Journal of Scientific Computing (2019) 80:1577–1606

https://doi.org/10.1007/s10915-019-00990-7

Analysis and Approximation of a Vorticity–Velocity–Pressure

Formulation for the Oseen Equations

V. Anaya1 · A. Bouharguane2 · D. Mora1,3 · C. Reales4 · R. Ruiz-Baier5 ·

N. Seloula6 · H. Torres7

Received: 29 August 2018 / Revised: 18 April 2019 / Accepted: 11 June 2019 / Published online: 22 June 2019
© The Author(s) 2019

Abstract
We introduce a family of mixed methods and discontinuous Galerkin discretisations designed
to numerically solve the Oseen equations written in terms of velocity, vorticity, and Bernoulli
pressure. The unique solvability of the continuous problem is addressed by invoking a global
inf-sup property in an adequate abstract setting for non-symmetric systems. The proposed
finite element schemes, which produce exactly divergence-free discrete velocities, are shown
to be well-defined and optimal convergence rates are derived in suitable norms. This mixed
finite element method is also pressure-robust. In addition, we establish optimal rates of
convergence for a class of discontinuous Galerkin schemes, which employ stabilisation. A
set of numerical examples serves to illustrate salient features of these methods.

Keywords Oseen equations · Vorticity-based formulation · Mixed finite elements · Exactly

divergence-free velocity · Discontinuous Galerkin schemes · Numerical fluxes · A priori
error bounds

Mathematics Subject Classification 65N30 · 65N12 · 76D07 · 65N15

1 Introduction

The Oseen equations stem from linearisation of the steady (or alternatively from the backward
Euler time-discretisation of the transient) Navier–Stokes equations. Of particular appeal to
us is their formulation in terms of fluid velocity, vorticity vector, and pressure. A diversity of
discretisation methods is available to solve incompressible flow problems using these three
fields as principal unknowns. Some recent examples include spectral elements [3,9,30], cell-
based pressure schemes [12], as well as stabilised and least-squares schemes [2,10] for
Navier–Stokes; also several mixed and augmented methods for Brinkman [4–6,8], and a
number of other discretisations specifically designed for Stokes flows [7,24,25,27,34].
Both the implementation and the analysis of numerical schemes for Navier–Stokes are typ-
ically based on the Oseen linearisation. A few related contributions (not only restricted to the
velocity–pressure formulation) include for instance [11], that presents a least-squares method
for Navier–Stokes equations with vorticity-based first-order reformulation, and whose analy-

Extended author information available on the last page of the article

123
1578 Journal of Scientific Computing (2019) 80:1577–1606

sis appeals to the elliptic theory of Agmon–Douglas–Nirenberg, and the proposed conforming
finite element methods exhibit optimal order of accuracy for diverse boundary conditions. We
also mention the non-conforming exponentially accurate least-squares spectral method for
Oseen equations proposed in [32], where a suitable preconditioner is also proposed. In [36]
the authors introduce a velocity–vorticity-pressure least-squares finite element method for
Oseen and Navier–Stokes equations with velocity boundary conditions. They derive error
estimates and reported a degeneracy of the convergence for large Reynolds numbers. A
least-squares minimisation problem based on the stress–velocity–pressure formulation was
introduced in [15]. The study shows that the corresponding homogeneous least-squares func-
tional is elliptic and continuous in suitable norms. Several first-order Oseen-type systems are
analysed in [17], also including vorticity and pressure in the formulation.
Discontinuous Galerkin (DG) methods have also been used to solve the Oseen problem,
as for example, in [20,21] for the case of Dirichlet boundary conditions. Compared with
conforming finite elements, discretisations based on DG methods have a number of attractive,
and well-documented features. These include high order accuracy, being amenable for hp-
adaptivity, relatively simple implementation on highly unstructured meshes, and superior
robustness when handling rough coefficients. We also mention the a priori error analysis
of hybridisable DG schemes introduced in [18] for the Oseen equations. The family of DG
methods we propose here has resemblance with those schemes, but concentrates on a three-
field formulation described below.
This paper is concerned with mixed non-symmetric variational problems which will be
analysed using a global inf-sup argument. To do this, we conveniently restrict the set of equa-
tions to the space of divergence-free velocities, and apply results from [26, Sect. 2] in order
to prove that the equivalent resulting non-symmetric saddle-point problem is well-posed. For
the numerical approximation, we first consider Raviart–Thomas elements of order k ≥ 0 for
the velocity field, Nédélec elements of order k for the vorticity, and piecewise polynomi-
als of degree k without continuity restrictions, for the Bernoulli pressure. We prove unique
solvability of the discrete problem by adapting the same tools utilised in the analysis of the
continuous problem. In addition, the proposed family of Galerkin finite element methods
turns out to be optimally convergent, under the common assumptions of enough regularity
of the exact solutions to the continuous problem. An appealing feature of this mixed finite
element method is that it produces exactly divergence-free approximations of the velocity by
construction; thus preserving, at the discrete level, an essential constraint of the governing
equations. Next, inspired by the methods presented in [21,22], we present another scheme
involving the discontinuous Galerkin discretisation of the curl-curl and grad-div operators.
An advantage is related to the robustness with respect to rough coefficients and the relax-
ation of inter-element continuity conditions. We prove the well-posedness of the DG scheme
and derive error estimates under appropriate regularity assumptions. The novelties of our
contribution are mainly related to the formulation of Oseen equations in terms of vorticity,
in considering H(div)-conforming velocities, in proposing mixed finite element and discon-
tinuous Galerkin schemes based on such formulation, and in establishing their convergence
properties.
We have structured the contents of the paper in the following manner. Notation-related
preliminaries are stated in the remainder of this section. We then present the model problem
as well as the three-field weak formulation and its solvability analysis in Sect. 2. The finite
element discretisation is constructed in Sect. 3, where we also derive the stability and conver-
gence bounds. In Sect. 4, we present the mixed DG formulation for the model problem. The
well-posedness of the method and the error analysis are established in the same section. We

123
Journal of Scientific Computing (2019) 80:1577–1606 1579

close in Sect. 5 with a set of computational tests that illustrate the properties of the proposed
numerical schemes in a variety of scenarios.
Let Ω be a bounded domain of R3 with Lipschitz boundary ∂Ω. Moreover, we assume
that ∂Ω admits a disjoint partition ∂Ω = Γ ∪ Σ. For any s ≥ 0, the symbol ·s,Ω denotes
the norm of the Hilbertian Sobolev spaces Hs (Ω) or Hs (Ω)3 , with the usual convention
H0 (Ω) := L2 (Ω). For s ≥ 0, we recall the definition of the space

 
Hs (curl; Ω) := θ ∈ Hs (Ω)3 : curl θ ∈ Hs (Ω)3 ,

that we will endow with the norm θ 2Hs (curl;Ω) = θ 2s,Ω + curl θ 2s,Ω , and will denote
H(curl; Ω) = H0 (curl; Ω).

2 Statement and Solvability of the Continuous Problem

2.1 Oseen Problem in Terms of Velocity–Vorticity–Pressure

A standard backward Euler time-discretisation of the classical Navier–Stokes equations, or a

linearisation of the steady version of the problem combined with standard curl-div identities,
leads to the following set of equations, known as the Oseen equations:

σ u − νΔu + curl u × β + ∇ p = f in Ω,
(2.1)
div u = 0 in Ω,

where ν > 0 is the kinematic fluid viscosity, σ > 0 is inversely proportional to the time-
step, β is an adequate approximation of velocity to be made precise below, and the vector
of external forces f also absorbs the contributions related to previous time steps, or to fixed
states in the linearisation procedure of the steady Navier–Stokes equations. As usual in this
context, in the momentum equation we have conveniently introduced the Bernoulli (also
known as dynamic) pressure p := P + 21 |u|2 , where P is the actual fluid pressure. √
The structure of (2.1) suggests to introduce the rescaled vorticity vector ω := ν curl u
as a new unknown. Furthermore, in this study we focus on the case of zero normal velocities
and zero tangential vorticity trace imposed on a part of the boundary Γ ⊂ ∂Ω, whereas a
non-homogeneous tangential velocity uΣ and a fixed Bernoulli pressure pΣ are set on the
remainder of the boundary Σ = ∂Ω \ Γ . Therefore, system (2.1) can be recast in the form

√ √
σu + ν curl ω + ν −1/2 ω × β + ∇ p = f , ω − ν curl u = 0, and div u = 0 in Ω,
u·n=0 and ω×n=0 on Γ ,
p = pΣ and u × n = uΣ on Σ, (2.2)

where n stands for the outward unit normal on ∂Ω. Should the boundary Σ have zero
measure, the additional condition ( p, 1)Ω,0 = 0 is required to enforce uniqueness of the
Bernoulli pressure. Note that even if the formulation and its analysis are restricted to the
case of constant coefficients, the theory might be extended to the case of piecewise constant
model coefficients.

123
1580 Journal of Scientific Computing (2019) 80:1577–1606

2.2 Defining a Weak Formulation

Let us introduce the following functional spaces

H := {v ∈ H(div; Ω) : v · n = 0 on Γ },
Z := {θ ∈ H(curl; Ω) : γt (θ ) = 0 on Γ }, and Q := L2 (Ω),

where γt is the tangential trace operator on Γ , defined by: γt (θ ) = θ × n. Let us endow

H and Q with their natural norms. For the space Z however, we consider the following
viscosity-weighted norm:
 1/2
θ Z := θ20,Ω + ν curl θ20,Ω .

From now on, we will assume that the data are regular enough: β ∈ L∞ (Ω)3 and f ∈
L2 (Ω)3 .We proceed to test (2.2) against adequate functions and to impose the boundary
conditions in such a manner that we end up with the following formulation: Find (u, ω, p) ∈
H × Z × Q such that

a(u, v) + b1 (v, ω) + b2 (v, p) + c(ω, v) = F(v) ∀v ∈ H,

b1 (u, θ ) − d(ω, θ ) = G(θ) ∀θ ∈ Z,
b2 (u, q) = 0 ∀q ∈ Q, (2.3)

where the bilinear forms a : H×H → R, b1 : H×Z → R, d : Z ×Z → R, b2 : H×Q → R,

c : Z × H → R, and the linear functionals F : H → R, and G : Z → R are specified as
follows
 
√
a(u, v) := σ u · v d x, b1 (v, θ ) := ν curl θ · v d x,
Ω Ω

b2 (v, q) := − q div v d x,
Ω
 
1
d(ω, θ ) := ω · θ d x, c(θ , v) := √ (θ × β) · v d x,
Ω ν Ω

√
F(v) := f · v d x − v · n, pΣ Σ , G(θ) := − νuΣ , θ Σ ,
Ω

for all u, v ∈ H, ω, θ ∈ Z, and q ∈ Q.

The continuity of these bilinear and linear functionals is stated in the following lemma,
whose proof is obtained by rather standard arguments. In particular, the estimate for c(·, ·)
is proven using the assumption on β and the fact that θ × β0,Ω ≤ 2β∞,Ω θ 0,Ω .

Lemma 1 The following estimates hold true,

|a(u, v)| ≤ σ uH vH , (2.4)

|b1 (v, θ )| ≤ vH θZ , (2.5)
|b2 (v, q)| ≤ vH q0,Ω , (2.6)
|d(ω, θ )| ≤ ωZ θ Z , (2.7)
2β∞,Ω
|c(θ , v)| ≤ √ θ Z vH , (2.8)
ν

123
Journal of Scientific Computing (2019) 80:1577–1606 1581

|F(v)| ≤ C( f 0,Ω +  pΣ 1/2,Σ )vH , (2.9)

|G(θ )| ≤ CuΣ −1/2,Σ θ Z , (2.10)
where C is a positive constant.

2.3 Solvability Analysis

In order to analyse the variational formulation (2.3), let us introduce the kernel of the bilinear
form b2 (·, ·) and its classical characterisation
X := {v ∈ H : b2 (v, q) = 0, ∀ q ∈ Q} = {v ∈ H : div v = 0 in Ω},
and let us recall that b2 satisfies the inf-sup condition:
|b2 (v, q)|
sup ≥ β2 q0,Ω ∀q ∈ Q, (2.11)
v∈H vH
v =0

with an inf-sup constant β2 > 0 only depending on Ω (see e.g. [26, Sect. 2.4.1]).
We will now address the well-posedness of (2.3). To that end, it is enough to study its
reduced counterpart, defined on X × Z: Find (u, ω) ∈ X × Z such that
a(u, v) + b1 (v, ω) + c(ω, v) = F(v) ∀v ∈ X,
b1 (u, θ ) − d(ω, θ ) = G(θ ) ∀θ ∈ Z. (2.12)
The equivalence between (2.3) and (2.12) is established in the following result, whose proof
follows [28, Sect. I.4.1] and it is basically a direct consequence of the inf-sup condition
(2.11).

Lemma 2 If (u, ω, p) ∈ H × Z × Q is a solution of (2.3), then u ∈ X and (u, ω) ∈ X × Z

also solve (2.12). Conversely, if (u, ω) ∈ X × Z is a solution of (2.12), then there exists a
unique p ∈ Q such that (u, ω, p) ∈ H × Z × Q solves (2.3).

The abstract setting that will permit the analysis of (2.3) is stated in the following general
result [26, Theorem 1.2].

Theorem 1 Let A : X × X → R be a bounded bilinear form and G : X → R a bounded

functional, both defined on the Hilbert space (X , ·, · X ). If there exists α > 0 such that
A(x, y)
sup ≥ α xX ∀ x ∈ X , (2.13)
y∈X \{0} yX

and
sup A(x, y) > 0 ∀ y ∈ X , (2.14)
x∈X
y =0

then there exists a unique solution x ∈ X to the problem

A(x, y) = G (y) ∀ y ∈ X .

Furthermore, there exists C > 0 (independent of x) such that

1
xX ≤ G  X  .
C

123
1582 Journal of Scientific Computing (2019) 80:1577–1606

Lemma 3 Let us assume that

2β2∞,Ω
< 1, (2.15)
νσ
and let us define the bilinear form A(·, ·) specified as
A((u, ω), (v, θ )) := a(u, v) + b1 (v, ω) + b1 (u, θ ) − d(ω, θ ) + c(ω, v).
Then, there exist α1 , α2 > 0 such that
|A((u, ω), (v, θ ))| ≤ α1 (u, ω)X (v, θ )X , (2.16)
and
A((u, ω), (v, θ ))
sup ≥ α2 (u, ω)X ∀ (u, ω) ∈ X , (2.17)
(v,θ )∈X (v, θ )X
(v,θ)=0

where X := X × Z, endowed with the corresponding product norm, is a Hilbert space.

Proof As a consequence of Lemma 1, we have that the bilinear form A(·, ·) is bounded and
therefore the condition (2.16) readily follows.
Concerning the satisfaction of the inf-sup condition (2.17), for a given (u, ω) ∈ X , we
can define
√
θ̃ := −ω ∈ Z, and ṽ := (u + ĉ ν curl ω) ∈ X,
where ĉ > 0 is a constant to be chosen later. We can then immediately assert that
  
√ √
A((u, ω), (ṽ, θ̃ )) = σ u · ṽ d x + ν curl ω · ṽ d x + ν curl θ̃ · u d x
Ω Ω Ω
1
− ω · θ̃ d x + √ (ω × β) · ṽ d x
Ω ν
 Ω
√
≥ σ u20,Ω + ĉ νσ u · curl ω d x
Ω

√
+ ν u · curl ω d x + ĉν curl ω20,Ω
Ω 
√ 1
− ν u · curl ω d x + ω20,Ω + √ (ω × β) · u d x
ν Ω
 Ω
+ ĉ (ω × β) · curl ω d x
Ω
σ
≥ σ u20,Ω − u20,Ω − ĉ2 σ ν curl ω20,Ω
4
2σ
+ ĉν curl ω20,Ω + ω20,Ω − u20,Ω
3
3β2∞,Ω β2∞,Ω
− ω20,Ω − ω20,Ω − 2ĉ2 σ ν curl ω20,Ω
2νσ 2νσ  
σ   2β 2
∞,Ω
= u20,Ω + ĉ 1 − 3ĉσ ν curl ω20,Ω + 1 − ω20,Ω ,
12 νσ

where we have used the bound ω × β0,Ω ≤ 2β∞,Ω ω0,Ω . Choosing ĉ = 1/(4σ ) and
employing (2.15), we arrive at
A((u, ω), (ṽ, θ̃ )) ≥ C(u, ω)2X ,

123
Journal of Scientific Computing (2019) 80:1577–1606 1583

with C independent of ν. On the other hand, by construction we realise that θ̃ Z = ωZ
and ṽ0,Ω ≤ C ĉ(u0,Ω + ωZ ), and consequently

A((u, ω), (v, θ )) A((u, ω), (ṽ, θ̃ ))

sup ≥ ≥ α2 (u, ω)X ∀(u, ω) ∈ X ,
(v,θ)∈X (v, θ )X (ṽ, θ̃ )X
(v,θ)=0

which finishes the proof. 



Lemma 4 Suppose that the bound (2.15) is satisfied. Then,

sup A((u, ω), (v, θ )) > 0 ∀ (v, θ ) ∈ X .
(u,ω)∈X
(u,ω) =(0,0)

Proof For all (v, θ ) ∈ X , we have that:


1
A((v, −θ ), (v, θ )) = σ v20,Ω + θ 20,Ω − √ (θ × β) · v d x
ν Ω
σ 2β2∞,Ω
≥ σ v20,Ω + θ20,Ω − v20,Ω − θ20,Ω
 2  νσ
σ 2β2∞,Ω
≥ v20,Ω + 1 − θ 20,Ω .
2 νσ



As a consequence of the previous lemmas, we have the following result.

Theorem 2 Let us assume (2.15). Then, the variational problem (2.12) admits a unique
solution (u, ω) ∈ X × Z. Moreover, there exists C > 0 such that
uH + ωZ ≤ C( f 0,Ω +  pΣ 1/2,Σ + uΣ −1/2,Σ ). (2.18)

Proof It suffices to verify the hypotheses of Theorem 1. First, we define the linear functional
G (v, θ ) := F(v) + G(θ ),
which is bounded on X × Z as a consequence of the last two estimates in Lemma 1. Thus,
the proof follows from Lemmas 3 and 4. 


The following result establishes the corresponding stability estimate for the Bernoulli
pressure.
Corollary 1 Let (u, ω) ∈ X × Z, be the unique solution of (2.12), with u and ω satisfying
(2.18). In addition, let p ∈ Q be the unique pressure provided by Lemma 2, so that (u, ω, p) ∈
H × Z × Q is the unique solution of (2.3). Then, there exists C > 0 such that
 p0,Ω ≤ C( f 0,Ω +  pΣ 1/2,Σ + uΣ −1/2,Σ ).

Proof Combining the inf-sup condition (2.11) with the first equation in (2.3) gives the bound
1 |b2 (v, p)| 1 |F(v) − a(u, v) − b1 (v, ω) − c(ω, v)|
 p0,Ω ≤ sup = sup ,
β2 v∈H vH β2 v∈H vH
v=0 v =0

which together with (2.18) and Lemma 1, complete the proof. 



123
1584 Journal of Scientific Computing (2019) 80:1577–1606

Remark 1 An alternative analysis for the nonsymmetric variational problem (2.3) can be car-
ried out using a fixed-point argument that allows a symmetrisation of the mixed structure.
The resulting weak form could then be analysed using classical tools for saddle-point prob-
lems, for instance, following the similar treatment in [16, Sect. 3.3]. Establishing inf-sup
conditions for the off-diagonal bilinear forms in the original nonsymmetric formulation is,
however, much more involved (see e.g. [33, Sect. 3]).

Remark 2 Assumption (2.15) holds provided one chooses σ appropriately. As this parameter
represents the inverse of the timestep, the aforementioned relation constitutes then a CFL-
type condition. We have observed experimentally that the bound for β is not sharp, but we
also stress that the same condition coincides with the hypotheses that yield solvability of
least-squares formulations for the Oseen problem analysed in [15,17]. Further investigations
would be then required.

3 Finite Element Discretisation

In this section we introduce a Galerkin scheme for (2.3) and analyse its well-posedness by
establishing suitable assumptions on the finite element subspaces involved. Error estimates
are also derived.

3.1 Defining the Discrete Problem

Let {Th (Ω)}h>0 be a shape-regular family of partitions of the polyhedral region Ω̄, by
tetrahedrons T of diameter h T , with mesh size h := max{h T : T ∈ Th (Ω)}. In what
follows, given an integer k ≥ 0 and a subset S of R3 , Pk (S) will denote the space of
polynomial functions defined locally in S and being of total degree ≤ k.
Moreover, for any T ∈ Th (Ω), we introduce the local Nédélec space

Nk (T ) := Pk (T )3 ⊕ Rk+1 (T ),

where Rk+1 (T ) is a subspace of Pk+1 (T )3 composed by homogeneous polynomials of degree

k + 1, and being orthogonal to x. With these tools, let us define the following finite element
subspaces:

Zh := {θ h ∈ Z : θ h |T ∈ Nk (T ) ∀T ∈ Th (Ω)}, (3.1)
Qh := {qh ∈ Q : qh |T ∈ Pk (T ) ∀T ∈ Th (Ω)}, (3.2)
Hh := {v h ∈ H : v h |T ∈ RTk (T ) ∀T ∈ Th (Ω)}, (3.3)

where RTk (T ) = Pk (T )3 ⊕ Pk (T )x is the Raviart–Thomas space defined locally in T ∈

Th (Ω).
The proposed Galerkin scheme approximating (2.3) reads as follows: Find (uh , ωh , ph ) ∈
Hh × Zh × Qh such that

a(uh , v h ) + b1 (v h , ωh ) + b2 (v h , ph ) + c(ωh , v h ) = F(v h ) ∀v h ∈ Hh ,

b1 (uh , θ h ) − d(ωh , θ h ) = G(θ h ) ∀θ h ∈ Zh ,
b2 (uh , qh ) = 0 ∀qh ∈ Qh . (3.4)

123
Journal of Scientific Computing (2019) 80:1577–1606 1585

3.2 Solvability and Stability of the Discrete Problem

The analysis of the Galerkin formulation will follow the same arguments exploited in the
continuous setting. Let us then consider the discrete kernel of b2 :
Xh := {v h ∈ Hh : b2 (v h , qh ) = 0, ∀ qh ∈ Qh } = {v h ∈ Hh : div v h ≡ 0 in Ω},
(3.5)
where the characterisation is indeed possible thanks to the inclusion div Hh ⊆ Qh . Moreover,
it is well-known that the following discrete inf-sup condition holds (see [26, Sect. 4.2]):
b2 (v h , qh )
sup ≥ β̃2 qh 0,Ω ∀qh ∈ Qh . (3.6)
v h ∈Hh v h H
v h  =0

We again resort to a reduced version of the problem, now defined on the product space
Xh × Zh . Find (uh , ωh ) ∈ Xh × Zh such that
a(uh , v h ) + b1 (v h , ωh ) + c(ωh , v h ) = F(v h ) ∀ v h ∈ Xh ,
b1 (uh , θ h ) − d(ωh , θ h ) = G(θ h ) ∀ θ h ∈ Zh , (3.7)
and its equivalence with (3.4) is once more a direct consequence of the inf-sup condition
(3.6).
Lemma 5 If (uh , ωh , ph ) ∈ Hh × Zh × Qh is a solution of (3.4), then uh ∈ Xh , and
(uh , ωh ) ∈ Xh × Zh is also a solution of (3.7). Conversely, if (uh , ωh ) ∈ Xh × Zh is a
solution of (3.7), then there exists a unique ph ∈ Qh such that (uh , ωh , ph ) ∈ Hh × Zh × Qh
is a solution of (3.4).
In order to establish the well-posedness of (3.7), we will employ the following discrete
version of Theorem 1.
Theorem 3 Assume (2.15). Let k ≥ 0 be an integer and let Xh and Zh be given by (3.5) and
(3.1), respectively. Then, there exists a unique (uh , ωh ) ∈ Xh × Zh solution of the discrete
scheme (3.7). Moreover, there exist positive constants Ĉ1 , Ĉ2 > 0 independent of h such that
uh H + ωh Z ≤ Ĉ1 ( f 0,Ω +  pΣ 1/2,Σ + uΣ −1/2,Σ ), (3.8)
and
u − uh H + ω − ωh Z ≤ Ĉ2 inf (u − v h H + ω − θ h Z ), (3.9)
(v h ,θ h )∈Xh ×Zh

where (u, ω) ∈ X × Z is the unique solution to (2.12).

Proof Let us define Xh := Xh × Zh and reuse the forms A(·, ·) and G (·) as in the proof of
Lemma 2. The next step consists in proving that A(·, ·) satisfies the discrete version of the
inf-sup conditions (2.13)–(2.14), as in Lemmas 3 and 4. In order to assert (2.13), we consider
(uh , ωh ) ∈ Xh , and define
√

ν
θ̃ h := −ωh ∈ Zh , and ṽ h := uh + curl ωh ∈ Xh .
4σ
Then, repeating exactly the same steps used in the proof of Lemma 3 the discrete version
of (2.13) follows. Regarding the discrete version of (2.14), we once again repeat the same
arguments given in the proof of Lemma 4. Finally, the Céa estimate follows from classical
arguments. 


123
1586 Journal of Scientific Computing (2019) 80:1577–1606

Remark 3 Note that since the discrete kernel satisfies Xh ⊂ X, the infimum in the Céa
estimate (3.9) can be taken over (v h , θ h ) ∈ Hh × Zh , which implies in particular that the
mixed finite element scheme does not depend on the pressure (see e.g. [26, pp. 58–59]).
Implications of pressure-robust schemes as this one are reviewed in e.g. [29].

We now state the stability and an adequate approximation property of the discrete pressure.

Corollary 2 Let (uh , ωh ) ∈ Xh ×Zh be the unique solution of (3.7), with uh and ωh satisfying
(3.8). In addition, let ph ∈ Q h be the unique discrete Bernoulli pressure provided by Lemma
5, so that (uh , ωh , ph ) ∈ Hh × Zh × Qh is the unique solution of (3.4). Then, there exist
positive constants C̄1 , C̄2 > 0, independent of h and ν, such that
 ph 0,Ω ≤ C̄1 ( f 0,Ω +  pΣ 1/2,Σ + uΣ −1/2,Σ ),
and
 p − ph 0,Ω ≤ C̄2 inf (u − v h H + ω − θ h Z +  p − qh 0,Ω ). (3.10)
(v h ,θ h ,qh )∈Hh ×Zh ×Q h

Proof The result follows using the same arguments considered in the proof of Corollary 1,
but using the discrete inf-sup condition (3.6). We omit further details. 


3.3 A Priori Error Estimates

Let us introduce for a given s > 1/2, the Nédeléc global interpolation operator Rh :
Hs (curl; Ω) ∩ Z → Zh . From [1, Proposition 5.6] we know that for all θ ∈ Hs (curl; Ω)
with s > 1/2, there exists C > 0 independent of h, such that
θ − Rh θ Z ≤ Ch min{s,k+1} θHs (curl;Ω) . (3.11)
On the other hand, for the Raviart–Thomas interpolation Πh : Hs (Ω)3 ∩ H → Hh , with
s > 0, we recall (see e.g. [26, Theorem 3.6 and Lemma 3.19]) that there exists C > 0,
independent of h, such that for all s > 0:
v − Πh vH ≤ Ch min{s,k+1} vHs (div;Ω) ∀v ∈ Hs (div; Ω) ∩ H. (3.12)
Finally we recall that the orthogonal projection from L2 (Ω) onto the finite element subspace
Qh , here denoted by ΠQ , satisfies the following error estimate for all s > 0:
q − ΠQ q0,Ω ≤ Ch min{s,k+1} qs,Ω ∀q ∈ Hs (Ω). (3.13)
The following result summarises the error analysis for our mixed finite element scheme
(3.4).
Theorem 4 Assume (2.15). Let k ≥ 0 be an integer and let Hh , Zh and Qh be given by (3.1),
(3.2), and (3.3). Let (u, ω, p) ∈ H × Z × Q and (uh , ωh , ph ) ∈ Hh × Zh × Qh be the unique
solutions to the continuous and discrete problems (2.3) and (3.4), respectively. Assume that
u ∈ Hs (Ω)3 , div u ∈ Hs (Ω), ω ∈ H1+s (Ω)3 and p ∈ Hs (Ω), for some s > 1/2. Then,
there exist positive constants Ĉ and C̃, independent of h, such that
u − uh H + ω − ωh Z ≤ Ĉh min{s,k+1} (uHs (div;Ω) + ωHs (curl;Ω) ),
and
 p − ph 0,Ω ≤ C̃h min{s,k+1} (uHs (div;Ω) + ωHs (curl;Ω) +  ps,Ω ).

123
Journal of Scientific Computing (2019) 80:1577–1606 1587

Proof The proof follows from (3.9), (3.10), and standard interpolation estimates satisfied by
the operators Rh , Πh and ΠQ (see (3.11), (3.12) and (3.13), respectively). 


4 Discontinuous Galerkin Method

In this section, we propose and analyse a DG method for (2.2). We provide solvability
and stability of the discrete scheme by introducing suitable numerical fluxes. A priori error
estimates are also derived.

4.1 Preliminaries

Apart from the definitions laid out at the beginning of Sect. 3, let us denote by Eh the set of
internal faces, by FhΣ the set of external faces on Σ and by FhΓ the set of external faces on
Γ . We set Fh = Eh ∪ FhΣ ∪ FhΓ . We denote by h e the diameter of each face e. Let T + and
T − be two adjacent elements of Th and let n+ (respectively n− ) be the outward unit normal
vector on ∂ T + (respectively ∂ T − ). For a vector field u, we denote by u± the trace of u from
the interior of T ± . We define jumps

[[v]]T := v + × n+ + v − × n− , [[v]] N := v + · n+ + v − · n− , [[q]] := q + n+ + q − n− ,

and averages
1 + 1 +
{{v}} := (v + v − ), {{q}} := (q + q − ),
2 2
and adopt the convention that for boundary faces e ∈ FhΣ ∪ FhΓ , we set [[v]]T = v × n,
[[v]] N = v · n, [[q]] = qn, {{v}} = v and {{q}} = q.
Now, for k ≥ 0, suitable finite dimensional spaces for vorticity and velocity that remove
the restriction of continuity are defined by:

Z̃h := {θ h ∈ L2 (Ω)3 : θ h |T ∈ Pk (T )3 ∀T ∈ Th },
H̃h := {v h ∈ L2 (Ω)3 : v h |T ∈ Pk+1 (T )3 ∀T ∈ Th },

and we remark that the space for pressure approximation will coincide with the one used in
Sect. 3, that is Q̃h := Qh .

4.2 Discrete Formulation and Solvability Analysis

Multiplying each equation in (2.2) by suitable functions, the resulting DG scheme consists
in finding (uh , ωh , ph ) ∈ H̃h × Z̃ h × Q̃h , such that for any test functions (v h , θ h , qh ) ∈
H̃h × Z̃ h × Q̃h and for all elements T in the partition Th
 
√
σ uh · v h d x + ν ωh · curl v h d x
T T
 
√ 1
+ ν 
ωh · (v h × n) ds + √ (ωh × β) · v h d x
ν T
 ∂T  
− ph div v h d x + 
ph v h · n ds = f · v h d x, (4.1)
T ∂T T

123
1588 Journal of Scientific Computing (2019) 80:1577–1606
  
√ √
ωh · θ h d x = ν uh · curl θ h d x + ν uωh · (θ h × n) ds,
 (4.2)
T T ∂T
 
p
− uh · ∇qh d x + 
uh · n q ds = 0, (4.3)
T ∂T

uw
p
where  h,uh , 
ωh and ph are numerical fluxes, which approximate the traces of uh , ωh
and ph on the boundary. The fluxes  uw
ωh and  h are related to the curl-curl operator and are
defined by
⎧ ⎧
⎪
⎨{{ωh }} + C11 [[u]]T , if e ∈ Eh , ⎪
⎨{{uh }}, if e ∈ Eh ,
ωh := ω+
 + C 11 (u +
× n + − u ), if e ∈ F Σ ,
Σ uw
 := n × u Σ , if e ∈ FhΣ ,
⎪
⎩
h h h h ⎪
⎩ +
ωh × n = 0, if e ∈ FhΓ , uh , if e ∈ FhΓ ,
(4.4)
p
whereas the fluxes 
uh and 
ph are associated with the grad-div operator and defined by
⎧
⎪
⎨{{uh }} + D11 [[ ph ]], if e ∈ Eh ,
uh := u+ + +
p − ), if e ∈ F Σ ,
 + D (
11 hp n + p Σ n
⎪
⎩ p
h h
uh · n = 0, if e ∈ FhΓ ,
⎧
⎪
⎨{{ ph }} + A11 [[uh ]] N , if e ∈ Eh ,
ph := pΣ , if e ∈ FhΣ , (4.5)
⎪
⎩ +
p + A11 u · n , if e ∈ FhΓ .
+ +

The parameters C11 , A11 and D11 are positive stabilisation parameters, and following [22]
we choose

c11 max{h −1
T+
, h −1
T−
}, if x ∈ ∂ T + ∪ ∂ T − ,
C11 (x) := −1 (4.6)
c11 h T , if x ∈ ∂ T ∩ Σ,

a11 max{h −1
T+
, h −1
T−
}, if x ∈ ∂ T + ∪ ∂ T − ,
A11 (x) := −1 (4.7)
a11 h T , if x ∈ ∂ T ∩ Γ ,

d11 max{h T + , h T − }, if x ∈ ∂ T + ∪ ∂ T − ,
D11 (x) := (4.8)
d11 h T , if x ∈ ∂ T ∩ Σ,

where c11 , d11 , a11 > 0. Moreover, we suppose that C11 (respectively D11 and A11 ) have
a uniform positive bound above and below denoted by C11 and C11 (respectively D11 , D11
and A11 , A11 ).
We then proceed to integrate by parts equations (4.1) and (4.3), and then summing up over
all T ∈ Th , we obtain the following DG scheme: Find (uh , ωh , ph ) ∈  Hh × 
Zh × 
Qh such
that

 h ),
a(uh , v h ) + b̃1 (v h , ωh ) + b̃2 (v h , ph ) + c(ωh , v h ) + j(uh , v h ) = F(v ∀v h ∈ 
Hh ,
d(ωh , θ h ) − b̃1 (uh , θ h ) = G̃(θ h ), ∀θ h ∈ 
Zh ,
e( ph , qh ) − 
b2 (uh , qh ) = L̃(qh ), ∀qh ∈ 
Qh ,
(4.9)

123
Journal of Scientific Computing (2019) 80:1577–1606 1589

where the forms a, c and d are the same in (2.3), while b̃1 , b̃2 , j and e are defined, respectively,
by:
  
√  √
b̃1 (uh , θ h ) := ν curl θ h · uh d x + ν {{uh }} · [[θ h ]]T ds,
T ∈T h T e
e∈Eh ∪FhΓ
   
b̃2 (v h , ph ) := − ph div v h d x + {{ ph }} · [[v h ]] N ds,
T ∈T h T e
e∈Eh ∪FhΓ
   
√
j(uh , v h ) := ν C11 [[uh ]]T · [[v h ]]T ds + A11 [[uh ]] N [[v h ]] N ds,
e e
e∈Eh ∪FhΣ e∈Eh ∪FhΓ
 
e( ph , qh ) := D11 [[ ph ]] · [[qh ]] ds.
e
e∈Eh ∪FhΣ

 G
In addition, the linear functionals F,  and L associated with the source terms are defined
as:
    
√
 h ) :=
F(v f · vh − pΣ (v · n) ds − ν C11 uΣ · (v h × n) ds ,
Ω e e
e∈FhΣ
√    
 h ) := − ν
G(θ uΣ · θ h ds and 
L(qh ) := D11 ( pΣ n) · qh n ds.
e e
e∈FhΣ e∈FhΣ

By integration by parts and as a consequence of the identity:

   
u · (θ × n) ds = − [[u]]T · {{θ }} ds + {{u}} · [[θ ]]T ds,
T ∈T h ∂T e∈Eh e e∈Fh e

it follows that the form b̃1 can be written as:

 
√  √ 
b̃1 (uh , θ h ) = ν curl uh · θ h d x + ν [[uh ]]T · {{θ h }} ds. (4.10)
T ∈T h T e
e∈Eh ∪FhΣ

Similarly, using the following identity:

   
p(v · n) ds = {{v}} · [[ p]] ds + [[v]] N {{ p}} ds,
T ∈T h ∂T e∈Eh e e∈Fh e

the form b̃2 can be recast, after integration by parts, as follows

   
b̃2 (v h , ph ) = v h · ∇ ph d x − {{v h }} · [[ ph ]] ds. (4.11)
T ∈T h T e
e∈Eh ∪FhΣ

To simplify the exposition of the analysis of the method, we will write the mixed scheme
(4.9) in the following equivalent form: Find (uh , ωh , ph ) ∈ 
Hh × 
Zh × 
Qh such that
A(uh , ωh , ph ; v h , θ h , qh ) = F (v h , θ h , qh ), ∀(v h , θ h , qh ) ∈ 
Hh × 
Zh × 
Qh , (4.12)
where
A(uh , ωh , ph ; v h , θ h , qh ) := a(uh , v h ) + b̃1 (v h , ωh ) − b̃1 (uh , θ h )

123
1590 Journal of Scientific Computing (2019) 80:1577–1606

+ b̃2 (v h , ph ) − 
b2 (uh , qh ) + c(ωh , v h )
+ j(uh , v h ) + d(ωh , θ h ) + e( ph , qh ),

and
 h ) + G(θ
F (v h , θ h , qh ) := F(v  h) + 
L(qh ).

Let us now show the existence and uniqueness of solution to formulation (4.9).

Proposition 1 Let k ≥ 0 be an integer. The DG method (4.9) with the numerical fluxes given
by (4.4)–(4.5) defines a unique approximate solution (uh , ωh , ph ) ∈ 
Hh × 
Zh × 
Qh provided
that
2β2∞
< 1. (4.13)
νσ
Proof Since the problem is linear and finite dimensional, it suffices to show that if f = 0,
pΣ = 0 and uΣ = 0, then (uh , ωh , ph ) = (0, 0, 0). To this end, take v h = uh , θ h = ωh and
qh = ph in (4.9), summing up the three equations, we obtain:

a(uh , uh ) + c(ωh , uh ) + d(ωh , ωh ) + j(uh , uh ) + e( ph , ph ) = 0.

It follows that

1
σ uh 20,Ω + ωh 20,Ω + |uh |2j + | ph |2e = − √ (ωh × β) · uh d x,
ν Ω

where we define
   
√
|uh |2j := ν C11 [[uh ]]2T ds + A11 [[uh ]]2N ds,
e e
e∈Eh ∪FhΣ e∈Eh ∪FhΓ
 
| ph |2e := D11 [[ ph ]]2 ds.
e
e∈Eh ∪FhΣ

Using Young’s inequality, we can assert that

2
σ uh 20,Ω + ωh 20,Ω + |uh |2j + | ph |2e ≤ √ β∞,Ω ωh 0,Ω uh 0,Ω
ν
2 σ
≤ β2∞,Ω ωh 20,Ω + uh 20,Ω .
νσ 2
Therefore, in particular we have that:

σ 2β2∞
uh 0,Ω + 1 −
2
ωh 20,Ω + | ph |2e ≤ 0,
2 νσ
which, owing to the assumption (4.13), implies that uh = 0, ωh = 0 and [[ ph ]] = 0 on Eh ,
ph = 0 on FhΣ . The first equation in (4.9) then becomes
 
v h · ∇ ph d x = 0, ∀v h ∈ 
Hh ,
T ∈T h T

and then ∇ ph = 0. Employing this result, together with [[ ph ]] = 0 on Eh , ph = 0 on FhΣ or

the fact that ph has zero mean value if Σ have zero measure, we conclude that ph = 0.  

123
Journal of Scientific Computing (2019) 80:1577–1606 1591

4.3 A Priori Error Bounds

Let us now present and discuss a priori error bounds for the proposed DG method. The proof
involves two steps. The first one consists in establishing an error estimate in the natural
semi-norm. In the second step, we prove the error estimate for the pressure in the L2 -norm.
For this, we introduce the following semi-norm | · |A :
|(u, ω, p)|2A := σ u20,Ω + ω20,Ω + |u|2j + | p|2e . (4.14)
For the analysis, we will also employ the following norm
√ 1
(u, p)s := νus+1,Ω + √  ps,Ω .
ν
We define eu = u − uh , eω = ω − ωh and e p = p − ph . Let us denote by Π  H (respectively
Π and Π ) the L 2 -projection onto 
H (P 3 ) (respectively 
Z (P 3 ) and 
Q (Pk )), and let
Z 
Q h k+1 h k h
us split the errors in the following manner
eu = ξ u + η u , eω = ξ ω + ηω and e p = ξ p + η p ,
where the numerical and approximation errors are defined by:
ξ u = u − Π
H u, ξ ω = ω − Π
Z ω, ξ p = p − Π
Q p,
ηu = Π 
H u − uh , ηω = Π
Z ω − ω h , η p = Π
Q p − ph .

We recall the following standard approximation properties (see for instance [19]).
Lemma 6 Let v ∈ H1+r (Ω), r ≥ 0. Let Π the projection operator such that Πv = v for all
v ∈ Pk (T ), k ≥ 0. Then we have
min{r ,k}+1
v − Πv0,T + h T |v − Πv|1,T ≤ Ch T vr +1,T , (4.15)
min{r ,k}+1/2
v − Πv0,∂ T ≤ Ch T vr +1,T . (4.16)
As a consequence, we have the following result.
Lemma 7 Let (u, ω, p) ∈ H × Z × Q be the unique solution to the continuous problem (2.3).
Assume that u ∈ H1+s (Ω)3 , ω ∈ Hs (Ω)3 and p ∈ Hs (Ω), for some s ≥ 1. Then, we have:
ξ u 0,Ω ≤ Ca h min{s, k+1}+1 (u, 0)s ,
ξ ω 0,Ω ≤ Cd h min{s, k}+1 (u, 0)s ,
|ξ u | j ≤ C j h min{s, k+1} (u, 0)s ,
|ξ p |e ≤ Ce h min{s, k}+1 (0, p)s ,
where Ca , Cd , C j and Ce are positive constants independent of the meshsize.

Proof The two first estimates are a simple consequence of (4.15). Next we can state that
⎛ ⎞1/2 ⎛ ⎞1/2
√  
|ξ u | j ≤ 2 ⎝ ν C11 ξ u 0,∂ T ⎠ + 2 ⎝
2
A11 ξ u 0,∂ T ⎠
2

T ∈T h T ∈T h

and ⎛ ⎞1/2

|ξ p |e ≤ 2 ⎝ D11 ξ p 20,∂ T ⎠ .
T ∈T h

123
1592 Journal of Scientific Computing (2019) 80:1577–1606

Recalling that C11 , A11 and D11 are as in (4.6)–(4.8) and using (4.16), we end up with the
bound   2 min{s,k+1} 1/2
|ξ u | j ≤ C j hT u2s+1,T ≤ C j h min{s,k+1} us+1,Ω ,
T ∈T h

and similarly we obtain

⎛ ⎞1/2

|ξ p |e ≤ Ce ⎝  p2s,T ⎠
2 min{s,k}+2
hT ≤ Ce h min{s,k}+1  ps,Ω ,
T ∈T h

where C j (respectively Ce ) depends on c11 , a11 and ν (respectively d11 ). 



We next concentrate on obtaining bounds for the forms b̃1 , b̃2 , c and j.
Lemma 8 Let (u, ω, p) ∈ H × Z × Q be the unique solution to the continuous problem (2.3).
Assume that u ∈ H1+s (Ω)3 , ω ∈ Hs (Ω)3 and p ∈ Hs (Ω), for some s ≥ 1. Then one has
|b̃1 (ξ u , θ h )| ≤ Cb1 h min{s,k+1} (u, 0)s θ h 0,Ω , ∀θ h ∈ 
Zh ,
|b̃1 (v h , ξ ω )| ≤ Cb1 h min{s,k}+1 (u, 0)s |v h | j , ∀v h ∈ 
Hh ,
|b̃2 (v h , ξ p )| ≤ Cb2 h min{s,k}+1 (0, p)s |v h | j , ∀v h ∈ 
Hh ,
|b̃2 (ξ u , qh )| ≤ Cb2 h min{s,k+1} (u, 0)s |qh |e , ∀qh ∈ 
Qh ,
| j(ξ u , v h )| ≤ C j h min{s, k+1} (u, 0)s |v h | j , ∀v h ∈ 
Hh ,
|e(ξ p , qh )| ≤ Ce h min{s, k}+1 (0, p)s |qh |e , ∀qh ∈ 
Qh ,
|c(ξ ω , v h )| ≤ Cω h min{s, k}+1 (u, 0)s v h 0,Ω , ∀v h ∈ 
Hh ,
where Cb1 , Cb2 , C j , Ce and Cω are positive constants independent of the meshsize.

Proof The bounds associated with the form b̃2 can be proved exactly with the same arguments
as [22, Sect. 3.3]. Let us now deal with the term b̃1 . We observe that due to the properties of
the L2 -projection onto Hh (Pk+1
3 ), we can write


ξ u · curl θ h d x = 0.
T

Using Cauchy–Schwarz’s inequality, we then readily obtain

⎛ ⎞1/2 ⎛ ⎞1/2
 
|b̃1 (ξ u , θ h )| ≤ C ⎝ νh −1
T ξ u 0,∂ T
2 ⎠ ⎝ h T θ h 20,∂ T ⎠ ,
T ∈T h T ∈T h

and the desired estimate follows from the inverse inequality and Lemma 7.
Similarly, using (4.10) and again the properties of the L2 -projection onto 
Zh (Pk3 ), we have

curl v h · ξ ω d x = 0,
T

thus
⎛ ⎞1/2
 ν
|b̃1 (v h , ξ ω )| ≤ C ⎝ ξ 2 ⎠
C11 ω 0,∂ T
T ∈T h

123
Journal of Scientific Computing (2019) 80:1577–1606 1593

⎛ ⎞1/2
   
⎜√ √ ⎟
⎝ ν C11 [[v h ]]2T ds + ν C11 v 2h ds ⎠
e∈Eh e e
e∈FhΣ
⎛ ⎞1/2
 ν
≤⎝ ξ 2 ⎠ |v h | j
C11 ω 0,∂ T
T ∈T h

and then, we simply have to use (4.6) and Lemma 6.

The estimates for the forms j and e are obtained in a similar way. Indeed, using once
again Cauchy–Schwarz’s inequality and Lemma 7, it follows that

√     
     
| j(ξ u , v h )| =  ν C11 ξ u T · [[v h ]]T ds + A11 ξ u N [[v h ]] N ds 
e e
e∈Eh ∪FhΣ e∈Eh ∪FhΓ

≤ |v h | j |ξ u | j ≤ C j h min{s, k+1} (u, 0)s |v h | j ,

and proceeding analogously as before, we get
|e(ξ p , qh )| ≤ |ξ p |e |qh |e ≤ Ce h min{s, k}+1 (0, p)s |qh |e .
Finally, concerning the term c(ξ ω , v h ) we can assert that
2
c(ξ ω , v h ) ≤ √ β∞ ξ ω 0,Ω v h 0,Ω ,
ν
and exploiting the previous bounds, we obtain the corresponding estimate with Cω =
2C
√ d β∞ .
ν



Theorem 5 Assume (4.13). Let k ≥ 0 be an integer and let (u, ω, p) ∈ H×Z×Q be the unique
solution to the continuous problem (2.3). Assume that u ∈ H1+s (Ω)3 , ω ∈ Hs (Ω)3 and
p ∈ Hs (Ω), for some s ≥ 1. Then, the mixed DG approximation (uh , ωh , ph ) ∈ 
Hh ×
Zh × 
Qh
defined by (4.9), satisfies the following a priori error bounds
|(eu , eω , e p )|A ≤ CA h min{s, k+1} (u, p)s , (4.17)
e p 0,Ω ≤ Ch min{s, k+1} (u, p)s , (4.18)
where CA and C are positive constants independent of the meshsize.

Proof We begin with the estimate (4.17). A direct application of the definition of the
A−seminorm in combination with Lemma 7 gives
|(ξ u , ξ ω , ξ p )|A ≤ Ch min{s, k+1} (u, p)s . (4.19)
Concentrating on the projection of the errors, we can exploit the Galerkin orthogonality to
obtain
|(η u , ηω , η p )|2A = A(η u , ηω , η p ; η u , ηω , η p ) − c(ηω , η u ) (4.20)
= A(η u , ηω , η p ; ξ u , ξ ω , ξ p ) − c(ηω , η u ).

Due to the orthogonality of the L2 -projections, we have that a(ξ u , η u ) = 0 and d(ξ ω , ηω ) =
0. Then, from the definition of the form A it follows that

A(η u , ηω , η p ; ξ u , ξ ω , ξ p ) = b̃1 (ξ u , ηω ) − b̃1 (η u , ξ ω )

123
1594 Journal of Scientific Computing (2019) 80:1577–1606

+ b̃2 (ξ u , η p ) − 
b2 (η u , ξ p ) + c(ξ ω , η u ) + j(ξ u , η u ) + e(ξ p , η p ).

Note that all these terms can be controlled using Lemma 8. Indeed we have

A(η u , ηω , η p ; ξ u , ξ ω , ξ p ) ≤ Cb1 ηω 0,Ω + (Cb1 + Cb2 + C j )|η u | j + (Cb2 + Ce )|η p |e

+ Cω η u 0,Ω h min{s, k+1} (u, p)s ,
≤ C1 |(η u , ηω , η p )|A h min{s, k+1} (u, p)s , (4.21)
Cω
where C1 = Cb1 + (Cb1 + Cb2 + C j ) + (Cb2 + Ce ) + σ . Next we only need to estimate
c(ηω , η u ) in (4.20). To do so we use Young’s inequality
2 2 σ
c(ηω , η u ) ≤ √ β∞,Ω ηω 0,Ω η u 0,Ω ≤ β2∞,Ω ηω 20,Ω + η u 20,Ω . (4.22)
ν νσ 2
Substituting (4.21) and (4.22) back into (4.20), we obtain the bound

σ 2β2∞
η  2
+ 1− ηω 20,Ω + |η u |2j + |η p |2e
2 u 0,Ω νσ
≤ C1 |(η u , ηω , η p )|A h min{s, k+1} (u, p)s ,

and thanks to assumption (4.13), we can arrive at

|(η u , ηω , η p )|A ≤ C  h min{s, k+1} (u, p)s , (4.23)

2β2
where C  = C1 (min{ 21 , 1 − νσ ∞ })−1 . The error estimate in (4.17) is then obtained by
combining estimates (4.19), (4.23) and using triangle inequality.
We now turn to the estimate of the L2 -norm of error in the pressure (4.18). Since e p ∈
L02 (Ω),
we can find z ∈ H01 (Ω)3 such that (see for instance [28, Chapter I, Corollary 2.4])

− e p divz d x ≥ κe p 20,Ω , z1,Ω ≤ e p 0,Ω , (4.24)
Ω

where κ > 0 is the inf-sup constant. Therefore, we infer from (4.12) that

κe p 20,Ω ≤ b̃2 (z, e p )

 
= b̃2 (z, e p ) + a(eu , z) + b̃1 (z, eω ) + c(eω , z)
− a(eu , z) − b̃1 (z, eω ) − c(eω , z)
= A(eu , eω , e p ; z, 0, 0) − a(eu , z) − b̃1 (z, eω ) − c(eω , z),

where we have used that j(eu , z) = 0 for z ∈ H01 (Ω)3 . Using the Galerkin orthogonality,
we obtain

A(eu , eω , e p ; z, 0, 0) = A(eu , eω , e p ; ξ z , 0, 0)
= A(ξ u , ξ ω , ξ p ; ξ z , 0, 0) + A(η u , ηω , η p ; ξ z , 0, 0).

Therefore

κe p 20,Ω ≤ |A(ξ u , ξ ω , ξ p ; ξ z , 0, 0)| + |A(η u , ηω , η p ; ξ z , 0, 0)|

+ |a(eu , z)| + |b̃1 (z, eω )| + |c(eω , z)|.

123
Journal of Scientific Computing (2019) 80:1577–1606 1595

Next, by the definition of the form A, we have the relation

|A(η u , ηω , η p ; ξ z , 0, 0)| ≤ |b̃1 (ξ z , ηω )| + |b̃2 (ξ z , η p )|
+ |c(ηω , ξ z )| + | j(η u , ξ z )| := T1 + T2 + T3 + T4 ,
and then we can write
κe p 20,Ω ≤ |A(ξ u , ξ ω , ξ p ; ξ z , 0, 0)| + T1 + T2 + T3 + T4 + T5 + T6 + T7 . (4.25)
The first term in the right-hand side of the above inequality can be easily estimated by
using Lemma 8. Indeed, it follows by choosing v h = ξ z that

|A(ξ u , ξ ω , ξ p ; ξ z , 0, 0)| ≤ Ch min{s,k+1} (u, p)s .

Let us now estimate each of the terms Ti , i = 1, . . . , 7 in (4.25). Using the properties of
the L2 −projection, Cauchy–Schwarz’s inequality and the inverse inequality, we obtain the
bounds
⎛ ⎞1/2 ⎛ ⎞1/2
√   √
T1 ≤ C ν ⎝ h −1
T ξ z 0,∂ T
2 ⎠ ⎝ h T ηω 20,∂ T ⎠ ≤ C νz1,Ω ηω 0,Ω .
T ∈T h T ∈T h
√
≤ C νz1,Ω |(η u , ηω , η p )|A
Then, using (4.23) and (4.24), we can deduce that
√
T1 ≤ C νh min{s, k+1} (u, p)s e p 0,Ω .

Furthermore, since T ξ z · ∇η p d x = 0, we get from (4.11) the following estimates

 
 
      
 
T2 =  {{ξ z }} · η p ds + (ξ z · n)η p ds 
 
e∈Eh e e∈FhΣ
e 
⎛ ⎞1/2
   
⎜ 1 1 ⎟
≤⎝ {{ξ z }}2 ds + (ξ z · n)2 ds ⎠ |η p |e
e∈Ehe D11 Σ e D 11
e∈Fh
⎛ ⎞1/2
 1
≤C⎝ ξ 2 ⎠ |η p |e
D11 z 0,∂ T
T ∈T h

≤ Cz1,Ω |(η u , ηω , η p )|A .

Then, using (4.23) and (4.24), we can infer that
T2 ≤ Ch min{s, k+1} (u, p)s e p 0,Ω .
Next, using Cauchy–Schwarz’s inequality and again (4.23) together with (4.24). we get
T3 = |c(ηω , ξ z )|
1
≤ √ Chβ∞ ηω 0,Ω z1,Ω
ν
≤ Cμh h min{s,k+1} (u, p)s e p 0,Ω ,

123
1596 Journal of Scientific Computing (2019) 80:1577–1606

hβ
where μh = √ ∞.
ν
Similarly, we have:

T4 = | j(η u , ξ z )|
≤ |η u | j |ξ z | j ≤ C|(η u , ηω , η p )|A z1,Ω ≤ Ch min{s, k+1} (u, 0)s e p 0,Ω .
The terms T5 , T6 and T7 can be readily estimated, much in the same way as before, using
the error bound in (4.17) and the fact that z ∈ H01 (Ω)3 .
Finally, the pressure estimate follows after putting all individual bounds back into (4.25).



Note that, differently from the conforming method introduced in Sect. 3, the discrete
velocity generated by scheme (4.9) is not necessarily divergence-free, and as a consequence
the method is not pressure-robust. A recent remedy in the context of DG methods can be
found in [31].

5 Numerical Tests

We present a set of examples to confirm numerically the convergence rates anticipated in

Theorems 4 and 5. We stress that whenever Γ = ∂Ω, the zero-mean condition enforcing
the uniqueness of the Bernoulli pressure is implemented using a real Lagrange multiplier
(which amounts to adding one row and one column to the corresponding matrix system).
Linear solves are performed with the direct method SuperLU.

5.1 Test 1: Experimental Convergence in 2D

For our first example we produce the error history associated with the proposed mixed finite
element and mixed DG approximations. Let us consider the following closed-form solutions
to the Oseen equations defined on the unit square domain Ω = (0, 1)2 :

sin(π x)2 sin(π y)2 cos(π y) √

u(x, y) = , ω(x, y) = ν curl u, p(x, y) = x 4 − y 4 .
− 3 sin(2π x) sin(π y)
1 3

The exact velocity has zero normal component on the whole boundary, and the exact vorticity
is employed to impose a non-homogeneous vorticity trace. In this example we are assuming
that Γ = ∂Ω, and the exact Bernoulli pressure fulfils the null-average condition. We consider
the model parameters ν = 0.1 and σ = 10, and the convecting velocity β is taken as the
exact velocity solution, which in particular satisfies the bound (2.15).
Test 1A On a sequence of uniformly refined meshes we compute errors between the exact
and approximate solutions, measured in the norms involved in the convergence analysis
of Sect. 3. The obtained error history is reported in Table 1, where the rightmost column
displays the ∞ −norm of the nodal values of the velocity divergence projected to the space
Qh , all approaching machine precision. The asymptotic O(h k+1 ) decay of the error observed
for each field variable confirms the overall optimal convergence predicted by Theorem 4.
Sample approximate solutions generated with the lowest order method on a coarse mesh are
portrayed in Fig. 1.
Test 1B With the purpose of verifying that the mixed finite element scheme is pressure-robust
and also performs well for different viscosity values, we modify the exact pressure to be

123
Journal of Scientific Computing (2019) 80:1577–1606 1597

Table 1 Test 1A: Error history (errors on a sequence of successively refined grids, convergence rates, and
divergence norms) associated with the mixed finite element method (3.4) using different polynomial degrees
DoF u − uh H rate ω − ωh Z rate  p − ph 0,Ω rate  div uh ∞

k=0
34 0.1357 – 1.2943 – 0.2002 – 1.0e−14
114 0.1129 0.2654 1.0072 0.3612 0.1219 0.7161 1.3e−14
418 0.0619 0.8655 0.5623 0.8405 0.0572 1.0910 1.3e−15
1602 0.0315 0.9763 0.2869 0.9707 0.0280 1.0311 1.2e−15
6274 0.0158 0.9952 0.1441 0.9937 0.0139 1.0072 1.0e−14
24,834 0.0079 0.9989 0.0721 0.9985 0.0069 1.0022 1.3e−15
98,818 0.0039 0.9997 0.0361 0.9996 0.0035 1.0000 1.7e−13
k=1
98 0.0980 – 0.8753 – 0.0624 – 2.4e−15
354 0.0337 1.5384 0.3448 1.3441 0.0173 1.8502 4.3e−14
1346 0.0094 1.8472 0.0979 1.8173 0.0038 2.1804 3.8e−13
5250 0.0024 1.9514 0.0255 1.9403 8.3e−04 2.1952 8.0e−14
20,738 6.4e−04 1.9873 0.0064 1.9835 1.9e−04 2.0791 1.2e−15
82,434 2.2e−04 1.9973 0.0016 1.9960 4.8e−05 2.0233 6.2e−15
328,706 3.8e−05 1.9992 4.1e−04 1.9992 1.2e−05 2.0064 1.4e−14
k=2
194 0.0563 – 0.5138 – 0.0237 – 1.7e−13
722 0.0078 2.8844 0.0893 2.7524 0.0024 3.2137 5.8e−14
2786 0.0011 2.9302 0.0121 2.9886 1.8e−04 3.1729 4.6e−14
10,946 1.3e−04 2.9872 0.0015 2.9873 1.4e−05 3.2661 4.3e−15
43,394 1.6e−05 2.9992 1.9e−04 3.0002 1.4e−06 3.2349 2.7e−15
172,802 2.1e−06 2.9981 2.3e−05 3.0014 1.6e−07 3.1252 4.9e−15
689,666 5.3e−07 2.9840 8.2e−06 3.0070 3.7e−08 2.9206 5.0e−14

(a) (b) (c)

Fig. 1 Experimental convergence in 2D. Lowest-order mixed finite element approximation of velocity mag-
nitude (a), vorticity (b), and Bernoulli pressure (c) on the unit square

p(x, y) = 1000(x 4 − y 4 ) and rerun Test 1A for decreasing values of ν (down to ν =1e-
6). Sample values of the error history are collected in Table 2, where we only show the
results for the lowest-order method. Even if the pressure error shows optimal convergence

123
1598 Journal of Scientific Computing (2019) 80:1577–1606

Table 2 Test 1B: Error history (errors on a sequence of successively refined grids, convergence rates, and
divergence norms) associated with the mixed finite element method (3.4) for k = 0, using a much larger
pressure, and for decreasing values of the viscosity
DoF h u − uh H rate ω − ωh Z rate  p − ph 0,Ω rate  div uh ∞

ν = 0.01
34 0.7071 0.1327 – 0.1709 – 191.8 – 1.76e−14
114 0.3536 0.1108 0.2601 0.1241 0.4615 107.3 0.8384 1.71e−13
418 0.1768 0.0615 0.8605 0.0610 1.0324 55.17 0.9597 5.22e−12
1602 0.0884 0.0313 0.9614 0.0293 1.0458 27.78 0.9943 4.23e−10
6274 0.0442 0.0157 0.9901 0.0144 1.0157 13.91 0.9975 3.84e−09
24834 0.0221 0.0079 0.9975 0.0072 1.0040 6.9591 0.9994 3.28e−08
98818 0.0111 0.0039 0.9994 0.0036 1.0012 3.4833 0.9998 6.46e−07
ν = 0.0001
34 0.7071 0.1327 – 0.01188 – 191.8 – 2.31e−14
114 0.3536 0.1107 0.2617 0.0061 0.9457 107.3 0.8384 1.65e−13
418 0.1768 0.0613 0.8599 0.0012 2.1252 55.17 0.9597 3.11e−12
1602 0.0884 0.0313 0.9606 0.0003 1.3765 27.78 0.9944 2.73e−10
6274 0.0442 0.0158 0.9899 0.0001 1.2588 13.91 0.9975 7.63e−09
24834 0.0221 0.0079 0.9974 7.36e−05 1.0832 6.9590 0.9994 1.64e−08
98818 0.0111 0.0039 0.9994 3.62e−05 1.0223 3.4833 0.9998 1.89e−07
ν = 0.000001
34 0.7071 0.1327 – 0.0011 – 191.8 – 2.73e−14
114 0.3536 0.1107 0.2617 0.0006 0.9621 107.3 0.8384 1.96e−13
418 0.1768 0.0613 0.8599 0.0001 1.4123 55.17 0.9597 3.98e−12
1602 0.0884 0.0313 0.9606 2.42e−05 1.2313 27.78 0.9932 1.39e−10
6274 0.0442 0.0157 0.9899 9.81e−06 1.0495 13.91 0.9975 1.06e−08
24834 0.0221 0.0079 0.9974 1.56e−06 1.5904 6.9591 0.9994 6.15e−07
98818 0.0111 0.0039 0.9994 4.99e−07 1.3651 3.4832 0.9998 2.35e−07

rates, the error values are quite large but these do not affect the error decay of the velocity,
suggesting pressure-robustness. We also see that smaller viscosity values do not pollute the
approximations. In view of (3.9) and Remark 3, perhaps a clearer numerical confirmation of
pressure-robustness can be observed by choosing exact velocity and exact vorticity belonging
to the finite element subspaces (for instance, we can use constant values u = (0, 0), ω = 0),
while keeping the exact pressure p(x, y) = x 4 − y 4 . As discussed in [29, Sect. 3], pressure
robustness of the method will imply that the velocity and vorticity errors will be practically
zero whereas the pressure error will decay with O(h k+1 ). Table 3 shows that this is precisely
the case, where we have taken ν = 0.01.
Test 1C An analogous example is now carried out to confirm numerically the convergence
rates of the DG methods defined by (4.9). The same model parameters and closed-form
solutions as in Test 1A are used, and the stabilisation constants in (4.6)–(4.8) take the values
a11 = c11 = σ and d11 = ν. The results collected in Table 4 indicate that the DG scheme
converges optimally when we measure errors in the energy A-seminorm (4.14) and in the
L2 −norm of the pressure. Here, however, we do not expect divergence-free approximate
velocities. Moreover, for the DG case we do not have pressure-robustness. Nevertheless, as

123
Journal of Scientific Computing (2019) 80:1577–1606 1599

Table 3 Test 1B: Error history (errors on a sequence of successively refined grids, convergence rate for p, and
divergence norms) associated with the mixed finite element method (3.4) for k = 0, 1, using exact velocity
and vorticity belonging to the finite element subspaces
DoF h u − uh H ω − ωh Z  p − ph 0,Ω rate  div uh ∞

k=0
34 0.7071 5.35e−16 9.88e−17 0.1918 – 1.03e−15
114 0.3536 6.15e−17 2.56e−17 0.1073 0.8384 1.78e−16
418 0.1768 7.61e−16 1.64e−16 0.0551 0.9597 2.53e−15
1602 0.0884 9.08e−15 3.11e−15 0.0277 0.9902 4.13e−14
6274 0.0442 1.90e−13 4.16e−14 0.0139 0.9975 2.59e−12
24834 0.0221 7.34e−12 6.41e−13 0.0069 0.9994 1.61e−10
98818 0.0111 1.98e−10 5.47e−12 0.0038 0.9998 9.79e−09
k=1
98 0.7071 2.34e−16 4.41e−16 0.0450 – 2.42e−16
354 0.3536 7.46e−15 1.36e−14 0.0120 1.8970 2.57e−14
1346 0.1768 7.04e−15 1.65e−14 0.0030 1.9752 3.73e−14
5250 0.0884 7.31e−14 4.94e−14 0.0007 1.9944 7.71e−13
20738 0.0442 8.57e−12 1.40e−13 0.0002 1.9982 2.62e−09
82434 0.0221 3.94e−12 5.02e−11 4.8e−05 2.0001 9.57e−08
328706 0.0111 1.98e−11 2.56e−11 1.2e−05 2.0000 5.67e−08

evidenced in Table 5, the errors in the A-norm and the pressure errors exhibit optimal decay
regardless of the viscosity values.

5.2 Test 2: Transient Flow in an Open Cavity

In this example we illustrate a more complex problem involving the non-stationary behaviour
of the flow in an open 2D cavity. The main compartment of the domain consists on a rectangle
(0, 1.2)×(0, 1) whereas smaller rectangles (0.25, 0.45)×(−0.1, 0) and (1.2, 1.3)×(0.7, 0.9)
play the role of inlet and outlet channels. The domain is discretised into an unstructured mesh
of 35433 triangular elements. Normal velocities and a compatible trace vorticity are imposed
on the whole boundary Γ = ∂Ω according to
⎧
⎪
⎨−75(x − 0.25)(0.45 − x) on y = −0.1,
u · n = 75(y − 0.7)(0.9 − y) on x = 1.3,
⎪
⎩
0 otherwise,
⎧ √
⎪
⎨75 √ ν(0.7 − 2x) on y = −0.1,
ω × n = −75 ν(1.6 − 2y) on x = 1.3,
⎪
⎩
0 otherwise,

that is, parabolic inlet and outlet profiles together with slip velocities elsewhere on ∂Ω.
We set a fluid viscosity of ν = 0.001 and use as initial velocity the solution adapted
from the previous test u0 (x, y) = [sin(π/1.3x)2 sin(π/1.1(y + 0.1))2 cos(π/1.1(y +
0.1)), − 13 sin(2/1.3π x) sin(π/1.1(y + 0.1))3 ]T . The parameter σ = 10 indicates a timestep

123
1600 Journal of Scientific Computing (2019) 80:1577–1606

Table 4 Test 1C: Error history associated to the DG method defined in (4.9) using increasing polynomial
degree
DoF h |(u − uh , ω − ωh , p − ph )|A rate  p − ph 0,Ω rate

k=0
65 0.7071 0.8031 – 0.4623 –
257 0.3536 0.4321 0.8893 0.2298 0.8384
1025 0.1768 0.2343 0.8827 0.1276 0.8438
4097 0.0884 0.1217 0.9448 0.0652 0.9668
16385 0.0442 0.0616 0.9822 0.0326 1.0010
65537 0.0221 0.0302 1.0231 0.0163 1.0005
k=1
145 0.7071 0.4787 – 0.1847 –
577 0.3536 0.1496 1.6760 0.0529 1.8033
2305 0.1768 0.0366 2.0297 0.0133 1.9914
9217 0.0884 0.0089 2.0413 0.0033 2.0220
36865 0.0442 0.0021 2.0320 0.0008 2.0361
134696 0.0221 0.0005 2.0034 0.0002 2.0049
k=2
257 0.7071 0.1882 – 0.0534 –
1025 0.3536 0.0319 2.5864 0.0111 2.2722
4097 0.1768 0.0042 2.9180 0.0013 3.0093
16385 0.0884 0.0005 3.0362 0.0002 3.1105
65537 0.0442 6.16e−5 3.0640 1.81e−5 3.1510
268049 0.0221 1.03e−5 2.9973 2.78e−6 3.0076

of Δt = 0.1, and a backward Euler discretisation implies that we take f = σ û, where
û denotes the velocity approximation at the previous iteration. The convective velocity
β = û therefore needs to be updated at each iteration. At least for the initial solution we
have that the convecting velocity satisfies the assumption (2.15). The simulation is run until
Tfinal = 4Δt and we present in Fig. 2 two snapshots of the numerical solutions at t = Δt and
t = Tfinal , computed with a second-order DG scheme, and using the stabilisation parameters
a11 = c11 = σ and d11 = ν. From the velocity plots (including a line integral convolution
visualisation), we can evidence the formation of a main vortex on the centre of the domain
plus smaller recirculation areas that emerge on the top left and bottom left corners, together
with a preferential path joining the inlet and outlet boundaries.

5.3 Test 3: Lid Driven Cavity Flow

For this classical benchmark problem we consider zero external forces and concentrate on the
case where flow recirculation occurs by Dirichlet conditions only. First we consider the two-
dimensional case, where on the top lid of the unit square (at y = 1) we set a unidirectional
velocity of unit magnitude, whereas no-slip velocity and zero tangential vorticity are imposed
on the remaining sides of the boundary. We set the parameters ν = 0.001, σ = 50 and employ
a structured mesh of 4096 elements. The initial velocity is computed from a Stokes solution
(setting both σ and β to zero). We compare the results obtained with our lowest-order FE

123
Journal of Scientific Computing (2019) 80:1577–1606 1601

Table 5 Test 1C: Error history associated to the DG method defined in (4.9) for the lowest-order case and
with decreasing viscosity
DoF h |(u − uh , ω − ωh , p − ph )|A rate  p − ph 0,Ω rate

ν = 0.01
65 0.7071 0.5059 – 0.2034 –
257 0.3536 0.2884 0.8105 0.1264 0.6864
1025 0.1768 0.1642 0.8126 0.0842 0.7585
4097 0.0884 0.0902 0.8632 0.0507 0.7929
16385 0.0442 0.0473 0.9315 0.0274 0.8878
65537 0.0221 0.0244 0.9514 0.0140 0.9606
ν = 0.0001
65 0.7071 0.3787 – 0.1963 –
257 0.3536 0.1882 1.0091 0.1102 0.8322
1025 0.1768 0.0796 1.2426 0.0562 0.9709
4097 0.0884 0.0363 1.1314 0.0285 0.9916
16385 0.0442 0.0176 1.0423 0.0141 0.9974
65537 0.0221 0.0087 1.0154 0.0071 0.9992
ν = 0.000001
65 0.7071 0.3434 – 0.1967 –
257 0.3536 0.1490 1.2025 0.1088 0.8488
1025 0.1768 0.0586 1.3461 0.0553 0.9764
4097 0.0884 0.0239 1.2931 0.0278 0.9933
16385 0.0442 0.0108 1.1437 0.0135 0.9982
65537 0.0221 0.0051 1.0654 0.0069 0.9996

(a) (b) (c)

(d) (e) (f)

Fig. 2 Test 2: Second-order DG approximation of the transient flow patterns in an open cavity after one
timestep (a–c) and after four time steps (d–f), using Δt = 0.1 and ν = 0.001

123
1602 Journal of Scientific Computing (2019) 80:1577–1606

(a) (b)

(c) (d) (e)

(f) (g) (h)

Fig. 3 Test 3. Line integral convolution visualisation of velocity for the 2D cavity flow benchmark with
ν = 0.001 (a), cuts of the vertical velocity and vorticity on the line y = 0.5 (b, c); and horizontal velocity and
vorticity on the line x = 0.5 (d, e). The circle markers indicate benchmark values from [13]. Lowest-order
approximation of velocity (f) for the 3D case with ν = 0.0025, shown at t = 20; and profile of the horizontal
velocity on the line y = 0.5, z = 0.5 (g), and of the vertical velocity on the line x = 0.5, y = 0.5 (h). The
asterisks indicate benchmark values from [23] and all numerical approximations for this test were obtained
with the lowest-order mixed finite element method

scheme against the benchmark data from [13] (produced with a spectral method applied to a
vorticity-based formulation). Figure 3a shows the generated velocity profile at t = 20, having
all the flow features expected for this regime. The solid lines in Fig. 3b–e portray cuts of the
solution on the mid-lines of the domain, whereas the circle markers indicate benchmark data.
The approximate vorticity has been rescaled with ν −1/2 to reflect the overall agreement with
the results reported in [13].
We also test the 3D implementation and formulation by conducting the same benchmark
on the unit cube Ω = (0, 1)3 . Again, boundary Σ is the top plate (defined by z = 1), where we
set tangential velocity of magnitude one, and on Γ = ∂Ω \ Σ we consider no-slip velocities

123
Journal of Scientific Computing (2019) 80:1577–1606 1603

(a) (b) (c)

(d) (e) (f)

(g) (h) (i)

Fig. 4 Test 4. Transients at early (left), moderate (middle column) and advanced (right) times, for the Kelvin–
Helmholtz instability problem computed with the first-order DG scheme (4.9). Velocity magnitude (a–c),
scalar vorticity (d–f), and Bernoulli pressure (g–i)

and zero tangential vorticity. The fluid viscosity is now ν = 0.0025 and a structured mesh of
58752 tetrahedral elements is employed. Once again we focus on a non-stationary regime with
a backward Euler scheme, now using σ = 10, and proceed to update the convective velocity
and the right-hand side using the velocity approximation at the previous time iteration. The
velocity field for a converged solution after 200 time steps is shown in Fig. 3f. We can observe
the expected asymmetric vortex forming parallel to the x z plane (also the generation of high
pressure near the corners where the Dirichlet velocity datum has a discontinuity). We have
also compared our results with the benchmark values obtained in [23] (using multiquadric
differential quadratures) for a Reynolds number of 400: the solid lines in Fig. 3g–h show
velocity profiles captured on the plane y = 0.5, concentrating on the vertical and horizontal
centrelines, where we also include the data from [23] (in asterisks) showing a reasonable
match (we have rotated the data, as in their tests the unit velocity is imposed on the face
y = 1).

5.4 Test 4: Kelvin–Helmholtz Mixing Layer

We close this section with a benchmark test related to the well-known vortex formation
mechanisms known as the Kelvin–Helmholtz instability problem. The setup of the test follows
the specifications in [35] (see also [14]), the transient Oseen equations are solved on the unit
square Ω = (0, 1)2 and the bottom and top walls constitute Γ , where we impose a free-
slip velocity condition and zero vorticity. The left and right walls are regarded as a periodic
boundary. The initial velocity is

123
1604 Journal of Scientific Computing (2019) 80:1577–1606

⎛
⎞
a x) + cos(w b x)] (2y−1) − (y−1/2)
2
⎜u ∞ tanh((2y − 1)/δ0 ) − cn u ∞ [cos(w∞ ∞ exp δ2 δ02 ⎟
u=⎜
⎝

0 ⎟,
⎠
cn u ∞ exp − (y−1/2)
2

δ2
[w∞
a sin(w a x) + w b sin(w b x)]
∞ ∞ ∞
0

with perturbation scaling cn = 0.001, reference velocity u ∞ = 1, w∞a = 8π, w b = 20π,

∞
¯
δ0 = 1/28. The characteristic time is t = δ0 /u ∞ , the Reynolds number is Re= 10000,
and the kinematic viscosity is ν = δ0 u ∞ /Re. We use a structured mesh of 128 seg-
ments per side, representing 131072 triangular elements, and we solve the problem using
our first-order DG scheme, setting again the stabilisation constants to a11 = c11 =
σ = 1/Δt and d11 = ν, where the timestep is taken as Δt = t¯/20. The specifi-
cation of this problem implies that the solutions will be quite sensitive to the initial
perturbations present in the velocity, which will amplify and consequently vortices will
appear. We proceed to compute numerical solutions until the dimensionless time t = 7,
and present in Fig. 4 sample solutions at three different simulation times. For visualisa-
tion purposes we zoom into the region 0.25 ≤ y ≤ 0.75, where all flow patterns are
concentrated.

Acknowledgements This work has been supported by CNRS through the PEPS programme; by CONICYT -
Chile through FONDECYT project 11160706; by DIUBB, Universidad del Bío-Bío through projects 165608–
3/R and 171508 GI/VC; by CONICYT-Chile through the project AFB170001 of the PIA Program: Concurso
Apoyo a Centros Científicos y Tecnológicos de Excelencia con Financiamiento Basal. by DIDULS, Uni-
versidad de La Serena through project PR17151; and by Unversidad de Córdoba through the programme
“Estrategias para la Sostenibilidad de Grupos de Investigación”, project FCB-03-17.

Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International
License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and repro-
duction in any medium, provided you give appropriate credit to the original author(s) and the source, provide
a link to the Creative Commons license, and indicate if changes were made.

References
1. Alonso, A., Valli, A.: An optimal domain decomposition preconditioner for low-frequency time harmonic
Maxwell equations. Math. Comput. 68, 607–631 (1999)
2. Amara, M., Capatina-Papaghiuc, D., Trujillo, D.: Stabilized finite element method for Navier–Stokes
equations with physical boundary conditions. Math. Comput. 76(259), 1195–1217 (2007)
3. Amoura, K., Azaïez, M., Bernardi, C., Chorfi, N., Saadi, S.: Spectral element discretization of the vorticity,
velocity and pressure formulation of the Navier–Stokes problem. Calcolo 44(3), 165–188 (2007)
4. Anaya, V., Bendahmane, M., Mora, D., Ruiz-Baier, R.: On a vorticity-based formulation for reaction-
diffusion-Brinkman systems. Netw. Heterog. Media 13(1), 69–94 (2018)
5. Anaya, V., Gatica, G.N., Mora, D., Ruiz-Baier, R.: An augmented velocity–vorticity–pressure formulation
for the Brinkman problem. Int. J. Numer. Methods Fluids 79(3), 109–137 (2015)
6. Anaya, V., Mora, D., Oyarzúa, R., Ruiz-Baier, R.: A priori and a posteriori error analysis for a mixed
scheme for the Brinkman problem. Numer. Math. 133(4), 781–817 (2016)
7. Anaya, V., Mora, D., Ruiz-Baier, R.: An augmented mixed finite element method for the vorticity–
velocity–pressure formulation of the Stokes equations. Comput. Methods Appl. Mech. Eng. 267, 261–274
(2013)
8. Anaya, V., Mora, D., Ruiz-Baier, R.: Pure vorticity formulation and Galerkin discretization for the
Brinkman equations. IMA J. Numer. Anal. 37(4), 2020–2041 (2017)
9. Azaïez, M., Bernardi, C., Chorfi, N.: Spectral discretization of the vorticity, velocity and pressure formu-
lation of the Navier–Stokes equations. Numer. Math. 104(1), 1–26 (2006)
10. Bochev, P.B.: Analysis of least-squares finite element methods for the Navier–Stokes equations. SIAM J.
Numer. Anal. 34(5), 1817–1844 (1997)

123
Journal of Scientific Computing (2019) 80:1577–1606 1605

11. Bochev, P.B.: Negative norm least-squares methods for the velocity–vorticity–pressure Navier–Stokes
equations. Numer. Methods PDEs 15(2), 237–256 (1999)
12. Bonelle, J., Ern, A.: Analysis of compatible discrete operator schemes for the Stokes equations on poly-
hedral meshes. IMA J. Numer. Anal. 35(4), 1672–1697 (2015)
13. Botella, O., Peyret, R.: Benchmark spectral results on the lid-driven cavity flow. Comput. Fluids 27(4),
421–433 (1998)
14. Burman, E., Ern, A., Fernández, M.A.: Fractional-step methods and finite elements with symmetric
stabilization for the transient Oseen problem. ESAIM Math. Model. Numer. Anal. 51(2), 487–507 (2017)
15. Cai, Z., Chen, B.: Least-squares method for the Oseen equation. Numer. Methods PDEs 32, 1289–1303
(2016)
16. Caucao, S., Mora, D., Oyarzúa, R.: A priori and a posteriori error analysis of a pseudostress-based mixed
formulation of the Stokes problem with varying density. IMA J. Numer. Anal. 36(2), 947–983 (2016)
17. Chang, C.L., Yang, S.-Y.: Analysis of the [L 2 , L 2 , L 2 ] least-squares finite element method for incom-
pressible Oseen-type problems. Int. J. Numer. Anal. Model. 4(3–4), 402–424 (2007)
18. Çeşmelioğlu, A., Cockburn, B., Nguyen, N.C., Peraire, J.: Analysis of HDG methods for Oseen equations.
J. Sci. Comput. 55(2), 392–431 (2013)
19. Ciarlet, P.: The Finite Element Method for Elliptic Problems. North-Holland, Amsterdam (1978)
20. Cockburn, B., Kanschat, G., Schötzau, D.: The local discontinuous Galerkin method for linear incom-
pressible fluid flow: a review. Comput. Fluids 34(4–5), 491–506 (2005)
21. Cockburn, B., Kanschat, G., Schötzau, D.: The local discontinuous Galerkin method for the Oseen equa-
tions. Math. Comput. 73(246), 569–593 (2003)
22. Cockburn, B., Kanschat, G., Schötzau, D., Schwab, C.: Local discontinuous Galerkin methods for the
Stokes system. SIAM J. Numer. Anal. 40(1), 319–343 (2002)
23. Ding, H., Shu, C., Yeo, K.S., Xu, D.: Numerical computation of three-dimensional incompressible vis-
cous flows in the primitive variable form by local multiquadric differential quadrature method. Comput.
Methods Appl. Mech. Eng. 195, 516–533 (2006)
24. Duan, H.-Y., Liang, G.-P.: On the velocity–pressure–vorticity least-squares mixed finite element method
for the 3D Stokes equations. SIAM J. Numer. Anal. 41(6), 2114–2130 (2003)
25. Dubois, F., Salaün, M., Salmon, S.: First vorticity–velocity–pressure numerical scheme for the Stokes
problem. Comput. Methods Appl. Mech. Eng. 192(44–46), 4877–4907 (2003)
26. Gatica, G.N.: A Simple Introduction to the Mixed Finite Element Method. Theory and Applications.
Springer Briefs in Mathematics. Springer, Cham (2014)
27. Gatica, G.N., Gatica, L.F., Márquez, A.: Augmented mixed finite element methods for a vorticity-based
velocity–pressure–stress formulation of the Stokes problem in 2D. Int. J. Numer. Methods Fluids 67(4),
450–477 (2011)
28. Girault, V., Raviart, P.A.: Finite Element Methods for Navier–Stokes Equations. Theory and Algorithms.
Springer, Berlin (1986)
29. John, V., Linke, A., Merdon, C., Neilan, M., Rebholz, L.G.: On the divergence constraint in mixed finite
element methods for incompressible flows. SIAM Rev. 59(3), 492–544 (2017)
30. Kreeft, J., Gerritsma, M.: Mixed mimetic spectral element method for Stokes flow: a pointwise divergence-
free solution. J. Comput. Phys. 240, 284–309 (2013)
31. Lederer, P.L., Lehrenfeld, C., Schöberl, J.: Hybrid discontinuous Galerkin methods with relaxed H(div)-
conformity for incompressible flows. Part I. SIAM J. Numer. Anal. 56(4), 2070–2094 (2018)
32. Mohapatra, S., Ganesan, S.: A non-conforming least squares spectral element formulation for Oseen
equations with applications to Navier–Stokes equations. Numer. Funct. Anal. Optim. 37(10), 1295–1311
(2016)
33. Nicolaides, R.A.: Existence, uniqueness and approximation for generalized saddle point problems. SIAM
J. Numer. Anal. 19, 349–357 (1982)
34. Salaün, M., Salmon, S.: Low-order finite element method for the well-posed bidimensional Stokes prob-
lem. IMA J. Numer. Anal. 35(1), 427–453 (2015)
35. Schroeder, P.W., John, V., Lederer, P.L., Lehrenfeld, C., Lube, G., Schöberl, J.: On reference solutions and
the sensitivity of the 2D Kelvin–Helmholtz instability problem. Comput. Math. Appl. 77(4), 1010–1028
(2019)
36. Tsai, C.-C., Yang, S.-Y.: On the velocity–vorticity–pressure least-squares finite element method for the
stationary incompressible Oseen problem. J. Comput. Appl. Math. 182(1), 211–232 (2005)

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.

123
1606 Journal of Scientific Computing (2019) 80:1577–1606

Affiliations

V. Anaya1 · A. Bouharguane2 · D. Mora1,3 · C. Reales4 · R. Ruiz-Baier5 ·

N. Seloula6 · H. Torres7

B R. Ruiz-Baier
ruizbaier@maths.ox.ac.uk
V. Anaya
vanaya@ubiobio.cl
A. Bouharguane
afaf.bouharguane@math.u-bordeaux.fr
D. Mora
dmora@ubiobio.cl
C. Reales
creales@correo.unicordoba.edu.co
N. Seloula
nour-elhouda.seloula@unicaen.fr
H. Torres
htorres@userena.cl
1 GIMNAP, Departamento de Matemática, Universidad del Bío-Bío, Concepción, Chile
2 Institut de Mathématiques de Bordeaux, CNRS UMR 5251, Université de Bordeaux,
33405 Talence, France
3 Centro de Investigación en Ingeniería Matemática (CI2MA), Universidad de Concepción,
Concepción, Chile
4 Departamento de Matemáticas y Estadísticas, Universidad de Córdoba, Montería, Colombia
5 Mathematical Institute, University of Oxford, Radcliffe Observatory Quarter, Woodstock Road,
Oxford OX2 6GG, UK
6 LMNO, CNRS UMR 6139, Université de Caen, 5186 Caen, France
7 Departamento de Matemáticas, Universidad de La Serena, La Serena, Chile

123