HAL Id: hal-00398782
https://hal.archives-ouvertes.fr/hal-00398782v3
Submitted on 29 Oct 2009
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of
sci-entific research documents, whether they are
pub-lished or not. The documents may come from
teaching and research institutions in France or
abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est
destinée au dépôt et à la diffusion de documents
scientifiques de niveau recherche, publiés ou non,
émanant des établissements d’enseignement et de
recherche français ou étrangers, des laboratoires
publics ou privés.
Cell centered Galerkin methods
Daniele Antonio Di Pietro
To cite this version:
Daniele A. Di Pietro
Institut Français du Pétrole, 1 & 4, avenue de Bois Préau, 92852 Rueil
Malmaison, Fran e
Abstra t
Inthisworkweproposeanewapproa htoobtainandanalyzelowestordermethods
for diusive problemsyielding at the sametime onvergen e rates and onvergen e
tominimalregularitysolutions.Theapproa hmergesideasfromFiniteVolumeand
dis ontinuousGalerkin methods ina uniedperspe tive.
Résumé
Dans ette note, on propose une nouvelle appro he pour obtenir et analyser des
méthodes d'ordre bas pour les problèmes diusifs. L'analyse permet d'estimer à
la fois le taux de onvergen e et de prouver la onvergen e vers des solutions à
régularitéminimale.Cetteappro he ombinelesavan éesré entesdanslesméthodes
de Volumes Finisetde Galerkine dis ontinues.
1 Introdu tion
Dans ette note on propose une nouvelle appro he pour obtenir et analyser
desméthodesd'ordrebaspourlesproblèmediusifs.Lesidéessontprésentées
pour le problème modèle
−∇·(ν∇u) = f
inΩ,
u = 0
on∂Ω,
(1) oùΩ
denote un ouvert polygonal borné,f ∈ L
2
(Ω)
et
ν
est un tenseur uni-formementelliptiqueet onstantpar mor eauxsur une partitionP
Ω
deΩ
. La méthode est omposée de plusieurs étapes : (i) pour toute fa e du maillage,on dénit un interpolateur de tra e onsistant pour des fon tions
susam-ment régulières; (ii) à partir des valeurs de tra e, on re onstruit un gradient
par maille qui est aussi onsistant; (iii) le gradient permet dénir un
sous-espa e de l'espa e des fon tions anes par mor eaux, qui est ensuite utilisé
Pour dessolutionssusammentrégulières, laformebilinéairedis rètes'étend
à l'espa e onténant la solution, e qui permet de béné ier des te hniques
d'analyse par estimation d'erreur propres aux méthodes dG. D'autre part,
les te hniques d'analyse par ompa ité typiques des VolumesFinis ( .f.,e.g.,
Eymard, Gallouët et Herbin [5℄) s'appliquent aussi en vertu des résultats
ré- emment obtenus pour les espa es polynomiaux par mor eaux dans [4℄. Par
simpli ité,onutiliseradans laprésentation l'interpolateurbary entrique
pro-posé dans [5℄. Cet interpolateur,simple à mettreen ÷uvre, ne tient toutefois
pas omptedes hétérogénéités du tenseur de diusion. Parmi lesautres hoix
possiblepour mieuxgérerlesopérateurshétérogène,onsignalel'interpolateur
en L introduit par Aavatsmark et al. [1℄ et ré emment analysé dans [2℄. A
titre d'exemple, la méthode sera appliquée aussi à un problème de
diusion-adve tion-réa tion.
2 Dis retization
Letting
a(u, v)
def
=
R
Ω
∇u·∇v
, the weak formulationof problem (1) readsFind
u ∈ H
1
0
(Ω)
s.t.a(u, v) =
Z
Ω
f v
for allv ∈ H
1
0
(Ω).
(2) LetH ⊂
R+
\ {0}
be a ountable set having 0 asan a umulation point and let{T
h
}
h∈H
dene an admissible polyhedral mesh family in the sense of [4℄ parametrizedbythe meshsizeh
and ompatiblewithP
Ω
.Leth ∈ H
begiven. For ea hT ∈ T
h
, letx
T
∈ T \ ∂T
be su h thatT
is star-shaped with respe t tox
T
(the ell enter)and,forallfa esF ∈ F
h
,setx
b
F
def
=
R
F
x/|F |
.Thesetof fa es of an elementT ∈ T
h
willbe denoted byF
T
h
, the set of boundary fa es byF
b
h
.Forea h internalfa ewe sele tanarbitrarybutxed dire tionforthe normaln
F
≡ n
T
1
,F
and denotebyT
1
the element out ofwhi hn
F
points.The diameters of the generi elementT ∈ T
h
and fa eF ∈ F
h
are denoted byh
T
andh
F
respe tively. Let VT
h
def
=
Rcard(T
h
)
, VF
h
def
=
Rcard(F
h
)
. We all a tra e interpolator
I
a linear bounded operatorfromVT
h
toX
⊂
VF
h
su hthat,forallV
T
h
∋ V = {v
T
}
T
∈T
h
, X∋ I(V) = {I
F
(V)}
F
∈F
h
is su h thatI
F
(V) = 0
ifF ∈ F
b
h
. To ease the presentation, we shall hen eforth refer to the bary entri interpolator of [5℄dened hereafter. Let, for ea h
F ∈ F
i
h
,I
F
h
⊂ T
h
be su h thatcard(I
h
) = d
anddenoteby{α
F
T
}
T
∈I
F
h
the familyofrealssu hthat
x
b
F
=
P
T
∈I
F
h
α
F
T
x
T
.The bary entri interpolator is dened as follows:fullled by the tra e interpolator for any fun tion
w
in a suitablespa eV
:∀T ∈ T
h
, ∀F ∈ F
h
T
,
|w(
x
b
F
) − I
F
(Π
VT
h
(w))| ≤ C
w
h
2
F
,
(4) whereΠ
VTh
: V →
VT
h
is the operator that maps every element
w ∈ V
ontoΠ
VTh
(w)
def
= {w(x
T
)}
T
∈T
h
. Forthe bary entri interpolator (3)we an takeV
def
= {w ∈ C
0
2
(Ω) ; ν∇w ∈ H(div; Ω)}.
The global regularity assumptions in
V
are only needed to prove optimal onvergen e ratesandstronglydependonthe hoi eoftheinterpolator.Whenν
isheterogeneous,theL-interpolatorintrodu edin[1℄andanalyzedin[2℄may be preferable, as onlylo alregularityinside ea hΩ
i
is required. Convergen e tominimalregularity solutionsthatare barely inH
1
0
(Ω)
is alsoobtained,but noestimate of the onvergen e rate is available inthis ase.The tra einterpolatoristhe keyingredient to onstru ta onsistent gradient
approximation. Forall
V
∈
VT
h
, welet ( .f.[5℄)G
T
(V)
def
=
1
|T |
X
F
∈F
T
h
|F | (I
F
(V) − v
T
) n
T,F
.
Fork ≥ 0
, letP
k
h
def
= {p
h
∈ L
2
(Ω) ; ∀T ∈ T
h
, p
h|T
∈
Pk
d
(T )}
. Consider the following approximation spa e:V
h
def
=
n
v
h
∈ P
h
1
; ∃V ∈
VT
h
, ∀T ∈ T
h
,v
h
(x
T
) = v
T
and∇v
h|T
= G
T
(V)
o
.
For every
v
h
∈ V
h
, there holdsv
h|T
(x) = v
T
+ G
T
(V)·(x − x
T
)
for allT ∈ T
h
andV
the element of VT
h
asso iated to
v
h
. The spa eV
h
is thus a subspa e ofP
1
h
with pie ewise anetest fun tionswhose value atx
T
ispres ribed and whose gradient depends on the values at neighbouring ell enters, on themesh and,possibly, ontheproblemdata viathe interpolator
I
.Wedenethe following norm onthe augmented spa eV (h)
def
= V + V
h
: Forallv ∈ V (h)
,kvk
2
h
def
=
X
T
∈T
h
k∇vk
2
[L
2
(T )]
d
+
X
F
∈F
h
1
h
F
kJvKk
2
L
2
(F )
,
where,for all
ϕ
su h that a(possibly two-valued) tra e isavailableonF
,JϕK
def
=
ϕ
|T
1
− ϕ
|T
2
forallF ∈ F
T
1
h
∩ F
T
2
h
,
ϕ
|T
forallF ∈ F
T
h
∩ F
h
b
.
The Poin aré inequality proved in [4℄ for the pie ewise broken ane spa e
Lemma 1 (Approximation estimate). For all
w ∈ V
, there existsC
w
inde-pendent of the meshsizeh
su h thatinf
w
h
∈V
h
kw − w
h
k
h
≤ C
w
h.
(5)Proof. Use (4)and pro eed asin [5,Theorem 4.2℄ toprove that
X
T
∈T
h
k∇w
h
− ∇wk
2
[L
2
(T )]
d
≤ C
w
h
2
.
Forthe jumpterm,observe that,sin e
w
is ontinuousa ross meshinterfa es anditvanisheson∂Ω
,Jw
h
K = Jw
h
− wK
forallx ∈ F
,F ∈ F
h
.UsingAgmon's tra e inequality together with Taylorexpansion yieldsthe desiredresult.ConsiderthefollowingbilinearforminspiredfromtheSIPGmethodofArnold:
Forall
(w, v
h
) ∈ V (h) × V
h
a
h
(w, v
h
)
def
=
Z
Ω
ν∇
h
w·∇
h
v
h
−
X
F∈F
h
Z
F
[{ν∇
h
w}· n
F
Jv
h
K + JwK{ν∇
h
v
h
}· n
F
]
+
X
F
∈F
h
η
h
F
Z
F
n
F
ν· n
F
JwKJv
h
K,
(6)where
∇
h
denotes the broken gradient,η > 0
is large enoughto ensure oer- ivity and{ϕ}
def
=
1
2
(ϕ
|T
1
+ ϕ
|T
2
)
for allF ∈ F
T
1
h
∩ F
T
2
h
,
ϕ
|T
for allF ∈ F
T
h
∩ F
h
b
.
The implementation does not require ubatures. Indeed, using linearity, the
termsinsquarebra ketsin(6) anberewrittenas
P
F
∈F
h
|F |{ν∇
h
u
h
}· n
F
Jv
h
(
x
b
F
)K
and
P
F
∈F
h
|F |{ν∇
h
v
h
}· n
F
Ju
h
(
x
b
F
)K
respe tivelyand the penaltyterm an be safelyapproximatedusingthemiddlepointrulebyP
F
∈F
h
η|F |
h
F
Ju
h
(
x
b
F
)KJv
h
(
x
b
F
)K
.
As the dis rete bilinear form
a
h
has been naturally extended toV (h) × V
h
, the usual dGanalysis te hniques an beapplied.We onsider the following approximation of (2):
Find
u
h
∈ V
h
s.t.a
h
(u
h
, v
h
) =
Z
Ω
f v
h
forall
v
h
∈ V
h
.
(7) Problem (7) has only one unknown per ell, as the elimination of fa eun-knowns has been arried out in the interpolation pro ess. Su h an operation
an be performed in pra ti e when, for a given
V
∈
VT
h
, the dependen e of
ea h
I
F
(V)
onV
islimitedtoafew ellvalues(whi h isthe ase, e.g.,forthe bary entri interpolator). On the other hand, when ea hI
F
(V)
potentially depends on all the values{v
T
}
T
∈T
h
(as it is the ase for the interpolator ob-tained by expressing fa e unknowns in terms of ell unknowns in the hybridNumeri al example for a pure diusion problem on a triangular mesh family (
u
= sin(πx) sin(πy)
,ν
=
1d
,f
= 2π
2
sin(πx) sin(πy)
).h
ku − u
h
k
L
2
(Ω)
orderku − u
h
k
h
order 1 /4 2.8891e-02 1.7460e-01 1 /8 6.9933e-03 2.05 8.1249e-02 1.10 1 /16 1.7562e-03 1.99 3.9300e-02 1.05 1 /32 4.4048e-04 2.00 1.9467e-02 1.01 1 /64 1.1098e-04 1.99 9.6534e-03 1.01 1 /128 2.7795e-05 2.00 4.8156e-03 1.00 1 /256 6.9559e-06 2.00 2.4051e-03 1.003 Main results and omments
Throughoutthis se tion,
{u
h
}
h∈H
willdenotethe sequen e of solutionsofthe dis reteproblem(7)onthe admissiblemeshfamily{T
h
}
h∈H
.Pro eedingasin Arnold et al. [3℄, one an prove the following result:Theorem 2 (Convergen e to regular solutions). Let
u ∈ H
1
0
(Ω)
solve (2) and assume, moreover, thatu ∈ V
. Then, there existsC
u
independent of the meshsizeh
su h thatku − u
h
k
h
≤ C
u
h.
Furthermore,the Aubin-Nits he tri kyields
ku − u
h
k
L
2
(Ω)
≤ C
u
h
2
.The above
estimates are numeri ally veried in Table 1. For all
v
h
∈ V
h
and for allF ∈ F
h
, dene the lifting operator solution of the following problem: Findr
F
(Jv
h
K) ∈ [P
h
0
]
d
su h thatZ
Ω
r
F
(Jv
h
K)·τ
h
=
Z
F
Jv
h
K{τ
h
}· n
F
∀τ
h
∈ [P
0
h
]
d
,
and setR
h
(Jv
h
K)
def
=
P
F
∈F
h
r
F
(Jv
h
K)
. Then, pro eeding asin [4, Theorem 2.2℄ one an prove that, for all sequen es{v
h
}
h∈H
bounded in thek·k
h
norm, ash → 0
, there isv ∈ H
1
0
(Ω)
su h thatv
h
→ v
inL
2
(Ω)
and∇
h
v
h
− R
h
(Jv
h
K)
weakly onvergesto∇v
in[L
2
(Ω)]
d
.Followingtheguidelinesof[4,Theorem 3.1℄
one an then prove that
Theorem 3 (Convergen e to minimalregularity solutions). Let
u
solve (2). Then,ash → 0
,(i)u
h
→ u
inL
2
(Ω)
,(ii)
∇
h
u
h
→ ∇u
and∇
h
u
h
− R
h
(Ju
h
K) → ∇u
in[L
2
(Ω)]
d
and (iii)P
F
∈F
h
h
−1
F
kJu
h
Kk
2
L
2
(F )
→ 0
.Some omments are of order. The derivation of the method outlined in 2
remainsvalidtaking inspiration fromany alternativedGmethod forproblem
(1) ( .f. [3℄ for a review). Moreover, the ideas of 2 virtually extend to any
problem for whi h a dG method has been devised. As an example, for the
diusion-adve tion-rea tion problem
Numeri al example for a diusion-adve tion-rea tion problem on a triangular
mesh family (
k·k
ν,β,h
is the asso iated energy norm,u
= sin(πx) sin(πy)
,ν
=
1d
,β
= (2π, 0)
,µ
= 0
,f
= 2π
2
(sin(πx) + cos(πx)) sin(πy)
).
h
ku − u
h
k
L
2
(Ω)
orderku − u
h
k
ν,β,h
order 1 /4 3.0457e-02 1.7902e-01 1 /8 7.4629e-03 2.03 8.3564e-02 1.10 1 /16 1.8935e-03 1.98 4.0205e-02 1.06 1 /32 4.7634e-04 1.99 1.9815e-02 1.02 1 /64 1.1978e-04 1.99 9.7981e-03 1.02 1 /128 2.9997e-05 1.99 4.8796e-03 1.00 1 /256 7.4900e-06 2.00 2.4344e-03 1.00 withβ ∈ C
1
(Ω)
and
µ ≥ 0
su h thatµ +
1
2
∇·β > 0
, the following bilinear formyieldsanoptimally onvergingapproximation( .f.Table2foranumeri alexample):
b
h
(w, v
h
)
def
= a
h
(w, v
h
) −
Z
Ω
wβ·∇
h
v
h
+
X
F
∈F
h
Z
F
w
↑
Jv
h
K +
Z
Ω
µwv
h
,
wherew
↑ def
= (β· n
F
){w} +
|β· n
2
F
|
JwK
on everyF ∈ F
i
h
and asw
↑ def
= (β· n)
⊕
w
on everyF ∈ F
b
h
. Finally, the ompa tness analysis of [4℄ alsoapply, so that valuable toolsfor the analysis of nonlinear problems are available.Referen es
[1℄I. Aavatsmark,G.T.Eigestad,B.T.Mallison,andJ.M.Nordbotten, A ompa t
multipointuxapproximationmethodwithimproved robustness,Numer.Methods
PartialDierential Equations 24(2008), no.5,13291360.
[2℄L. Agélas, D. A. Di Pietro, and J. Droniou, The G method for heterogeneous
anisotropi diusion on
general meshes, M2ANMath.Model. Numer.Anal. (2009), To appear.Preprint
availableat http://hal.ar hives-ouvertes.fr/hal-0034273 9/fr.
[3℄D. N. Arnold, F. Brezzi, B. Co kburn, and D. Marini, Unied analysis of
dis ontinuous Galerkin methods for ellipti problems, SIAMJ. Numer.Anal. 39
(2002), no.5,17491779.
[4℄D. A. Di Pietro and A. Ern, Dis rete fun tional analysis tools for
dis ontinuous Galerkin methods with appli ation to the in ompressible
Navier-Stokes equations, Math. Comp. (2009), To appear. Preprint available at
http://hal.ar hives-ouvertes.fr/hal-002 78925 /fr/.
[5℄R. Eymard, T. Gallouët, and R. Herbin, Dis retization of heterogeneous and
anisotropi diusion problems on general non- onforming meshes. SUSHI:
a s heme using stabilization and hybrid interfa es, Preprint available at