CANADIAN APPLIED

MATHEMATICS QUARTERLY
Volume 8, Number 1, Spring 2000

ON A MIXED FINITE ELEMENT METHOD

FOR THE p-LAPLACIAN

MOHAMED FARHLOUL AND HASSAN MANOUZI

ABSTRACT. In this article we propose a mixed method for

the p-Laplacian problem. We improve the result obtained in
[6] by showing that the error estimates obtained, for a mixed
finite element method, are valid in three dimensions without
restrictions. The key of this improvement is the use of a more
appropriate functional setting. This allows also to improve
the previously known error estimates.

1. Introduction. Let Ω be a bounded open set of RN , N = 2 or 3,

with a Lipschitz boundary Γ. The p-Laplace equation is the quasilinear
(degenerate) partial differential equation:


−∇ · (|∇u|p−2 ∇u) = f in Ω,
(1.1)
u = 0 on Γ,

where 1 < p < ∞, ∇u is the vector gradient of u, and

N 
 2 1/2
∂u
|∇u| = .
i=1
∂xi

This type of equation appears in many mathematical models of

physical processes: nonlinear diffusion and filtration, see Philip [14];
power-law materials, see Atkinson and Champion [1]; and viscoelastic
materials, see Le Tallec [9], for example.
The approximation of (1.1) by finite elements has been studied in sev-
eral works (see Glowinski and Marrocco [8], Ciarlet [5], Tyukhtin [19]
and Barrett and Liu [3]). Further results for more general problems,
including (1.1), are given in Liu and Barrett [10], [11].
In Farhloul [6] we have studied a mixed finite element method for
problem (1.1) in the case 1 < p < 2. This mixed finite element method
is based on the introduction of σ = |∇u|p−2 ∇u as a new unknown and
its approximation by the lowest degree Raviart-Thomas finite element
c
Copyright
2000 Rocky Mountain Mathematics Consortium

67
68 M. FARHLOUL AND H. MANOUZI

(cf. Thomas [18], Raviart and Thomas [15] and Nedelec [13]). We
have proved in [6] error estimates in both two and three dimensions
but in three dimensions we had a restriction on p, 6/5 ≤ p < 2 (see
Theorem 3.2 in [6]).
It is the purpose of this paper to give a proof of error estimates
which are valid in three dimensions without restrictions on p, i.e., for
all p > 1. The basic idea is to replace the functional setting in [6] by
a more appropriate one. More precisely, in the case 1 < p < 2, we
prove an optimal error bound for u, and we improve the one obtained
in Baranger et al. [2] for σ. Furthermore, we extend our analysis to
the case p > 2 by proving the same error bound for u as the one
obtained in Glowinski and Marrocco [8] and improving the one obtained
in Glowinski and Marrocco [8] for σ (see also Ciarlet [5]). We point out
that all these estimates are obtained under minimal regularity of the
solution of problem (1.1) (compare with the results obtained in Barrett
and Liu [3], and Theorem 3.2 in [6]). Finally, for some numerical results
for a nonlinear Stokes model using the mixed finite element proposed
here, we refer to Manouzi and Farhloul [12].

2. Mixed formulation. The mixed formulation is based on the

inversion of the relation σ = |∇u|p−2 ∇u, i.e.,

∇u = |σ|q−2 σ

where (1/p) + (1/q) = 1. With this purpose the problem (1.1) is

decomposed into a first order system as follows
 q−2
 |σ| σ = ∇u in Ω,

(2.1) ∇·σ+f =0 in Ω,


u=0 on Γ.

The weak formulation appropriate for the mixed method is



 A(σ) · τ dx + ∇ · τ u dx = 0 ∀ τ ∈ X,

Ω Ω
(2.2)


 ∇ · σv dx + f v dx = 0 ∀ v ∈ M,
Ω Ω
 MIXED FINITE ELEMENT METHOD FOR p-LAPLACIAN 69

where

(2.3) A(σ) = |σ|q−2 σ,

(2.4) X = {τ ∈ (Lq (Ω))N ; ∇ · τ ∈ Lq (Ω)}, M = Lp (Ω),

and f ∈ Lq (Ω).
We denote the norm of X by τ X = ( τ q0,q,Ω + ∇·τ q0,q,Ω )1/q , where
τ 0,q,Ω = ( Ω |τ |q dx)1/q , and the norm of M by v M = v 0,p,Ω .

Remark 2.1. The difference between the functional setting here and
the one in [6] is that in [6] we have considered ∇ · τ ∈ L2 (Ω) and
M = L2 (Ω).
To prove the existence and the uniqueness of the solution of (2.2),
we shall use the following abstract result (see [7, p. 61, Remark 4.2],
Sandri [17]).

Lemma 2.1. Let (X, · X ) and (M, · M ) be two reflexive Banach

spaces. Let (X  , · X
) and (M  , · M
) be their corresponding
dual space. Let B : X → M  be a linear continuous operator and
B  : M → X  the dual operator of B. Let V = ker (B) be the kernel
of B; we denote by V 0 ⊂ X  the polar set of V , i.e., V 0 = {x ∈
X  , x , v = 0, ∀ v ∈ V } and Ḃ : (X/V ) → M  the quotient operator
associated with B. The three following properties are equivalent:
(i) β > 0 exists such that

Bv, r
inf sup ≥ β,
r∈M v∈X r M v X

(ii) B  is an isomorphism from M onto V 0 and

Br X
≥β r M ∀ r ∈ M,

(iii) Ḃ is an isomorphism from (X/V ) onto M  and

Ḃ v̇ M
≥ β v̇ (X/V ) ∀ v̇ ∈ (X/V ).
70 M. FARHLOUL AND H. MANOUZI

Now let us prove the following result concerning the inf-sup condition.

Proposition 2.1. There exists a positive constant β such that

Ω
∇ · τ vdx
(2.5) inf sup ≥ β.
v∈M τ ∈X v M τ X

Proof. Let v ∈ M and w denote the solution of the p-Laplace problem

−∇ · (|∇w|p−2 ∇w) = |v|p−2 v in Ω,
w=0 on Γ.

From Ciarlet [5] and Glowinski and Marrocco [8], we know that this
problem has a unique solution w ∈ W01,p (Ω) and there exists a constant
C such that
w 1,p,Ω ≤ C v 0,p,Ω ,
where
p p 1/p
w 1,p,Ω =( w 0,p,Ω + ∇w 0,p,Ω ) .

Therefore, τ ∗ = −|∇w|p−2 ∇w is in the space X and there exists a

constant C such that
q q ∗ q
τ∗ X = τ∗ 0,q,Ω + ∇ · τ 0,q,Ω
= ∇w p0,p,Ω + v p0,p,Ω
≤ C v p0,p,Ω .

p−1 p
Hence τ ∗ X ≤ C v 0,p,Ω , and Ω
∇ · τ ∗ v dx = v 0,p,Ω . Thus, (2.5)
follows immediately.

As a consequence of Lemma 2.1 and Proposition 2.1, we have the

following result.

Theorem 2.1. Problem (2.2) has a unique solution (σ, u) ∈ X × M

and there exists a constant C > 0 such that
q−1
σ X + u M ≤ C( f 0,q,Ω + f 0,q,Ω ).
 MIXED FINITE ELEMENT METHOD FOR p-LAPLACIAN 71

Proof. The proof of this result is the same as the one of Theorem 2.1
in [6].

3. Discrete problem and error analysis. Let Th be a triangula-

tion of Ω into closed triangles (in the case N = 2) or tetrahedrons (in
the case N = 3). The elements of Th are assumed to be regular in the
usual sense (cf. Ciarlet [5]). We define the finite element spaces

(3.1) Xh = {τh ∈ X; τh |K ∈ RT0 (K), ∀ K ∈ Th },

(3.2) Mh = {vh ∈ M ; vh |K ∈ P0 (K), ∀ K ∈ Th },

where P0 (K) denotes the space of constants and

RT0 (K) = (P0 (K))N + xP0 (K), x = (x1 , . . . , xN ).

The finite element space Xh is nothing else than the approximation

of the space X by the lowest degree element of Raviart-Thomas (cf.
Raviart and Thomas [15], Nedelec [13], Brezzi and Fortin [4], Roberts
and Thomas [16]).
Then the finite element approximation to problem (2.2) is given by
(σh , uh ) ∈ Xh × Mh satisfying


 A(σh ) · τh dx + ∇ · τh uh dx = 0 ∀ τh ∈ Xh ,

Ω Ω
(3.3)


 ∇ · σh vh dx + f vh dx = 0 ∀ vh ∈ Mh .
Ω Ω

We introduce the Raviart-Thomas projection (cf. [13], [15]), Πh

defined from X ∩ (W 1,s (Ω))N , s > 1, onto Xh by

(3.4) (τ − ΠK τ ) · n ds = 0 ∀ e ∈ ∂K, ∀ K ∈ Th
e

where ΠK τ = Πh τ |K and n denotes the unit outward normal to the

boundary ∂K of K.
The operator Πh is well defined for all τ ∈ (W 1,s (Ω))N , s > 1.
Indeed, for every K ∈ Th , τ |K ∈ (W 1,s (K))N and thus τ |∂K ∈
72 M. FARHLOUL AND H. MANOUZI

(W 1−1/s,s (∂K))N ⊂ (Ls (∂K))N ⊂ (L1 (∂K))N . Therefore, e

τ · n ds
has sense for every e ⊂ ∂K.
Moreover (cf. [4], [6], [13], [15]),

(3.5)
∇ · (Πh τ −τ )vh dx = 0, ∀ τ ∈ X ∩ (W 1,s (Ω))N , ∀ vh ∈ Mh ,
Ω
(3.6) τ − Πh τ 0,s,Ω ≤ Ch|τ |1,s,Ω , ∀ τ ∈ (W 1,s (Ω))N , s > 1.

Remark 3.1. In the case q > 2, i.e., 1 < p < 2, the operator Πh is
well defined for all τ ∈ X (cf. Brezzi and Fortin [4]) and (3.5) is valid
for all τ ∈ X.
Before giving the error estimates for problem (3.3), let us first prove
the following result.

Proposition 3.1. There exists a constant C independent of h such

that

Ω
∇ · τh vh dx
(3.7) inf sup ≥ C.
vh ∈Mh τh ∈Xh vh M τh X

Proof. Let vh ∈ Mh and τ ∗ = −|∇w|p−2 ∇w where w is the solution

of the p-Laplace problem


−∇ · (|∇w|p−2 ∇w) = |vh |p−2 vh in Ω,
w=0 on Γ.

This problem has a unique solution w ∈ W01,p (Ω), and there exists a
constant C > 0 such that w 1,p,Ω ≤ C vh 0,p,Ω . Thus, τ ∗ ∈ X and
τ ∗ X ≤ C vh p−1
0,p,Ω . Now, let τ̃ = ∇w̃, where w̃ is the solution of the
Dirichlet problem

∆w̃ = ∇ · τ ∗ in Ω,
w̃ = 0 on Γ.
Since ∇ · τ ∗ ∈ Lq (Ω), we have τ̃ ∈ (W 1,q (Ω))N , there exists a constant
C such that τ̃ 1,q,Ω ≤ C ∇ · τ ∗ 0,q,Ω , and ∇ · τ̃ = ∇ · τ ∗ in Ω.
 MIXED FINITE ELEMENT METHOD FOR p-LAPLACIAN 73

Finally, let τ̃h = Πh τ̃ . Due to (3.5) and (3.6), we have

p
∇ · τ̃h vh dx = ∇ · τ̃ vh dx = ∇ · τ ∗ vh dx = vh 0,p,Ω
Ω Ω Ω

and
p−1
τ̃h X ≤ C( τ̃ 1,q,Ω + ∇ · τ̃ 0,q,Ω ) ≤ C ∇ · τ∗ 0,q,Ω ≤ C vh 0,p,Ω .

Therefore, (3.7) follows immediately.

Due to (3.7), the existence and uniqueness of the solution of problem

(3.3) are obtained as in Theorem 2.1.
We are now in a position to state the error estimates for problem
(3.3). We begin by the case 2 ≤ p < ∞.

Theorem 3.1. Suppose that 2 ≤ p < ∞. Let (σ, u) be the solution

of (2.2) and (σh , uh ) the solution of (3.3). If (σ, u) ∈ (W 1,q (Ω))N ×
W 1,p (Ω), then there exists a constant C, independent of h, such that
(3.8) σ − σh 0,q,Ω ≤ Chq/2 ,
(3.9) u − uh 0,p,Ω ≤ Chq−1 ,
where (1/p) + (1/q) = 1.

Proof. To prove the error estimate (3.8), first observe that, from the
second equation of (2.2) and of (3.3), respectively, and (3.5),

∇ · (Πh σ − σh )vh dx = 0 ∀ vh ∈ Mh .
Ω

On the other hand, by using the first equation of (2.2) and (3.3),
respectively, and this last relation, we get

(A(σ) − A(σh )) · (σh − Πh σ) dx = 0,

Ω

and then

(3.10) (A(σ) − A(σh )) · (σ − σh ) dx

Ω

= (A(σ) − A(σh )) · (σ − Πh σ) dx.

Ω
74 M. FARHLOUL AND H. MANOUZI

Now, using the fact that (cf. Sandri [17])

2
σ − σh 0,q,Ω
(A(σ) − A(σh )) · (σ − σh ) dx ≥ C 2−q
Ω σ 0,q,Ω + σh 2−q
0,q,Ω

+C |A(σ) − A(σh )||σ − σh | dx,

Ω

(A(σ) − A(σh )) · (σ − Πh σ) dx
Ω
1/p
≤C |A(σ) − A(σh )||σ − σh | dx σ − Πh σ 0,q,Ω ,
Ω

the inequality of Young: for ε > 0, a ≥ 0 and b ≥ 0,

 q
1 1 b
ab ≤ (εa)p + ,
p q ε

and (3.10), we get

2
σ − σh 0,q,Ω
(3.11) 2−q + |A(σ) − A(σh )||σ − σh | dx
σ 0,q,Ω + σh 2−q
0,q,Ω Ω
q
≤ C σ − Πh σ 0,q,Ω .

Thus,
2 2−q 2−q q
σ − σh 0,q,Ω ≤ C( σ 0,q,Ω + σh 0,q,Ω ) σ − Πh σ 0,q,Ω .

This last estimate, with (3.6), gives us

q/2
σ − σh 0,q,Ω ≤ Chq/2 |σ|1,q,Ω ,

which is the desired result.

It remains to prove the error estimate (3.9). First, observe that, from
the first equation of (2.2) and (3.3), respectively,

∇ · τh (Ph u − uh ) dx = (A(σh ) − A(σ)) · τh dx ∀ τh ∈ Xh ,

Ω Ω
 MIXED FINITE ELEMENT METHOD FOR p-LAPLACIAN 75

where Ph u is the projection of u onto ΠK∈Th P0 (K).

Thus, using this last relation and the inf-sup condition (3.7), we get

Ω
∇ · τh (Ph u − uh )dx
C P h u − uh 0,p,Ω ≤ sup
τh ∈Xh τh X
Ω
(A(σh ) − A(σ)) · τh dx
≤ sup .
τh ∈Xh τh X

Therefore,
1/p
P h u − uh 0,p,Ω ≤C |A(σ) − A(σh )||σ − σh | dx
Ω

due to the fact that (cf. Sandri [17]) for all τh ∈ Xh ,

(A(σ) − A(σh )) · τh dx
Ω
1/p
≤C |A(σ) − A(σh )||σ − σh | dx τh 0,q,Ω .
Ω

On the other hand, by using (3.11) and this last estimate, we get

q/p
P h u − uh 0,p,Ω ≤ C σ − Πh σ 0,q,Ω .

Finally, error estimate (3.9) is a consequence of this last one, (3.6), the
triangle inequality and the following standard error estimation

u − Ph u 0,p,Ω ≤ Ch|u|1,p,Ω .

Remark 3.2. In the proof of Theorem 3.1, we have used the fact that
σh 0,q,Ω ≤ C with C a constant independent of h. The proof of this
bound is analogous to the one for σ in Theorem 2.1, due to the inf-sup
condition (3.7).
The error estimation (3.9) is the same as the one obtained in Glowin-
ski and Marrocco [8], while the error estimation (3.8) is better than
the one obtained in Glowinski and Marrocco [8].
76 M. FARHLOUL AND H. MANOUZI

In the case 1 < p < 2, we get better error estimates. Furthermore,

these error estimates are valid without restrictions on p, i.e., for all
p ∈ ]1, 2[. This is the main difference between Theorem 3.2 in [6] and
the following one.

Theorem 3.2. Suppose that 1 < p < 2. Let (σ, u) be the solution
of (2.2) and (σh , uh ) the solution of (3.3). If (σ, u) ∈ (W 1,q (Ω))N ×
W 1,p (Ω), then there exists a constant C independent of h such that

(3.12) σ − σh 0,q,Ω ≤ Ch2/q ,

(3.13) u − uh 0,p,Ω ≤ Ch.

Proof. As in the proof of Theorem 3.1, we have

(A(σ) − A(σh )) · (σ − σh ) dx = (A(σ) − A(σh )) · (σ − Πh σ) dx.

Ω Ω

Now, from (cf. Sandri [17])

(3.14) (A(σ) − A(σh )) · (σ − σh ) dx

Ω
q
≥ C{ σ − σh 0,q,Ω + (|σ| + |σh |)q−2 |σ − σh |2 dx},
Ω

1/2
A(σ) − A(σh ) 0,p,Ω ≤C (|σ| + |σh |)q−2 |σ − σh |2 dx
(3.15) Ω
(q−2)/2
·[ σ 0,q,Ω + σh 0,q,Ω ] ,

and using Young’s inequality, we obtain for all ε > 0,

q
σ − σh 0,q,Ω + (|σ| + |σh |)q−2 |σ − σh |2 dx
Ω

≤ε (|σ| + |σh |)q−2 |σ − σh |2 dx

Ω
C q−2 2
+ [ σ 0,q,Ω + σh 0,q,Ω ] σ − Πh σ 0,q,Ω ,
ε
 MIXED FINITE ELEMENT METHOD FOR p-LAPLACIAN 77

where C is a constant independent of h and ε. Using this last estimate

and Remark 3.2, we get

q
(3.16) σ − σh 0,q,Ω + (|σ| + |σh |)q−2 |σ − σh |2 dx
Ω
2
≤ C σ − Πh σ 0,q,Ω .

Then (3.12) is a consequence of this last estimation and (3.6).

For the error estimation (3.13) we have (see the proof of Theorem 3.1)

P h u − uh 0,p,Ω ≤ C A(σ) − A(σh ) 0,p,Ω ,

and, using (3.15) and (3.16), we obtain

P h u − uh 0,p,Ω ≤ C σ − Πh σ 0,q,Ω .

Therefore, (3.13) is a consequence of this last inequality, (3.6), the

triangle inequality and the standard error estimate on the operator Ph .

Remark 3.3. The error estimate (3.12) for σ is better than the one
obtained in Baranger et al. [2] for a quasi-Newtonian flow, and the
error estimation (3.13) for u is optimal. Furthermore, these error
estimates are valid with minimal regularity of the solution of problem
(1.1) (compare with Theorem 3.2 in [6], and the results obtained in
Barrett and Liu [3]).

REFERENCES

1. C. Atkinson and C.R. Champion, Some boundary-value problems for the

equation ∇ · (|∇φ|N ∇φ) = 0, Quart. J. Mech. Appl. Math. 37 (1984), 401 419.
2. J. Baranger, K. Najib and D. Sandri, Numerical analysis of a three-fields model
for a quasi-Newtonian flow, Comput. Methods Appl. Mech. Engrg. 109 (1993),
281 292.
3. J.W. Barrett and W.B. Liu, Finite element approximation of the p-Laplacian,
Math. Comp. 61 (1993), 523 537.
4. F. Brezzi and M. Fortin, Mixed and hybrid finite element methods, Springer-
Verlag, New York, 1991.
5. P.G. Ciarlet, The finite element method for elliptic problems, North-Holland,
Amsterdam, 1978.
78 M. FARHLOUL AND H. MANOUZI

6. M. Farhloul, A mixed finite element method for a nonlinear Dirichlet problem,

IMA J. Numer. Anal. 18 (1998), 121 132.
7. V. Girault and P.A. Raviart, Finite element for Navier-Stokes equations,
theory and algorithms, Springer Ser. Comput. Math. 5, Springer-Verlag, New York,
1986.
8. R. Glowinski and A. Marrocco, Sur l’approximation par éléments finis d’ordre
un, et la résolution, par pénalisation-dualité, d’une classe de problèmes de Dirichlet
non linéaires, RAIRO, R-2, 9e année (1975), 41 76.
9. P. Le Tallec, Numerical analysis of viscoelastic problems, Rech. Math. Appl.
15, Masson, Springer-Verlag, 1990.
10. W.B. Liu and J.W. Barrett, Quasi-norm error bounds for the finite element
approximation of some degenerate quasilinear elliptic equations and variational
inequalities, M2 AN 28 (1994), 725 744.
11. , Finite element approximation of some degenerate monotone quasi-
linear elliptic systems, SIAM J. Numer. Anal. 33 (1996), 88 106.
12. H. Manouzi and M. Farhloul, Mixed finite element analysis of a non-linear
three-fields Stokes model, IMA J. Numer. Anal., to appear.
13. J.C. Nedelec, Mixed finite elements in R3 , Numer. Math. 35 (1980), 315 341.
14. J.R. Philip, N -diffusion, Austral. J. Phys. 14 (1961), 1 13.
15. P.A. Raviart and J.M. Thomas, A mixed finite element method for 2-nd order
elliptic problems, Lecture Notes in Math. 606 (1977), 292 315.
16. J.E. Roberts and J.M. Thomas, Mixed and hybrid finite element methods,
in Handbook of numerical analysis, Vol. II, Finite element methods, Part 1 (P.G.
Ciarlet and J.L. Lions, eds.), North-Holland, Amsterdam, 1989.
17. D. Sandri, Sur l’approximation numérique des écoulements quasi-Newtoniens
dont la viscosité suit la loi puissance oú la loi de Carreau, M2 AN 27 (1993),
131 155.
18. J.M. Thomas, Sur l’analyse numérique des méthodes d’éléments finis hybrides
et mixtes, Thèse, Université Pierre et Marie Curie, Paris, 1977.
19. V.B. Tyukhtin, Sur la vitesse de convergence des méthodes d’approximation
de la solution des problèmes variationnels unilatéraux, Vestnik Leningrad Univ.,
Math. Mec. Astronom. 3 (1983), 36 43 (en russe).

D épartement de Mathématiques et de Statistique, Université de Monc-

ton, Moncton, NB, E1A 3E9 Canada
E-mail address: farhlom@[Link]

D épartement de Mathématiques et de Statistique, Université Laval,

Québec, G1K 7P4 Canada