4.20) ≤ Cenc Munc ∥αh ∥1,T nc ∥v h ∥1,T nc ,

h h
|chy (ubh , y bh , sh ) − chy (uah , y bh , sh )| ≤ Cenc ∥αh ∥1,T nc y bh 1,T nc

∥sh ∥1,T nc
h h h
(4.21) ≤ Cenc Mync ∥αh ∥1,T nc ∥sh ∥1,T nc ,

h h
h
(4.22) |d (y ah , v h ) −d h
(y bh , v h )| ≤ CF ∥β h ∥1,T nc ∥v h ∥1,T nc .
h h
Choosing v h = αh ∈ X cr
h and sh = β h in (4.18) and exploiting (4.7), (4.5) and (4.6) gives
2
(4.23) αanc ∥αh ∥1,T nc ≤ |ah (y ah ; uah , αh ) − ah (y ah ; ubh , αh )| + |ch (ubh , ubh , αh ) − ch (uah , ubh , αh )|
h
+ |dh (y ah , αh ) − dh (y bh , αh )|,
2
(4.24) α2nc ∥β h ∥1,T nc ≤ |chy (ubh , y bh , β h ) − chy (uah , y bh , β h )|.
h
Invoking the above established bounds in (4.23) and (4.24), we get
αanc ∥αh ∥1,T nc ≤ γν M̃ + CF ∥β h ∥1,T nc + Cenc Munc ∥αh ∥1,T nc ,

h h h
α2nc ∥β h ∥1,T nc ≤ Cenc Mync ∥αh ∥1,T nc .

h h
Adding the above two inequalities and simplifying gives the desired uniqueness of the discrete solutions.
The stability result (4.17) can be derived analogously using (4.3) and the fact that
nc nc
C e My

nc nc nc
αa − γν M̃ + CF + Ce Mu ∥αh ∥1,T nc ≤ ∥U 1 − U 2 ∥Lr (Ω) ∥αh ∥L r−1
r ,
α2nc h (Ω)
which completes the proof of (4.17).

4.1.4. A priori error analysis. Let Πh : Q −→ Q0h be such that (Πh p − p, qh ) = 0 for all p ∈ Q
and qh ∈ Q0h denote the pressure interpolation operator, Πnc defined in (4.2) denote the nonconforming
interpolation operator and without loss of generality let P e denote the standard edge projection on to
piecewise constants on e. Then under adequate regularity assumptions, the following approximation
properties hold (cf. [26, 27, 28]):
√ √
∥u − Πnc u∥1,T nc ≤ C(h3/2+δ σ + h1/2+δ ν2 ) ∥u∥3/2+δ,Ω ,
h
√
(4.25) ∥y − Πnc y∥1,T nc ≤ Ch3/2+δ σ̄ ∥y∥3/2+δ,Ω , ∥p − Πh p∥0,Ω ≤ Ch1/2+δ ∥p∥1/2+δ,Ω ,
h
2
1/2
∥v − P e v∥0,e ≤ ChK |v|1,K , ∀ v ∈ H 1 (K) or, v ∈ H 1 (K)

, e ∈ E(Th ).
Theorem 4.8. Let us consider the discrete lifting (4.10) with sufficiently small data (4.13). Assume
(4.15) such that (4.16) holds. Also, for some U ∈ Lr (Ω), let (u, p, y) and (uh , ph , y h ) be the solutions
of (2.2) and (4.1), respectively. Furthermore, for δ ∈ (0, 12 ) and u ∈ H 10 (Ω) ∩ H 3/2+δ (Ω), p ∈
2
L20 (Ω) ∩ H 1/2+δ (Ω), y ∈ H 3/2+δ (Ω) . Then under (2.9) and (4.16), there exists a positive constant C
independent of the mesh parameter h such that
∥u − uh ∥1,T nc + ∥y − y h ∥1,T nc ≤ Ch1/2+δ ∥u∥3/2+δ,Ω + ∥y∥3/2+δ,Ω ,

(4.26)
h h
∥p − ph ∥0,Ω ≤ Ch1/2+δ ∥u∥3/2+δ,Ω + ∥y∥3/2+δ,Ω + ∥p∥1/2+δ,Ω .

(4.27)
Proof. Let us decompose the errors into
eu := (u − χ) + (χ − uh ) = êu + ẽu , ey := (y − θ) + (θ − y h ) = êy + ẽy

(4.28)
ep := (p − ψ) + (ψ − ph ) = êp + ẽp ,
where, (χ, ψ, θ) ∈ V cr 0 cr
h × Qh × Mh . Now an application of integration by parts implies that (u, p, y)
satisfies the following equation:
X Z
(4.29) ah (y; u, v h ) + ch (u, u, v h ) + bh (v h , p) − dh (y, v h ) = ν(T )∇u · nv h ds
K∈Th ∂K
+ (U , v h )r,r′ ∀ v h ∈ V cr
h .
0
Writing a discrete analogue of (4.29) and subtracting the result, yields the following:
(4.30) ah (y; u, v h ) − ah (y h ; uh , v h ) + ch (u, u, v h ) − ch (uh , uh , v h ) + bh (v h , p − ph )

X Z
− dh (y, v h ) + dh (y h , v h ) = ν(T )∇u · n v h ds.
K∈Th ∂K
Furthermore, we can obtain the following:
(4.31) bh (u − uh , qh ) = 0,
X Z
(4.32) ahy (y, sh ) − ahy (y h , sh ) + chy (u, y, sh ) − chy (uh , y h , sh ) = D∇y · n sh ds.
K∈Th ∂K
Now, using the error decomposition in (4.28), testing (4.30) with v h = ẽu and (4.32) with sh = ẽy and
rearranging the terms, we end up with the following error equations:
(4.33) ah (y; ẽu , ẽu ) + ch (uh , ẽu , ẽu ) = T1 + T2 + T3 + T4 ,

(4.34) ah (ẽy , ẽy ) + chy (uh , ẽy , ẽy ) = I1 + I2 + I3 ,
where
T1 := ah (y h ; uh , ẽu ) − ah (y; uh , ẽu ),

T2 := ch (χ, u, ẽu ) − ch (u, u, ẽu ) + ch (uh , u, ẽu ) − ch (χ, u, ẽu ) − ch (uh ; êu , ẽu ),

X Z
T3 := dh (y h , ẽu ) − dh (y, ẽu ), T4 := ν(T )∇u · n ẽu ds,
K∈Th ∂K
I1 := −ahy (êy , ẽy ), I2 := −chy (êu , y, ẽy ) − chy (ẽu , y, ẽy ) − chy (uh , êy , ẽy ),
X Z
I3 := D∇y · n ẽy ds.
K∈Th ∂K
Now we focus our attention on finding appropriate bounds for these terms. Following steps analogous to
(4.19), (4.20) and (4.22), we can get the following bounds:
|T1 | ≤ γν M̃ ∥y − y h ∥1,T nc ∥ẽu ∥1,T nc ,

h h
|T2 | ≤ Cenc Mu (∥êu ∥1,T nc + ∥ẽu ∥1,T nc ) ∥ẽu ∥1,T nc + Cenc Munc ∥êu ∥1,T nc ∥ẽu ∥1,T nc ,
h h h h h
|T3 | ≤ CF (∥êy ∥1,T nc + ∥ẽy ∥1,T nc ) ∥ẽu ∥1,T nc ,

h h h
|I1 | ≤ Ĉanc ∥êy ∥1,T nc ∥ẽy ∥1,T nc ,

h h
|I2 | ≤ Cenc My (∥êu ∥1,T nc + ∥ẽu ∥1,T nc ) ∥ẽy ∥1,T nc + Cenc Munc ∥êy ∥1,T nc ∥ẽy ∥1,T nc .
h h h h h
The consistency error terms T4 and I3 which arise due to the nonconforming approximation are bounded
by using the edge projection P e and [29, (10.3.9)], as follows:
X Z X Z
|T4 | = ν(T )∇u · ne [[ẽu ]] ds = ν(T ) (∇u · ne − P e (∇u · ne )) [[ẽu ]] ds
e∈ETh e e∈ETh e
X 1/2 X 1/2
2 2
≤ ν2 |e|−1 ∥∇u · ne − P e (∇u · ne ) ∥0,e |e| ∥[[ẽu ]]∥0,e ,
e∈ETh e∈ETh
X 1/2 X 1/2
2
≤ Cν2 |e|−1 ∥∇u · ne − P e (∇u · ne ) ∥0,e h2K |ẽu |21,K
e∈ETh K∈Th
X 1/2
2
≤ Cν2 h |e|−1 ∥∇u · ne − P e (∇u · ne ) ∥0,e ∥ẽu ∥1,T nc ,
h
e∈ETh
X 1/2
2
|I3 | ≤ C Ĉa h |e|−1 ∥∇y · ne − P (∇y · ne ) ∥0,e ∥ẽy ∥1,T nc .
h
e∈ETh
Inserting the bounds into (4.33) and (4.34) and applying (4.5), (4.6) and (4.7) yields
αanc ∥ẽu ∥1,T nc ≤ γν M̃ (∥êy ∥1,T nc + ∥ẽy ∥1,T nc ) + Cenc Mu (∥êu ∥1,T nc + ∥ẽu ∥1,T nc ) + Cenc Munc ∥êu ∥1,T nc
h h h h h h
X 1/2
2
(4.35) + CF (∥êy ∥1,T nc + ∥ẽy ∥1,T nc ) + Cν2 h |e|−1 ∥∇u · ne − P e (∇u · ne ) ∥0,e ,
h h
e∈ETh
α2nc Ĉanc Cenc My

∥êu ∥1,T nc + ∥ẽu ∥1,T nc + Cenc Munc ∥êy ∥1,T nc

∥ẽy ∥1,T nc ≤ ∥êy ∥1,T nc +
h h h h h
X 1/2
2
(4.36) + C Ĉa h |e|−1 ∥∇y · ne − P e (∇y · ne ) ∥0,e .
e∈ETh
Substituting (4.36) in (4.35) gives

Cenc Mync
αanc − nc nc

γ ν M̃ + C F + C e M u ∥ẽu ∥1,T nc
α2nc h
h X 1/2
2
≤ C ∥êu ∥1,T nc + ∥êy ∥1,T nc + h |e|−1 ∥∇u · ne − P e (∇u · ne ) ∥0,e
h h
e∈ETh
X 1/2 i
2
+h |e|−1 ∥∇y · ne − P e (∇y · ne ) ∥0,e .
e∈ETh
Choosing χ = Πnc u, θ = Πnc y and invoking the approximation properties given in (4.25) straightforwardly
leads to the estimate (4.26). To derive the pressure estimate, we exploit the discrete inf-sup condition
(4.8) and Lemma 4.1 as follows:
!
1 bh (v h , ep ) bh (v h , êp ) 1 bh (v h , ep ) 1
(4.37) ∥ẽp ∥0,Ω ≤ sup + ≤ sup + ∥êp ∥0,Ω .
β̃ v h ∈V h \{0} ∥v h ∥1,T nc ∥v h ∥1,T nc β̃ v h ∈V h \{0} ∥v ∥
h 1,T nc β̃
h h h
For any v h ∈ V cr
h , (4.30) implies
bh (v h , ph − p) = ah (y; u, v h ) − ah (y h ; uh , v h ) + ch (u, u, v h ) − ch (uh , uh , v h )

X Z
h h
(4.38) − d (y, v h ) + d (y h , v h ) − ν(T )∇u · n v h ds.
K∈Th ∂K
Utilizing (4.19), (4.20), (4.22) and analogous steps used to bound |T4 |, we get

|bh (v h , ph − p)| ≤ ∥v h ∥1,T nc γν M̃ ∥ey ∥1,T nc + Cenc Munc ∥eu ∥1,T nc + CF ∥ey ∥1,T nc
h h h h
X 1/2
2
(4.39) + Cν2 h |e|−1 ∥∇u · ne − P e (∇u · ne ) ∥0,e .
e∈ETh
Thus the estimate (4.27) follows by using (4.39) in (4.37), choosing ψ = Πh p and using (4.26) and (4.25).
5. Numerical analysis of the adjoint equation. The goal of this section is to propose a nonconforming
discretization of the adjoint equation (2.12) and establish the well-posedness and a priori estimates of the
resulting discrete system. In this section, for some U ∈ Lr (Ω), we use the following notations:
uh := uh (U ), ph := ph (U ), y h := y h (U ), φh := φh (U ), ξh := ξh (U ), η h := η h (U ).
5.1. Nonconforming method. For some U ∈ U ad and uh := uh (U ) ∈ V cr cr

h , y h := y h (U ) ∈ Mh ,
0
cr0 0 cr0 cr0 0 cr0

find (φh , ξh , η h ) ∈ V h × Qh × Mh for all (v h , qh , sh ) ∈ V h × Qh × Mh such that
 h
 a (y h ; φh , v h ) − ch (uh , φh , v h ) + ch (v h , uh , φh ) + b(v h , ξh ) − chy (v h , y h , η h ) = (uh − ud , v h ),

(5.1) bh (φh , qh ) = 0,

ay (η h , sh ) − chy (uh , η h , sh ) + (νT (Th )∇h uh : ∇h φh , sh ) − ((F y (y h ))⊤ φh , sh ) = (y h − y d , sh ),
 h
where, the discrete bilinear and trilinear forms are defined as in Section 4.1.
5.1.1. Discrete stability properties. The stability properties discussed in Section 4.1.1 hold
alongwith the following properties which can be derived by applying Cauchy-Schwarz, Hölder’s and
discrete Poincaré’s inequalities and (4.3),
|(uh − ud , v h )| ≤ C̄dnc ∥uh − ud ∥0,Ω ∥v h ∥1,T nc ,

h
|(y h − y d , sh )| ≤ Ĉdnc ∥y h − y d ∥0,Ω ∥sh ∥1,T nc ,

h
|((F y (y h ))⊤ φh , sh )| ≤ CFy Cpnc ∥φh ∥1,T nc ∥sh ∥1,T nc ,

h h
X
⊤
|(νT (Th )∇h uh : ∇h φh , sh )| ≤ ∥(νT (Th )) ∥∞,K ∥∇φh ∥0,K ∥∇uh sh ∥0,K ,
K∈Th
≤ CνT Cenc ∥φh ∥1,Thnc ∥∇h uh ∥L4 (Ω) ∥sh ∥1,T nc .

h
5.1.2. Well-posedness of discrete solutions. We have the following equivalence result similar
to Theorem 4.2.
Lemma 5.1. If (φh , ξh , η h ) ∈ V cr 0
h × Qh × Mh
0 cr0
solves (5.1), then for some U ∈ U ad , (uh , y h ) ∈
cr0
X cr
h × Mh satisfies
(
ah (y h , φh , v h ) − ch (uh , φh , v h ) + ch (v h , uh , φh ) + chy (v h , y h , η h ) = (uh − ud , v h ),
(5.2)
ahy (η h , sh ) − chy (uh , η h , sh ) + (νT (Th )∇h uh : ∇h φh , sh ) − ((F y (y h ))⊤ φh , sh ) = (y h − y d , sh ).
cr0
Conversely, if (φh , η h ) ∈ X cr
h × Mh is a solution of the reduced problem (4.9), then for some U ∈ U ad
there exists ξh ∈ Q0h such that (φh , ξh , η h ) is a solution of (5.1).
The following lemma can be proven along the lines of Lemma 4.4 and following the steps analogous
to the proof of Theorem 4.7.
Lemma 5.2. Let (ūh , ȳ h ) ∈ X cr cr
h × Mh be the states associated with the control Ū ∈ U ad and assume
(4.15). Then the solution (φh , η h ) ∈ X h × Mcr
cr
h
0
of (5.2), satisfies the following a priori estimate:

nc
(5.3) ∥φh ∥1,T nc ≤ Cφ ∥uh − ud ∥0,Ω + ∥y h − y d ∥0,Ω := Mφnc ,
h

(5.4) ∥η h ∥1,T nc ≤ Cηnc ∥uh − ud ∥0,Ω + ∥y h − y d ∥0,Ω := Mηnc ,
h
nc
where, Cφ and Cηnc are positive constants such that
Cenc Mync
(5.5) αanc > nc CνT M̃ + CFy + Cenc Munc .
α2
Furthermore, let (φh (U i ), η h (U i )) be solutions of (5.2) corresponding to U i ∈ Lr (Ω) for i = 1, 2, then

the following stability result holds:
∥φh (U 1 ) − φh (U 2 )∥1,T nc + ∥η h (U 1 ) − η h (U 2 )∥1,T nc

h h

(5.6) ≤ C ∥uh (U 1 ) − uh (U 2 )∥1,T nc + ∥y h (U 1 ) − y h (U 2 )∥1,T nc .
h h
We rewrite the discrete adjoint equation (5.2) as follows:
(5.7) B nc ((φh , η h ), (v h , sh )) = f nc (v h , sh ),
where,
B nc ((φh , η h ), (v h , sh )) := ah (y h , φh , v h ) − ch (uh , φh , v h ) + ch (v h , uh , φh )
+ ahy (η h , sh ) − chy (uh , η h , sh ) + chy (v h , y h , η h ) − ((F y (y h ))⊤ φh , sh )
+ (νT (Th )∇h uh : ∇h φh , sh ),
nc
f (v h , sh ) := (uh − ud , v h ) + (y h − y d , sh ).
The boundedness of B nc and f nc follows from the boundedness properties discussed in Section 4.1.1
and 5.1.1, assumption (4.15) and Lemma 4.4. The well-posedness of the adjoint equation (5.7) is a direct
consequence of the Lax-Milgram theorem:
Theorem 5.3. Let (ūh , ȳ h ) ∈ X cr cr
h × Mh be the states associated with the control Ū ∈ U ad and
nc cr0
assume (4.15). For every f , there exists unique (φh , η h ) ∈ X cr h × Mh such that (5.7) holds with
2 2 2
|||(v h , sh )|||nc := ∥v h ∥1,T nc + ∥sh ∥1,T nc .
h h
Proof. It is enough to show the coercivity of B nc . Choosing v h = φh , sh = η h , and using the

continuity properties discussed in Section 5.1.1, Lemmas 4.1 and 4.4 as follows:
2 2 2
B nc ((φh , η h ), (φh , η h )) ≥ αanc ∥φ∥1,T nc − Cenc Munc ∥φh ∥1,T nc + α2nc ∥η h ∥1,T nc
h h h
Cenc Mync 2 2
C
Fy 2 2

− ∥φh ∥1,T nc + ∥η h ∥1,T nc − ∥φh ∥1,T nc + ∥η h ∥1,T nc
2 h h 2 h h
CνT Cenc M̃ 2 2

− ∥φh ∥1,T nc + ∥η h ∥1,T nc
2 h h
2 2 2
nc
≥ (αa − M1 ) ∥φh ∥1,T nc + (α2 − M2 ) ∥η h ∥1,T nc ≥ C2nc |||(φh , η h )|||nc
nc
h h
such that αanc ≥ M1 and α2nc ≥ M2 where,
Cenc Mync CFy Cν C nc M̃

M2 := + + T e and M1 := Cenc Munc + M2 .
2 2 2
5.1.3. A priori error analysis.

Theorem 5.4. Assume (4.15) with M̃ > 0 such that (4.16). Also, for some U ∈ Lr (Ω), (u, y) ∈
2
H 10 (Ω)∩ H 3/2+δ (Ω) × H 3/2+δ (Ω) and (uh , y h ) ∈ V cr
cr
h × Mh , let (φ, ξ, η) and (φh , ξh , η h ) be the
0
solutions of (2.21) and (5.1), respectively with the regularity of (φ, ξ, η) given in Theorem 2.13. Then
under (5.5) and for φh ∈ V crh ∩W
0 1,3
(Thnc ), there exists positive constants C independent of the mesh
parameter h such that
∥φ − φh ∥1,T nc + ∥η − η h ∥1,T nc ≤ Ch1/2+δ ∥φ∥3/2+δ,Ω + ∥η∥3/2+δ,Ω ,

(5.8)
h h
∥ξ − ξh ∥0,Ω ≤ Ch1/2+δ ∥φ∥3/2+δ,Ω + ∥η∥3/2+δ,Ω + ∥ξ∥1/2+δ,Ω ,

(5.9)
Proof. Let us decompose the errors into
eφ := (φ − χ) + (χ − φh ) = êφ + ẽφ , eη := (η − θ) + (θ − η h ) = êη + ẽη ,

(5.10)
eξ := (ξ − ψ) + (ψ − ξh ) = êξ + ẽξ ,
where, (χ, ψ, θ) ∈ V cr 0 cr
h × Qh × Mh . Now an application of integration by parts implies that (φ, ξ, η)
satisfies the following equation:
ah (y, φ, v h ) − ch (u, φ, v h ) + ch (v h , u, φ) + bh (v h , ζ) − chy (v h , y, η) = (u − ud , v h )

XZ
(5.11) + ν(Th )∇φ · n v h ds,
K∈T ∂K
ahy (η, sh ) − chy (u, η, sh ) + (νT (T )∇h u : ∇h φ, sh ) − ((F y (y))⊤ φ, sh ) = (y − y d , sh )

X Z
(5.12) + D∇η · n sh ds.
K∈Th ∂K
Now similar to the proof of Theorem 4.8, we subtract (5.11) and (5.12) from their respective discrete
counterparts, testing the resulting equations with v h = ẽφ and sh = ẽη , respectively and rearranging the
terms using the error decomposition in (5.10) yields the following error equations:
(5.13) ah (y h , ẽφ , ẽφ ) − ch (uh , ẽφ , ẽφ ) = Ta + Tb + Tc + Td + Te + Tf ,

(5.14) ahy (ẽη , ẽη ) − chy (uh , ẽη , ẽη ) = Ia + Ib + Ic + Id + Ie ,
where
Ta := ah (y h ; φh , ẽφ ) − ah (y; φh , ẽφ ), Tb := −ch (uh , êφ , ẽφ ) + ch (uh − u, φ, ẽφ ), Tc := ch (ẽφ , uh , ẽφ ),
Td := chy (ẽφ , y − y h , η) + chy (ẽφ , y h , êη ) + chy (ẽφ , y h , ẽη ), Te := (uh − u, ẽφ ),
X Z
Tf := ν(T )∇φ · n ẽφ ds, Ia := −chy (u, êη , ẽη ) − chy (u − uh , η h , ẽη ), Id := (y − y h , ẽη ),
K∈Th ∂K
Ib := (νT (T )∇h (u − uh ) : ∇h φ, ẽη ) + ((νT (T ) − νT (Th ))∇h uh : ∇h φh , ẽη ) + (νT (T )∇h uh : ∇h eφ , ẽη ),
X Z
⊤ ⊤ ⊤
Ic := (F y (y) (φ − φh ), ẽη ) + (((F y (y)) − ((F y (y h )) ) φh , ẽη ), Ie := D∇η · n ẽη ds.
K∈Th ∂K
Applying the properties of the bilinear and trilinear forms discussed in Sections 4.1.1 and 5.1.1, and
tackling the consistency error terms similarly to T4 and I3 in the proof of Theorem 4.8, we arrive at the
following inequalities:
αanc ∥ẽφ ∥1,T nc ≤ γν M̃ Cenc ∥y − y h ∥1,T nc + Cenc (Munc ∥êφ ∥1,Thnc + ∥u − uh ∥1,T nc ) + Cenc Munc ∥ẽφ ∥1,T nc
h h h h
+ Cenc (Mη
∥y − y h ∥1,T nc + Cenc Mync (∥êη ∥1,T nc
+ ∥ẽη ∥1,T nc )) + ∥u − uh ∥1,T nc
h h h h
X 1/2
2
(5.15) + Cν2 h |e|−1 ∥∇φ · ne − P e (∇φ · ne ) ∥0,e ,
e∈ETh
α2nc ∥ẽη ∥1,T nc ≤ Cenc ((Munc∥êη ∥1,T nc + Mηnc ∥u − uh ∥1,T nc ) + ∥y − y h ∥1,T nc + CνT Mφ ∥u − uh ∥1,T nc )
h h h h h
+ Cenc M̃ ∥T − Th ∥1,T nc + CνT M̃ ∥êφ ∥1,T nc + CνT M̃ ∥ẽφ ∥1,T nc + Mφnc ∥y − y h ∥1,T nc

h h h h
X 1/2
nc nc −1 2
(5.16) + CFy Ce (∥êφ ∥1,T nc + ∥ẽφ ∥1,T nc ) + C Ĉa h |e| ∥∇η · ne − P e (∇η · ne ) ∥0,e .
h h
e∈ETh
Substituting (5.16) in (5.15) gives

Cenc Mync
αanc − nc nc

C ν M̃ + C F + C e M u ∥ẽφ ∥1,T nc
α2nc T y
h
h
≤ C ∥êφ ∥1,T nc + ∥êη ∥1,T nc + ∥êu ∥1,T nc + ∥êy ∥1,T nc
h h h h
X 1/2 X 1/2 i
2 2
+h |e|−1 ∥∇φ · ne − P e (∇φ · ne ) ∥0,e +h |e|−1 ∥∇η · ne − P e (∇η · ne ) ∥0,e .
e∈ETh e∈ETh
Choosing χ = Πnc φ, θ = Πnc η and invoking the approximation properties given in (4.25) and Theorem
4.8 lead to the estimate (5.8). To derive the adjoint pressure estimate we exploit the discrete inf-sup
condition (4.8) similar to the proof of Theorem 4.8 as follows:
1 bh (v h , eξ ) 1
(5.17) ∥ẽξ ∥0,Ω ≤ sup + ∥êξ ∥0,Ω .
β̃ v h ∈V h \{0} ∥v ∥
h 1,T nc β̃
h
For any v h ∈ V hcr0 , we have
bh (v h , ξh − ξ) = ah (y h ; eφ , v h ) + ch (uh , eu , v h ) + ch (v h , uh , eφ )
X Z
(5.18) − cy (v h , y h , eη ) − ν(Th )∇φ · n v h ds.
K∈Th ∂K
Utilizing the continuity properties of and analogous steps used to bound |T4 |, we get
h
|bh (v h , ξh − ξ)| ≤ ∥v h ∥1,T nc Canc ∥eφ ∥1,T nc + Cenc Munc ∥eφ ∥1,T nc + Cenc Mync ∥eη ∥1,T nc
h h h h
X 1/2 i
2
(5.19) + Cν2 h |e|−1 ∥∇φ · ne − P e (∇φ · ne ) ∥0,e .
e∈ETh
The estimate (5.9) follows by using (5.19) in (5.17), choosing ψ = Πh ξ and using (5.8) and (4.25).
6. Numerical analysis of the control problem. The aim of this section is to propose a discretization
for the control variable and establish a priori error estimates for the resulting fully discretized control
problem.
6.1. Control discretization. Let U h ⊂ U denote the discrete control space which is defined as,
U h := {U h ∈ U : U h |K ∈ P0 (K) ∀ K ∈ Th } .
The set of admissible controls is defined as U h,ad := U h ∩ U ad . From here on, we use the following
notations, uh := uh (U h ), ph := ph (U h ), y h := y h (U h ), φh := φh (U h ), ξh := ξh (U h ), η h := η h (U h ).
The finite dimensional approximate optimal control problem is defined by
1 2 1 2 λ 2
(6.1) min J(uh , y h , U h ) = ∥uh − ud ∥0,Ω + ∥y h − y d ∥0,Ω + ∥U h ∥0,Ω .
U h ∈U h,ad 2 2 2
The discrete problem (6.1) admits an optimal solution follows from arguments analogous to the proof
of [2, Theorem 4.4]. The optimize-then-discretize approach results in the following discrete optimality
system under piecewise constant discretization of control for the unknowns (ūh , ph , ȳ h , φ̄h , ξh , η̄ h , Ū h ) ∈
V cr 0 cr cr0 0 cr0 D
h × Qh × Mh × V h × Qh × Mh × U h,ad satisfying ȳ h |Γ = y h , which can be used to determine
0
the discrete optimal control:
a (ȳ h ; ūh , v h ) + ch (ūh , ūh , v h ) + bh (v h , ph ) = dh (ȳ h , v h ) + (Ū h , v h )r,r′ ∀ v h ∈ V cr

 h
h ,
0


(6.2) bh (ūh , qh ) = 0 ∀ qh ∈ Q0h ,

ahy (ȳ h , sh ) + chy (ūh , ȳ h , sh ) = 0 ∀ sh ∈ Mcr

h ,
0
 h

a (ȳ h , φ̄h , v h ) − ch (ūh , φ̄h , v h ) + ch (v h , ūh , φ̄h ) + b(v h , ξh )

− chy (v h , ȳ h , η̄ h ) = (ūh − ud , v h ) ∀ sh ∈ V cr

h ,
 0



(6.3) bh (φ̄h , qh ) = 0 ∀ qh ∈ Q0h ,

ahy (η̄ h , sh ) − chy (ūh , η̄ h , sh ) + (νT (T̄h )∇h ūh : ∇h φ̄h , sh )






− ((F y (ȳ h ))⊤ φ̄h , sh ) = (ȳ h − y d , sh ) ∀ sh ∈ Mcr h ,
0

and

(6.4) λŪ h + φ̄h , U h − Ū h ≥ 0 ∀ U h ∈ U h,ad .
6.2. A priori error analysis. We assume in this section that the constraints Uaj , Ubj are real
constants. This along with the regularity of adjoint variable φ̄ and projection formula (2.15) imply that
optimal control Ū ∈ H 1 (Ω). Let Π0 : L2 (Ω) −→ U h be defined locally as follows:
(6.5) (Π0 U , v)K = (U , v)K ∀ v ∈ U h .
Notice that Π0 U ad ⊂ U h,ad and has the approximation property (cf. [30]),
(6.6) ∥Π0 U − U ∥0,K ≤ CΠ hK ∥U ∥1,K .
Now, we prove the existence of an appropriate solution Ū h in any neighborhood of Ū .
Lemma 6.1. Let (Ū h )h>0 be any sequence of solutions to (6.1). Then there exists a weakly-converging
subsequence (still indexed by h), and if the subsequence (Ū h )h>0 is converging weakly to (Ū ), then Ū is
a solution of (1.1) and
(6.7) ¯
lim J(ūh , ȳ h , Ū h ) = J(ū, ȳ, Ū ) =: J.
h→0
Moreover, the weak convergence of (Ū h )h>0 to Ū and (6.7) implies

(6.8) lim ∥Ū h − Ū ∥0,Ω = 0.
h→0
Proof. Note that (Ū h )h>0 is bounded in U h,ad , and the bound is independent of h, since by the
construction it is contained in U ad . Then there exists a subsequence still indexed by h, that converges
weakly to Ū in L2 (Ω) due to the Banach-Alaoglu theorem. Moreover, Ū ∈ U ad follows from Mazur’s
theorem and the continuity and convexity of J implies (see [31, Lemma 4.2] or [17, Theorem 2.12])
(6.9) J¯ ≤ J(ū, ȳ, Ū ) ≤ lim inf J(ūh , ȳ h , Ū h ).
h→0
Let W̄ be a solution of (1.1). Then we have ∥W̄ − Π0 W̄ ∥0,Ω ≤ Ch due to (6.6). Thus, (see [31, Lemma
4.2, part (i)]) limh→0 J(ūh , ȳ h , Π0 W̄ ) = J(ū, ȳ, W̄ ) = J¯ due to Theorem 4.8 and (6.6). Moreover, Π0 W̄
is admissible for (6.1), and thus J(ūh , ȳ h , Ū h ) ≤ J(ūh , ȳ h , Π0 W̄ ). Passing the limit in the last inequality,
we obtain
(6.10) ¯
lim inf J(ūh , ȳ h , Ū h ) ≤ lim sup J(ūh , ȳ h , Ū h ) ≤ lim sup J(ūh , ȳ h , Π0 W̄ ) = J(ū, ȳ, W̄ ) = J.
h→0 h→0 h→0
By (6.9) and (6.10), we arrive at (6.7). In order to prove (6.8) we first observe that
∥Ū h − Ū ∥20,Ω = ∥Ū h ∥20,Ω + ∥Ū ∥20,Ω − 2⟨Ū h − Ū , Ū ⟩ − 2∥Ū ∥20,Ω .
Now using (6.7) which implies that ∥Ū h ∥20,Ω → ∥Ū ∥20,Ω and weak convergence of (Ū h )h>0 to Ū , we
conclude (6.8).
Theorem 6.2. Let Ū be a locally optimal control in the sense of the Definition 2.8 and satisfy the
first order necessary and second order sufficient optimality conditions. Let (ū, p, ȳ, Ū ) be the solution to
the optimal control problem (1.1)-(1.2), which together with (φ̄, ξ, η̄) satisfies the continuous optimality
system (2.20)-(2.22). Further, let (ūh , ph , ȳ h , φ̄h , ξh , η̄ h , Ū h ) be the solution of the corresponding discrete
optimality system (6.2)-(6.4). Then under the hypothesis of Theorems 4.8 and 5.4, there exists a positive
constant C independent of the mesh parameter h such that the following estimates hold:
∥Ū − Ū h ∥Lr (Ω) ≤ Ch1/2+δ ∥Ū ∥1,Ω + ∥φ̄∥3/2+δ,Ω + ∥η̄∥3/2+δ,Ω ,

(6.11)
∥ū − ūh ∥1,T nc + ∥ȳ − ȳ h ∥1,T nc ≤ Ch1/2+δ ∥Ū ∥1,Ω + ∥φ̄∥3/2+δ,Ω + ∥η̄∥3/2+δ,Ω
h h

(6.12) + ∥ū∥3/2+δ,Ω + ∥ȳ∥3/2+δ,Ω ,
∥p − ph ∥0,Ω ≤ Ch1/2+δ ∥Ū ∥1,Ω + ∥φ̄∥3/2+δ,Ω + ∥η̄∥3/2+δ,Ω

(6.13) + ∥ū∥3/2+δ,Ω + ∥ȳ∥3/2+δ,Ω + ∥p∥1/2+δ,Ω ,
≤ Ch1/2+δ ∥Ū ∥1,Ω + ∥φ̄∥3/2+δ,Ω + ∥η̄∥3/2+δ,Ω ,

(6.14) ∥φ̄ − φ̄h ∥1,T nc + ∥η̄ − η̄ h ∥1,T nc
h h
∥ξ − ξh ∥0,Ω ≤ Ch1/2+δ ∥Ū ∥1,Ω + ∥φ̄∥3/2+δ,Ω + ∥η̄∥3/2+δ,Ω + ∥ξ∥1/2+δ,Ω .

(6.15)
Proof. From (3.6) we know that

2 2 2
(6.16) LU U (ū, ȳ, Ū , Λ̄) [h] = LU U (w̄, Λ̄) [h] = λ ∥h∥0,Ω .
Now we perturb LU U (ū, ȳ, Ū , Λ̄) only with respect to Ū , and since (ū, ȳ, Λ̄) depend on Ū , thus, by an
abuse of notation we write, LU U (Ū ). Choosing h = Ū h − Ū ∈ L2 (Ω) in (6.16) and applying mean value
theorem for some 0 < Θ < 1 (see [20, Theorem 7.9-1 (c)]), yields
2 2
(1 − Θ)2 λ ∥h∥0,Ω = LU U (Ū + Θh) [h] = LU (Ū h ) − LU (Ū ) [h] ,

= λŪ h + φ(Ū h ), h − λŪ + φ̄, h .
Thus we obtain the following relation

2
(1 − Θ)2 λ Ū h − Ū

(6.17) 0,Ω
= λŪ + φ̄, Ū − Ū h − λŪ h + φ(Ū h ), Ū − Ū h .
The continuous variational inequality (2.13) for U = Ū h ∈ U h,ad ⊂ U ad , and the discrete variational
inequality (6.4) for U h = Π0 Ū give the following relation:

(6.18) λŪ h + φ̄h , Π0 Ū − Ū h ≥ 0 ≥ λŪ + φ̄, Ū − Ū h .
Using (6.18) in (6.17) and simplifying results to

2
(1 − Θ)2 λ Ū h − Ū

0,Ω
≤ λŪ h + φ̄h , Π0 Ū − Ū h − λŪ h + φ(Ū h ), Ū − Ū h

(6.19) = φ̄h − φ(Ū h ), Ū − Ū h + λ Ū h , Π0 Ū − Ū + φ̄h , Π0 Ū − Ū .
The second term on the right hand side of (6.19) vanishes due to (6.5) and the third term can be rewritten
using Π0 as follows:

(6.20) φ̄h , Π0 Ū − Ū = φ̄h − Π0 φ̄h , Π0 Ū − Ū ≤ ∥φ̄h − Π0 φ̄h ∥0,Ω Π0 Ū − Ū 0,Ω .
The first term on the right hand side of (6.19) is bounded using (4.3), Young’s inequality as follows:

φ̄h − φ̄(Ū h ), Ū − Ū h ≤ φ̄h − φ̄(Ū h ) 0,Ω Ū − Ū h 0,Ω
2 2
(6.21) ≤ (1/2λ(1 − Θ)2 ) φ̄h − φ(Ū h ) 0,Ω
+ ((1 − Θ)2 λ/2) Ū − Ū h 0,Ω
.
Plugging (6.20) and (6.21) in (6.19) and applying (4.3), (6.6), (5.8) gives
2 2
((1 − Θ)2 λ/2) Ū h − Ū 0,Ω
≤ ∥φ̄h − Π0 φ̄h ∥0,Ω Π0 Ū − Ū 0,Ω
+ (1/2λ(1 − Θ)2 ) φ̄h − φ̄(Ū h ) 0,Ω
2
≤ CΠ h ∥φ̄h ∥1,T nc Π0 Ū − Ū 0,Ω
+ (Cenc 2 /2λ(1 − Θ)2 ) φ̄h − φ̄(Ū h ) 1,T nc
h h

2 2
≤ C̃ h ∥φ̄h ∥1,T nc Π0 Ū − Ū 0,Ω
+ h1+2δ ∥φ̄∥3/2+δ,Ω + ∥η̄∥3/2+δ,Ω .
h
where C̃ = max CΠ , (Cenc 2 /2λ(1 − Θ)2 ) is a positive constant. Inferring from (5.3) the following

∥φ̄h ∥1,T nc ≤ Cφ ∥ūh ∥0,Ω + ∥ud ∥0,Ω < C,
h
and employing (6.6), we reach the following estimate:
≤ Ch1/2+δ ∥Ū ∥1,Ω + ∥φ̄∥3/2+δ,Ω + ∥η̄∥3/2+δ,Ω .

(6.22) Ū h − Ū 0,Ω
Now we show that using (SSC) and (6.22), we can achieve the same convergence rate for the control
variable in the neighbourhood dictated by (SSC) which states, there exists ε > 0 and σ > 0 such that
2 2 2 2 2
Luu (w̄, Λ̄) [ζ] + Lyy (w̄, Λ̄) [µ] + LU U (w̄, Λ̄) [h] = Lww (w̄, Λ̄) [(ζ, µ, h)] ≥ σ ∥h∥Lr (Ω) ,
with h = U − Ū ∀ U ∈ U ad , hi = 0 on Ωε,i for i = 1, . . . , n. Observing from (3.6) that

2 2 2 2
LU U (ū, ȳ, Ū , Λ̄) [h] = LU U (w̄, Λ̄) [h] = λ ∥h∥0,Ω ≥ λCe ∥h∥Lr (Ω) ,
we conclude that
2 2
(6.23) LU U (ū, ȳ, Ū , Λ̄) [h] ≥ σ̃ ∥h∥Lr (Ω) , where σ̃ = min {σ, λCe } is a positive constant.
Following (6.17) we can obtain

2
LU (Ū h ) − LU (Ū ) [h] = LU U (Ū + Θh) [h]
2 2 2 2
≥ LU U (Ū + Θh) − LU U (Ū ) [h] + σ̃ ∥h∥Lr (Ω) = ∥h∥L2 (Ω) λ(Θ2 − 2Θ) + σ̃ ∥h∥Lr (Ω) ,
2
where in the last inequality we used (6.23) after adding and subtracting LU U (Ū ) [h] . This leads to
2 2
σ̃ ∥h∥Lr (Ω) ≤ LU (Ū h ) − LU (Ū ) [h] + λ(2Θ − Θ2 ) ∥h∥0,Ω .

(6.24)
The first term on the right hand side is treated analogous to (6.17) and then applying (6.22) leads to
control estimate (6.11). The estimates in (6.12) and (6.14) can be readily derived using the stability
results (4.17) and (5.6) and the a priori estimates (4.26), (5.8) and (6.11) after decomposing them as
follows:
ū − ūh = ū − uh (Ū ) + uh (Ū ) − ūh and ȳ − ȳ h = ȳ − y h (Ū ) + y h (Ū ) − ȳ h ,

φ̄ − φ̄h = φ̄ − φh (Ū ) + φh (Ū ) − φ̄h and η̄ − η̄ h = η̄ − η h (Ū ) + η h (Ū ) − η̄ h .
To derive the pressure estimates we decompose them as follows:
∥p − ph ∥0,Ω ≤ ∥p(U ) − ph (U )∥0,Ω + ∥ph (U ) − ph ∥0,Ω ,

(6.25)
∥ξ − ξh ∥0,Ω ≤ ∥ξ(U ) − ξh (U )∥0,Ω + ∥ξh (U ) − ξh ∥0,Ω .
Now the estimates (6.13) and (6.15) can be derived using (4.27) and (5.9), respectively on the first term(s)
on the right hand side of (6.25), and the remaining terms can be estimated similar to (4.37) and (5.17),
respectively.
7. Numerical Experiments. In this section we provide a set of examples which verify the convergence
rates proved in Theorem 6.11, and present two application oriented examples. A primal dual active set
strategy as a semi-smooth Newton (SSN) method (see for more details [16, 32, 33]) is employed to
determine the optimal control from the discrete optimality system (6.2)-(6.4). The resulting linearized
system is solved using the linear solver MUMPS in open source finite element library FEniCS [34]. The
essential boundary conditions are implemented using the Dirichlet boundary class feature of FEniCS and
the zero mean value for the pressure approximation is tackled using a Lagrange multiplier approach.
7.1. Example 1: Accuracy test. Let the computational domain be Ω := [0, 1]×[0, 1] and consider
a sequence of uniformly refined meshes on Ω with mesh parameter h. We take the buoyancy term of
the form F (y) = (T + Nr S)g, where Nr is the solutal to the thermal buoyancy ratio, and the viscosity
is chosen to be of exponential form, that is, ν(T ) = ν2 exp(−T ). The other parameters are defined as
follows: g = (0, 1)⊤ , K −1 = σI, D = 1000I. We use the following smooth functions as the closed form
solutions of (1.1)-(1.2) in our accuracy test:
⊤
u(x, y) = (sin(πx) cos(πy), − cos(πx) sin(πy)) , p(x, y) = cos(πx) exp(y), T (x, y) = 0.5 + 0.5 cos(xy),
S(x, y) = 0.1 + 0.3 exp(xy), φ(x, y) = sin(πx)2 sin(πy) cos(πy), − sin(πy)2 sin(πx) cos(πx) ,

ζ(x, y) = cos(πy) exp(x), η T (x, y) = 0.5 cos(xy)x(x − 1)y(y − 1), η S (x, y) = 0.5 exp(xy)x(x − 1)y(y − 1),

1
Uj (x, y) = min 0.25, max −0.1, − φj (x, y) , λ = 1.
λ
The boundary conditions, additional external forces and source terms are prescribed according to the
above prescribed manufactured solutions. The errors for the state variables and control in their natural
norms and the corresponding convergence rates are denoted as follows:
eu = ∥u − uh ∥1,T nc , ep = ∥p − ph ∥0,Ω , eT = ∥T − Th ∥1,T nc , eS = ∥S − Sh ∥1,T nc ,

h h h
eU 1 = ∥U1 − Uh1 ∥Lr (Ω) , eU 2 = ∥U2 − Uh2 ∥Lr (Ω) , rate = log(e(·) /ẽ(·) )(log(h/h̃))−1 ,
where e, ẽ denote the errors generated on two consecutive mesh refinements of mesh size h and h̃,
respectively. The error in control is measured in the x and y components separately and the errors in
adjoint variables are denoted by eφ , eζ , eηT , eηS , which are defined analogously to the state variable errors.
The errors and convergence rates focusing on cases where the viscosity and permeability coefficients scale
differently, changing from Flow to Darcy regimes are reported in Table 1, where It denotes the number
of iterations the semi-smooth Newton algorithm takes to reach the absolute tolerance of 10−6 and χ(·)
denotes the degrees of freedom associated with the space (·). The tabulated errors confirm the theoretical
bounds of Theorem 6.11 and that the discrete velocities are divergence free.
Flow regime (ν2 = 1, σ = 1)

χV cr eu rate ∥divuh ∥∞ eφ rate ∥divφh ∥∞ χQ0 ep rate eξ rate It
h h
66 2.7363 - 2.6E-15 2.2562 - 2.4E-15 18 0.3532 - 0.4113 - 6
240 1.7111 0.68 4.4E-15 1.0684 1.08 1.3E-15 72 0.1845 0.94 0.2175 0.92 6
912 0.9578 0.84 3.5E-15 0.5089 1.07 1.5E-15 288 0.0892 1.05 0.0984 1.14 6
3552 0.5048 0.92 5.9E-15 0.2452 1.05 2.9E-15 1152 0.0431 1.05 0.0462 1.09 5
14016 0.2584 0.97 1.3E-14 0.1200 1.03 7.9E-15 4608 0.0212 1.02 0.0228 1.02 5
55680 0.1307 0.98 2.9E-14 0.0593 1.02 1.8E-14 18432 0.0105 1.01 0.0117 1.0 5
eT rate eS rate eη T rate eηS rate χU h,ad eU 1 rate eU 2 rate
1.6757 - 2.8265 - 0.9276 - 1.3509 - 18 0.0785 - 0.0826 -
0.8509 0.98 1.4152 1.0 0.4754 0.96 0.7101 0.93 72 0.0321 1.29 0.308 1.42
0.4281 0.99 0.7080 1.0 0.2392 0.99 0.3597 0.98 288 0.0151 1.08 0.0148 1.06
0.2144 1.0 0.2144 1.0 0.1197 1.0 0.1804 1.0 1152 0.0075 1.01 0.0074 0.99
0.1073 1.0 0.1770 1.0 0.0599 1.0 0.0903 1.0 4608 0.0037 1.02 0.0037 1.01
0.0536 1.0 0.0885 1.0 0.0299 1.0 0.0452 1.0 18432 0.0018 1.0 0.0018 1.0
Stokes regime (ν2 = 10, σ = 0)
h h
66 3.6250 - 1.33E-15 4.5523 - 4.7E-15 18 0.8079 - 1.2520 - 5
240 1.8689 0.96 4.4E-15 2.4415 0.9 1.3E-15 72 0.4301 0.91 0.7032 0.83 5
912 0.9427 0.99 2.9E-15 1.2424 0.97 1.7E-15 288 0.2159 0.99 0.3409 1.04 5
3552 0.4726 1.0 5.7E-15 0.6234 0.99 3.3E-15 1152 0.1077 1.0 0.1670 1.03 5
14016 0.2365 1.0 1.3E-14 0.3118 1.0 7.9E-15 4608 0.0538 1.0 0.0829 1.01 5
55680 0.1182 1.0 2.7E-14 0.1559 1.0 1.8E-14 18432 0.0269 1.0 0.0414 1.0 5
1.6757 - 2.8265 - 0.9276 - 1.3509 - 18 0.0597 - 0.0595 -
0.8509 0.98 1.4152 1.0 0.4754 0.96 0.7101 0.93 72 0.0279 1.09 0.0279 1.09
0.4281 0.99 0.7080 1.0 0.2392 0.99 0.3597 0.98 288 0.0145 0.95 0.0145 0.95
0.2144 1.0 0.3541 1.0 0.1197 1.0 0.1804 1.0 1152 0.0074 0.97 0.0074 0.97
0.1073 1.0 0.1770 1.0 0.0599 1.0 0.0903 1.0 4608 0.0037 1.0 0.0037 1.0
0.0536 1.0 0.0885 1.0 0.0299 1.0 0.0452 1.0 18432 0.0018 1.0 0.0018 1.0
Darcy regime (ν2 = 1, σ = 10000)
h h
66 16.401 - 8.8E-16 15.457 - 5.5E-16 18 87.600 - 42.188 - 4
240 7.3961 1.15 1.8E-14 7.6961 1.01 1.3E-15 72 56.757 0.63 44.241 -0.07 4
912 3.4361 1.11 3.5E-15 3.7037 1.06 1.5E-15 288 31.281 0.86 29.079 0.61 4
3552 1.6473 1.06 6.6E-15 1.7601 1.07 3.4E-15 1152 16.265 0.94 15.941 0.87 4
14016 0.8052 1.03 1.3E-14 0.8466 1.06 8.8E-15 4608 8.2737 0.98 8.2298 0.95 4
55680 0.3978 1.02 2.7E-14 0.4133 1.03 1.9E-14 18432 4.1700 0.99 4.1642 0.98 4
1.6757 - 2.8265 - 0.9276 - 1.3509 - 18 0.0617 - 0.0617 -
0.8509 0.98 1.4152 1.0 0.4754 0.96 0.7101 0.93 72 0.0285 1.11 0.0285 1.11
0.4281 0.99 0.7080 1.0 0.2392 0.99 0.3597 0.98 288 0.0146 0.96 0.0146 0.96
0.2144 1.0 0.2144 1.0 0.1197 1.0 0.1804 1.0 1152 0.0074 0.98 0.0074 0.98
0.1073 1.0 0.1770 1.0 0.0599 1.0 0.0903 1.0 4608 0.0037 1.01 0.0037 1.01
0.0536 1.0 0.0885 1.0 0.0299 1.0 0.0452 1.0 18432 0.0019 1.0 0.0019 1.0
Table 1: Experimental errors and order of convergence in state, adjoint and control variables in their natural
norms under Flow, Stokes and Darcy regimes, respectively.
Remark 7.1. In the Darcy regime to handle the reaction domination we

R amodify our scheme by adding
the following penalty in the discrete bilinear form ah (·; ·, ·):
P 0
e∈E(Th ) e he ν2 [[u h ]] : [[v h ]] ds, with the
√
stabilization parameter a0 = 10 σ and calculate the error in the modified natural norm
X 2 2
X 1
∥v∥1,T nc = σ ∥v∥0,K + ν2 ∥∇v∥0,K + ∥[[v]]∥20,e , ∀ v ∈ H 1 (Th ).
h he
K∈Th e∈E(Th )
7.2. Example 2: Optimal control of Soret and Dufour effects in a porous cavity. Using
the dimensionless variables described in [1, Section 5.2], we can write the equations describing transport
phenomena in a square porous cavity with thermal (T ) and concentration (C) diffusion in the form
(1.2) and it’s optimal control will be described by (1.1)-(1.2). Set K = DaI, ν(T ) = 1 and F (y) =
(GrT T + GrC C)g, where g = (0, −1)⊤ points in the direction of gravity with y = (T, C)⊤ . The diffusion
coefficient D is given by, D = [Rk /P r Du; Sr 1/Sc] , where Rk is the thermal conductivity ratio,
GrT , GrC are the thermal and solutal Grashof numbers respectively, Da = κ/H 2 is the Darcy number,
P r = ν/α the Prandtl number, Sc = ν/DC the Schmidt number, and the ratio Le = Sc/P r is the Lewis
number. We perform a computational test to show that (1.1)-(1.2) steers the velocity (u), Temperature
(T ) and Concentration (C) to the desired states ud , Td and Cd , respectively. The computational domain,
Ω is the unit square with no-slip velocity conditions on Γ. Temperature and concentration are kept at
T0 , C0 and T1 , C1 at the right and left walls, respectively, T0 < T1 and C0 < C1 . The horizontal walls are
adiabatic and impermeable. The corresponding adjoint equation also has no-slip velocity conditions on Γ
and the adjoint temperature and concentration have no-slip boundary conditions on the right and left walls
with adiabatic and impermeable horizontal walls. The mesh used to discretise the computational domain
has 150 × 150 elements and the relative tolerance used for the termination of the semi-smooth Newton
algorithm is 1−6 . For the values λ = 1, Rk = 1.0, Le = 10, Le = 0.8, N = 1, T0 , C0 = −1, T1 , C1 = 1,
we set the Rayleigh value Ra = GrT P rDa as 100 and the Soret and Dufour parameters as Sr = 0 and
Du = 0.1, respectivley (cf. [35, 36]). The optimal solution profiles steering towards the desired states
ud = 0, Td = 0, Cd = 0 and Da = 10−3 , Uaj = −0.005, Ubj = 0.005 and Da = 10−7 , Uaj = −5−6 , Ubj = 56
are reported in Figures 1 and 2, respectively.
Acknowledgments. The authors greatly acknowledge the funding from SERB-CRG India (Grant
Number CRG/2021/002569). The first author thanks Indian Institute of Technology, Roorkee, India
where majority of the work was done during his stay there.
REFERENCES
[1] R. Bürger, P. E. Méndez, and R. Ruiz-Baier, “On H(div)-conforming methods for double-diffusion equations in porous
media,” SIAM Journal on Numerical Analysis, vol. 57, no. 3, pp. 1318–1343, 2019.
[2] J. Tushar, A. Khan, and M. T. Mohan, “Optimal control of stationary doubly diffusive flows on two and three
dimensional bounded lipschitz domains: A theoretical study,” arXiv:2308.02178, 2023.
[3] S. A. Lorca and J. L. Boldrini, “Stationary solutions for generalized boussinesq models,” Journal of Differential
equations, vol. 124, no. 2, pp. 389–406, 1996.
[4] A. Allendes, G. R. Barrenechea, and C. Naranjo, “A divergence-free low-order stabilized finite element method for a
generalized steady state Boussinesq problem,” Computer Methods in Applied Mechanics and Engineering, vol. 340,
pp. 90–120, 2018.
[5] M. Alvarez, G. N. Gatica, B. Gomez-Vargas, and R. Ruiz-Baier, “New mixed finite element methods for natural
convection with phase-change in porous media,” Journal of Scientific Computing, vol. 80, no. 1, pp. 141–174,
2019.
[6] R. Oyarzúa, T. Qin, and D. Schötzau, “An exactly divergence-free finite element method for a generalized Boussinesq
problem,” IMA Journal of Numerical Analysis, vol. 34, no. 3, pp. 1104–1135, 2014.
[7] R. Temam, Navier-Stokes equations. Theory and numerical analysis. Studies in Mathematics and its Applications,
Vol. 2, North-Holland Publishing Co., Amsterdam-New York-Oxford, 1977.
Fig. 1: Optimal velocity, Temperature, concentration and control profiles, respectively under Da = 10−3 and
Uaj = −0.005, Ubj = 0.005. The SSN algorithm terminates in 7 iterations.
Fig. 2: Optimal velocity, Temperature, concentration and control profiles, respectively under Da = 10−7 and
Uaj = −5−6 , Ubj = 5−6 . The SSN algorithm terminates in 8 iterations.
[8] S. Ravindran, “Design of optimal boundary control for nonstationary doubly diffusive convective flow,” in 2012
American Control Conference (ACC), pp. 1442–1447, IEEE, 2012.
[9] H.-C. Lee and O. Y. Imanuvilov, “Analysis of Neumann boundary optimal control problems for the stationary
Boussinesq equations including solid media,” SIAM Journal on Control and Optimization, vol. 39, no. 2, pp. 457–
477, 2000.
[10] H.-C. Lee, “Analysis and computational methods of Dirichlet boundary optimal control problems for 2D Boussinesq
equations,” vol. 19, pp. 255–275, 2003. Challenges in computational mathematics (Pohang, 2001).
[11] K. Ito and S. S. Ravindran, “Optimal control of thermally convected fluid flows,” SIAM Journal on Scientific
Computing, vol. 19, no. 6, pp. 1847–1869, 1998.
[12] F. Tröltzsch and D. Wachsmuth, “Second-order sufficient optimality conditions for the optimal control of Navier-Stokes
equations,” ESAIM. Control, Optimisation and Calculus of Variations, vol. 12, no. 1, pp. 93–119, 2006.
[13] S. C. Brenner, “Forty years of the Crouzeix-Raviart element,” Numerical Methods for Partial Differential Equations.
An International Journal, vol. 31, no. 2, pp. 367–396, 2015.
[14] B. Cockburn, G. Kanschat, and D. Schotzau, “A locally conservative LDG method for the incompressible Navier-Stokes
equations,” Mathematics of Computation, vol. 74, no. 251, pp. 1067–1095, 2005.
[15] T. Kashiwabara, I. Oikawa, and G. Zhou, “Penalty method with Crouzeix-Raviart approximation for the Stokes
equations under slip boundary condition,” ESAIM. Mathematical Modelling and Numerical Analysis, vol. 53,
no. 3, pp. 869–891, 2019.
[16] M. Hintermüller, K. Ito, and K. Kunisch, “The primal-dual active set strategy as a semismooth Newton method,”
SIAM Journal on Optimization, vol. 13, no. 3, pp. 865–888 (2003), 2002.
[17] F. Tröltzsch, Optimal control of partial differential equations: theory, methods, and applications, vol. 112. American
Mathematical Soc., 2010.
[18] M. Dauge, “Stationary Stokes and Navier-Stokes systems on two- or three-dimensional domains with corners. I.
Linearized equations,” SIAM Journal on Mathematical Analysis, vol. 20, no. 1, pp. 74–97, 1989.
[19] G. Savaré, “Regularity results for elliptic equations in Lipschitz domains,” Journal of Functional Analysis, vol. 152,
no. 1, pp. 176–201, 1998.
[20] P. G. Ciarlet, Linear and nonlinear functional analysis with applications. Society for Industrial and Applied
Mathematics, Philadelphia, PA, 2013.
[21] D. Schötzau and L. Zhu, “A robust a-posteriori error estimator for discontinuous Galerkin methods for convection-
diffusion equations,” Applied Numerical Mathematics. An IMACS Journal, vol. 59, no. 9, pp. 2236–2255, 2009.
[22] M. Ainsworth, “Robust a posteriori error estimation for nonconforming finite element approximation,” SIAM Journal
on Numerical Analysis, vol. 42, no. 6, pp. 2320–2341, 2005.
[23] A. Buffa and C. Ortner, “Compact embeddings of broken Sobolev spaces and applications,” IMA Journal of Numerical
Analysis, vol. 29, no. 4, pp. 827–855, 2009.
[24] O. A. Karakashian and W. N. Jureidini, “A nonconforming finite element method for the stationary Navier-Stokes
equations,” SIAM Journal on Numerical Analysis, vol. 35, no. 1, pp. 93–120, 1998.
[25] S. C. Brenner, L.-Y. Sung, and Y. Zhang, “A quadratic C 0 interior penalty method for an elliptic optimal control
problem with state constraints,” in Recent developments in discontinuous Galerkin finite element methods for
partial differential equations, vol. 157 of IMA Vol. Math. Appl., pp. 97–132, Springer, Cham, 2014.
[26] V. Girault, G. Kanschat, and B. Rivière, “Error analysis for a monolithic discretization of coupled Darcy and Stokes
problems,” Journal of Numerical Mathematics, vol. 22, no. 2, pp. 109–142, 2014.
[27] A. Ern and J.-L. Guermond, Theory and practice of finite elements, vol. 159 of Applied Mathematical Sciences.
Springer-Verlag, New York, 2004.
[28] V. John, G. Matthies, F. Schieweck, and L. Tobiska, “A streamline-diffusion method for nonconforming finite
element approximations applied to convection-diffusion problems,” Computer Methods in Applied Mechanics and
Engineering, vol. 166, no. 1-2, pp. 85–97, 1998.
[29] S. C. Brenner and L. R. Scott, The mathematical theory of finite element methods, vol. 15 of Texts in Applied
Mathematics. Springer, New York, third ed., 2008.
[30] E. Casas and F. Tröltzsch, “Error estimates for linear-quadratic elliptic control problems,” in Analysis and optimization
of differential systems (Constanta, 2002), pp. 89–100, Kluwer Acad. Publ., Boston, MA, 2003.
[31] N. Arada, E. Casas, and F. Tröltzsch, “Error estimates for the numerical approximation of a semilinear elliptic control
problem,” Computational Optimization and Applications, vol. 23, pp. 201–229, 2002.
[32] K. Ito and K. Kunisch, Lagrange multiplier approach to variational problems and applications, vol. 15 of Advances in
Design and Control. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 2008.
[33] J. C. De los Reyes, Numerical PDE-constrained optimization. SpringerBriefs in Optimization, Springer, Cham, 2015.
[34] M. Alnæs, J. Blechta, J. Hake, A. Johansson, B. Kehlet, A. Logg, C. Richardson, J. Ring, M. E. Rognes, and G. N.
Wells, “The fenics project version 1.5,” Archive of numerical software, vol. 3, no. 100, 2015.
[35] B. Goyeau, J.-P. Songbe, and D. Gobin, “Numerical study of double-diffusive natural convection in a porous cavity
using the darcy-brinkman formulation,” International Journal of Heat and Mass Transfer, vol. 39, no. 7, pp. 1363–
1378, 1996.
[36] A. Çı bık and S. Kaya, “Finite element analysis of a projection-based stabilization method for the Darcy-Brinkman
equations in double-diffusive convection,” Applied Numerical Mathematics. An IMACS Journal, vol. 64, pp. 35–
49, 2013.
.
View publication stats

Optimal Control of Stationary Doubly Diffusive Flows On Two and Three Dimensional Bounded Lipschitz Domains: Numerical Analysis

Uploaded by

Copyright:

Available Formats

Optimal Control of Stationary Doubly Diffusive Flows On Two and Three Dimensional Bounded Lipschitz Domains: Numerical Analysis

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Optimal Control of Stationary Doubly Diffusive Flows On Two and Three Dimensional Bounded Lipschitz Domains: Numerical Analysis

Uploaded by

Copyright:

Available Formats

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

OPTIMAL CONTROL OF STATIONARY DOUBLY DIFFUSIVE FLOWS ON TWO AND

Preprint · March 2024

Arbaz Khan Manil T. Mohan

SEE PROFILE SEE PROFILE

The user has requested enhancement of the downloaded file.

AMS subject classifications. 65N30, 65N50, 74F99, 74A50, 76S05.

1. Introduction. We consider the following bilaterally constrained flow control problem:

a(y; u, v) + c(u, u, v) + b(v, p) = d(y, v) + (U , v)r,r′ ∀ v ∈ H 10 (Ω),

The involved forms in (2.2) are defined as follows:

Let X denote the kernel of b(·, ·), namely,

X := v ∈ H 10 (Ω) : b(v, q) = 0 ∀ q ∈ L20 (Ω) = v ∈ H 10 (Ω) : div v = 0 in Ω .

(2.4) c(w, v, v) = 0, c(w, u, v) = −c(w, v, u), and c(u, v, w) = ((∇v)⊤ w, u).

Furthermore, the following equivalence result holds (weak sense):

(2.6) y 1 ∈ [H 1 (Ω)]2 with y 1 |Γ = y D .

(2.7) ∥u∥1,Ω ≤ Cu y D 1/2,Γ + ∥U ∥Lr (Ω) := Mu , ∥y∥1,Ω ≤ Cy y D 1/2,Γ + ∥U ∥Lr (Ω) := My .

the following stability result also holds under (2.9):

where, Cφ and Cη are positive constants provided

where p11 + p12 = 3−2δ

Now an application of (2.7) and Theorem 2.5 results to

∥φ̄∥3/2+δ,Ω ≤ ∥F ∥−1/2+δ,Ω ≤ C(ū, ȳ, y D , Ū , ud , y d ).

Applying fractional GN after choosing p2 = 6 implies p1 = 6/(3 + 2δ) gives

∥(1/ν(T ))(∇ū)⊤ φ̄∥−1/2−δ,Ω ≤ C∥ū∥1−δ,Ω ∥φ̄∥1,Ω ≤ C∥ū∥1,Ω ∥φ̄∥1,Ω .

∥φ̄∥3/2−δ,Ω ≤ ∥F ∥−1/2−δ,Ω ≤ C(ū, ȳ, y D , Ū , ud , y d ).

∥(1/ν(T ))(∇ū)⊤ φ̄∥−1/2+δa ,Ω ≤ C∥ū∥W 1,p1 (Ω) ∥φ̄∥Lp2 (Ω) ,

∥(1/ν(T ))(∇ū)⊤ φ̄∥−1/2+δa ,Ω ≤ C∥ū∥3/2+δa ,Ω ∥φ̄∥1,Ω .

∥(1/ν(T ))∇ν(T̄ ) · ∇ū∥−1/2+δa ,Ω ≤ C ∥ū∥3/2−δa ,Ω T̄ 3/2+2δa ,Ω

∥(1/ν(T ))∇ν(T̄ ) · ∇ū∥−1/2+δb ,Ω ≤ C ∥ū∥3/2+δa ,Ω T̄ 3/2+(δb −δa ),Ω

which is finite, leading to the desired regularity.

L : X × [H 1 (Ω)]2 × L2 (Ω) × X × [H01 (Ω)]2 −→ R such that

In the next Lemma, we discuss the Fréchet-differentiability of L.

together with the adjoint state Λ̄ = (φ̄, η̄)

|Luu (w̄, Λ̄) [ζ 1 , ζ 2 ] | ≤ CL1 ∥ζ 1 ∥1,Ω ∥ζ 2 ∥1,Ω ,

|Lyy (w̄, Λ̄) [µ1 , µ2 ] | ≤ CL2 ∥µ1 ∥1,Ω ∥µ2 ∥1,Ω ,

(3.4) Luu (w̄, Λ̄)[ζ 1 , ζ 2 ] = (ζ 1 , ζ 2 ) − c(ζ 1 , ζ 2 , φ̄) − c(ζ 2 , ζ 1 , φ̄),

where µi := (µTi , µSi ) and the following estimates follow:

and the boundedness follows from Lemma 2.11 and,

(3.9) ≤ CFyyy Mφ ∥µ1 ∥1,Ω ∥µ2 ∥1,Ω ∥µ3 ∥1,Ω ,

holds for all admissible triplets w = (u, y, U ) with U − Ū Ls (Ω)

(3.12) J(w̄) = L(w̄, Λ̄) and J(w) = L(w, Λ̄).

Taylor-expansion of the Lagrangian around w̄ yields

(3.15) LU (w̄, Λ̄)(U − Ū ) ≥ ε U − Ū L1,Ωε

Then U − Ū = (U − Ũ ) + (Ũ − Ū ), such that h := Ũ − Ū = 0 on Ωε,i which is the assumption of (SSC).

In order to apply (SSC), we split ζ = u − ū and µ = y − ȳ as follows: ζ = uh + ur and µ = y h + y r .

(3.19) |⟨Ru , v⟩| ≤ C∥U − Ū ∥2Lr (Ω) ∥v∥1,Ω .

Similarly, we can get the following estimates for Ry :

Substituting (3.23)-(3.25) in (3.22) results to

Using Lemma 3.1 along with (3.18) and (3.21) gives

Focusing on z̃, by definition of our constructed control Ũ , (U − Ũ )i = 0 on Ω\Ωε,i and (Ũ − Ū )i = 0 on

(3.33) z̃ 2 ≤ ∥U − Ũ ∥L1 (Ω) ∥U − Ũ ∥Ls (Ω) = U − Ū L1,Ωε

4.1.1. Discrete boundedness properties. The space V cr h

Lemma 4.1. There holds

|ah (y h ; uh , v h )| ≤ Canc ∥uh ∥1,T nc ∥v h ∥1,T nc ∀ uh , v h ∈ V cr

|ch (wh , uh , v h )| ≤ Cenc ∥wh ∥1,T nc ∥uh ∥1,T nc ∥v h ∥1,T nc ∀ wh , uh , v h ∈ V cr

(4.3) ∥v∥Lr (Ω) ≤ Cenc ∥v∥1,T nc ∀ v ∈ H 1 (Th ).

Using the inverse inequality,

Moreover, the following coercivity properties of ah and ahy also hold:

(4.10) y h = y h,0 + y h,1 with y h,0 ∈ Mcr cr D

∥uh ∥1,T nc ≤ Cunc yD + ∥U ∥Lr (Ω) := Munc , ≤ Cync yD