J Sci Comput (2011) 49:383–401

DOI 10.1007/s10915-011-9469-3

Two-Grid Method for Nonlinear Reaction-Diffusion

Equations by Mixed Finite Element Methods

Luoping Chen · Yanping Chen

Received: 8 July 2010 / Accepted: 25 January 2011 / Published online: 9 February 2011
© Springer Science+Business Media, LLC 2011

Abstract In this paper, we investigate a scheme for nonlinear reaction-diffusion equations

using the mixed finite element methods. To linearize the mixed method equations, we use the
two-grid algorithm. First, we solve the original nonlinear equations on the coarse grid, then,
we solve the linearized problem on the fine grid used Newton iteration once. It is shown that
the algorithm can achieve asymptotically optimal approximation as long as the mesh sizes
1
satisfy H = O(h 2 ). As a result, solving such a large class of nonlinear equations will not
much more difficult than the solution of one linearized equation.

Keywords Two-grid method · Reaction-diffusion equations · Mixed finite element

methods

1 Introduction

The two-grid method, proposed by Xu [20, 21] as a discretization method for nonsym-
metric, indefinite and nonlinear partial differential equations, is based on the fact that the
non-symmetry, indefiniteness and nonlinearity behaving like low frequencies are governed
by the coarse grid and the related high frequencies are governed by some linear symmetric
positive-definition operators. The basic idea of the two-grid method is to solve a complicated
problem (non-symmetric indefinite or nonlinear, etc.) on the coarse grid and then solve a
simple symmetric positive or linearized problem on the fine grid. Here, we shall apply this
scheme to the mixed finite element methods for nonlinear parabolic equations.

L. Chen
School of Mathematical and Computational Science, Sun Yat-Sen University, Guangzhou 510275,
Guangdong, P.R. China
e-mail: [email protected]

Y. Chen ()
School of Mathematical Science, South China Normal University, Guangzhou 520631, Guangdong,
P.R. China
e-mail: [email protected]
384 J Sci Comput (2011) 49:383–401

In this paper, we consider the following nonlinear reaction-diffusion equations:

∂p
− ∇ · (K(p)∇p) = f (p), (x, t) ∈  × J, (1.1)
∂t
with initial condition

p(x, 0) = p 0 (x), x ∈ , (1.2)

and boundary condition

K(p)∇p · ν = 0, (x, t) ∈ ∂ × J, (1.3)

where  ∈ R2 is a bounded and convex domain with C 1 boundary ∂, ν is the unit exterior
normal to ∂, J = (0, T ], K(p):  × R → R2×2 is a symmetric positive definite tensor.
Equation (1.1) can be rewritten as followings
∂p
+ ∇ · u = f (p), (1.4)
∂t
K(p)−1 u + ∇p = 0. (1.5)

Reaction-diffusion equations have received a great deal of attention, motivated by their

widespread occurrence in models of hydrologic, biology and bio-geochemical phenomena
[14, 15, 22]. Typical examples include the modeling of groundwater through porous media
[10]. In this case, p denotes the fluid pressure, u is the Darcy velocity of the flow and f (p)
models the external flow rate. Here, for brevity, we drop the dependence of variable x in
f (x, p).
Mixed finite element methods have been found to be very important for solving the par-
abolic partial differential equations [13, 16]. For example, there are many applications of
mixed finite element methods to miscible displacement problems that describe two-phase
flow in petroleum reservoir [10]. In a mixed finite element formulation, both the pressure
and the flux, or displacements and stresses, are approximated simultaneously.
To linearize the resulting discrete equations, we use the two-grid method which was
analyzed by Xu [20, 21] for Galerkin procedures applied to nonlinear elliptic equations.
In 1994, Chen and Huang [3] presented a multilevel iterative method for solving the finite
element equations of non-linear singular two-point problems.
For non-linear parabolic equations, this approach was applied to the mixed finite ele-
ment method by Dawson and Wheeler [11] with f dependent of p, ∇p. The case that f
is independent of p is treated by Dawson and Wheeler [12] by two-grid difference scheme.
Moreover, Wu and Allen [19] established and analyzed a two-step algorithm by using the
two-grid idea for the semi-linear reaction-diffusion equations with the expanded mixed fi-
nite element method. Based on this work, in order to improve accuracy of the algorithm, in
our work [5], we proposed a three steps algorithm using the correction idea from reference
[21]. Our continued work [6] presented two efficient algorithms of the two-grid methods for
the approximation of two-dimensional semi-linear reaction-diffusion equations using the
expanded mixed finite element method which was first developed by Arbogast et al. [1]. We
designed another three-step algorithm with two Newton iterations and four-step algorithm
using Newton iteration twice and correction technique once. We also analyze the error es-
timation of the two-grid solutions in the sense of Lp norm in [4] and analyze a much more
complex equation in [7].
In this paper, we consider a much more complicated nonlinear reaction-diffusion equa-
tion with matrix K dependent on p by the efficient two-grid algorithm. The two-grid ideas
J Sci Comput (2011) 49:383–401 385

from Xu’s work [20, 21] are mainly applied on the standard finite element method, we ex-
tended the ideas to the mixed finite element method. The main algorithm in our literature
involves two steps: first, solving a small non-linear system on the coarse space and then
solving a large linear system on the fine grid with one Newton iteration. We obtain the er-
ror estimates for our two-grid algorithm both theoretically and numerically. It is shown that
coarse space can be coarser and it can achieve asymptotically optimal approximation as long
1
as the mesh sizes satisfy H = O(h 2 ). The key ingredient of our error analysis is that we use
two important superconvergence properties to analyze the two-step algorithm.
The remainder of the article is organized as follows: In Sect. 2, we give some notations
and projection operators. The mixed method for the nonlinear reaction-diffusion equations
(1.1)–(1.3) and the corresponding convergence analysis will be presented in Sect. 3. The
main algorithm and its error estimates will be discussed in Sect. 4. Section 5 is devoted to
the presentation of a numerical example showing the effectiveness of our method.

2 Some Notation and Projection Operators

In this paper, we assume that the function f :  ¯ × R → R is a triple continuously dif-

ferentiable function with bounded derivatives to the third order and κ(p)  K(p)−1 is a
square-integrable, symmetric, uniformly positive-definite tensor defined on . Then, there
exists positive constants K∗ and K ∗ , such that

K∗ |y|2 ≤ y T κ(p)y ≤ K ∗ |y|2 , ∀p ∈ L2 ().

Let L2 () be the set of square-integrable functions on  and (L2 ())2 the space of
two-dimensional vectors which have all components in L2 (), with usual norm  · 2 . Fur-
thermore, let (· , ·) denote the L2 inner product, scalar and vector, and (· , ·)∂ present the
L2 (∂) inner product with norm  · ∂ .
We shall m,p
also use αthe pstandard Sobolev space W m() with am,2 norm  · m,p given by
p
φm,p = |α|≤m D φLp () . For p = 2, we define H () = W (),  · m =  · m,2 ,
 ·  =  · 0,2 , and  · 0,∞ =  · L∞ .
Further, for some integer k ≥ 0, we assume that the solution function (p, u) of (1.4)–(1.5)
has the following regularity:

p ∈ L2 (J, W k+2,4 ()) ∩ L2 (J, W k+1,∞ ()), u ∈ (L2 (J, W k+1,4 ()))2 . (2.1)

We also assume uL∞ (J ;L∞ ) ≤ k1 .

To analyze the discretization on a time interval (0, T ), let N > 0, t = T /N, and t n =
n t , and set

φ n − φ n−1
φ n = φ(·, t n ), ∂t φ n = ,
t
 N  12

φl 2 ((0,T );X) = tφ n 2X ,
n=1

φl ∞ ((0,T );X) = max φ n X ,

1≤n≤N

 T  12
φL2 ((0,T );X) = φ(·, t)2X dt .
0
386 J Sci Comput (2011) 49:383–401

Let H (div, ) be the space of vectors in (L2 ())2 which have divergence in L2 () with
1
norm uH (div,) ≡ (u2 + ∇ · u2 ) 2 .
Finally, set
V = H (div, ) ∩ {v · ν = 0}, W = L2 ().
Let h denote a quasi-uniform partition of  into rectangles or triangles with the partition
step h. We form V h and Wh , discrete subspace of V and W , using standard mixed finite
element space such as RT [17] spaces of order k, RTk , or Brazzi-Douglas-Marini [2] spaces
of order k, BDMk . In addition, the following inclusion holds for the RT spaces or BDM
spaces
∇ · v h ∈ Wh , ∀v h ∈ V h .
2
Let Qh denote the L projection defined by

(φ, wh ) = (Qh φ, wh ), ∀wh ∈ Wh , (2.2)

for any φ ∈ L2 (), and

(φ, v h ) = (Qh φ, v h ), ∀v h ∈ V h , (2.3)

for any vector valued function φ ∈ (L2 ())2 .

Throughout the following, assume that 1 < q ≤ ∞. The L2 projection operator
has the following stable and approximation properties [2]: for φ ∈ W k+1,q () (or φ ∈
(W k+1,q ())2 ),

Qh φ0,q ≤ Cφ0,q , 2 ≤ q < ∞, (2.4)

φ − Qh φ0,q ≤ Cφr,q h , r
0 ≤ r ≤ k + 1. (2.5)

Associate with the standard mixed finite element spaces, we define the Fortin projection
h : (H 1 ())2 → V h such that for q ∈ H (div, ),

(∇ · h q, wh ) = (∇ · q, wh ), ∀wh ∈ Wh . (2.6)

The following approximation properties [2] hold for projection h:

1
q − h q0,q ≤ Cqr,q hr , < r ≤ k + 1, (2.7)
q
∇ · (q − h q)0,q ≤ C∇ · qr,q hr , 0 ≤ r ≤ k + 1. (2.8)

We also assume that

v h ∞ ≤ Cv h h−1 , ∀v h ∈ V h . (2.9)

Now, we recall the discrete Gronwall’s lemma (see, e.g., [18]):

Lemma 2.1 Assume that kn is a non-negative sequence, and that the sequence φn satisfies
⎧
⎪ ϕ0 ≤ g0 ,
⎪
⎨

n−1 
n−1 (2.10)
⎪
⎪ ≤ + + ks ϕs , n ≥ 1.
⎩ ϕ n g0 p s
s=0 s=0
J Sci Comput (2011) 49:383–401 387

Then ϕn satisfies
⎧
⎪ ϕ1 ≤ g0 (1 + k0 ) + p0 ,
⎪
⎨
n−1 
n−2 n−1 (2.11)
⎪
⎪
⎩ ϕn ≤ g0 (1 + ks ) + ps (1 + kr ) + pn−1 , n ≥ 2.
s=0 s=0 r=s+1

Moreover, if g0 ≥ 0 and pn ≥ 0 for n ≥ 0, it follows

   n−1 

n−1 
ϕn ≤ g0 + ps exp ks , n ≥ 1. (2.12)
s=0 s=0

3 Mixed Finite Element Discretization

Set the following two variables: the pressure p and the flux u = −K(p)∇p and we get that
κ(p)u = −∇p . In this notation, ∇ · u = −∇ · (K(p)∇p).
We suppose that the symmetric, positive definite tensor κ(p) is triple continuously differ-
entiable with bounded derivatives up to the second order on , i.e., there exists M1 , M2 > 0,
such that κp 0,∞ ≤ M1 , κpp 0,∞ ≤ M2 .
Now, we define the weak form of the nonlinear parabolic equations (1.1)–(1.3) as fol-
lows: find (p, u) ∈ W × V such that
 
∂p
, w + (∇ · u, w) = (f (p), w), ∀w ∈ W, (3.1)
∂t
(κ(p)u, v) − (∇ · v, p) = 0, ∀v ∈ V . (3.2)

The full discretization of (3.1)–(3.2) at t = t n can be defined as follows: find (phn , unh ) ∈
Wh × V h such that
 
phn − phn−1
, wh + (∇ · unh , wh ) = (f (phn ), wh ), ∀wh ∈ Wh , (3.3)
t

(κ(phn )unh , v h ) − (∇ · v h , phn ) = 0, ∀v h ∈ V h . (3.4)

At the time t = t n , we rewrite (1.1)–(1.3) as follows:

 
∂p n
, w + (∇ · un , w) = (f (p n ), w), ∀w ∈ W, (3.5)
∂t
(κ(p n )un , v) − (∇ · v, p n ) = 0, ∀v ∈ V . (3.6)

Define an elliptic-mixed finite method projection (Rh p, Rh u) of the solution of the dif-
ferential equations (1.1)–(1.3) into the finite-dimensional space Wh × V h given by

(∇ · (u − Rh u), wh ) + λ(p − Rh p, wh ) = 0, wh ∈ Wh , (3.7)

(κ(p)(u − Rh u), v h ) − (∇ · v h , p − Rh p) + (γ1 (p − Rh p), v h ) = 0, vh ∈ V h , (3.8)
388 J Sci Comput (2011) 49:383–401

where γ1 = ∂κ
∂p
(x, p)u, the constant λ will be assumed to be sufficiently large constant that

(κ(p)ζ, ζ ) + λ(η, η) + (γ1 ζ, η) ≥ C(ζ 2 + η2 ), ∀ζ ∈ V , η ∈ W, (3.9)

where C ≥ 0 is independent of t ∈ J .
Set
d = Qh p − Rh p, α = p − Rh p, β = u − Rh u.
Estimates for α, β and div β are given in [8] and [9] and are presented in the following as
a lemma without proof. A duality lemma from [9] is also present as Lemma 3.1. Here and
below, duality is understood to be with respect to H k () rather than H0k (). If φ ∈ L2 (),
then
(φ, ψ)
φ−k = φ−k,2, = sup , k ≥ 0.
0=ψ∈H k ψk

Assume that a1 ∈ L∞ (J, H k+1 ()). Then the domain  is said to be s + 2 regular if the
Dirichlet problem

−∇ · (a1 ∇φ) + b1 ∇φ + λφ = ψ, x ∈ ,
φ = 0, x ∈ ∂,

is uniquely solvable for ψ ∈ L2 () and if

φk+2 ≤ Cψk ,

for all ψ ∈ H k ().

First, we need to prove the superconvergence property of d. To get this result, we need
to introduce two lemmas with their proofs found in [8] and [9].

Lemma 3.1 Let ζ ∈ H 1 (), f ∈ H 1 and g ∈ L2 (), let z satisfy the relations

(κζ, v) − (∇ · v, z) + (γ1 z, v) = f (v), v ∈V,

(∇ · ζ, w) + (λz, w) = g(w), w ∈ W.

Let the index k of Wh × V h be at least one and let 0 ≤ s ≤ k − 1. If  is s + 2 regular and

h is sufficiently small, then

z−s ≤ C(hs+1 ζ  + hs+2 ∇ · ζ  + f −s−1

+ g−s−2 + hs+1 f  + hs+2 g). (3.10)

If k is non-negative and  is s + 2 regular, then, for h sufficiently small

z−k ≤ Chk+1 (ζ  + ∇ · ζ  + f  + g) + f −k−1 + g−k−2 . (3.11)

Lemma 3.2 For t ∈ J and for h sufficiently small,

p − Rh p ≤ Chk+1 pk+1 , (3.12)

u − Rh u ≤ Ch k+1
uk+1 , (3.13)
∇ · (u − Rh u) ≤ Chk+1 ∇ · uk+1 . (3.14)
J Sci Comput (2011) 49:383–401 389

Using Lemmas 3.1 and 3.2, we derive our estimate in the following.

Lemma 3.3

d ≤ Ch(β + h∇ · β + p − Qh p). (3.15)

Proof From the elliptic mixed finite element projection (3.7), (3.8) and the definition of L2
projection, we get that

(κ(p)β, v h ) − (∇ · v h , d) + (γ1 d, v h ) = (γ1 (Qh p − p), v h ), vh ∈ V h ,

(∇ · β, wh ) − (λd, wh ) = 0, wh ∈ Wh .

Then from the Lemma 3.1, we get that

d ≤ C(hβ + h2 ∇ · β + hQh p − p)

≤ Ch(β + h∇ · β + p − Qh p),

which proves (3.15). 

Remark 3.1 As in the argument above, we can also derive the estimates for the time deriva-
tives of d and α:
(Qh p − Rh p)t and (p − Rh p)t ,
just by differentiating equations (3.7) and (3.8) with respect to time t .

Therefore, combine the results in Lemmas 3.2 and 3.3, we obtain the following lemma.

Lemma 3.4 Let (p, u) ∈ W × V be the solution of the differential problem (3.1)–(3.2) and
let (Rh p, Rh u) ∈ Wh × V h be its elliptic-mixed projection define in (3.7)–(3.8). Then, for
1 ≤ r ≤ k + 1,

Qh p − Rh p + (Qh p − Rh p)t  ≤ Cpr+1 hr+1 , (3.16)

p − Rh p + (p − Rh p)t  ≤ Cpk+1 hk+1 . (3.17)

The estimate for p − Rh p0,∞ can be found in [8], where the error estimates of this
type first proved. Here, we present the result in the following as a lemma without proof.

Lemma 3.5 There exists a positive constant C, independent of h, such that

p − Rh p0,∞ ≤ C(pk+1,∞ + pk+2 )hk+1 . (3.18)

Next, we can also investigate another superconvergence phenomenon between the full
discrete solution and the elliptic-mixed finite method projection.
First, we have to introduce an useful result which plays an important role in the proof of
superconvergence property.

Lemma 3.6 Let g be a piecewise smooth function on the partition Th . If ḡ(p) is the average
of g(p) on each element e of the Th and ∇g0,∞ ≤ K, then

|(g(p)θ, ψ) − (ḡ(p)θ, ψ)| ≤ CKhθ 0 ψ0 . (3.19)

390 J Sci Comput (2011) 49:383–401

Now, let us present one of our main results here.

Lemma 3.7 Let (phn , unh ) ∈ Wh × V h be the solution of the mixed finite element equations
(3.3)–(3.4) and (Rh p n , Rh un ) be the elliptic projection solution for n ≥ 1. If the regularity
assumptions (2.1) hold, time step t ≤ (12 + 2f 1,∞ + M)−1 with M = M12 k12 /K∗ , and
the initial function

ph0 = Rh p 0 , (3.20)

then for 1 ≤ m ≤ N , we have

1
Rh p m − phm  + κ 2 (ph )(Rh u − uh )l 2 ≤ C(hk+2 + t). (3.21)
(0,t m ;L2 )

Proof At the time t = t n , we rewrite (3.7)–(3.8) as follows:

(∇ · (un − Rh un ), wh ) + λ(p n − Rh p n , wh ) = 0, (3.22)

(κ(p )(u − Rh u ), v h ) − (∇ · v h , p − Rh p ) + (γ1 (p − Rh p ), v h ) = 0,
n n n n n n n
(3.23)

subtract (3.3)–(3.4), (3.22)–(3.23) from (3.5)–(3.6), and let ξ n = Rh p n − phn , ηn =

Rh un − unh , we have

(∂t ξ n , wh ) + (∇ · ηn , wh ) = (T1 , wh ) + (T3 , wh ) + (T2 , wh ), (3.24)

(κ(phn )ηn , v h ) − (∇ · v h , ξ n ) = (K1 , v h ) + (K2 , v h ) + (K3 , v h ), (3.25)

where

T1 = f (p n ) − f (phn ),
T2 = ∂t Rh p n − ptn ,
T3 = λ(Qh p n − Rh p n ),
K1 = (κ(phn ) − κ(p n ))(Rh un − un ),
K2 = κ  (p n )un ξ n ,
1
K3 = κ  (p ∗ )un (p n − phn )2 .
2
Let wh = ξ n , v h = ηn in (3.24)–(3.25) respectively, then adding (3.24), (3.25) to get

(∂t ξ n , ξ n ) + (κ(phn )ηn , ηn ) = (T1 + T2 + T3 , ξ n )

+ (K1 + K2 + K3 , ηn ). (3.26)

First, we bound the right-hand side of the relation (3.26). Because of the smoothness of
f (p) and the superconvergence property in Lemma 3.3, we see that

T1 = f (p n ) − f (phn )
= f (p n ) − f (Qh p n ) + f (Qh p n ) − f (Rh p n ) + f (Rh p n ) − f (phn ), (3.27)
J Sci Comput (2011) 49:383–401 391

then, we bound the right hand side of (3.27) as follows:

|(f (Qh p n ) − f (Rh p n ), ξ n )| ≤ f 1,∞ Qh p n − Rh p n ξ n 

≤ Ch2k+4 + ξ n 2 , (3.28)
|(f (Rh p n
) − f (phn ), ξ n )| ≤ f 1,∞ ξ  , n 2
(3.29)
|(f (p n ) − f (Qh p n ), ξ n )| ≤ |(fp (p n )(Qh p n − p n ), ξ n )|
 
 f 2,∞ n 
+  (p − Qh p ) , ξ .
n 2 n 
(3.30)
2

Set θ n = p n − Qh p n . Now, by using Lemma 3.6 with g(p) = fp (p n ), definition of (2.2) and
its approximation property (2.5), we have

|(f (p n ) − f (Qh p n ), ξ n )| ≤ Chf 1,∞ θ n ξ n  + Cf 2,∞ θ n 20,4 ξ n 

≤ Ch2k+4 + ξ n 2 . (3.31)

Hence, from (3.27)–(3.31), we conclude that

|(T1 , ξ n )| ≤ Ch2k+4 + (2 + f 1,∞ )ξ n 2 . (3.32)

Observe that

T2 = ∂t Rh p n − ptn = ∂t (Rh p n − p n ) + ∂t p n − ptn = ∂t p n − ptn − ∂t α n , (3.33)

where α n = p n − Rh p n . It is simple to see

 
 1  tn 
 n 
|(∂t p − pt , ξ )| = 
n n n
(t − t )ptt (·, t)dt, ξ 
n−1
 t t n−1 
  n 
 t 
 n 
≤ ptt (·, t)dt, ξ 
 t n−1 
 tn
≤ t ptt (·, t)2 dt + ξ n 2 , (3.34)
t n−1

and

|(∂t α n , ξ n )| = |(∂t θ n , ξ n ) + (∂t d n , ξ n )|

= |(∂t d n , ξ n )| ≤ ∂t d n ξ n 
  tn 
 
 
≤ ( t)−1 dt (·, s)ds  ξ n 
 t n−1 
 tn
≤ ( t)−1 dt (·, s)dsξ n 
t n−1

≤ Ch2k+4 + ξ n 2 . (3.35)
392 J Sci Comput (2011) 49:383–401

Then, from (3.33)–(3.35), we derive

  
tn
|(T2 , ξ n )| ≤ C t ptt (·, s)2 ds + h2k+4 + 2ξ n 2 . (3.36)
t n−1

T3 is bounded directly from (3.16) as:

|(T3 , ξ n )| ≤ Ch2k+4 + ξ n 2 . (3.37)

A direct calculation of K1 shows that

|(K1 , ηn )| = |((κ(phn ) − κ(p n ))(Rh un − un ), ηn )|

= |(κ(phn ) − κ(Rh p n ) + κ(Rh p n ) − κ(p n ))(Rh u − u), ηn )|
= |((κ(phn ) − κ(Rh p n ))(Rh un − un ), ηn )|
+ |((κ(Rh p n ) − κ(p n ))(Rh un − un ), ηn )|
≤ Cξ n 0,∞ β n ηn  + Cα n 0,∞ β n ηn 
≤ Ch−1 ξ n β n ηn  + Cα n 0,∞ β n ηn . (3.38)

In order to get the superconvergence of (3.38), we need to introduce a lemma with its
proof to be followed later.

Lemma 3.8 Let

S(h) = sup ξ(t)h−1 .

0≤t≤T

Then

S(h) = O(h),

and consequently

S(h) → 0 as h → 0.

Proof To be followed.
The lemma is consistently with our goal to show that

ξ n  ≤ S(h)h = O(hk+2 ),

using this result, we bound (3.38) as follows:

|(K1 , ηn )| ≤ Ch2 β n 2 + Cα n 20,∞ β n 2 + 1 ηn 2

≤ Ch2k+4 + 1 ηn 2 . (3.39)

For K2 , K3 , we have

M12 k12 n 2 2 n 2
|(K2 , ηn )| = |(κ  (p n )un ξ n , ηn )| ≤ ξ  + η  , (3.40)
22 2
J Sci Comput (2011) 49:383–401 393
 
 1 
|(K3 , ηn )| =  κ  (p ∗ )un (p n − phn )2 , ηn 
2
= |(κ̃un (α n )2 , ηn ) + (2κ̃un (α n )(ξ n ), ηn ) + (κ̃un (ξ n )2 , ηn )|
≤ C(α n 0,∞ α n ηn  + α n 0,∞ ξ n ηn  + ξ n 0,∞ ξ n ηn )
≤ C(α n 0,∞ α n ηn  + α n 0,∞ ξ n ηn  + h2 ξ n ηn )
≤ Ch4k+4 + 1 η2 . (3.41)

Associate to the left hand side of (3.26), we notice that

1
|(∂t ξ n , ξ n )| ≥ (ξ n 2 − ξ n−1 2 ), (3.42)
2 t

and
1
|(κ(phn )ηn , ηn )| = κ 2 (phn )ηn 2 . (3.43)

Then, from the relations (3.32), (3.36), (3.37), (3.39)–(3.43), and choose 1 sufficiently
small, we easily get
 
1 1 M 2 k2
(ξ n 2 − ξ n−1 2 ) + κ 2 (phn )ηn 2 ≤ Ch2k+4 + 6 + f 1,∞ + 1 1 ξ n 2
2 t 22
 tn
2
+ t ptt (·, s)2 ds + ηn .
t n−1 2

Multiply 2 t on both side of the inequality and sum over from n = 0 to m to get


m
1
ξ m 2 − ξ 0 2 + 2 tκ 2 (phn )ηn 2
n=0


m  
M 2k2
≤ Ch2k+4 + 2 t 6 + f 1,∞ + 1 1 ξ n 2
n=0
22


m  tn 
m
+2 ( t)2 ptt (·, s)2 ds + 2 tηn 2 . (3.44)
n=0 t n−1 n=0

M12 k12 −1
Choose 2 = K∗ , time step t satisfying t ≤ (12 + 2f 1,∞ + K∗
) and that
Rh p 0 = p 0 , then, use the discrete Gronwall’s lemma to get
1
ξ m  + κ 2 (ph )ηl 2 ≤ C(hk+2 + t). (3.45)
(0,t m ;L2 )

This theorem follows if we prove the Lemma 3.8.

Proof of Lemma 3.8 For t = 0, we have h−1 ξ(0) = 0 < . For some  > 0, assume that
∗
t,h = inf{t : h−1 ξ(t) ≥ } ≤ T .
It follows from the similar argument above that there exists a constant C such that
∗
ξ(t) ≤ C hk+2 on [0, t,h ]. Thus, for h sufficiently small, h−1 ξ(t) ≤ C hk+1 ≤ 12  on
394 J Sci Comput (2011) 49:383–401

∗ ∗
[0, t,h ], then, as a consequence, no such t,h can exist. Since h < 1, then hk+1 ≤ h for k ≥ 0,
therefore, S(h) → 0 as h → 0. For k = 0, S(h) = O(h), for k ≥ 1, S(h) = o(h), so, we can
conclude that k ≥ 0, S(h) = O(h). 

From Lemma 3.7, it is easy to get the following theorem.

Theorem 3.1 Let (phn , unh ) ∈ Wh × V h be the solution of the mixed finite element equation
(3.3)–(3.4), for n ≥ 1. If the regularity assumption (2.1) holds,

t ≤ (12 + 2f 1,∞ + M)−1 and ph0 = Rh p0 , (3.46)

with M = M12 k12 /K∗ , then for 1 ≤ m ≤ N

1
p m − phm  + κ 2 (ph )(u − uh )l 2 ≤ C( t + hk+1 ). (3.47)
(0,t m ;L2 )

4 The Two-Grid Algorithm and Error Analysis

In this section, we shall present the main algorithm of the paper. The fundamental ingredient
of the algorithm is another mixed finite element space V H × WH (⊂ V h × Wh ) defined
on a coarser quasi-uniform triangulations or rectangulations of . The algorithm has two
steps:
Step 1 On the coarse grid TH , solve the following nonlinear system for (pH
n
, unH ) ∈
WH × V H :
n
(∂t pH , wH ) + (∇ · unH , wH ) = (f (pH
n
), wH ), wH ∈ WH , (4.1)
n
(κ(pH )unH , v H ) − (pH
n
, ∇ · v H ) = 0, vH ∈ V H . (4.2)

Step 2 On the fine grid Th , compute (Phn , U nh ) ∈ Wh × V h to satisfy the following linear
system:

(∂t Phn , wh ) + (∇ · U nh , wh ) = (f (pH

n
) + f  (pH
n
)(Phn − pH
n
), wh ), wh ∈ Wh ,
(4.3)
(κ  (pH
n
)unH (Phn − pH
n
) + κ(pH
n
)U nh , v h ) = (Phn , ∇ · v h ), vh ∈ V h . (4.4)

In the following analysis, we need the estimate of p n − pH

n
0,q .

0
Lemma 4.1 If we choose pH = RH p 0 , the regularity assumptions (2.1) hold, then for 1 ≤
n ≤ N, 2 ≤ q ≤ ∞ and t ≤ (12 + 2f 1,∞ + M)−1 with M = M12 k12 /K∗ , we have

p n − pH
n
0,q ≤ C( t + H k+1 ).

Proof Using approximation property (2.5), Lemmas 3.5, 3.7 and the inverse inequality,

p n − pH
n
0,q ≤ p n − RH p n 0,q + RH p n − pH
n
0,q
≤ C{p n k+1,q H k+1 + H 2(1/q−1/2) RH p n − pH
n
}
J Sci Comput (2011) 49:383–401 395

≤ C{H k+1 + H 2/q−1 (H k+2 + t)}

≤ C(H k+1 + t),

which yields the lemma. 

Theorem 4.1 Let (Phn , U nh ) ∈ Wh × V h be the solution of the two-grid method of (4.3)–
(4.4) for solving the mixed finite element scheme (3.1)–(3.2). If the regularity assumptions
(2.1) hold, time step satisfying t ≤ (4 + 2f 1,∞ + M)−1 , with M = M12 k12 /K∗ and initial
function
Ph0 = Qh p 0 , 0
pH = RH p 0 ,
then, for 1 ≤ m ≤ N , we have
1
p m − Phm  + κ 2 (pH )(u − U h )l 2 ≤ C( t + hk+1 + H 2k+2 ). (4.5)
(0,t m ;L2 )

Proof Subtract equations (4.3)–(4.4) from (3.5)–(3.6) and use the projection definition (2.2),
(2.6), let ρ n = Qh p n − Phn , σ n = h un − U nh , then, we have:

(∂t ρ n , wh ) + (∇ · σ n , wh ) = (∂t p n − ptn , wh ) + (F1 , wh ), (4.6)

(F2 , v h ) − (ρ , ∇ · v h ) = 0,
n
(4.7)

where

F1 = f (p n ) − f (pH
n
) − f  (pH
n
)(Phn − pH
n
),
F2 = κ(p n )un − κ  (pH
n
)unH (Phn − pH
n
) + κ(pH
n
)U nh .

Note that

F2 = κ(p n )un − κ  (pH

n
)unH (Phn − pH
n
) − κ(pH
n
)U nh
= −κ(p n )( hu
n
− un ) + κ(pH
n
)( hu
n
− U nh )
+ κ  (pH
n
)(ρ n + (p n − Qh p n ))unH + κ  (pH
n
)(p n − pH
n
)( hu
n
− unH )
1
+ κ  (p ∗ )(p n − pH
n 2
) hu
n
.
2
Then, we can rewrite (4.6)–(4.7) as follows:

(∂t ρ n , wh ) + (∇ · σ n , wh ) = (F1 , wh ) + (∂t p n − ptn , wh ), (4.8)

n
(κ(pH )σ n , v h ) − (∇ · v h , ρ ) = (F3 , v h ),
n
(4.9)

where

F3 = κ(p n )( hu
n
− un ) − κ  (pH
n
)(ρ n + (p n − Qh p n ))unH
1
− κ  (pH
n
)(p n − pH
n
)( hu
n
− unH ) − κ  (p ∗ )(p n − pH
n 2
) hu
n
.
2
Let wh = ρ n , v h = σ n in (4.8), (4.9) respectively, then add (4.8), (4.9) to get

(∂t ρ n , ρ n ) + (κ(pH
n
)σ n , σ n ) = (∂t p n − ptn + F1 , ρ n ) + (F3 , σ n ). (4.10)
396 J Sci Comput (2011) 49:383–401

First, consider F1 on the right hand side of (4.10). Use the Taylor expansion to get
1
f (p n ) = f (pH
n
) + f  (pH
n
)(p n − pH
n
) + f  (p ∗ )(p n − pH
n
),
2
where p ∗ is some value between pH
n
and p n . Replace it in F1 , we can estimate F1 as follows:

1
|(F1 , ρ n )| = |(f  (pH
n
)(p n − Phn ), ρ n ) + (f  (p ∗ )(p n − pH
n 2
) , ρ n )|
2
≤ f 1,∞ p n − Phn ρ n  + Cf 2,∞ p n − pH  + ρ2
n 4

≤ f 1,∞ p n − Qh p n + Qh p n − Phn ρ n  + CH 4k+4 + ρ n 2

≤ C(h2k+2 + H 4k+4 ) + (2 + f 1,∞ )ρ n 2 . (4.11)

As to each term of F3 , we have following results:

ϑ1 n 2
|(κ(p n )( hu
n
− un ), σ n )| ≤ K ∗  hu
n
− un  2 + σ 
2
ϑ1 n 2
≤ Ch2k+2 + σ  , (4.12)
2
|(κ  (pH
n
)(ρ n + (p n − Qh p n ))unH , σ n )| ≤ C|(ρ n , σ n )| + C|((p n − Qh p n )unH , σ n )|
≤ Cρ n σ n  + CunH L∞ p n − Qh p n σ n 
δ ϑ
≤ ρ n 2 + σ n 2 + Ch2k+2 , (4.13)
2ϑ 2
where δ  = M1 2 unH 2∞ ≤ M12 k12 .
From Lemma 4.1, we know that

p n − pH
n
0,∞ ≤ C( t + H k+1 ),

then, we have the following estimate:

(κ  (pH
n
)(p n − pH
n
)( hu
n
− unH ), σ n ) ≤ M1 p n − pH
n
0,∞  hu
n
− un σ n 
+ M1 p n − pH
n
0,∞ un − unH σ n 
≤ Ch2k+2 H 2k+2 + ϑ1 σ n 2 + C(H 4k+4 + ( t)2 )
≤ C(H 4k+4 + ( t)2 ) + ϑ1 σ n 2 , (4.14)
(κ  (p ∗ )(p n − pH
n 2
) hu
n n 2 
, σ n ) = ((p n − pH ) κ (p ∗ )( hu
n
− un + un − unH ), σ n )
n 2 
+ ((p n − pH ) κ (p ∗ )(un − un + unH ), σ n ).

Use the Höder’s inequality to derive

n 2 
|((p n − pH ) κ (p ∗ )(un − un + unH ), σ n )|
ϑ1 n 2
≤ M22 unH 2L∞ p n − pH 0,4 +
n 4
σ 
2
ϑ1 n 2 ϑ1
≤ Cp n − pH 0,4 +
n 4
σ  ≤ CH 4k+4 + σ n 2 , (4.15)
2 2
J Sci Comput (2011) 49:383–401 397

n 2 
|((p n − pH ) κ (p ∗ )( hu
n
− un ), σ n )|
≤ M2 (p n − pH
n 2
)( hu
n
− un )σ n 
ϑ1 n 2
≤ CH −2 p n − pH 
n 4
hu
n
− un 2∞ + σ 
2
ϑ1
≤ h2k H 4k+2 + σ 2 , (4.16)
2
n 2 
|((p n − pH ) κ (p ∗ )(un − unH ), σ n )|
≤ M2 (p n − pH ) (u − unH )σ n 
n 2 n

ϑ1 n 2
≤ M2 H −2 p n − pH 0,4 H −2 un − unH 2 +
n 4
σ 
2
ϑ1 n 2
≤ CH 6k+2 + σ  . (4.17)
2
From (4.12)–(4.17), we can bound F3 as follows:

ϑ n 2 δ
|(F3 , σ n )| ≤ C(h2k+2 + H 4k+4 ) + σ  + ρ n 2 + 3ϑ1 σ n 2 . (4.18)
2 2ϑ
Use the same method of (3.42)–(3.43) for the left hand side, combine the estimations (3.34),
(4.11), (4.18) and choose ϑ1 sufficiently small, then we have

1 1
(ρ n 2 − ρ n−1 2 ) + K 2 (pH
n
)σ n 2
2 t
 tn  
δ
≤ t ptt (·, t) dt +
2
+ 2 + f 1,∞ ρ n 2
t n−1 2ϑ
+ C(h2k+2 + H 4k+4 ).

Multiply 2 t on both side of this inequality and sum from n = 1 to m. Notice that ρ 0 = 0,
M12 k12
then, choose ϑ = K∗ , t ≤ (4 + M + 2f ∞ )−1 with M = K∗
, and use the discrete
Gronwall’s lemma, we deduce
1
ρ m  + κ 2 (pH )σ l 2 ≤ C( t + hk+1 + H 2k+2 ).
(0,t m ;L2 )

The theorem can be derived with simple use of the triangle inequality. 

5 Numerical Example

In this section, we will demonstrate the efficiency of our algorithm proposed in Sect. 4 with
a numerical example. Here, we consider the following initial-boundary valued problem of
the reaction-diffusion type:

∂p
− ∇ · (K(p)∇p) = f (p), ∀(x, t) ∈  × J, (5.1)
∂t
p(x, 0) = 0, ∀(x, t) ∈  × {t = 0}, (5.2)
398 J Sci Comput (2011) 49:383–401

Fig. 1 Exact solution at t = 1

Fig. 2 The mixed finite element solution at t = 1

K(p)∇p · ν = 0, ∀(x, t) ∈ ∂ × J, (5.3)

 exp(p) 0 
where  = [0, 1]2 , J = [0, 1], K(p) = 0 exp(p)
, f (p) = p 3 + g(x, t) and g(x, t) is so
chosen that p(x, t) = t 2 (x1 (x1 − 1)2 + x2 (x2 − 1)2 ) is the exact solution.
2 2

The domain  is uniformly divided by the triangulations of mesh size H and h, respec-
tively. V h is the Raviart-Thomas space with index k = 0. J is also uniformly divided so that
1
t is a constant. We choose H = 14 , h = 16 1
to satisfy H = h 2 . We also choose t = H 2
to satisfy the conditions presented in Theorems 3.1 and 4.1 associated with this example.
We demonstrate our results with a Table 1 and Figs. 1, 2, 3 and 4. The mixed finite element
solutions ph , uh are computed directly and the convergence results p n − phn L2 () at dif-
ferent time level are listed in the second column of Table 1. The two-grid solutions Phn , U nh
are computed by the algorithm discussed in Sect. 4 and the approximate results are listed in
J Sci Comput (2011) 49:383–401 399

Fig. 3 The two-grid solution at t = 1

Fig. 4 Error of u − uh V and u − U h V at different time levels

the fourth column of the table. We also compare the computing time that used to solve the
nonlinear reaction-diffusion equations by the MFEM and the two-grid method in the table.
From the numerical results represented in the figures and the table above, we find that
the two-grid algorithm achieves the same proximation for nonlinear equations as the mixed
1
finite element method with the mesh sizes satisfying H = O(h 2 ). Therefore, we verify the
theoretical results in Sect. 4 numerically.

6 Conclusion

In this paper, we present a two-grid algorithm for nonlinear reaction-diffusion equations

discretized by the mixed finite element methods. The key ingredient of the two-grid method
400 J Sci Comput (2011) 49:383–401

Table 1 Error of MFEM (pn − phn ), Two-Grid method (pn − Phn )

Time level pn − phn L2 Computing time pn − Phn L2 Computing time
(MFEM) (Two-Grid)

1 2.629963e−004 4.625 s 2.629964e−004 2.469 s

2 5.247857e−004 3.469 s 5.247862e−004 2.078 s
3 7.893769e−004 4.609 s 7.893789e−004 2.110 s
4 1.058404e−003 4.500 s 1.058409e−003 2.188 s
5 1.333209e−003 4.500 s 1.333223e−003 2.140 s
6 1.615090e−003 4.500 s 1.615119e−003 2.157 s
7 1.905305e−003 4.500 s 1.905362e−003 2.140 s
8 2.205070e−003 4.500 s 2.205172e−003 2.140 s
9 2.515550e−003 5.625 s 2.515719e−003 2.141 s
10 2.837853e−003 5.625 s 2.838116e−003 2.156 s
11 3.173032e−003 5.625 s 3.173421e−003 2.141 s
12 3.522085e−003 5.625 s 3.522635e−003 2.219 s
13 3.885961e−003 5.625 s 3.886706e−003 2.203 s
14 4.265561e−003 5.641 s 4.266533e−003 2.203 s
15 4.661744e−003 5.625 s 4.662965e−003 2.203 s
16 5.075335e−003 6.860 s 5.076815e−003 2.219 s

in this paper is that we use one Newton iteration on the fine grid. We show that when the
1
coarse grid and the fine grid satisfy H = O(h 2 ), the two-grid algorithm can achieve the
same accuracy of the mixed finite element solution. Generally speaking, different aspects
of a complex problem can be treated by spaces of different scales. For the problem we
studied in this paper, a very coarse grid space is sufficient for the nonlinear problem that
is dominated by its linear part. The two-grid method studied in this paper provides a new
approach to take the advantage of some nice properties hidden in a complex problem. In our
future work, we will consider the more complicated two-grid algorithms for (1.1)–(1.3) and
give some numerical experiments for these algorithms.

Acknowledgements This work is supported by the Foundation for Talent Introduction of Guangdong
Provincial University, Guangdong Province Universities and Colleges Pearl River Scholar Funded Scheme
(2008), National Science Foundation of China (10971074).

References

1. Arbogast, T., Wheeller, M.F., Yotov, I.: Mixed finite elements for elliptic problems with tensor coeffi-
cients as cell-centered finite differences. SIAM J. Numer. Anal. 32, 828–852 (1997)
2. Brezzi, F., Douglas, J. Jr., Marini, L.D.: Two families of mixed finite elements for second order elliptic
problems. Numer. Math. 47, 217–235 (1985)
3. Chen, Y., Huang, Y.: A multilevel method for finite element solutions for singular two-point boundary
value problems. Nat. Sci. J. Xiangtan Univ., 16, 23–26 (1994) (in Chinese)
4. Chen, Y., Li, L.: Lp error estimates of two-grid schemes of expanded mixed finite element methods.
Appl. Math. Comput. 209, 197–205 (2009)
5. Chen, Y., Huang, Y., Yu, D.: A two-grid method for expanded mixed finite-element solution of semilinear
reaction-diffusion equations. Int. J. Numer. Methods Eng. 57, 193–209 (2003)
6. Chen, Y., Liu, H., Liu, S.: Analysis of two-grid methods for reaction-diffusion equations by expanded
mixed finite element methods. Int. J. Numer. Methods Eng. 69, 408–422 (2007)
J Sci Comput (2011) 49:383–401 401

7. Chen, Y., Luan, P., Lu, Z.: Analysis of two-grid methods for nonlinear parabolic equations by expanded
mixed finite element methods. Adv. Appl. Math. Mech. 1, 1–15 (2009)
8. Douglas, J. Jr., Roberts, J.E.: Mixed finite element methods for second order elliptic problems. Mat.
Appl. Comput. 1, 91–103 (1982)
9. Douglas, J. Jr., Roberts, J.E.: Global estimates for mixed finite element methods for second order elliptic
equations. Math. Comput. 44, 39–52 (1985)
10. Douglas, J. Jr., Ewing, R.E., Wheeler, M.F.: The approximation of the pressure by a mixed method in
the simulation of miscible displacement. RAIRO. Anal. Numèr. 17, 17–33 (1983)
11. Dawson, C.N., Wheeler, M.F.: Two-grid method for mixed finite element approximations of non-linear
parabolic equations. Contemp. Math. 180, 191–203 (1994)
12. Dawson, C.N., Wheeler, M.F., Woodward, C.S.: A two-grid finite difference scheme for nonlinear par-
abolic equations. SIAM J. Numer. Anal. 35, 435–452 (1998)
13. Garcia, M.F.: Improved error estimates for mixed finite element approximations for nonlinear parabolic
equations: the continuously-time case. Numer. Methods Partial Differ. Equ., 10, 129–149 (1994)
14. Huyakorn, P.S., Pinder, G.F.: Computational Methods in Surface Flow. Academic Press, New York
(1983)
15. Murray, J.: Mathematical Biology, 2nd edn. Springer, New York (1993)
16. Squeff, M.C.J.: Superconvergence of mixed finite elements for parabolic equations. Math. Model. Nu-
mer. Anal. 21, 327–352 (1987)
17. Raviart, P.A., Thomas, J.M.: A mixed finite element method for 2nd order elliptic problems. In: Math.
Aspects of the Finite Element Method. Lecture Notes in Math., vol. 606, pp. 292–315. Springer, Berlin
(1977)
18. Thomée, V.: Galerkin Finite Element Methods for Parabolic Problems. Springer, Berlin (1997)
19. Wu, L., Allen, M.B.: A two-grid method for mixed finite element solution of reaction-diffusion equa-
tions. Numer. Methods Partial Differ. Equ. 15, 317–332 (1999)
20. Xu, J.: A novel two-grid method for semilinear equations. SIAM J. Sci. Comput. 15, 231–237 (1994)
21. Xu, J.: Two-grid discretization techniques for linear and non-linear PDEs. SIAM J. Numer. Anal. 33,
1759–1777 (1996)
22. Zhu, J., Zhang, Y.T., Stuart, A.N., Mark, A.: Application of discontinuous Galerkin methods for reaction
diffusion systems in developmental biology. J. Sci. Comput. 40, 391–418 (2009)