Applied Numerical Mathematics 137 (2019) 19–33

Contents lists available at ScienceDirect

Applied Numerical Mathematics

www.elsevier.com/locate/apnum

Error analysis of weak Galerkin finite element methods for

time-dependent convection–diffusion equations ✩
Shenglan Xie a,b , Peng Zhu c,∗ , Xiaoshen Wang d
a
Department of Mathematics, Shanghai Normal University, Shanghai, China
b
Nanhu College, Jiaxing University, Jiaxing, Zhejiang 314001, China
c
College of Mathematics, Physics and Information Engineering, Jiaxing University, Jiaxing, Zhejiang 314001, China
d
Department of Mathematics and Statistics, University of Arkansas at Little Rock, Little Rock, AR 72204, USA

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

Article history: In this paper, a fully discrete weak Galerkin (WG) finite element method is presented for
Received 29 August 2018 solving time-dependent convection-dominated diffusion equations with convection term
Received in revised form 4 December 2018 in non-conservation form. WG method and Crank–Nicolson scheme are used for spatial
Accepted 4 December 2018
and temporal discretization, respectively. Stability and error analysis are established on
Available online 6 December 2018
polygonal mesh. Numerical examples are presented to verify our theoretical findings and
Keywords: show the efficiency of the method.
Time-dependent convection–diffusion © 2018 IMACS. Published by Elsevier B.V. All rights reserved.
equation
Weak Galerkin finite element method
Crank–Nicolson scheme
Error estimate

1. Introduction

Let  ⊂ R2 be a bounded polygonal domain with boundary ∂ , T > 0 be the final time. In this work, we will present
and analyze a weak Galerkin finite element method for the following time-dependent convection–diffusion problem
⎧
⎨ ∂t u − ε u + β · ∇ u + γ u = f in Q T :=  × (0, T ],
u=0 on  T := ∂  × (0, T ], (1.1)
⎩
u (x, 0) = u 0 (x) in ,
where ε > 0 is a small parameter. For well-posedness of the differential equation, we assume that β , γ , and f are suffi-
ciently smooth, β ∈ [ W 1,∞ ()]2 , and γ − 12 ∇ · β ≥ γ0 > 0 for some constant γ0 (cf., e.g., [23]).
Convection-dominated diffusion equations appear in several branches of applied physics, such as fluid flows, semiconduc-
tor device simulation and so on. The analysis and numerical solution of convection-dominated diffusion equations attracts
substantial attention. As is well known, the standard Galerkin discretization leads to unsatisfied numerical solutions with
non-physical oscillations. To avoid these spurious oscillations, several stabilization methods have been presented over the
years, including SUPG methods [4,12,13,29], continuous interior penalty (CIP) method [5,24], local projection stabilization
(LPS) method [1,15,23], and discontinuous Galerkin methods [10,11,21,22]. Several alternatives were compared in [2,13].
Although a large amount of work has been devoted to the numerical solution of convection-dominated equations, there

✩
The research of Zhu was supported in part by Zhejiang Provincial Natural Science Foundation of China (Grant No. LY19A010008).
* Corresponding author.
E-mail addresses: [email protected] (S. Xie), [email protected] (P. Zhu), [email protected] (X. Wang).

https://doi.org/10.1016/j.apnum.2018.12.005
0168-9274/© 2018 IMACS. Published by Elsevier B.V. All rights reserved.
20 S. Xie et al. / Applied Numerical Mathematics 137 (2019) 19–33

is still no method could be considered to be completely satisfactory. For example, the popular SUPG method suffers from
producing overshoots and undershoots near the layer region.
Recently, a novel finite element framework, the so-called Weak Galerkin (WG) method, was first developed by Wang
and Ye [26,27] for numerical solution of second order elliptic equations. With the new concepts of weak function and weak
gradient, the WG methods introduce degrees of freedom on each element as well as each edge in the partition, and allow
the use of totally discontinuous functions as basis functions. In addition, the WG method has a simple and parameter-free
formulation and the flexibility of using general polytopal meshes. The WG method has been developed for different types of
partial differential equations, such as second order elliptic problems, biharmonic problems, Stokes problems, and interface
problems, etc.; see, e.g., [16–20,26–28]. Most recently, two different robust WG methods were presented and analyzed for
steady convection–diffusion–reaction equations in [6] and [14], respectively. In both papers, the convection terms are in
conservation form, i.e., ∇ · (β u ).
The goal of this work is to develop a robust WG method for unsteady convection-dominated diffusion equations with
convection term in non-conservation form β · ∇ u. This kind of equations often arise in the linearization process of Navier–
Stokes equation. In this paper, a generalized weak directional derivative β · ∇ w u related to β · ∇ u is introduced and an
upwind-type stabilization term is introduced. A fully discrete scheme is obtained by Crank–Nicolson scheme for time deriva-
tive and WG technique for space discretization.
The rest of this article is organized as follows. In Section 2, semi-discrete WG FEM is proposed. Stability and error
analysis of the semi-discrete WG scheme are established. Crank–Nicolson time discrete WG scheme and its stability and
convergence results are found in Section 3. Numerical experiments are presented in Section 4 to support the theoretical
results. In the end, some concluding remarks are given in Section 5.
In the sequel, with C we shall denote a generic positive constant independent of ε , time step size τ and mesh step
size h. And it may not necessarily be the same at each occurrence.

2. Semi-discrete weak Galerkin finite element method

Let Th be a partition of the domain  consisting of elements which are closed and simply connected polygons in R2
satisfying a set of shape regularity assumptions [27]. Let Eh be the set of all edges in Th , and Eh0 = Eh \∂  be the set of all
interior edges. Denote by h K the diameter for every element K ∈ Th and h = max K ∈Th h K the mesh size for Th .
For a given integer k ≥ 1, let V h be the weak Galerkin finite element space associated with Th defined as follows:

V h = { v = { v 0 , v b } : v 0 | K ∈ Pk ( K ), v b |e ∈ Pk (e ), e ⊂ ∂ K , K ∈ Th }, (2.1)

and

V h0 = { v : v ∈ V h , v b = 0 on ∂ }, (2.2)

where Pk is the space of polynomials of total degree k or less. We would like to emphasize that any function v = { v 0 , v b } ∈
V h has a single value v b on each edge e ∈ Eh , which is not necessary to be the trace of v 0 on e. The discontinuity between
elements allows us to apply Pk spaces on polytopal elements, which brings convenience in many aspects in implementation.
For any v = { v 0 , v b }, the weak gradient ∇ w v ∈ [Pk−1 ( K )]2 is defined on K as the unique polynomial satisfying

(∇ w v , τ ) K = −( v 0 , ∇ · τ ) K +  v b , τ · n ∂K ∀τ ∈ [Pk−1 ( K )]2 , (2.3)

where n is the unit outward normal vector to ∂ K , and (·, ·) K and ·, · ∂ K are standard L inner products on K and ∂ K ,
2

respectively.
For any v = { v 0 , v b }, we define the weak directional derivative β · ∇ w v ∈ Pk ( K ) related to β · ∇ v on K as the unique
polynomial satisfying

(β · ∇ w v , w ) K = −( v 0 , ∇ · (β w )) K +  v b , β · nw ∂K ∀ w ∈ Pk ( K ). (2.4)

We introduce four global projections Q 0 , Q b , Q h and Qh . They are element-wise defined L projections detailed as 2

follows. For each element K ∈ Th , Q 0 : L 2 ( K ) → Pk ( K ) and Q b : L 2 (e ) → Pk (e ) are the L 2 projections onto the associated
local polynomial spaces, Qh : [ L 2 ( K )]d → [Pk−1 ( K )]d is the L 2 projection onto the local weak gradient space. Finally, we
define a projection operator Q h u = { Q 0 u , Q b u } ∈ V h for the true solution u.
For simplicity, we adopt the following notations,

 
( v , w )Th = (v , w )K = v wdx,
K ∈Th K ∈Th K

 
v , w ∂ Th = v , w ∂K = v wds.
K ∈Th K ∈Th∂ K
 S. Xie et al. / Applied Numerical Mathematics 137 (2019) 19–33 21

For v = { v 0 , v b } and w = { w 0 , w b } in V h , we define three bilinear forms as

A( v , w ) := Ad ( v , w ) + Ac ( v , w ), (2.5)
Ad ( v , w ) := ε (∇ w v , ∇ w w )Th + sd ( v , w ), (2.6)
Ac ( v , w ) := (β · ∇ w v , w 0 )Th + (γ v 0 , w 0 ) + sc ( v , w ), (2.7)

where sc (·, ·) and sd (·, ·) are two stabilizer terms corresponding to convection term and diffusion term, respectively, i.e.,

sc ( v , w ) = β · n( v 0 − v b ), w 0 − w b ∂+ K ,
K ∈Th

sd ( v , w ) = εh−K 1  v 0 − v b , w 0 − w b ∂K ,
K ∈Th

and

∂+ K = {x ∈ ∂ K : β(x) · n(x) ≥ 0}.

The semi-discrete WG finite element scheme for (1.1) is as follows: find u h (t ) = {u 0,h (t ), u b,h (t )} ∈ V h0 , 0 < t ≤ T such that

(∂t u 0,h , v 0 ) + A(uh , v ) = ( f , v 0 ),
(2.8)
u h (0) = Q h u 0 ,

for all v = { v 0 , v b } ∈ V h0 .
Accordingly, we define an energy norm | · | in V h0 : for any v ∈ V h0 ,
 
| v |2 = ε ∇w v 2
K + |β · n|1/2 ( v 0 − v b ) 2
∂K + v0 2
+ sd ( v , v ). (2.9)
K ∈Th K ∈Th

We next show the coercivity of the bilinear form.

Lemma 2.1. There exists a positive constant α , independent of ε and h, such that for v ∈ V h0

A( v , v ) ≥ α | v |2 . (2.10)

Proof. Let v , w ∈ V h0 . It follows from the definition of the weak divergence (2.4) and integration by parts that

(β · ∇ w v , w 0 )Th = −( v 0 , ∇ · (β w 0 ))Th +  v b , β · nw 0 ∂ Th
= (β · ∇ v 0 , w 0 )Th − β · n( v 0 − v b ), w 0 ∂ Th (2.11)

and

(β · ∇ w w , v 0 )Th = −( w 0 , ∇ · (β v 0 ))Th +  w b , β · nv 0 ∂ Th
= −( w 0 , ∇ · (β v 0 ))Th +  w b , β · n( v 0 − v b ) ∂ Th , (2.12)
where in the last equality we use the facts that β · nv b , w b ∂ Th = 0.
Summing (2.11) and (2.12), and let v = w, we have

1 1
(β · ∇ w v , v 0 )Th = − (∇ · β v 0 , v 0 )Th − β · n( v 0 − v b ), v 0 − v b ∂ Th . (2.13)
2 2
Using (2.13), we have

1 1
Ac ( v , v ) = ((γ − ∇ · β) v 0 , v 0 ) − β · n( v 0 − v b ), v 0 − v b ∂ Th + sc ( v , v )
2 2
1
≥ (γ0 v 0 , v 0 ) + |β · n|( v 0 − v b ), v 0 − v b ∂ Th ,
2
which together with

Ad ( v , v ) = ε (∇ w v , ∇ w v )Th + sd ( v , v )

yields (2.10) with α = min{γ0 , 12 }. 2

22 S. Xie et al. / Applied Numerical Mathematics 137 (2019) 19–33

Theorem 2.1 (Stability). Assume that f (t ) ∈ L 2 ( Q T ). The semi-discrete solution u h (t ) defined in (2.8) satisfies
⎛ ⎞
t
u 0,h (t ) 2
≤ C ⎝ u 0,h (0) 2
+ f (s) 2
ds⎠ , ∀t ∈ (0, T ],
0

where the constant C > 0 is independent of t and h, and dependent exponentially on the final time T .

Proof. Taking v = u h in (2.8), we find

1 d 2
u 0,h + A(uh , uh ) = ( f , u 0,h ).
2 dt
It follows from (2.10) that

d 2 2 2
u 0,h ≤ 2( f , u 0,h ) ≤ f + u 0,h .
dt
Integrating the above inequality with respect to t, we find

t t
2 2 2 2
u 0,h (t ) ≤ u 0,h (0) + f (s) ds + u 0,h (s) ds.
0 0

By Gronwall inequality, we complete the proof. 2

Lemma 2.2. Let u be the solution of the problem (1.1). Then for v ∈ V h0 ,

−ε (u , v 0 ) = ε (∇ w Q h u , ∇ w v )Th − a (u , v ), (2.14)

(β · ∇ u , v 0 ) = (β · ∇ w Q h u , v 0 )Th − b (u , v ), (2.15)
(γ u , v 0 ) = (γ Q 0 u , v 0 ) − c (u , v ), (2.16)

where

a (u , v ) = ε (∇ u − Qh ∇ u ) · n, v 0 − v b ∂ Th ,

b (u , v ) = (u − Q 0 u , ∇ · (β v 0 ))Th − u − Q b u , β · n( v 0 − v b ) ∂ Th ,

c (u , v ) = −(γ u − γ Q 0 u , v 0 ).

Proof. The equation (2.14) follows from Lemma 3.4 of [14]. And obviously the equation (2.16) holds. Now we consider the
equation (2.15). It follows from integration by parts and the definition of the weak divergence (2.4) that

(β · ∇ u , v 0 )Th = − (u , ∇ · (β v 0 ))Th + u , β · nv 0 ∂ Th

= − ( Q 0 u , ∇ · (β v 0 ))Th + ( Q 0 u − u , ∇ · (β v 0 ))Th +  Q b u , β · nv 0 ∂ Th + u − Q b u , β · nv 0 ∂ Th

= (β · ∇ w ( Q h u ), v 0 )Th − b (u , v ).

The proof is completed. 2

In contrast to consistent finite element method such as conforming FEM and discontinuous Galerkin method, we do not
have the Galerkin orthogonality for WG method. So we need to derive an error equation for eh (t ), which will be used in
error analysis below.

Lemma 2.3. For a fixed time t ∈ (0, T ], let eh (t ) = Q h u (t ) − u h (t ) ∈ V h0 . Then, for any v ∈ V h0 we have

(∂t e 0,h , v 0 ) + A(eh , v ) = (u , v ), (2.17)

where

(u , v ) = a (u , v ) + b (u , v ) + c (u , v ) + sc ( Q h u , v ) + sd ( Q h u , v ). (2.18)
 S. Xie et al. / Applied Numerical Mathematics 137 (2019) 19–33 23

Proof. Testing the first equation in (1.1) by v = { v 0 , v b } ∈ V h0 , we arrive at

(∂t u , v 0 ) − ε (u , v 0 ) + (β · ∇ u , v 0 ) + (γ u , v 0 ) = ( f , v 0 ).
Using Lemma 2.2 and the definition of projector Q 0 , we find

(∂t Q 0 u , v 0 ) + ε (∇ w Q h u , ∇ w v )Th + (β · ∇ w Q h u , v 0 )Th + (γ Q 0 u , v 0 )

= ( f , v 0) + a (u , v ) + b (u , v ) + c (u , v ).

Adding sc ( Q h u , v ) and sd ( Q h u , v ) to both side of the above equation gives

(∂t Q 0 u , v 0 ) + A( Q h u , v ) = ( f , v 0 ) + (u , v ). (2.19)
Subtracting the first equation of (2.8) from (2.19) yields the error equation (2.17). The proof is completed. 2

The following trace inequality will be frequently used in our error estimates.
(Trace inequality) For any function ϕ ∈ H 1 ( K ), the following trace inequality holds true [27]:

ϕ 2
e ≤ Ch−
K
1
ϕ 2
K + hK ∇ϕ 2
K ). (2.20)

Lemma 2.4. ([14]) Let u be the solution of the problem (1.1) and Th be a finite element partition of  satisfying the shape regularity
assumptions A1–A4 descripted in [27]. Then, the L 2 projections Q 0 and Qh satisfy

( u − Q 0u 2
K + h2K ∇(u − Q 0 u ) 2
K) ≤ Ch2(s+1) u 2
s +1 , 0 ≤ s ≤ k,
K ∈Th

2
( ∇ u − Qh ∇ u K + h2K |∇ u − Qh ∇ u |21, K ) ≤ Ch2s u 2
s +1 , 0 ≤ s ≤ k.
K ∈Th

Lemma 2.5. Let u be the solution of the problem (1.1). Then for v ∈ V h0 ,
1
| a (u , v )| ≤ C ε 2 hk |u |k+1 | v |,
1
| b (u , v )| ≤ Chk+ 2 |u |k+1 | v |,
| c (u , v )| ≤ Chk+1 |u |k+1 | v |,
1
|sc ( Q h u , v )| ≤ Chk+ 2 |u |k+1 | v |,
1
|sd ( Q h u , v )| ≤ C ε 2 hk |u |k+1 | v |.

Proof. The estimate of a (u , v ), c (u , v ), sc ( Q h u , v ) and sd ( Q h u , v ) can be obtained from Lemma 3.7 of [14].
Next we consider the estimate of b (u , v ). Let β K = | K1 | (β, 1) K . The first term of b (u , v ) can be estimated as

(u − Q 0 u , ∇ · (β v 0 ))Th = (u − Q 0 u , (∇ · β) v 0 )Th + (u − Q 0 u , β · ∇ v 0 )Th


= {(u − Q 0 u , (∇ · β) v 0 ) K + (u − Q 0 u , (β − β K ) · ∇ v 0 ) K }
K ∈Th

≤C u − Q 0u K( v0 K + β − βK L∞ (K ) ∇ v0 K)
K ∈Th

≤ Chk+1 |u |k+1 v 0 . (2.21)

For the second term of b (u , v ) , it follows from the definition of Q b and the Cauchy–Schwarz inequality that

u − Q b u , β · n ( v 0 − v b ) ∂ Th ≤C u − Q b u ∂ K |β · n|1/2 ( v 0 − v b ) ∂ K
K ∈Th

⎛ ⎞1/2 ⎛ ⎞1/2
  
≤C⎝ u − Q 0u 2⎠
e
⎝ |β · n|1/2 ( v 0 − v b ) 2
∂K
⎠
K ∈Th e ⊂∂ K K ∈Th

≤ Chk+1/2 |u |k+1/2 | v |, (2.22)

24 S. Xie et al. / Applied Numerical Mathematics 137 (2019) 19–33

where we have used the fact u − Q b u e ≤ u − Q 0u e in the second inequality. Combining (2.21) and (2.22) yields

| b (u , v )| ≤ |(u − Q 0 u , ∇ · (β v 0 ))Th | + |u − Q b u , β · n( v 0 − v b ) ∂ Th |

k+1/2
≤ Ch |u |k+1/2 | v |.
The proof is completed. 2

We shall prove the following error estimate of the semi-discrete WG scheme (2.8).

Theorem 2.2 (Semi-discrete error estimate). Let u and u h be the solutions to the problem (1.1) and the semi-discrete WG scheme (2.8),
respectively. For a fixed time t ∈ (0, T ], let eh (t ) = Q h u (t ) − u h (t ) ∈ V h0 . Then, the following estimates hold

t t
2 2 2k
e 0,h (t ) + α |eh (s) | ds ≤ C (ε + h)h |u (s)|k2+1 ds. (2.23)
0 0

Proof. For each t ∈ (0, T ], taking v = eh in (2.17), we have

(∂t e 0,h , e 0,h ) + A(eh , eh ) = (u , eh ).

From (2.10), (2.18), and Lemma 2.5, we find

1 d 2
e 0,h + α |eh |2 ≤ C (ε 1/2 + h1/2 )hk |u |k+1 |eh |
2 dt
1
≤ C (ε + h)h2k |u |k2+1 + α |eh |2 ,
2
i.e.,
d 2
e 0,h + α |eh |2 ≤ C (ε + h)h2k |u |k2+1 . (2.24)
dt
Integrating (2.24) with respect to t and noting eh (0) = 0, we conclude (2.23). 2

3. Full-discrete WG finite element scheme

In order to obtain a fully discretization of (2.8), we consider an uniform mesh for time variable t and define

t n = nτ , n = 0, 1 , · · · , N ,
n+1/2
where τ = T /N is the time-step, and N is a positive integer. Denote unh := u h (tn ) = {un0,h , unb,h } and u h = 12 (unh+1 + unh ).
Applying the Crank–Nicolson scheme to the semi-discrete approximation (2.8), we obtain the Crank–Nicolson WG method
for the problem (1.1): find unh = {un0,h , unb,h } ∈ V h0 such that

n +1 n+1/2
1
τ (u 0,h − u 0,h , v 0 ) + A(u h
n
, v ) = ( f n+1/2 , v 0 ),
(3.1)
u h0 = Q hu , 0

for all v ∈ V h0 and n = 0, 1, · · · , N − 1, where f n+1/2 = 12 ( f (tn+1 ) + f (tn )).

We are going to prove that the Crank–Nicolson scheme is unconditionally stable with respect to the L 2 norm. It is useful
to introduce the following notation: for any function φ ∈ L 2 (), define

(φ, v h )
φ −1,h := sup ,
v h ∈ V h , v h =0 | v h |

which is a norm on V h . Clearly, φ −1,h ≤ φ for all φ ∈ L 2 ().

Theorem 3.1 (Stability). Assume that the map t → f (t ) is bounded in [0, T ]. Then unh defined in (3.1) satisfies

un0,h ≤C u 0
+ sup f (t ) , n = 0, 1 , · · · , N ,
t ∈[0, T ]

where the constant C > 0 is independent of N , τ and h, and dependent exponentially on the final time T .
 S. Xie et al. / Applied Numerical Mathematics 137 (2019) 19–33 25

n+1/2
Proof. Taking v = u h in (3.1). It is easily verified that

1 1 n+1/2 n+1/2 n+1/2

un0+
,h
1 2
− un0,h 2
+ τ A( u h , uh ) = τ ( f n+1/2 , u 0,h ).
2 2
By the coercivity (2.10), we find
τ
un0+
,h
1 2
− un0,h 2
≤ f n+1/2 2−1,h .
2α
Now let m be a fixed index, 1 ≤ m ≤ N . Summing up above inequalities from n = 0 to n = m − 1, we find
m −1
τ  n+1/2
um
0,h
2
≤ u 00,h 2
+ f 2
−1,h (3.2)
2α
n =0

and the theorem follows. 2

Now we turn to the convergence analysis. We prove an error estimate between Q h u (tn ) and the fully-discrete one unh
for any fixed h. Firstly, we derive an error equation for enh := Q h u (tn ) − unh , which will be used in error analysis below. For
simplicity, we denote
1 1
un+1/2 = (u (tn+1 ) + u (tn )), and f n+1/2 = ( f (tn+1 ) + f (tn )).
2 2
The error equation for enh reads as follows.

Lemma 3.1. For v = { v 0 , v b } ∈ V h0 , we have

1 n+1/2
(en0+ 1 n
,h − e 0,h , v ) + A(eh , v ) = (rn , v 0 ) + (un+1/2 , v ), (3.3)
τ
where r n = τ1 (un+1 − un ) − ∂t un+1/2 .

Proof. From the first equation in (1.1), we obtain

∂t un+1/2 − ε un+1/2 + β · ∇ un+1/2 + γ un+1/2 = f n+1/2 . (3.4)

Testing (3.4) by v = { v 0 , v b } ∈ V h0 , we arrive at

(∂t un+1/2 , v 0 ) − ε (un+1/2 , v 0 ) + (β · ∇ un+1/2 , v 0 ) + (γ un+1/2 , v 0 ) = ( f n+1/2 , v 0 ).

Using Lemma 2.2, we find

(∂t un+1/2 , v 0 ) + ε (∇ w Q h un+1/2 , ∇ w v )Th + (β · ∇ w Q h un+1/2 , v 0 )Th + (γ Q 0 un+1/2 , v 0 )

= ( f n+1/2 , v 0 ) + a (u
n+1/2
, v) + b (u
n+1/2
, v) + c (u
n+1/2
, v ).
n+1/2 n+1/2
Adding sc ( Q h u , v ) and sd ( Q h u , v ) to both side of the above equation gives

(∂t un+1/2 , v 0 ) + A( Q h un+1/2 , v ) = ( f n+1/2 , v 0 ) + (un+1/2 , v ),

which can be rewritten as
1
( Q 0 un+1 − Q 0 un , v 0 ) + A( Q h un+1/2 , v ) = (rn , v 0 ) + ( f n+1/2 , v 0 ) + (un+1/2 , v ). (3.5)
τ
Subtracting the first equation of (3.1) from (3.5) yields the error equation (3.3). The proof is completed. 2

Now the error estimate of the full discrete WG scheme (3.1) reads as follows.

Theorem 3.2 (full-discrete error estimate). Let u ∈ H k+1 () and unh be the solutions of problem (1.1) and the full discrete WG scheme
(3.1), respectively. For a fixed 0 < n ≤ N , let enh = Q h u (tn ) − unh ∈ V h0 . Then there exists a constant C such that

n −1
 tn
j +1/2
en0,h 2 + τα |eh 2
| ≤ Cτ 4
uttt (s) 2
ds + C (ε + h)h2k |u |k2+1 ,
j =0 0

for each n = 1, · · · , N , where |u |k+1 = max0≤ j ≤n |u (t j )|k+1 .

26 S. Xie et al. / Applied Numerical Mathematics 137 (2019) 19–33

n+1/2
Proof. Taking v = eh in (3.3) and by coercivity (2.10), we find
1 n+1/2 n+1/2 n+1/2
( en0+
,h
1 2
− en0,h 2
) + α |eh |2 = (rn , e0,h ) + (un+1/2 , eh ),
2τ
or equivalently
n+1/2 n+1/2 n+1/2
en0+
,h
1 2
− en0,h 2
+ 2τ α |eh |2 = 2τ (rn , e0,h ) + 2τ (un+1/2 , eh ). (3.6)
A calculation yields
1 1
r n = (u (tn+1 ) − u (tn )) − (∂t u (tn+1 ) + ∂t u (tn ))
τ 2

t n +1
1
= (tn+1 − s)(tn − s)uttt (s)ds,
2τ
tn

and thus
⎡ ⎤2
 
t n +1
1
rn 2
= ⎣ (tn+1 − s)(tn − s)uttt (s)ds⎦ dx
2τ
 tn
⎡ t n +1 ⎤
  
t n +1
1
≤ ⎣ (tn+1 − s) (tn − s) ds
2 2
uttt (s)ds⎦ dx
2
4τ 2
 tn tn


t n +1
1
≤ τ3 uttt (s) 2
ds.
120
tn

Then the first term on the right hand side of (3.6) can be estimated as
n+1/2 n+1/2 n+1/2
2τ (r n , e 0,h ) ≤ 2τ r n −1,h |eh | ≤ 2τ rn |eh |
2τ 1 n+1/2
≤ rn 2
+ τ α |eh |2
α 2

t n +1
1
≤ Cτ 4
uttt (s) 2
ds + τ α |enh+1/2 |2 . (3.7)
2
tn

By Lemma 2.5 and Young’s inequality, we obtain the estimate of the second term in the right hand side of (3.6) as follows
n+1/2 n+1/2
2τ (un+1/2 , eh ) ≤ 2τ C (ε 1/2 + h1/2 )hk |un+1/2 |k+1 |eh |
2
C 1
≤ τ (ε + h)h2k |un+1/2 |k2+1 + ατ |enh+1/2 |2 . (3.8)
α 2
Combining (3.6), (3.7), and (3.8) yields
n+1/2
en0+
,h
1 2
− en0,h 2
+ τ α |eh |2

t n +1

≤ Cτ 4
uttt (s) 2
ds + C τ (ε + h)h2k |un+1/2 |k2+1 .
tn

Since eh0 = 0, for any fixed index m, 1 ≤ m ≤ N , by summing up from n = 0 to n = m − 1, we finally obtain


m −1 tm 
m −1
n+1/2
em
0,h
2
+ τα |eh 2
| ≤ Cτ 4
uttt (s) 2
ds + C τ (ε + h)h2k |un+1/2 |k2+1
n =0 0 n =0

tm
≤ Cτ 4 uttt (s) 2
ds + C (ε + h)h2k |u |k2+1 ,
0

where |u |k+1 = max0≤n≤m |u (tn )|k+1 . The proof is completed. 2

 S. Xie et al. / Applied Numerical Mathematics 137 (2019) 19–33 27

Fig. 1. Demo of polygonal mesh.

Remark 1. In short, Theorem 3.2 states that the Crank–Nicolson WG scheme (3.1) achieves accuracy of O (τ 2 + (ε 1/2 +
h1/2 )hk ) in L 2 norm. For convection-dominated case, ε  h, the error estimate is O (τ 2 + hk+1/2 ). The same result can be
found in John and Novo [12], where the SUPG finite element discretization is adopted for solving evolutionary convection–
diffusion–reaction equations.

4. Numerical examples

In this section, we test numerically the performance of the proposed Crank–Nicolson WG FEM for solving initial boundary
value problems (1.1). Examples of different types have been tested to confirm the efficiency of the method. In all computa-
tions we use polygonal meshes generated by PolyMesher [25]. Fig. 1 shows two polygonal meshes with 8 × 8 and 16 × 16
elements, respectively.
The error for the Crank–Nicolson WG-FEM solution is measured in the following norms

en0,h 2
= |en0,h |2 dx,
K ∈Th K

N
 −1
j +1/2
|eh |2T = τ |eh |2 ,
j =0

j +1/2 j j +1
where N τ = T , eh = (eh + eh )/2 and the triple-bar norm | · | is defined in (2.9).
Example 1: Smooth solution. This example is adopted from [12]. Consider problem (1.1) with  = (0, 1)2 , T = 1, different
values of ε , β = (1, −1) T , and γ = 1. The initial data u 0 and the source term f is chosen so that the exact solution is

u (x, y , t ) = exp(sin(2π t )) sin(2π x) sin(2π y ).

We first study the temporal convergence property of the Crank–Nicolson WG scheme (3.1). The simulations are con-
ducted on a polygonal mesh with 32 × 32 elements and polynomials of order k = 3 for different time step-size of
τ = 1/16, 1/32, 1/64. We show in Table 1 the L 2 error and order of convergence for the WG solution uh at the final
time. It is observed that the accuracy in L 2 norm is second-order in time.
Table 2 shows the errors of WG scheme in L 2 norm and | · | T norm and their convergence rates of the initial boundary
value problem for ε = 1 in the diffusion-dominated regime and ε = 10−8 in the convection-dominated regime. The simu-
lations are conducted on polygonal meshes with N × N elements for N = 4, 8, 16, 32 and polynomials of order k = 1, 2 for

Table 1
Example 1. History of convergence of the L 2 error of u h at the final time computed with P3
element.

τ ε = 1.0 ε = 10−8
en0,h Rate en0,h Rate
1/16 8.14E−03 – 9.41E−03 –
1/32 2.34E−03 1.80 2.25E−03 2.06
1/64 5.97E−04 1.97 5.43E−04 2.05
28 S. Xie et al. / Applied Numerical Mathematics 137 (2019) 19–33

Table 2
Example 1. History of en0,h and |eh | T and their convergence rates at the final time for P1 and P2 elements, fixing τ = 10−5 .
k N ε = 1.0 ε = 10−8
en0,h Rate |eh | T Rate en0,h Rate |eh | T Rate
4 2.85E−01 – 4.47E+00 – 9.09E−02 – 7.61E−01 –
8 9.02E−02 1.67 2.94E+00 0.60 2.18E−02 2.06 2.99E−01 1.35
1
16 2.45E−02 1.88 1.67E+00 0.82 4.96E−03 2.14 1.03E−01 1.54
32 6.51E−03 1.91 8.84E−01 0.91 1.29E−03 1.94 3.54E−02 1.54
4 1.35E−01 – 2.76E+00 – 1.90E−02 – 2.37E−01 –
8 1.43E−02 3.24 6.05E−01 2.19 2.63E−03 2.86 4.14E−02 2.51
2
16 1.74E−03 3.04 1.40E−01 2.11 3.67E−04 2.84 6.53E−03 2.67
32 2.21E−04 2.98 3.11E−02 2.17 4.87E−05 2.91 1.15E−03 2.50

Table 3
Example 2. History of convergence of the L 2 error of u h at the final time computed with P3
element on 64 × 64 polygonal mesh.

τ ε = 10−2 ε = 10−3
en0,h Rate en0,h Rate
0.08 1.32E−02 – 3.74E−02 –
0.04 2.91E−03 2.18 1.17E−02 1.68
0.02 7.26E−04 2.00 3.07E−03 1.93
0.01 1.82E−04 2.00 7.69E−04 2.00

Table 4
Example 2. History of en0,h and |eh | T errors and convergence rates, using Pk element, fixing τ = h(k+1)/2 , k = 1, 2.
k N ε = 1.0e−2 ε = 1.0e−3
en0,h Rate |eh | T Rate en0,h Rate |eh | T Rate
4 4.30E−02 – 1.84E−01 – 9.71E−02 – 2.13E−01 –
8 2.15E−02 0.99 7.03E−02 1.39 5.71E−02 0.77 9.27E−02 1.20
1 16 7.38E−03 1.54 3.27E−02 1.10 2.39E−02 1.26 3.44E−02 1.43
32 2.00E−03 1.88 2.09E−02 0.65 7.29E−03 1.71 1.13E−02 1.61
64 5.36E−04 1.90 1.16E−02 0.84 1.88E−03 1.96 3.67E−03 1.62
4 2.12E−02 – 6.88E−02 – 5.58E−02 – 1.01E−01 –
8 3.61E−03 2.55 1.73E−02 1.99 1.31E−02 2.09 2.26E−02 2.15
2 16 4.67E−04 2.95 3.96E−03 2.13 1.88E−03 2.81 3.86E−03 2.55
32 6.10E−05 2.93 9.24E−04 2.10 2.37E−04 2.98 6.76E−04 2.51
64 7.92E−06 2.95 2.16E−04 2.09 2.97E−05 3.00 1.20E−04 2.50

fixing time step-size of τ = 10−5 . It is observed that WG scheme with Pk (k = 1, 2) element has optimal convergence order
of O (hk+1 ) in L 2 norm for both ε = 1 and ε = 10−8 . On the other hand, for Pk (k = 1, 2) element, |eh | T converges at O (hk ),
and O (hk+1/2 ) for ε = 1 and ε = 10−8 , respectively.

Example 2: Rotating Gaussian pulse. This example was considered in [21]. Let  = (− 12 , 12 ) × (− 12 , 12 ), and β = (−4 y , 4x),
γ = f = 0. The finial time is T = π /4. The initial condition is given by
 
(x − xc )2 + ( y − y c )2
u 0 (x, y ) = exp − .
2σ 2
In the numerical experiments, we take xc = −0.2, y c = 0, and σ = 0.1. The exact solution is
 
2σ 2 (x̄ − xc )2 + ( ȳ − y c )2
u (x, y , t ) = exp − ,
2σ + 4ε t
2 2σ 2 + 4ε t
where x̄ = x cos(4t ) + y sin(4t ) and ȳ = −x sin(4t ) + y cos(4t ). The Dirichlet boundary condition is deduced from the exact
solution. Furthermore, we consider two different values of diffusion coefficient ε = 10−2 and ε = 10−3 .
Table 3 shows the L 2 error and order of convergence for the WG solution u h at the final time for ε = 10−2 and ε = 10−3 .
The WG solutions are computed on a polygonal mesh with 64 × 64 elements and polynomials of order k = 3 for different
time step-size of τ = 0.08, 0.04, 0.02, 0.01. Again the second-order of temporal accuracy is observed.
The L 2 and | · | T norm errors and their corresponding convergence rates are displayed in Table 4. The numerical
solutions are computed on polygonal meshes with N × N elements for N = 4, 8, 16, 32 and polynomials of order k(k = 1, 2)
for time step-size of τ = h(k+1)/2 . The convergence properties in the previous example are also observed in this example.
Fig. 2 presents the WG solutions, the exact solutions and their errors at final time t = π /4 for ε = 10−2 and ε = 10−3 .
 S. Xie et al. / Applied Numerical Mathematics 137 (2019) 19–33 29

Fig. 2. Example 2. WG solution, exact solutions and their error eh = u h − Q h u at final time for P3 element on a polygonal mesh with 64 × 64 elements,
τ = 0.01. (a), (c), (e) for ε = 10−2 , and (b), (d), (f) for ε = 10−3 .

Example 3: Boundary layer. This example is adopted from [8]. Consider problem (1.1) with  = (0, 1)2 , T = 1, different
values of ε , β = (1, 1) T , and γ = 0. The initial data u 0 and the source term f is chosen so that the exact solution is

u (x, y , t ) = (1 − e −t )ϕ (x)ϕ ( y ),
30 S. Xie et al. / Applied Numerical Mathematics 137 (2019) 19–33

Table 5
Example 3. History of convergence of the L 2 error of u h at the final time computed with P3
element on 64 × 64 polygonal mesh.

τ ε = 1.0 ε = 10−8
en0,h Rate en0,h Rate
0.08 9.42E−07 – 1.09E−04 –
0.04 2.36E−07 2.00 2.73E−05 2.00
0.02 6.01E−08 1.97 6.95E−06 1.97

Table 6
Example 3. History of L 2 errors and convergence rates, using Pk element, fixing τ = h(k+1)/2 , k = 1, 2.
k N ε = 1.0 ε = 10−8
en0,h Rate en0,h Rate
4 5.13E−04 – 1.93E−03 –
8 1.24E−04 2.05 4.82E−04 2.00
1
16 3.12E−05 1.99 1.30E−04 1.89
32 7.82E−06 1.99 3.29E−05 1.99
4 3.66E−05 – 2.51E−04 –
8 4.25E−06 3.11 3.51E−05 2.84
2
16 5.18E−07 3.04 4.38E−06 3.00
32 6.70E−08 2.95 5.45E−07 3.01

where ϕ (s) = m1 + m2 s + e −(1−s)/ε , m1 = −e −1/ε and m2 = −1 − m1 . This problem has two boundary layers along x = 1 and
y = 1.
Table 5 demonstrates the history of convergence of the L 2 error of WG solution u h at the final time computed with P3
element on a 64 × 64 polygonal mesh for time step-size τ = 0.08, 0.04, 0.02. Again observe that the accuracy in L 2 norm is
second-order in time.
In Table 6, we list the L 2 errors and orders of convergence for WG solution u h computed on polygonal meshes with N × N
elements for N = 4, 8, 16, 32. Again, it is observed that WG scheme with Pk (k = 1, 2) element has optimal convergence order
of O (hk+1 ) in L 2 norm for both ε = 1 and ε = 10−8 .
Fig. 3 shows the figures of the WG solutions, the exact solutions and their errors eh at final time for ε = 10−2 and
ε = 10−8 . Observe that the numerical solutions do not exhibit spurious oscillations near the boundary layers x = 1 and
y = 1.

Example 4: Interior layer. This example was considered in [7]. Consider problem (1.1) with  = (0, 1)2 , T = 1, ε = 10−8 ,
β = (−(x − 1/2)(3 + t + x2 y 2 ), −( y − 1/2)(xyt + 2e xy ))T , and γ = 3 + t cos(xy ). The initial data u 0 = 0 and the source term
f is

t (2 + sin(x + y )), if x ≤ 1/2,
f (x, y , t ) =
t 2 (1 − xy ), if x > 1/2.
The exact solution of this problem is unknown.
Fig. 4 shows the used mesh and the WG solution computed by P1 element on a polygonal mesh with 64 × 64 elements
at the final time t = 1. From it, we clearly see that there is no spurious oscillations. Also observe that there is an interior
layer at x = 1/2.

Example 5: The contaminant transport problem. This example is adopted from [21], which was also discussed in [3]. Con-
sider the transport of the contaminant concentration with kinematic viscosity ε in the domain  = (0, 10) ×(−1.25, 1.25)
driven by a convective velocity field β = (1 − exp(θ x) cos(2π y ), 2θπ exp (θ x) sin(2π y )), where θ = Re /2 − Re 2 /4 + 4π 2
and Re = 100. The same contaminant concentration w 0 is injected into the flow field after every two seconds T = 2, while
the contaminant concentration u is being transposed downstream. The contaminant transport over the time horizon of
T f = 10 s is considered. The mathematical model of this physical process is the following time-dependent convection–
diffusion equation
⎧
⎪
⎨ ∂t u + β · ∇ u − ε u = 0, in  × (( j − 1) T , jT ],
u (x, t ) = 0, on ∂  × (( j − 1) T , jT ], (4.1)
⎪
⎩ u (x, ( j − 1) T ) = u 0 , in .
The initial condition is given by

0 w0, if j = 1,
u =
w 0 + u (( j − 1) T ), if j = 2, · · · , 5,
 S. Xie et al. / Applied Numerical Mathematics 137 (2019) 19–33 31

Fig. 3. Example 3. WG solution, exact solutions and their error eh = u h − Q h u at final time for P3 element on a polygonal mesh with 64 × 64 elements,
τ = 0.01. (a), (c), (e) for ε = 10−2 , and (b), (d), (f) for ε = 10−8 .

where initial contaminant concentration is given by

2 + y2 2 +( y −0.5)2 2 +( y +0.5)2
− (x−1) − (x−1) − (x−1)
w0 = e 0.52 +e 0.52 +e 0.52 .
Fig. 5 shows the approximated contaminant concentration u h at the time t = T , 2T , 3T , 4T , 5T on a polygonal mesh with
32 × 32 elements for ε = 10−2 . In this computation, P3 element is used and time step-size τ = 0.025.
32 S. Xie et al. / Applied Numerical Mathematics 137 (2019) 19–33

Fig. 4. Example 4. The WG solution using P1 element on a polygonal mesh with 64 × 64 elements for ε = 10−8 , t = 1.

Fig. 5. Example 5. The WG solution using P3 element on polygonal mesh with 32 × 32 elements for ε = 10−2 .
 S. Xie et al. / Applied Numerical Mathematics 137 (2019) 19–33 33

5. Conclusions

In this work, we proposed a fully discrete scheme, based on weak Galerkin method in space and Crank–Nicolson method
in time, for time-dependent convection-dominated diffusion equation. The stability and the convergence of semi-discrete
and full-discrete WG schemes are proved respectively. The analysis in this paper can be extended to a more general time
discretization, such as discontinuous Galerkin time stepping method [9].

References

[1] N. Ahmed, G. Matthies, Numerical study of SUPG and LPS methods combined with high order variational time discretization schemes applied to
time-dependent linear convection–diffusion–reaction equations, J. Sci. Comput. 67 (2016) 988–1018.
[2] M. Augustin, A. Caiazzo, A. Fiebach, J. Fuhrmann, V. John, A. Linke, R. Umla, An assessment of discretizations for convection-dominated convection–
diffusion equations, Comput. Methods Appl. Mech. Eng. 200 (47–48) (2011) 3395–3409.
[3] O. Bashir, K. Willcox, O. Ghattas, B. van Bloemen Waander, J. Hill, Hessian-based model reduction for large-scale systems with initial-condition inputs,
Int. J. Numer. Methods Eng. 73 (6) (2008) 844–868.
[4] E. Burman, Consistent SUPG-method for transient transport problems: stability and convergence, Comput. Methods Appl. Mech. Eng. 199 (2010)
1114–1123.
[5] E. Burman, A. Ern, Continuous interior penalty hp-finite element methods for advection and advection–diffusion equations, Math. Comput. 76 (2007)
1119–1140.
[6] G. Chen, M. Feng, X. Xie, A robust WG finite element method for convection–diffusion–reaction equations, J. Comput. Appl. Math. 315 (2017) 107–125.
[7] C. Clavero, J. Vigo-Aguiar, Numerical approximation of 2D time dependent singularly perturbed convection–diffusion problems with attractive or re-
pulsive turning points, Appl. Math. Comput. 317 (2018) 223–233.
[8] A. Das, S. Natesan, Higher-order convergence with fractional-step method for singularly perturbed 2D parabolic convection–diffusion problems on
Shishkin mesh, Comput. Math. Appl. 75 (2018) 2387–2403.
[9] S. Franz, G. Matthies, A unified framework for time-dependent singularly perturbed problems with discontinuous Galerkin methods in time, Math.
Comput. 87 (2018) 2113–2132.
[10] G. Fu, C.-W. Shu, Analysis of an embedded discontinuous Galerkin method with implicit–explicit time-marching for convection–diffusion problems, Int.
J. Numer. Anal. Model. 14 (2017) 477–499.
[11] P. Houston, C. Schwab, E. Süli, Discontinuous hp-finite element methods for advection–diffusion–reaction problems, SIAM J. Numer. Anal. 39 (6) (2002)
2133–2163.
[12] V. John, J. Novo, Error analysis of the SUPG finite element discretization of evolutionary convection–diffusion–reaction equations, SIAM J. Numer. Anal.
49 (2011) 1149–1176.
[13] V. John, E. Schmeyer, Finite element methods for time-dependent convection–diffusion–reaction equations with small diffusion, Comput. Methods Appl.
Mech. Eng. 198 (2008) 475–494.
[14] R. Lin, X. Ye, S. Zhang, P. Zhu, A weak Galerkin finite element method for singularly perturbed convection–diffusion–reaction problems, SIAM J. Numer.
Anal. 56 (2018) 1482–1497.
[15] G. Matthies, P. Skrzypacz, L. Tobiska, A unified convergence analysis for local projection stabilisations applied to the Oseen problem, ESAIM: M2AN 41
(2007) 713–742.
[16] L. Mu, J. Wang, X. Ye, A weak Galerkin finite element method for biharmonic equations on polytopal meshes, Numer. Methods Partial Differ. Equ. 30
(2014) 1003–1029.
[17] L. Mu, J. Wang, X. Ye, Weak Galerkin finite element method for the Helmholtz equation with large wave number on polytopal meshes, IMA J. Numer.
Anal. 35 (2015) 1228–1255.
[18] L. Mu, J. Wang, X. Ye, Weak Galerkin finite element method for second-order elliptic problems on polytopal meshes, Int. J. Numer. Anal. Model. 12
(2015) 31–53.
[19] L. Mu, J. Wang, X. Ye, S. Zhang, A weak Galerkin finite element method for the Maxwell equations, J. Sci. Comput. 65 (2015) 363–386.
[20] L. Mu, J. Wang, X. Ye, S. Zhao, A new weak Galerkin finite element method for elliptic interface problems, J. Comput. Phys. 325 (2016) 157–173.
[21] N.C. Nguyen, J. Peraire, B. Cockburn, An implicit high-order hybridizable discontinuous Galerkin method for linear convection–diffusion equations,
J. Comput. Phys. 228 (2009) 3232–3254.
[22] I. Oikawa, Hybridized discontinuous Galerkin method for convection–diffusion problems, Jpn. J. Ind. Appl. Math. 31 (2) (2014) 335–354.
[23] H.-G. Roos, M. Stynes, L. Tobiska, Robust numerical methods for singularly perturbed differential equations, in: Convection–Diffusion–Reaction and
Flow Problems, second edition, in: Springer Series in Computational Mathematics, vol. 24, Springer-Verlag, Berlin, 2008.
[24] F. Schieweck, On the role of boundary conditions for CIP stabilization of higher order finite elements, Electron. Trans. Numer. Anal. 32 (2008) 1–16.
[25] C. Talischi, G.H. Paulino, A. Pereira, I.F.M. Menezes, PolyMesher: a general-purpose mesh generator for polygonal elements written in Matlab, Struct.
Multidiscip. Optim. 45 (3) (2012) 309–328.
[26] J. Wang, X. Ye, A weak Galerkin finite element method for second-order elliptic problems, J. Comput. Appl. Math. 241 (2013) 103–115.
[27] J. Wang, X. Ye, A Weak Galerkin mixed finite element method for second-order elliptic problems, Math. Comput. 83 (2014) 2101–2126.
[28] J. Wang, X. Ye, A weak Galerkin finite element method for the Stokes equations, Adv. Comput. Math. 42 (2016) 155–174.
[29] G. Zhou, How accurate is the streamline diffusion finite element method?, Math. Comput. 66 (217) (1997) 31–44.