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

net/publication/338499981

Superconvergent mixed finite element methods for the pseudostress-velocity

formulation of the Oseen equation

Preprint · January 2020

CITATIONS READS

0 196

2 authors:

Xi Chen Yuwen Li
Pennsylvania State University Zhejiang University
13 PUBLICATIONS 40 CITATIONS 46 PUBLICATIONS 317 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Yuwen Li on 18 January 2020.

The user has requested enhancement of the downloaded file.

 Noname manuscript No.
(will be inserted by the editor)

Superconvergent mixed finite element methods for

the pseudostress-velocity formulation of the Oseen
equation

Xi Chen · Yuwen Li
arXiv:2001.02805v1 [math.NA] 9 Jan 2020

Received: January 10, 2020 / Accepted: date

Abstract We presents a priori and superconvergence error estimates of mixed

finite element methods for the pseudostress-velocity formulation of the Oseen
equation. In particular, we derive supercloseness estimates on the velocity and
a priori error estimates under unstructured grids. We also obtain superclose-
ness estimates on the pseudostress under certain structured grids. With the
help of supercloseness and postprocessing procedures, postprocessed numerical
velocity and pseudostress are shown to be superconvergent. We finally present
numerical experiments to verify the theoretical results and illustrate the effec-
tiveness of the superconvergent recovery-based a posteriori error estimator.
Keywords pseudostress, superconvergence, supercloseness, Oseen equation,
Stokes equation, postprocessing
Mathematics Subject Classification (2000) 65N30, 65N15

1 Introduction

In fluid mechanics, the Oseen equations describe the flow of a viscous and in-
compressible fluid at small Reynolds numbers. Instead of dropping the advec-
tive term completely, the Oseen approximation linearizes the advective accel-
eration term u ·∇u to b·∇u, where b represents the velocity at large distance.
As a result it provides a lowest-order solution that is uniformly valid every-
where in the flow field. Since the Oseen equations partly account for the inertia
terms (at large distance), they have better approximation in the far field while
keeping the same order of accuracy as Stokes approximation near the body,
see, e.g., [41]. Mathematically speaking, the Oseen equation can be viewed as

X. Chen: Department of Mechanical Engineering, Pennsylvania State University, University

Park, PA 16802. Email:[email protected]
BY. Li: Department of Mathematics, Pennsylvania State University, University Park, PA
16802. Email:[email protected]
2 X. Chen, Y. Li

a linearized Navier–Stokes equation arising from fixed point iteration. It is the

purpose of this paper to shed light on a priori and superconvergence analysis
of pseudostress-velocity mixed methods for the Navier–Stokes equation.
Let Ω ⊂ Rd be a bounded Lipschiz domain. Let b, f : Ω → Rd , c : Ω → R
and g : ∂Ω → Rd be given data. Let u : Ω → Rd be the velocity field and
p : Ω → R denote the pressure subject to the constraint
Z
pdx = 0. (1.1)
Ω

The Oseen equation under the Dirichlet boundary condition is

−∆u + b · ∇u + cu + ∇p = f in Ω, (1.2a)
div u = 0 in Ω, (1.2b)
u = g on ∂Ω. (1.2c)

As a default, vectors such as b, f , u, g are always arranged as columns unless

confusion arises. In (1.2a), ∇u is the Jacobian matrix of u, and we adopt the
convention b · ∇u := (∇u)b. When b = 0 (resp. b = 0, c = 0), (1.2) reduces to
the Brinkman equation (resp. Stokes equation). Numerical analysis of Oseen,
Brinkman, or Stokes equations based on the velocity-pressure formulation is
extensive, see, e.g., [37,9,38,34] for conforming mixed methods, [21,24,26,35]
for hybridized discontinuous Galerkin methods, [48,55,45] for weak Galerkin
methods, and [1] for virtual element methods.
Let I ∈ Rd×d denote the identity matrix. Define the pseudostress
σ := ∇u − pI

which is not symmetric; compared with the symmetric stress

1
σ S := (∇u + (∇u)T ) − pI.
2
In a series of papers [20,19,18], Cai et al. proposed the pseudostress-velocity
formulation of the Stokes and Navier–Stokes equations, which allows classi-
cal Poisson-type mixed discretization by using Raviart–Thomas or Brezzi–
Douglas–Marini elements. Let tr σ denote the trace of the matrix σ and
1
Aσ := σ − (tr σ)I
d
denote the deviatoric part of σ. Direct calculation shows that the Oseen equa-
tion (1.2) is equivalent to the following pseudostress-velocity formulation
Aσ = ∇u,
(1.3)
− div σ + b · Aσ + cu = f
subject to the constraints
Z
u|∂Ω = g, tr σdx = 0.
Ω
Superconvergent mixed methods for the Oseen equation 3

Here div σ is the row-wise divergence of the matrix σ. Comparing to exten-

sive numerical results on velocity-pressure formulation of Stokes-related equa-
tions, the numerical analysis of pseudostress-based methods is restricted to
classical mixed methods [19,36], virtual element methods [16,17], discontinu-
ous Galerkin methods [51] and adaptive mixed methods [23,22,40,42] for the
Stokes or Brinkman equation. Very few work is devoted to the mixed formu-
lation of the Oseen equation (1.3), see, e.g., [50] for a lowest order upwinded
mixed method on rectangular meshes.
Given a vector space X, let [X]n denote the Cartesian product of n copies
of X and [X]n×n the space of n × n matrices whose components are contained
in X. Let H(div, Ω) = {v ∈ [L2 (Ω)]d : div v ∈ L2 (Ω)} be the usual H(div)
space. Let

V = [L2 (Ω)]d ,
Z
2 d×d
Σ = {τ ∈ [L (Ω)] : τi ∈ H(div, Ω), 1 ≤ i ≤ d, tr τ dx = 0},
Ω

where τi is the i-th row of τ . For scalar-, vector-, or matrix-valued functions, we

use (·, ·) and h·, ·i to denote the L2 (Ω)- and L2 (∂Ω)-inner product, respectively.
The variational formulation of (1.3) seeks σ ∈ Σ and u ∈ V such that

(Aσ, τ ) + (div τ , u) = hg, τ ni, τ ∈ Σ,

(1.4)
−(div σ, v) + (b · Aσ, v) + (cu, v) = (f , v), v∈V,

where n is the outward normal to ∂Ω.

Let Th be a conforming simplicial or cubical partition of Ω that is shape-
regular in the sense that
rK
max := Cratio < ∞
K∈Th ρK
for some absolute constant Cratio , where rK , ρK are radii of circumscribed and
inscribed spheres of K. For each K ∈ Th , let hK denote the diameter of K
and h := maxK∈Th hK < 1 the mesh size of Th . Given an integer r ≥ 0, let
Σe r (K), Vr (K) be suitable finite-dimensional vector spaces defined on K and

e h := {τ ∈ H(div, Ω) : τ |K ∈ Σ
Σ e r (K) for K ∈ Th },
Vh := {v ∈ L2 (Ω) : v|K ∈ Vr (K) for K ∈ Th }.

For instance, when Th is a simplicial partition, set

e r (K) := [Pr (K)]d + Pr (K)x,
Σ Vr (K) := Pr (K), (1.5)

where Pr (K) is the space of polynomials of degree at most r on K, and x =

eh ×
(x1 , . . . , xd )T denotes the position vector. In this case, the corresponding Σ
Vh is the classical Raviart–Thomas (RT) [52] space. Another possible choice is
e r (K) := [Pr+1 (K)]d ,
Σ Vr (K) := Pr (K). (1.6)
4 X. Chen, Y. Li

In this case, the corresponding Σe h × Vh becomes the Brezzi–Douglas–Marini

(BDM) [15] space.
Given a hypercube K ⊂ Rd , let Qr (K) denote the space of polynomials on
K of degree at most r in xi with 1 ≤ i ≤ d. For a cubical mesh Th , the pair
Σe h × Vh can be the cubical RT element using the shape functions

e r (K) := [Qr (K)]d + xQr (K),

Σ Vr (K) := Qr (K). (1.7)

Let Πe h : [H 1 (Ω)]d → Σ e h be the canonical interpolation onto Σ

e h de-
e 2
termined by the degrees of freedom of Σh . Let Ph : L (Ω) → Vh be the
L2 -projection onto Vh , i.e.,

(Ph v, w) = (v, w), v ∈ L2 (Ω), w ∈ Vh . (1.8)

Throughout the rest, we assume the commutativity property

e h = Ph div,
div Π (1.9)

which is satisfied by at least the aforementioned RT, BDM, and cubical RT el-
ements. Readers are referred to e.g., [15,13,14] for other cubical or rectangular
elements that satisfy (1.9). Let Σh be the matrix version of Σ e h , i.e.,

e h for 1 ≤ i ≤ d}.
Σh := {τ ∈ Σ : τi ∈ Σ

Let Vh be the vector-valued broken piecewise polynomial space

Vh := {v = (v1 , v2 , . . . , vd )T ∈ V : vi ∈ Vh for 1 ≤ i ≤ d}.

The mixed method for (1.3) seeks (σh , uh ) ∈ Σh × Vh satisfying

(Aσh , τ ) + (div τ , uh ) = hg, τ ni, τ ∈ Σh ,

(1.10)
−(div σh , v) + (b · Aσh , v) + (cuh , v) = (f , v), v ∈ Vh .

In this paper, we shall develop a priori and superconvergence error esti-

mates for (1.10). To the best of our knowledge, even a priori error estimates of
(1.10) is not available in literature. We emphasize that the presence of lower
order terms is a major challenge in error analysis of mixed methods, see, e.g.,
[30,28] for elliptic equations with lower order terms and [5] for the perturbed
Hodge–Laplace equation. For instance, due to the convection term b · Aσ and
possibly negative coefficient c, the construction of the discrete inf-sup condi-
tion is far from obvious. Hence we shall use the duality argument to derive a
priori error estimates, which also yields improved decoupled error estimates.
The main contribution of this paper is summarized as follows.
• We develop a supercloseness estimate on kPh u − uh k, which reduces to a
result in [23] for the Stokes equation (b = 0, c = 0).
Superconvergent mixed methods for the Oseen equation 5

• For the mixed method (1.10), we establish optimal a priori error estimates
in the energy-norm and L2 -norm using duality argument. It is noted that
mixed methods for the scalar elliptic equation
− div(A∇u + bu) + cu = f (1.11)
have been analyzed in [30,28], where A is a uniformly elliptic matrix-valued
function. In contrast, the deviatoric operator A in (1.2) is singular, which is
a major difficulty in error analysis. We overcome this difficulty by deriving
sharp estimates on k div(Πh σ − σh )k−1 and kPh u − uhk for closing a loop
in a priori error estimation, see Section 2 for details.
• We establish supercloseness estimates on kΠh σ − σh k for the lowest order
triangular and general rectangular RT elements. Even for the Stokes equa-
tion, such estimates are not known in literature. In [43], Li obtained similar
estimates of mixed methods for the elliptic equation (1.11). The novelty
of supercloseness analysis here is two-fold. First, (1.4) with a uniformly
elliptic A is still different from the mixed formulation of (1.11). Second,
the singular deviatoric operator A is handled in the framework of finite
element exterior calculus, see Section 3 for details.
Built upon the aforementioned supercloseness estimates and available post-
processing operators in literature (see, e.g., [59,23,8,57,7]), we obtain the re-
covery superconvergence in Theorems 4.1 and 4.2. Regarding superconvergence
results of velocity-pressure methods for the Stokes-type equation, readers are
referred to e.g., [54,56,27] for superconvergence by two-grid L2 -projections,
[49,32,44] for superconvergent recovery of lowest-order methods, and [26,35]
for superconvergent hybridized discontinuous Galerkin methods.
The rest of this paper is organized as follows. In section 2, we develop a apri-
ori error estimates of the mixed method (1.10) and supercloseness estimate on
kPh u − uh k. In Section 3, we develop supercloseness estimate on kΠh σ − σh k.
Section 4 is devoted to superconvergent postprocessing procedures. Numerical
results for smooth, singular, and convection-dominated problems are presented
in Section 5.

2 A priori error estimates

Let k·kr denote the k·kH r (Ω) -norm, |·|r the |·|H r (Ω) -semi norm, and k·k := k·k0
the k · kL2 (Ω) -norm. Let |||·||| denote the H(div)-norm,

|||τ |||2 := kτ k2 + k div τ k2 .

In this section, we shall derive a priori error estimates in Theorem 2.1. The
same analysis implies that the mixed method (1.10) admits a unique solution
provided h is sufficiently small. Throughout the rest, A . B means that A ≤
CB, where C > 0 is a generic constant dependent solely on b, c, Ω, Cratio .
The operator A is singular and satisfies
(Aτ1 , τ2 ) = (τ1 , Aτ2 ) = (Aτ1 , Aτ2 ), τ1 , τ2 ∈ [L2 (Ω)]d×d . (2.1)
6 X. Chen, Y. Li

In addition, it holds that (see [20,2])

kτ k . kτ kA + k div τ k−1 , τ ∈ Σ, (2.2)
1
where kτ kA := (Aτ , τ ) 2 . (2.2) is a key ingredient in the analysis of pseudostress-
based methods. Define Πh : [H 1 (Ω)]d×d → Σh by
Z 
Πh τ := Π ehτ − 1 e h τ dx I, τ ∈ Σ,
tr Π (2.3)
d|Ω| Ω

where Πe h is applied to each row of τ . By abuse of notation, we also use

Ph : [L2 (Ω)]d → Vh to denote the component-wise L2 -projection. It follows
from (1.9) that
div Πh τ = Ph div τ , τ ∈ [H 1 (Ω)]d×d . (2.4)
For convenience, we introduce the interpolation errors
ρh := u − Ph u, ζh := σ − Πh σ,
which can be easily estimated by (see [18])
kv − Ph vk . hs |v|s , (2.5a)
s
kτ − Πh τ k . h |τ |s , (2.5b)
k div(τ − Πh τ )k . hs | div τ |s , (2.5c)
R
where 0 ≤ s ≤ r + 1, Ω tr τ dx = 0, and v, τ satisfy the regularity indicated
by the right hand sides. For the BDM element (1.6), it holds that
kτ − Πh τ k . hs |τ |s , 0 ≤ s ≤ r + 2. (2.6)
The essential errors to be estimated are
eh := Ph u − uh , ξh := Πh σ − σh .
Subtracting (1.10) from (1.4), we obtain the error equation
(A(σ − σh ), τh ) + (div τh , u − uh ) = 0, (2.7a)
−(div(σ − σh ), vh ) + (b · A(σ − σh ), vh ) + (c(u − uh ), vh ) = 0, (2.7b)
for all (τh , vh ) ∈ Σh × Vh .
e ∈ Σ and
To estimate keh k, we consider the dual problem of (1.4): Find σ
e ∈ V such that
u
(Ae e ) − (b · Aτ , u)
σ , τ ) + (div τ , u e = 0, τ ∈ Σ, (2.8a)
−(div σ u, v) = (eh , v) v ∈ V .
e , v) + (ce (2.8b)
Throughout the rest, it is assumed that (2.8) admits the elliptic regularity
ke
σ k1 + ku
ek1 . keh k. (2.9)
The next lemma is a supercloseness estimate which is crucial for both a priori
and superconvergence error analysis.
Superconvergent mixed methods for the Oseen equation 7

Lemma 2.1 It holds that

keh k . h |||σ − σh ||| + ku − uh k + kρh k .

Proof Taking v = eh in (2.8b), we obtain

keh k2 = −(div σ
e , eh ) + (ce
u, eh ). (2.10)

Using the definition (1.8) and the property (2.4), we have

(div σ e , Ph u − u + u − uh )
e , eh ) = (Ph div σ
(2.11)
= (Ph div σ e , u − uh ).
e , u − uh ) = (div Πh σ

(2.7a) with τh = Πh σ
e implies that

e , u − uh ) = −(A(σ − σh ), Πh σ
(div Πh σ e)
(2.12)
= (A(σ − σh ), σ e ) − (A(σ − σh ), σ
e − Πh σ e ).

Using (2.1) and setting τe = σ − σh in (2.8a), we have

(A(σ − σh ), σ
e ) = (Aeσ , σ − σh )
= −(div(σ − σh ), ue) + (b · A(σ − σh ), u
e)
= −(div(σ − σh ), ue − Ph u e ) + (b · A(σ − σh ), u
e ) − (div(σ − σh ), Ph u e ).

It then follows from the previous equation and (2.7b) with vh = Ph u

e that

(A(σ − σh ), σ
e ) = −(div(σ − σh ), ue − Ph ue ) + (b · A(σ − σh ), u
e)
(2.13)
e ) − (c(u − uh ), Ph u
− (b · A(σ − σh ), Ph u e ).

As a result of (2.11), (2.12) and (2.13), we obtain

e , eh ) = (A(σ − σh ), σ
(div σ e − Πh σ e ) + (div(σ − σh ), ue − Ph ue)
(2.14)
− (b · A(σ − σh ), u
e − Ph ue ) + (c(u − uh ), Ph u
e ).

On the other hand, the second term on the right hand side of (2.10) is

(ce u, Ph u − u + u − uh )
u, eh ) = (ce
(2.15)
u − Ph (ce
= (ce u), Ph u − u) + (c(u − uh ), u
e ).

Finally with (2.14) and (2.15), the error in (2.10) can be written as

keh k2 = (A(σ − σh ), Πh σ
e −σ e ) − (div(σ − σh ), u
e − Ph u
e)
+ (b · A(σ − σh ), u e) + (c(u − uh ), u
e − Ph u e − Ph ue) (2.16)
u − Ph (ce
+ (ce u), Ph u − u).

Using (2.16), (2.5), and the Cauchy–Schwarz inequality, we obtain

keh k2 . hke ek1 |||σ − σh ||| + ku − uh k + kρh k .
σk1 kσ − σh k + hku

Combining the above estimate with (2.9) completes the proof. ⊓

⊔
8 X. Chen, Y. Li

Lemma 2.1 is a supercloseness estimate, i.e., Ph u and uh are much closer

than the distance predicted by standard a priori error estimates. However, a
priori error estimates on kσ − σh k and k div(σ − σh )k are not known at the
moment. When deriving a priori error estimates, we need the L2 and negative
norm estimates of div ξh .
Lemma 2.2 It holds that
k div ξh k . kζh k + kξh k + keh k + hkρh k, (2.17a)


k div ξh k−1 . h |||ζh ||| + |||ξh ||| + kρh k + keh k. (2.17b)
Lemme 2.2 is also essential for proving the supercloseness estimate in Theorem
3.1 and the proof is postponed in the end of this section. Throughout the rest,
we say h is sufficiently small provided h ≤ h0 , where h0 is a constant dependent
solely on Ω, b, c, Cratio . Now we are in a position to present the main result
in this paper.
Theorem 2.1 For sufficiently small h, it holds that
ku − uh k . kρh k + h|||ζh |||, (2.18a)
|||σ − σh ||| . |||ζh ||| + hkρh k, (2.18b)
kσ − σh k . kζh k + hk div ζh k + hkρh k. (2.18c)
Proof Using Lemma 2.1 and the triangle inequality
ku − uh k ≤ kρh k + keh k,
|||σ − σh ||| ≤ |||ζh ||| + |||ξh |||,
we obtain

keh k . h |||ζh ||| + |||ξh ||| + kρh k + hkeh k. (2.19)
Plugging (2.17a) into (2.19) and kicking hkeh k with sufficiently small h into
the left hand side yields


keh k . h |||ζh ||| + kξh k + kρh k . (2.20)
On the other hand, with help of (2.7a) with τh = ξh , we obtain
(Aξh , ξh ) = (A(Πh σ − σ), ξh ) − (div ξh , u − uh )
(2.21)
= (A(Πh σ − σ), ξh ) − (div ξh , Ph u − uh ).
A combination of (2.2), (2.21), with a Young’s inequality then shows that
kξh k2 . (Aξh , ξh ) + k div ξh k2−1
. kζh kkξh k + k div ξh kkeh k + k div ξh k2−1
ε2 ε−2 ε2 ε−2
≤ kξh k2 + kζh k2 + k div ξh k2 + keh k2 + k div ξh k2−1 ,
2 2 2 2
or equivalently,
kξh k ≤ C ∗ εkξh k + ε−1 kζh k + εk div ξh k + ε−1 keh k + k div ξh k−1 ), (2.22)
Superconvergent mixed methods for the Oseen equation 9

where ε > 0, C ∗ is independent of ε. Collecting (2.22), Lemma 2.2 and using

the triangle inequality, we deduce that


kξh k ≤ C ∗ ε kξh k + kζh k + keh k + hkρh k


+ ε−1 kζh k + keh k + h|||ζh ||| + hkξh k + hkρh k + keh k .
Then we could choose ε and h to be small enough to obtain


kξh k ≤ C kζh k + hk div ζh k + keh k + hkρh k . (2.23)

Plugging (2.23) into (2.20) with sufficiently small h leads to

keh k ≤ Ch |||ζh ||| + kρh k . (2.24)

Therefore we close the loop. As a result of (2.23), (2.24), (2.17a), we obtain

kξh k ≤ C kζh k + hk div ζh k + hkρh k ,

(2.25)
k div ξh k ≤ C kζh k + hk div ζh k + hkρh k .

Theorem 2.1 then follows from (2.24), (2.25), and the triangle inequality. ⊓
⊔

For sufficiently smooth (σ, u), Theorem 2.1 with (2.5) yields


ku − uh k . hr+1 |u|r+1 + h|σ|r+1 + h| div σ|r+1 , (2.26a)


|||σ − σh ||| . hr+1 |σ|r+1 + | div σ|r+1 + h|u|r+1 . (2.26b)

When Σh is based on the BDM element (1.6), the following improved error
estimate follows from (2.18c) and (2.6).


kσ − σh k . hr+2 |σ|r+2 + |u|r+1 . (2.27)

The well-posedness (1.10) follows from the same analysis in the proof of
Theorem 2.1.
Corollary 2.1 For sufficiently small h, the mixed method (1.10) has a unique
solution.

Proof To establish the existence and uniqueness of the solution to (1.10), it

suffices to show the uniqueness because of linearity. Suppose (σh , uh ) and
(σ̂h , ûh ) are both solutions to (1.10), then we have the error equation

(Aeσ , τh ) + (div τh , eu ) = 0, τh ∈ Σh ,
(2.28)
−(div eσ , vh ) + (b · Aeσ , vh ) + (ceu , vh ) = 0, vh ∈ Vh ,
where
eσ := σh − σ̂h , eu := uh − ûh .
Consider the dual problem
(Ae e ) − (b · Aτ , u
σ , τ ) + (div τ , u e ) = 0, τ ∈ Σ,
(2.29)
−(div σ u, v) = (eu , v), v ∈ V .
e , v) + (ce
10 X. Chen, Y. Li

It then follows from (2.28), (2.29) and the same analysis in Lemma 2.1 that

keu k . h|||eσ |||, (2.30)

provided h is sufficiently small. The same argument for proving Lemma 2.2
yields
k div eσ k . keσ k + keu k,
(2.31)
k div eσ k−1 . h|||eσ ||| + keu k.
Following the proof of (2.23) and using (2.30), (2.31), we obtain

keσ k . keu k, (2.32)

when h is small enough. Finally combining (2.30), (2.32), (2.31) yields

keu k . hkeu k.

Hence eu = 0 and eσ = 0 provided h is sufficiently small. ⊓

⊔
div ξh
Proof (Proof of Lemma 2.2) Let vh = k div ξh k . Using (2.4) and (2.7b), we
arrive at
k div ξh k = (div ξh , vh ) = (Ph div σ − div σh , vh ) = (div(σ − σh ), vh )
= (b · A(σ − σh ), vh ) + (c(Ph u − uh ), vh ) + ((c − ch )(u − Ph u), vh ).
(2.33)
where ch is the piecewise average of c w.r.t. Th . (2.17a) then follows from
(2.33), the Cauchy–Schwarz and triangle inequalities.
To prove (2.17b), let v ∈ [H01 (Ω)]d with kvk1 = 1. It follows from (2.4),
(1.8), and (2.7b) that

(div ξh , v) = (Ph div σ − div σ, v) + (div σ − div σh , v)

= (Ph div σ − div σ, v − Ph v) + (div σ − div σh , v − Ph v)
(2.34)
+ (b · A(σ − σh ), v) + (b · A(σ − σh ), Ph v − v)
+ (c(u − Ph u), Ph v) + (c(Ph u − uh ), Ph v).

Using (2.1) and (2.7a), we can rewrite (b · A(σ − σh ), v) as

(b · A(σ − σh ), v)
= (A(σ − σh ), vbT ) = (A(σ − σh ), A(vbT ))
= (A(σ − σh ), Πh A(vbT )) + (A(σ − σh ), (I − Πh )A(vbT )) (2.35)
T T
= −(div Πh A(vb ), u − uh ) + (A(σ − σh ), (I − Πh )A(vb ))
= −(div Πh A(vbT ), Ph u − uh ) + (A(σ − σh ), (I − Πh )A(vbT )).

On the other hand,

(c(u − Ph u), Ph v) = ((c − ch )(u − Ph u), Ph v). (2.36)

Superconvergent mixed methods for the Oseen equation 11

Combining (2.34) with (2.35) and (2.36) and using (2.5), we obtain

k div ξh k−1 = sup (div ξh , v)

kvk1 =1

(2.37)
. h kPh div σ − div σk + |||σ − σh ||| + ku − Ph uk
+ (1 + k div Πh A(vbT )k)keh k.

It follows from (2.4) and elementary calculation that

k div Πh A(vbT )k = kPh div A(vbT )k ≤ k div A(vbT )k ≤ Ckvk1 ≤ C.

Therefore using the previous estimate and (2.37), we obtain

k div ξk−1 . hk div Πh σ − div σk + h|||σ − σh ||| + hkρh k + keh k.

The proof is complete. ⊓

⊔

3 Superconvergence on pseudostress

Using Lemma 2.1 and (2.26), we obtain the supercloseness estimate on eh =

Ph u − uh .


keh k . hr+2 |σ|r+1 + | div σ|r+1 + |u|r+1 . (3.1)
In this section, we shall prove a supercloseness estimate on ξh = Πh σ − σh .
To that end, Σh is endowed with the Λ-inner product

Λ(τ1 , τ2 ) := (Aτ1 , τ2 ) + (div τ1 , div τ2 ), τ1 , τ2 ∈ Σh .

Let Zh := ker(div |Σh ) and Zh⊥ be the orthogonal complement of Zh w.r.t. Λ(·, ·)
in Σh . We shall decompose ξh ∈ Σh by the discrete Helmholtz decomposition

Σh = Zh ⊕Λ Zh⊥ , (3.2)

which can be analyzed using tools in finite element exterior calculus (FEEC),
see, e.g., [3,4]. In the theory of FEEC, an essential ingredient is the bounded
projections that commute with exterior differentiation. In particular, there
exist projections π eh : [L2 (Ω)]d → Σe h and Qh : L2 (Ω) → Vh (see, e.g., [3,25])
such that
eh |Σe h = id, Qh |Vh = id,
π
eh τ = Qh div τ for τ ∈ H(div, Ω),
div π (3.3)
2 d
ke
πh τ k . kτ k for τ ∈ [L (Ω)] ,
where id is the identity operator. Starting from π eh , Qh , one can easily obtain
commuting bounded projections onto Σh and Vh . Let
Z 
1
eh τ −
πh τ := π tr π
eh τ dx I,
d|Ω| Ω
12 X. Chen, Y. Li

where πeh is applied to each row of τ . Similarly, Qh can applied to each com-
ponent of a vector-valued function in [L2 (Ω)]d . Using the property of πeh in
(3.3), we have

πh |Σh = id, Qh |Vh = id, (3.4a)

div πh τ = Qh div τ for τ ∈ Σ, (3.4b)
kπh τ k . kτ k for τ ∈ [L2 (Ω)]d . (3.4c)

Then with the help of πh , Qh , we obtain the following lemma for estimating
the decomposition component living in Zh⊥ .

Lemma 3.1 It holds that

kτh kA . k div τh k−1 , τh ∈ Zh⊥ .

Proof Let φ ∈ [H01 (Ω)]d be the weak solution to

∆φ = div τh .

Then τ := ∇φ satisfies div τ = div τh . Using it and (3.4b), (3.4a), we obtain

div(τh − πh τ ) = div τh − Qh div τ

= div τh − Qh div τh = 0,

i.e., τh − πh τ ∈ Zh . Recall that k · k−1 denotes the H −1 (Ω)-norm. The natural

H 1 -elliptic regularity of ∆φ = div τh yields

kτ k ≤ kφk1 . k div τh k−1 . (3.5)

Hence using τh − πh τ ∈ Zh , τh ∈ Zh⊥ , (3.5), and (3.4c), we have

kτh k2A = Λ(τh − πh τ , τh ) + (Aπh τ , τh )

= (Aπh τ , τh ) ≤ kπh τ kA kτh kA
. kτ kkτh kA . k div τh k−1 kτh kA .

The proof is complete. ⊓

⊔

Remark 1 It follows from the discrete abstract Poincaré inequality (see [4],
Theorem 3.6) that

kτh kA . k div τh k for τh ∈ Zh⊥ .

Hence Lemma 3.1 is an improved discrete Poincaré inequality. The improve-

ment is achieved by utilizing the special property of the divergence operator.
Superconvergent mixed methods for the Oseen equation 13

We emphasize that the improved estimate of ξh is dependent on mesh

structure, type of finite elements, and quite technical. For simplicity of pre-
sentation, first let Σh × Vh be based on the lowest order RT element (1.5)
with r = 0 in R2 although the estimates can be generalized in several ways,
see the end of this section. For a scalar-valued function v and a vector-valued
function v = (v1 , v2 )T , let
curl v = (−∂x2 v, ∂x1 v)T , rot v = ∂x1 v2 − ∂x2 v1 .
The row-wise curl is defined as
 
(curl v1 )T
curl v = .
(curl v2 )T
It is noted that rot v = tr curl v. Let
Wh = {w ∈ H 1 (Ω) : w|K ∈ P1 (K) for K ∈ Th },
Z
Wh = {w ∈ [Wh ]2 : rot wdx = 0}.
Ω

Then we obtain the following discrete sequence in R2 .

curl div
Wh −−→ Σh −−→ Vh → 0. (3.6)
R R
Since rot wdx = Ω tr curl wh dx = 0 for all wh ∈ Wh , we have curl Wh ⊂
Σh .
The supercloseness estimate of ξh does not hold on general unstructured
grids. In [43,7], Li et al. derived several supercloseness estimates on the vec-
tor variable in mixed methods for elliptic equations using RT elements under
certain mildly structured grids. Readers are referred to [31,33,11,12] for simi-
lar results on Poisson’s equation under rectangular, h2 -uniform quadrilateral,
and uniform triangular grids. To avoid lengthy descriptions of different mesh
structures, we focus on the following piecewise uniform grids.

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

Fig. 1 Piecewise uniform grid on a square

14 X. Chen, Y. Li

Definition 3.1 We say Th is a uniform grid provided each pair of adjacent

triangles (two triangles sharing a common edge) form an exact parallelogram.
Let {Ωi }Ni=1 be a fixed polygonal partition of Ω. We say Th is a piecewise
uniform grid provided Th is aligned with {Ωi }N
i=1 and Th |Ωi is a uniform grid
for each 1 ≤ i ≤ N .
For instance, let Ω = [0, 1]2 be the unit square that is split into the 19 triangles
{Ωi }19
i=1 given in Figure 1 (left). After 3 consecutive uniform quad-refinement,
we obtain the triangulation in Figure 1 (right), which is a piecewise uniform
grid (w.r.t. {Ωi }19
i=1 ). The piecewise uniform mesh structure is only used in
the next technical lemma.

Lemma 3.2 Let D ∈ R2×2 be a constant matrix. Let Π e h be the canonical

e h , the lowest order RT element space. For a piecewise
interpolation onto Σ
2
uniform Th and v ∈ W∞ (Ω), wh ∈ Wh , it holds that

e h v), curl wh ) . h2 | log h| 21 kvkW 2 (Ω) k curl wh k.

(D(v − Π ∞

In [43], Li proved Lemma 3.2 on mildly structured grids when D = I. For

the general anisotropic D, we could not find a complete proof in literature. In
the appendix, we give a detailed proof of Lemma 3.2 using the technique in
Theorem 3.2 of [11] and Lemma 3.4 of [39]. Now we present a supercloseness
estimate for ξh , the second main result, in the next theorem.

Theorem 3.1 Let Ω be simply-connected and Th be a piecewise uniform grid.

For (1.10) using the lowest order RT element and sufficiently small h, it holds
that
1
kξh k . h2 | log h| 2 (kσkW∞
2 (Ω) + kuk2 ).

Proof When Ω is simply connected, (3.6) is an exact sequence, i.e., curl Wh =

Zh . It then follows from the exactness and the discrete Helmholtz decomposi-
tion (3.2) that
ξh = curl φh ⊕Λ ηh , (3.7)
where φh ∈ Wh , ηh ∈ Zh⊥ . Using Lemma 3.1 and div curl φh = 0, we obtain

kηh kA . k div ηh k−1 = k div ξh k−1 . (3.8)

It remains to estimate k curl φh kA . Note that

(A curl φh , ηh ) = Λ(curl φh , ηh ) = 0.

Then using the orthogonality, (2.7a), and (2.1), we have

k curl φh k2A = (A(Πh σ − σh ), curl φh )

= (Πh σ − σ, A curl φh ) = (Πe h σ − σ, A curl φh )

e h σ, curl φh ) + 1 (σ − Π
= −(σ − Π e h σ, (rot φh )I).
2
Superconvergent mixed methods for the Oseen equation 15

Recall that σi denotes the i-th row of σ. Let e1 = (1, 0), e2 = (0, 1), and
φh = (φh,1 , φh,2 )T . It follows from the previous equation and direct calculation
that
2
X
k curl φh k2A = e h σi , curl φh,i )
−(σi − Π
i=1
1 e h σi , (∂x1 φh,2 )ei ) − 1 (σi − Πe h σi , (∂x2 φh,1 )ei )
+ (σi − Π (3.9)
2 2
X 2
= (Dij (σi − Π e h σi )T , curl φh,j ),
i,j=1

where
 1     1  
−2 0 00 0 2 −1 0
D11 = , D12 = 1 , D21 = , D22 = .
0 −1 2 0 00 0 − 21

Then using (3.9), Lemma 3.2 and (2.2) with τ = curl φh , we have
1
k curl φh k2A . h2 | log h| 2 kσkW∞
2 (Ω) k curl φh k

1
(3.10)
. h2 | log h| 2 kσkW∞
2 (Ω) k curl φh kA .

Combining (2.2) with (3.8), (3.10), (2.17b), (3.1), we obtain

kξh k . kξh kA + k div ξh k−1

. k curl φh kA + kηh kA + k div ξh k−1
1

. h2 | log h| 2 kσkW∞
2 (Ω) + kuk2 .

The proof is complete. ⊓

⊔

The supercloseness estimate on kξh k can be generalized on rectangular

meshes. Throughout the rest of this section, let Th be a rectangular mesh,

Wh = {w ∈ H 1 (Ω) : w|K ∈ Qr+1 (K) for K ∈ Th },

Z
2
Wh = {w ∈ [Wh ] : rot wdx = 0},
Ω

and Σh × Vh be based on the rectangular RT element (1.7) with d = 2. When

Ω is simply-connected, we still have the discrete exact sequence (3.6). Similarly
to the case of triangular grids, we need the uniform mesh structure.
Definition 3.2 We say Th is a uniformly rectangular mesh if all the rectangles
of Th are of the same shape and size.
For wh ∈ Wh with (curl wh ) · n|∂Ω = 0, Theorem 5.1 of [33] implies

e h v), curl wh ) . hr+2 kvkr+2 k curl wh k,

(D(v − Π
16 X. Chen, Y. Li

where v ∈ [H r+2 (Ω)]2 . The previous estimate is not true when (curl wh )·n 6= 0
on ∂Ω. As claimed in the remark in Example 6.2 of [33], it holds that
e h v), curl wh ) ≤ Chr+1.5 k curl wh k,
(D(v − Π wh ∈ Wh (3.11)
for some absolute constant C independent of h. Following the same argument
in the proof of Theorem 3.1 and replacing Lemma 3.2 by (3.11), we actually
obtain the supercloseness estimate
kΠh σ − σh k = O(hr+1.5 ).
Similar estimates may hold on h2 -uniform quadrilateral grids described in [33],
which is beyond the scope of this paper.

4 Postprocessing

We have derived supercloseness estimates in Lemma 2.1 and Theorem 3.1.

However, those results can not be directly used because Ph u and Πh σ are not
known in practice. To extract superconvergence information from the small-
ness of eh and ξh , one may design easy-to-compute postprocessed solutions u∗h
and σh∗ . For instance, following the idea of [53], a element-by-element postpro-
cessing procedure for uh in the pseudostress-velocity formulation of the Stokes
equation is proposed in [23]. In particular, let
Pr+1 (K)/Pr (K) := {vh ∈ Pr+1 (K) : (vh , wh )K = 0 for all wh ∈ Pr (K)}.
where (·, ·)K is the L2 (K)-inner product. For each K ∈ Th , the postprocessed
solution u∗h |K ∈ Pr+1 (K) is the unique solution to
(u∗h , vh )K = (uh , vh )K ,
(4.1)
(∇u∗h , ∇vh )K = (σh , ∇vh )K + (ph , div vh )K , vh ∈ Pr+1 (K)/Pr (K),
where ph := − d1 tr σh ≈ p. Since the analysis of the mapping uh 7→ u∗h is
independent of equations, we can combine the analysis in Theorem 4.1 of [23]
for the Stokes equation with the supercloseness on kPh u − uh k in Theorem
2.1 to obtain the following postprocessing superconvergence estimate.
Theorem 4.1 For sufficiently small h, it holds that


ku − u∗h k . hr+2 kukr+2 + |σ|r+1 + | div σ|r+1 .
Proof Let û be the L2 -projection of u onto
{vh : vh |K ∈ [Pr+1 (K)]d for all K ∈ Th }.
It has been shown in the proof of Theorem 4.1 in [23] that
h−1 k(I − Ph )(û − u∗h )k . h−1 kPh (û − u∗h )k
! 12
X (4.2)
2
+ kσ − σh k + k∇(û − u)kL2 (K) .
K∈Th
Superconvergent mixed methods for the Oseen equation 17

Using the triangle inequality, Ph û = Ph u, Ph u∗h = uh , and (4.2), we obtain

ku − u∗h k ≤ ku − ûk + k(I − Ph )(û − u∗h )k + kPh (û − u∗h )k

. hr+2 |u|r+2 + kPh u − uh k + hkσ − σh k.

We finally conclude the proof from the previous estimate with (3.1) and
(2.26b). ⊓
⊔

Postprocessing technique on the scalar variable in mixed methods for Poisson’s

equation can be found in e.g., [15,10,53,46].
The postprocessing procedure σh 7→ σh∗ can be derived from existing post-
processing operator Reh . In particular, R
eh : Σ
e h → Y ⊂ [L2 (Ω)]2 is a linear
mapping onto the space Y of suitable piecewise polynomials and satisfies

Reh c = c, c ∈ R2 , (4.3a)
eh vk . kvk, v ∈ Σ
kR eh , (4.3b)
kτ − R ehτ k . h
eh Π 1+α
kτ kW∞
2 (Ω) , (4.3c)

e h is the lowest
for some positive constant α and sufficiently smooth τ . When Σ
e
order RT element space, Rh that satisfies (4.3) is given in e.g., [11,7]. The
simple nodal averaging [58] and superconvergent patch recovery [59] are also
possible choices.
For τ ∈ Σh , let
Z 
Rh τ := Reh τ − 1 eh τ I,
tr R
2|Ω| Ω

where Reh is applied to each row of τ . We have the following super-approximation

result of Rh .
2
Lemma 4.1 Assume (4.3) holds. For τ ∈ Σ ∩ [W∞ (Ω)]2×2 , we have

kτ − Rh Πh τ k . h1+α kτ kW∞
2 (Ω) .

e h τ . Then
Proof (4.3a) implies Rh I = O and thus Rh Πh τ = Rh Π

ehτ
τ − Rh Πh τ = τ − Rh Π
Z
(4.4)
=τ −R eh τ − 1
eh Π eh Π
tr(τ − R e h τ )dx,
2|Ω| Ω
R
where we use Ω tr τ dx = 0. Lemma 4.1 then follows from (4.4), the Cauchy–
Schwarz inequality, and (4.3c). ⊓
⊔

Combining Lemma 4.1 and Theorem 3.1, we obtain the following super-
convergence estimate by postprocessing.
18 X. Chen, Y. Li

Theorem 4.2 Let σh be the solution to (1.10) based on the lowest order RT el-
ement and σh∗ := Rh σh . Assume the assumptions in Theorem 3.1 and Lemma
4.1 hold. We have
1

kσ − σh∗ k . min(h2 | log h| 2 , h1+α ) kσkW∞
2 (Ω) + kuk2 .

Proof Using the triangle inequality, Theorem 3.1, and Lemma 4.1, we have

kσ − σh∗ k ≤ kσ − Rh Πh σk + kRh (Πh σ − σh )k

1

. min(h2 | log h| 2 , h1+α ) kσkW∞
2 (Ω) + kuk2 ,

which completes the proof. ⊓

⊔

Remark 2 Built upon Theorem 4.2, we can recover very accurate numerical
symmetric stress σhS∗ and pressure p∗h . In fact, let

1 ∗ 1
σhS∗ := (σ + σh∗T ), p∗h := − tr σh∗ .
2 h d

Recall σ S is the symmetric stress. We conclude that σh∗S superconverges to

σ S and p∗h superconverges to p in the L2 -norm from Theorem 4.2 and

1 1
kσ S − σh∗S k ≤ kσ − σh∗ k + kσ T − σh∗T k = kσ − σh∗ k,
2 2
1 1
kp − p∗h k ≤ k tr(σ − σh∗ )k ≤ √ kσ − σh∗ k.
d d

5 Experiments

In this section, the postprocessing procedure σh 7→ σh∗ in Theorem 4.2 is based

on the polynomial preserving recovery R eh for the lowest order RT element de-
scribed in [7]. In the first experiment, we verify our a priori and superconver-
gence error estimates. We also test the adaptivity performance of the recovery
a posteriori error estimator in the other two numerical experiments.

Table 1 Convergence of the lowest order RT element

nt ku − uh k keh k ku − u∗h k kσ − σh k kξh k kσ − σh∗k

19 3.210e-1 3.851e-2 9.334e-2 1.372 2.505e-1 2.113

76 1.620e-1 1.027e-2 2.438e-2 6.849e-1 7.172e-2 6.108e-1
304 8.121e-2 2.621e-3 6.176e-3 3.418e-1 1.957e-2 1.724e-1
1216 4.063e-2 6.598e-4 1.550e-3 1.707e-1 5.242e-3 4.353e-2
4864 2.032e-2 1.653e-4 3.880e-4 8.530e-2 1.390e-3 1.088e-2
19456 1.016e-2 4.136e-5 9.703e-5 4.270e-2 3.659e-4 2.719e-3
order 9.990e-1 1.990 1.994 1.001 1.904 1.961
Superconvergent mixed methods for the Oseen equation 19

Table 2 Convergence of the lowest order BDM element

nt ku − uh k keh k ku − u∗h k kσ − σh k kξh k
19 3.192e-1 1.779e-2 6.316e-02 4.480e-1 4.755e-1
76 1.618e-1 6.026e-3 1.645e-02 1.249e-1 1.180e-1
304 8.118e-2 1.636e-3 4.155e-03 3.216e-2 2.928e-2
1216 4.062e-2 4.181e-4 1.042e-03 8.104e-3 7.288e-3
4864 2.032e-2 1.051e-4 2.606e-04 2.031e-3 1.818e-3
19456 1.016e-2 2.631e-5 6.515e-05 5.081e-4 4.541e-4
order 9.986e-1 1.964 1.996 1.987 2.005

Problem 1: Consider the Oseen equation (1.2) on the unit square Ω = [0, 1]2
with the smooth solutions
 
sin(π(x1 + x2 ))
u= ,
− sin(π(x1 + x2 ))
p = x1 + x2 − 1.

Set c = 0, b = (cos(x2 ), sin(x1 ))T , g = u|∂Ω = 0. f is computed from u and b.

We start with the initial partition in Figure 1 (left). A sequence of piecewise
uniform meshes is obtained by uniform quad-refinement, i.e., dividing each
triangle into four similar subtriangles by connecting the midpoints of each
edge. Numerical results are presented in Tables 1 and 2, where nt is short for
“number of triangles”. The order of convergence is computed from the error
quantities in those tables by least squares without using the data in the first
rows.
The numerical rates of convergence coincide with a priori error estimates
(2.26), (2.27) and the supercloseness estimates in (3.1) and Theorem 3.1. It
is noted that for the lowest order BDM element, kξh k ≈ O(h2 ), which is pre-
dicted by a priori error estimates and thus not supersmall. Since the recovery
procedure in [7] provides the super-approximation rate α = 1 in (4.3c) and
Lemma 4.1, the recovery superconvergence estimate for the lowest order RT
1
element predicted by Theorem 4.2 is kσ − σh∗ k = O(h2 | log h| 2 ), numerically
confirmed by the last column in Table 1.
Problem 2: Consider (1.2) on the L-shaped domain Ω = [−1, 1]2 \([0, 1] ×
[−1, 0]) with the smooth pressure p = x1 + x2 and singular velocity
 α 
r sin(αθ)
u= α ,
r cos(αθ)

where α = 32 , (r, θ) is the polar coordinate w.r.t. (0, 0). Set c = 0, b = (1, 2)T
and g = u|∂Ω . Direct calculation shows that div u = 0 and f = (1, 1)T +b·∇u.
In this experiment, we use the classical adaptive feedback loop (cf. [6,29,47])

SOLVE → ESTIMATE → MARK → REFINE

to obtain a sequence of adaptively refined grids {Thℓ }ℓ≥0 and numerical solu-
tions ({σhℓ , uhℓ )}ℓ≥0 . In particular, the algorithm starts from the initial grid
20 X. Chen, Y. Li

1 1

0.5 0.5

0 0

-0.5 -0.5

-1 -1
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1

Fig. 2 Adaptive mesh for the singular problem

100

10-1
error

10-2
||u-u h ||=O(h1.035 )
|| -h
||=O(h1.085 )
10-3 *
||u-u h ||=O(h2.098 )
|| - *h
||=O(h1.775 )

10-4
102 nt 103 104

Fig. 3 Convergence history of the adaptive method

Th0 presented in Figure 2(left). The module ESTIMATE computes the supercon-
vergent recovery-based error indicator

12
Eℓ (K) := kσh∗ℓ − σhℓ k2L2 (K) + ku∗hℓ − uhℓ k2L2 (K)

for each triangle K ∈ Thℓ . The module MARK then selects a collection of trian-
gles
Mℓ := {K ∈ Tℓ : Eℓ (K) ≥ 0.7 max Eℓ (K ′ )} (5.1)
′ K ∈Tℓ

to be refined by local quad-refinement. To remove the newly created hanging

nodes, minimal number of neighboring elements of Mℓ are bisected and the
next level triangulation Thℓ+1 is generated. See Figure 2(right) for an adap-
tively refined triangulation.
It can be observed from Figure 3 that (σhℓ , uhℓ ) optimally converges to
(σ, u). In addtion, the errors kσ − σh∗ℓ k and ku − u∗hℓ k are apparently super-
Superconvergent mixed methods for the Oseen equation 21

convergent to 0. Let
X
1
1
Eℓ := Eℓ (K)2 2 = kσh∗ℓ − σhℓ k2 + ku∗hℓ − uhℓ k2 2 ,
K∈Thℓ

21
Eℓ := kσ − σhℓ k2 + ku − uhℓ k2 .

Due to the observed superconvergence phenomena and the triangle inequality

21
|Eℓ − Eℓ | ≤ kσ − σh∗ℓ k2 + ku − u∗hℓ k2 ,

the recovery-based error estimator Eℓ is asymptotically exact, i.e.,

Eℓ
lim = 1.
ℓ→∞ Eℓ

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

Fig. 4 (left) Initial grid. (right) Adaptive grid, 9683 triangles.

Fig. 5 Velocity field of the convection-dominated problem

22 X. Chen, Y. Li

Problem 3: Consider the Oseen equation (1.2) on the unit square Ω = [0, 1]2
with c = 0, b = (500, 1)T , g = 0, f = 5000(x2 , −x1 )T . In the end, we use
the same adaptive algorithm in Problem 2 to solve this convection-dominated
example. The initial grid is given in Figure 4 (left). The marked set in (5.1) is
replaced by
Mℓ = {K ∈ Tℓ : Eℓ (K) ≥ 0.3 max Eℓ (K ′ )}.
′ K ∈Tℓ

Due to the convection coefficient b = (500, 1)T , the exact solution u near
the axis x1 = 1 is rapidly changing to preserve the homogeneous Dirichlet
boundary condition. It can be observed from Figures 4 and 5 that the adaptive
mixed method is able to capture the boundary layer without using a uniformly
fine grid.

Appendix

f2
f1

NE NE
3 4
NE P P N
E P2
P1 1 2
P E,1 E P E,2 E1 P E2

Fig. 6 A uniform grid of Ωi .

Proof (Proof of Lemma 3.2) Let {Ωi }N i=1 be a polygonal partition of Ω and
Th |Ωi be a uniform grid for each i. Let f1 , f2 be fixed unit normals to two
f1T
edges of an arbitrary but fixed triangle in Th |Ωi . Let S = and recall
f2T
e1 = (1, 0), e2 = (0, 1). We have
Z Z
e
[D(v − Πh v)] · curl wh dx = S −T [D(v − Π e h v)] · S curl wh dx
Ωi Ωi
2 Z
X 2
X (5.2)
= [ej S −T e h v)](fjT curl wh )dx :=
D(v − Π Ij .
j=1 Ωi j=1

For j = 1, 2, let Ejo and Ej∂ be the set of interior (inside Ωi ) and boundary
(on ∂Ωi ) edges orthogonal to fj , respectively. Let NE denote the region which
Superconvergent mixed methods for the Oseen equation 23

is the union of triangles sharing E as an edge. To estimate I1 , let Ωi be

partitioned into parallelograms NE with E ∈ E1o and boundary triangles NE
with E ∈ E1∂ , see Figure 6, where E1 , E2 ∈ E1∂ , E3 , E4 ∈ E1o . For any E ∈ E1o ,
note that f1T curl wh is single-valued across E and thus constant on NE . For
E ∈ E1∂ , let PE,1 and PE,2 be the two endpoints of E and hE denote the length
of E. Then
X Z
I1 = [e1 S −T D(v − Πe h v)](f1T curl wh )dx
NE
E∈E1o ∪E1∂
X Z
= (f1T curl wh )e1 S −T D e h v)dx
(v − Π
E∈E1o NE (5.3)
X Z


+ h−1 −T e h v)dx wh (PE,2 ) − wh (PE,1 )
(v − Π
E e1 S D
NE
E∈E1∂

:= I11 + I12 .
When NE is an exact parallelogram, it is shown in Equation (3.15) of [11] that
Z
(v − Πe h v)dx . h3 |v|H 2 (N ) . (5.4)
E
NE

Using (5.4) and the Cauchy-Schwarz inequality, we have

X Z
I11 . −1
h k curl wh kL2 (NE ) e h v)dx
(v − Π
E∈E1o NE
X (5.5)
. h2 k curl wh kL2 (NE ) |v|H 2 (NE ) . h2 k curl wh k|v|2 .
E∈E1o

Let N denote the collection of endpoints of edges in E1∂ ,

N1 := {P ∈ N : P is not a corner of ∂Ωi },
and N2 := N \N1 . For instance, N2 = {P1 , P2 } in Figure 6. For P ∈ N1 , let
E1 , E2 be the two boundary edges sharing P and ωP be the union of three
triangles having P as a vertex. For P ∈ N2 , let E denote the unique edge in
E1∂ having P as a vertex. Rearranging the summation in I12 , we have
Z Z !
X
−1 −T e e
I12 = h e1 S D
E1 (v − Πh v)dx − (v − Πh v)dx wh (P )
P ∈N1 N E1 N E2
X Z
+ h−1
E e1 S
−T
D e h v)dxwh (P ).
(v − Π
P ∈N2 NE
(5.6)
For P ∈ N1 , Equation (3.15) in [39] shows that
Z Z
e
(v − Πh v)dx − e h v)dx . h3 |v|H 2 (ω ) .
(v − Π (5.7)
P
N E1 N E2
24 X. Chen, Y. Li

R
Without loss of generality, let Ω
wh dx = 0. Then the discrete Sobolev and
Poincaré inequalities yield
1 1
kwh kL∞ (Ω) . | log h| 2 kwh k1 . | log h| 2 k curl wh k. (5.8)
Standard interpolation error estimate yields
Z
(v − Π e h v)dx . h2 |v|H 1 (N ) . (5.9)
E
NE

It then follows from (5.6)–(5.9), and the Cauchy–Schwarz inequality that

X X

I12 . h2 |v|H 2 (ωP ) + h|v|H 1 (NE ) kwh kL∞ (Ω)
P ∈N1 P ∈N2
X 1
X 1
1
2
. h |v|W∞
2 (Ω) |ωP | 2 + h|v|W∞
1 (Ω) |NE | 2 | log h| 2 k curl wh k
P ∈N1 P ∈N2
1
2
. h kvkW∞
2 (Ω) | log h| k curl wh k.
2

(5.10)
We finally conclude the proof from (5.2), (5.5), (5.10), the same analysis for
I2 and all pieces in the partition {Ωi }N
i=1 . ⊓
⊔

References

1. Antonietti, P.F., Beirão da Veiga, L., Mora, D., Verani, M.: A stream virtual element
formulation of the Stokes problem on polygonal meshes. SIAM J. Numer. Anal. 52(1),
386–404 (2014). DOI 10.1137/13091141X. URL https://doi.org/10.1137/13091141X
2. Arnold, D.N., Douglas Jr., J., Gupta, C.P.: A family of higher order mixed finite
element methods for plane elasticity. Numer. Math. 45(1), 1–22 (1984). DOI
10.1007/BF01379659. URL https://doi.org/10.1007/BF01379659
3. Arnold, D.N., Falk, R.S., Winther, R.: Finite element exterior calculus, homologi-
cal techniques, and applications. Acta Numer. 15, 1–155 (2006). DOI 10.1017/
S0962492906210018. URL https://doi.org/10.1017/S0962492906210018
4. Arnold, D.N., Falk, R.S., Winther, R.: Finite element exterior calcu-
lus: from Hodge theory to numerical stability. Bull. Amer. Math. Soc.
(N.S.) 47(2), 281–354 (2010). DOI 10.1090/S0273-0979-10-01278-4. URL
https://doi.org/10.1090/S0273-0979-10-01278-4
5. Arnold, D.N., Li, L.: Finite element exterior calculus with lower-order terms.
Math. Comp. 86(307), 2193–2212 (2017). DOI 10.1090/mcom/3158. URL
https://doi.org/10.1090/mcom/3158
6. Babuška, I., Rheinboldt, W.C.: Error estimates for adaptive finite element computa-
tions. SIAM J. Numer. Anal. 15(4), 736–754 (1978). DOI 10.1137/0715049. URL
https://doi.org/10.1137/0715049
7. Bank, R.E., Li, Y.: Superconvergent recovery of Raviart–Thomas mixed finite elements
on triangular grids. J. Sci. Comput. 81, 1882–1905 (2019)
8. Bank, R.E., Xu, J.: Asymptotically exact a posteriori error estimators. I. Grids with
superconvergence. SIAM J. Numer. Anal. 41(6), 2294–2312 (2003). DOI 10.1137/
S003614290139874X. URL https://doi.org/10.1137/S003614290139874X
9. Boffi, D., Brezzi, F., Fortin, M.: Mixed finite element methods and applications, Springer
Series in Computational Mathematics, vol. 44. Springer, Heidelberg (2013). DOI
10.1007/978-3-642-36519-5
10. Bramble, J.H., Xu, J.: A local post-processing technique for improving the accuracy in
mixed finite-element approximations. SIAM J. Numer. Anal. 26(6), 1267–1275 (1989).
DOI 10.1137/0726073
Superconvergent mixed methods for the Oseen equation 25

11. Brandts, J.H.: Superconvergence and a posteriori error estimation for triangular mixed
finite elements. Numer. Math. 68(3), 311–324 (1994). DOI 10.1007/s002110050064.
URL https://doi.org/10.1007/s002110050064
12. Brandts, J.H.: Superconvergence for triangular order k = 1 Raviart-Thomas mixed
finite elements and for triangular standard quadratic finite element methods. Appl.
Numer. Math. 34(1), 39–58 (2000). DOI 10.1016/S0168-9274(99)00034-3. URL
https://doi.org/10.1016/S0168-9274(99)00034-3
13. Brezzi, F., Douglas Jr., J., Durán, R., Fortin, M.: Mixed finite elements for second
order elliptic problems in three variables. Numer. Math. 51(2), 237–250 (1987). DOI
10.1007/BF01396752. URL https://doi.org/10.1007/BF01396752
14. Brezzi, F., Douglas Jr., J., Fortin, M., Marini, L.D.: Efficient rectangular mixed
finite elements in two and three space variables. RAIRO Modél. Math.
Anal. Numér. 21(4), 581–604 (1987). DOI 10.1051/m2an/1987210405811. URL
https://doi.org/10.1051/m2an/1987210405811
15. Brezzi, F., Douglas Jr., J., Marini, L.D.: Two families of mixed finite elements for second
order elliptic problems. Numer. Math. 2(47), 217–235 (1985)
16. Cáceres, E., Gatica, G.N.: A mixed virtual element method for the pseudostress-velocity
formulation of the Stokes problem. IMA J. Numer. Anal. 37(1), 296–331 (2017). DOI
10.1093/imanum/drw002. URL https://doi.org/10.1093/imanum/drw002
17. Cáceres, E., Gatica, G.N., Sequeira, F.A.: A mixed virtual element method for the
Brinkman problem. Math. Models Methods Appl. Sci. 27(4), 707–743 (2017). DOI
10.1142/S0218202517500142. URL https://doi.org/10.1142/S0218202517500142
18. Cai, Z., Tong, C., Vassilevski, P.S., Wang, C.: Mixed finite element methods
for incompressible flow: stationary Stokes equations. Numer. Methods Partial
Differential Equations 26(4), 957–978 (2010). DOI 10.1002/num.20467. URL
https://doi.org/10.1002/num.20467
19. Cai, Z., Wang, C., Zhang, S.: Mixed finite element methods for incompressible flow:
stationary Navier-Stokes equations. SIAM J. Numer. Anal. 48(1), 79–94 (2010). DOI
10.1137/080718413. URL https://doi.org/10.1137/080718413
20. Cai, Z., Wang, Y.: A multigrid method for the pseudostress formulation of Stokes prob-
lems. SIAM J. Sci. Comput. 29(5), 2078–2095 (2007). DOI 10.1137/060661429. URL
https://doi.org/10.1137/060661429
21. Carrero, J., Cockburn, B., Schötzau, D.: Hybridized globally divergence-free LDG meth-
ods. I. The Stokes problem. Math. Comp. 75(254), 533–563 (2006). DOI 10.1090/
S0025-5718-05-01804-1. URL https://doi.org/10.1090/S0025-5718-05-01804-1
22. Carstensen, C., Gallistl, D., Schedensack, M.: Quasi-optimal adaptive pseudostress ap-
proximation of the Stokes equations. SIAM J. Numer. Anal. 51(3), 1715–1734 (2013)
23. Carstensen, C., Kim, D., Park, E.J.: A priori and a posteriori pseudostress-velocity
mixed finite element error analysis for the Stokes problem. SIAM J. Numer. Anal.
49(6), 2501–2523 (2011)
24. Cesmelioglu, A., Cockburn, B., Nguyen, N.C., Peraire, J.: Analysis of HDG meth-
ods for Oseen equations. J. Sci. Comput. 55(2), 392–431 (2013). DOI 10.1007/
s10915-012-9639-y. URL https://doi.org/10.1007/s10915-012-9639-y
25. Christiansen, S.H., Winther, R.: Smoothed projections in finite element exterior calcu-
lus. Math. Comp. 77(262), 813–829 (2008)
26. Cockburn, B., Sayas, F.J.: Divergence-conforming HDG methods for Stokes flows.
Math. Comp. 83(288), 1571–1598 (2014). DOI 10.1090/S0025-5718-2014-02802-0. URL
https://doi.org/10.1090/S0025-5718-2014-02802-0
27. Cui, M., Ye, X.: Superconvergence of finite volume methods for the Stokes equations.
Numer. Methods Partial Differential Equations 25(5), 1212–1230 (2009). DOI 10.1002/
num.20399. URL https://doi.org/10.1002/num.20399
28. Demlow, A.: Suboptimal and optimal convergence in mixed finite element methods.
SIAM J. Numer. Anal. 39(6), 1938–1953 (2002). DOI 10.1137/S0036142900376900
29. Dörfler, W.: A convergent adaptive algorithm for Poisson’s equation. SIAM
J. Numer. Anal. 33(3), 1106–1124 (1996). DOI 10.1137/0733054. URL
https://doi.org/10.1137/0733054
30. Douglas Jr., J., Roberts, J.E.: Global estimates for mixed methods for second order
elliptic equations. Math. Comp. 44(169), 39–52 (1985). DOI 10.2307/2007791. URL
https://doi.org/10.2307/2007791
26 X. Chen, Y. Li

31. Durán, R.: Superconvergence for rectangular mixed finite elements. Nu-
mer. Math. 58(3), 287–298 (1990). DOI 10.1007/BF01385626. URL
https://doi.org/10.1007/BF01385626
32. Eichel, H., Tobiska, L., Xie, H.: Supercloseness and superconvergence of sta-
bilized low-order finite element discretizations of the Stokes problem. Math.
Comp. 80(274), 697–722 (2011). DOI 10.1090/S0025-5718-2010-02404-4. URL
https://doi.org/10.1090/S0025-5718-2010-02404-4
33. Ewing, R.E., Liu, M.M., Wang, J.: Superconvergence of mixed finite element approxi-
mations over quadrilaterals. SIAM J. Numer. Anal. 36(3), 772–787 (1999)
34. Falk, R.S., Neilan, M.: Stokes complexes and the construction of stable finite elements
with pointwise mass conservation. SIAM J. Numer. Anal. 51(2), 1308–1326 (2013).
DOI 10.1137/120888132. URL https://doi.org/10.1137/120888132
35. Fu, G., Jin, Y., Qiu, W.: Parameter-free superconvergent H(div)-conforming HDG
methods for the Brinkman equations. IMA J. Numer. Anal. 39(2), 957–982 (2019).
DOI 10.1093/imanum/dry001. URL https://doi.org/10.1093/imanum/dry001
36. Gatica, G.N., Gatica, L.F., Márquez, A.: Analysis of a pseudostress-based mixed
finite element method for the Brinkman model of porous media flow. Nu-
mer. Math. 126(4), 635–677 (2014). DOI 10.1007/s00211-013-0577-x. URL
https://doi.org/10.1007/s00211-013-0577-x
37. Girault, V., Raviart, P.A.: Finite element methods for Navier-Stokes equations, Springer
Series in Computational Mathematics, vol. 5. Springer-Verlag, Berlin (1986). DOI
10.1007/978-3-642-61623-5. Theory and algorithms
38. Guzmán, J., Neilan, M.: Conforming and divergence-free Stokes elements on gen-
eral triangular meshes. Math. Comp. 83(285), 15–36 (2014). DOI 10.1090/
S0025-5718-2013-02753-6. URL https://doi.org/10.1090/S0025-5718-2013-02753-6
39. Hu, J., Ma, L., Ma, R.: Optimal Superconvergence Analysis for the Crouzeix-Raviart
and the Morley elements. arXiv e-prints arXiv:1808.09810 (2018)
40. Hu, J., Yu, G.: A unified analysis of quasi-optimal convergence for adaptive mixed finite
element methods. SIAM J. Numer. Anal. 56(1), 296–316 (2018)
41. Kundu, P.K., Dowling, D.R., Tryggvason, G., Cohen, I.M.: Fluid mechanics, 6th edn.
Academic Press (2015)
42. Li, Y.: Quasi-optimal adaptive mixed finite element methods for controlling natural
norm errors. arXiv e-prints arXiv:1907.03852 (2019)
43. Li, Y.W.: Global superconvergence of the lowest-order mixed finite element on mildly
structured meshes. SIAM J. Numer. Anal. 56(2), 792–815 (2018). DOI 10.1137/
17M112587X. URL https://doi.org/10.1137/17M112587X
44. Liu, H., Yan, N.: Superconvergence analysis of the nonconforming quadrilateral linear-
constant scheme for Stokes equations. Adv. Comput. Math. 29(4), 375–392 (2008).
DOI 10.1007/s10444-007-9054-3. URL https://doi.org/10.1007/s10444-007-9054-3
45. Liu, X., Li, J., Chen, Z.: A weak Galerkin finite element method for the Oseen equations.
Adv. Comput. Math. 42(6), 1473–1490 (2016). DOI 10.1007/s10444-016-9471-2. URL
https://doi.org/10.1007/s10444-016-9471-2
46. Lovadina, C., Stenberg, R.: Energy norm a posteriori error estimates for mixed fi-
nite element methods. Math. Comp. 75(256), 1659–1674 (2006). DOI 10.1090/
S0025-5718-06-01872-2
47. Morin, P., Nochetto, R.H., Siebert, K.G.: Data oscillation and convergence of adaptive
FEM. SIAM J. Numer. Anal. 38(2), 466–488 (2000)
48. Mu, L., Wang, J., Ye, X.: A stable numerical algorithm for the Brinkman equations by
weak Galerkin finite element methods. J. Comput. Phys. 273, 327–342 (2014). DOI
10.1016/j.jcp.2014.04.017. URL https://doi.org/10.1016/j.jcp.2014.04.017
49. Pan, J.: Global superconvergence for the bilinear-constant scheme for the Stokes prob-
lem. SIAM J. Numer. Anal. 34(6), 2424–2430 (1997). DOI 10.1137/S0036142995286167.
URL https://doi.org/10.1137/S0036142995286167
50. Park, E.J., Seo, B.: An upstream pseudostress-velocity mixed formulation for the Oseen
equations. Bull. Korean Math. Soc. 51(1), 267–285 (2014). DOI 10.4134/BKMS.2014.
51.1.267
51. Qian, Y., Wu, S., Wang, F.: A mixed discontinuous Galerkin method with symmetric
stress for Brinkman problem based on the velocity-pseudostress formulation. arXiv
e-prints arXiv:1907.01246 (2019)
 Superconvergent mixed methods for the Oseen equation 27

52. Raviart, P.A., Thomas, J.M.: A mixed finite element method for 2nd order elliptic
problems. In: Mathematical aspects of finite element methods (Proc. Conf., Consiglio
Naz. delle Ricerche (C.N.R.), Rome, 1975), pp. 292–315. Lecture Notes in Math., Vol.
606 (1977)
53. Stenberg, R.: Postprocessing schemes for some mixed finite elements. RAIRO Modél.
Math. Anal. Numér. 25(1), 151–167 (1991). DOI 10.1051/m2an/1991250101511. URL
https://doi.org/10.1051/m2an/1991250101511
54. Wang, J., Ye, X.: Superconvergence of finite element approximations for the Stokes
problem by projection methods. SIAM J. Numer. Anal. 39(3), 1001–1013 (2001). DOI
10.1137/S003614290037589X. URL https://doi.org/10.1137/S003614290037589X
55. Wang, J., Ye, X.: A weak Galerkin finite element method for the stokes equations.
Adv. Comput. Math. 42(1), 155–174 (2016). DOI 10.1007/s10444-015-9415-2. URL
https://doi.org/10.1007/s10444-015-9415-2
56. Ye, X.: Superconvergence of nonconforming finite element method for the Stokes equa-
tions. Numer. Methods Partial Differential Equations 18(2), 143–154 (2002). DOI
10.1002/num.1036.abs. URL https://doi.org/10.1002/num.1036.abs
57. Zhang, Z., Naga, A.: A new finite element gradient recovery method: superconver-
gence property. SIAM J. Sci. Comput. 26(4), 1192–1213 (2005). DOI 10.1137/
S1064827503402837. URL https://doi.org/10.1137/S1064827503402837
58. Zienkiewicz, O.C., Zhu, J.Z.: A simple error estimator and adaptive procedure for prac-
tical engineering analysis. Internat. J. Numer. Methods Engrg. 24(2), 337–357 (1987).
DOI 10.1002/nme.1620240206. URL https://doi.org/10.1002/nme.1620240206
59. Zienkiewicz, O.C., Zhu, J.Z.: The superconvergent patch recovery and a pos-
teriori error estimates. I. The recovery technique. Internat. J. Numer. Meth-
ods Engrg. 33(7), 1331–1364 (1992). DOI 10.1002/nme.1620330702. URL
https://doi.org/10.1002/nme.1620330702

View publication stats