HAL Id: hal-00908250
https://hal.archives-ouvertes.fr/hal-00908250
Submitted on 22 Nov 2013
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.
Homogenization and concentration for a diffusion equation with large convection in a bounded domain
Grégoire Allaire, Irina Pankratova, Andrey Piatnitski
To cite this version:
Grégoire Allaire, Irina Pankratova, Andrey Piatnitski. Homogenization and concentration for a diffu-
sion equation with large convection in a bounded domain. Journal of Functional Analysis, Elsevier,
2012, 262 (1), pp.300-330. �10.1016/j.jfa.2011.09.014�. �hal-00908250�
CENTRE DE MATHÉMATIQUES APPLIQUÉES
UMR CNRS 7641
91128PALAISEAU CEDEX (FRANCE).Tél: 0169 33 4600. Fax: 01 69 3346 46
http://www.map.polytehnique.fr/
Homogenization and
onentration for a diusion
equation with large onvetion in
a bounded domain
Grégoire Allaire, Irina Pankratova,
Andrey Piatnitski
R.I. 713 May 2011
equation with large onvetion in a bounded domain
G. Allaire
∗
I. Pankratova
†
A. Piatnitski
‡
May 24, 2011
Abstrat
Weonsiderthehomogenizationofanon-stationaryonvetion-diusionequa-
tionposedinabounded domainwithperiodially osillating oeientsandho-
mogeneous Dirihlet boundary onditions. Assuming that the onvetion term
is large, we give the asymptoti prole of the solution and determine its rate
of deay. In partiular, it allows us to haraterize the hotspot, i.e., the pre-
ise asymptotiloationofthesolutionmaximumwhihlieslose to thedomain
boundaryandisalsothepointofonentration. Duetotheompetitionbetween
onvetion and diusionthe position of the hot spot is not always intuitive as
exemplied insome numerial tests.
Keywords: Homogenization, onvetion-diusion, loalization.
1 Introdution
The goal of the paper is to study the homogenization of a onvetion-diusion
equation with rapidly periodially osillating oeients dened in a bounded
domain. Namely,weonsider thefollowing initial boundary problem:
∂ t u ε (t, x) + A ε u ε (t, x) = 0,
in(0, T ) × Ω, u ε (t, x) = 0,
on(0, T ) × ∂Ω, u ε (0, x) = u 0 (x), x ∈ Ω,
(1.1)
∗
Eole Polytehnique, Route de Salay, 91128 Palaiseau Cedex Frane (al-
lairemap.polytehnique.fr)
†
Narvik University College, Postbox 385, 8505 Narvik, Norway; Eole Polytehnique, Route de
Salay,91128PalaiseauCedex Frane(pankratovamap.polytehnique.fr)
‡
NarvikUniversityCollege,Postbox385,8505Narvik,Norway;LebedevPhysialInstitute RAS,
Leninskiave.,53,119991Mosow,Russia(andreysi.lebedev.ru)
where
Ω ⊂ R d
is a bounded domain with a Lipshitz boundary∂Ω
,u 0
belongsto
L 2 (Ω)
andA ε
isanoperator dened byA ε u ε = − ∂
∂x i
a ij x ε
∂u ε
∂x j
+ 1 ε b j x
ε ∂u ε
∂x j ,
where we employ the onvention of summation over repeated Latin indies. As
usual
ε
, whih denotes the period of theoeients, is a small positive param-eter intended to tend to zero. Note the large saling in front of the onvetive
term whih orresponds to the onvetive and diusive terms having both the
same order of magnitude at the small sale
ε
(this is a lassial assumption inhomogenization [5℄, [12℄, [13 ℄, [21 ℄). We make thefollowing assumptions on the
oeients oftheoperator
A ε
.(H1) The oeients
a ij (y), b j (y)
are measurable bounded funtions dened ontheunitell
Y = (0, 1] d
,thatisa ij , b j ∈ L ∞ (Y )
. Moreover,a ij (y), b j (y)
areY
-periodi.(H2) The
d × d
matrixa(y)
is uniformlyellipti, that is thereexistsΛ > 0
suhthat, for all
ξ ∈ R d
andfor almost ally ∈ Ω
,a ij (y)ξ i ξ j ≥ Λ | ξ | 2 .
For the large onvetion term we do not suppose that the eetive drift (the
weightedaverageof
b
denedbelowby(2.4))iszero,northatthevetoreldb(y)
isdivergene-free. Someadditional assumptionsonthesmoothnessandompat
supportoftheinitialdata
u 0
willbemadeinSetion2afterintroduingauxiliaryspetral ell problems. Inview of (H1) and (H2), for any
ε > 0
, problem(1.1)hasa unique weak solution
u ε ∈ L ∞ [0, T ; L 2 (Ω)] ∩ L 2 [0, T ; H 1 (Ω)]
(see [6 ℄).Ourmaingoalistodesribetheasymptotibehaviorofthesolution
u ε (t, x)
ofproblem (1.1) as
ε
goesto zero. There are of ourse manymotivations to study suh a problem (one of them being the transport of solutes in porous media[17 ℄). However, if (1.1) is interpreted as the heat equation in a uid domain
(theuidveloitybeinggivenby
ε −1 b(x/ε)
),we anparaphrasethefamous hotspot onjeture of J. Rauh [23℄, [7℄, [10 ℄, and ask a simple question in plain
words. Iftheinitialtemperature
u 0
hasitsmaximuminsidethedomainΩ
,whereshall this maximum or hot spot go astime evolves ? Morepreisely, we want
to answer this question asymptotially as
ε
goes to zero. Theorem 2.1 (andthe disussion following it) gives a omplete answer to this question. The hot
spot is a onentration point
x c
, loated asymptotially lose to the boundary∂Ω
(see Figure 1), whih maximizes the linear funtionΘ · x
onΩ
where thevetor parameter
Θ
is determined as an optimal parameter in an auxiliary ellproblem(seeLemma2.1). Surprisingly
Θ
isnotsomeaverageoftheveloityeldbut is the result of an intriate interation between onvetion and diusion in
the periodiity ell (even in the ase of onstant oeients ; see the numerial
examples of Setion 7). Furthermore, Theorem 2.1 gives the asymptoti prole
of thesolution, whih isloalizedintheviinityof thehotspot
x c
,intermsofa homogenized equation withan initial ondition thatdependson thegeometry
of thesupportof theinitial data
u 0
.Beforewe explainour resultsingreaterdetails,we brieyreviewprevious re-
sultsintheliterature. Intheasewhenthe vetor-eld
b(y)
issolenoidalandhaszeromean-value, problem(1.1)hasbeen studiedbythelassialhomogenization
methods (see, e.g, [8℄, [25 ℄). In partiular, the sequene of solutions is bounded
in
L ∞ [0, T ; L 2 (Ω)] ∩ L 2 [0, T ; H 1 (Ω)]
and onverges, asε → 0
, to the solutionofan eetive or homogenized problem in whih there is no onvetive term. For
general vetor-elds
b(y)
, and if the domainΩ
is the whole spaeR d
, the on-vetion might dominate the diusion and we annot expeta usual onvergene
of thesequene ofsolutions
u ε (t, x)
inthexed spatialreferene frame. Rather,introduingaframeofmovingoordinates
(t, x − ¯ bt/ε)
,wheretheonstantvetor¯ b
istheso-alled eetive drift(oreetive onvetion)whihis dened by(2.4)asaweightedaverageof
b
,itisknownthatthetranslatedsequeneu ε (t, x − ¯ bt/ε)
onverges to the solution of an homogenized paraboli equation [5 ℄, [13 ℄. Note
that the notion ofeetive drift wasrst introdued in[21℄. Ofourse, theon-
vergene inmoving oordinates annotworkinabounded domain. Thepurpose
of thepresent work is to study theasymptoti behavior of(1.1) inthease of a
bounded domain
Ω
.Bearingthesepreviousresultsinmind,intuitively,itislearthatinabounded
domain the initial prole should move rapidly in the diretion of the eetive
drift
¯ b
untilitreahesthe boundary,and thendissipatedue to thehomogeneous Dirihletboundary ondition,ast
grows. Sine theonvetion termis large,thedissipation inreases, as
ε → 0
,sothat thesolution asymptotially onverges to zero at nite time. Indeed, introduing a resaled (short) timeτ = ε −1 t
, werewrite problem(1.1) inthe form
∂ τ u ε − ε div a ε ∇ u ε
+ b ε · ∇ u ε = 0,
in(0, ε −1 T) × Ω, u ε (τ, x) = 0,
on(0, ε −1 T ) × ∂Ω,
u ε (0, x) = u 0 (x), x ∈ Ω.
(1.2)
Applying the lassial two-sale asymptoti expansion method [8℄,one an show
that, for any
τ ≥ 0 Z
Ω
| u ε (τ, x) − u 0 (τ, x) | 2 dx → 0, ε → 0,
where the leading term of the asymptotis
u 0
satises the following rst-orderequation
∂ τ u 0 (τ, x) + ¯ b · ∇ u 0 (τ, x) = 0,
in(0, + ∞ ) × Ω, u 0 (τ, x) = 0,
on(0, + ∞ ) × ∂Ω ¯ b , u 0 (0, x) = u 0 (x), x ∈ Ω,
(1.3)
with
¯ b
being the vetor of eetive onvetion dened by(2.4). Here∂Ω ¯ b
is thesubset of
∂Ω
suh that¯ b · n < 0
wheren
stands for the exteriorunit normal on∂Ω
. One an onstrut higher order terms in the asymptoti expansion foru ε
.Thisexpansionwillontainaboundarylayerorretorintheviinityof
∂Ω \ ∂Ω ¯ b
.A similarproblem inamore general settinghasbeen studied in[9℄.
Thesolution ofproblem (1.3)an befound expliitly,
u 0 (τ, x) =
u 0 (x − ¯ bτ ),
for(τ, x)
suh thatx, (x − ¯ bτ ) ∈ Ω,
0,
otherwise,
whih shows that
u 0
vanishes after a nite timeτ 0 = O(1)
. In the originaloordinates
(t, x)
wehaveZ
Ω
| u ε (t, x) − u 0 (x − ε −1 ¯ b t) | 2 dx → 0, ε → 0.
Thus, for
t = O(ε)
theinitial prole ofu ε
moveswith theveloityε −1 ¯ b
until itreahestheboundaryof
Ω
andthendissipates. Furthermore,anynitenumberof termsinthe two-saleasymptotiexpansion ofu ε (τ, x)
vanishforτ ≥ τ 0 = O(1)
and thus for
t ≥ t 0
with an arbitrary smallt 0 > 0
. On the other hand, ifu 0
is positive, thenby the maximum priniple,
u ε > 0
for allt
. Thus, the methodof two-saleasymptoti expansioninthis short-time salingisunable toapture
the limit behaviour of
u ε (t, x)
for positive time. The goal of the present paperis therefore to perform a more deliate analysis and to determine the rate of
vanishingof
u ε
,asε → 0
.Thehomogenizationofthespetralproblemorrespondingto(1.1)inabounded
domain for a general veloity
b(y)
was performed in [11℄, [12 ℄. Interestingly enough theeetivedriftdoesnotplayanyroleinsuhaasebutratherthekeyparameter is another onstant vetor
Θ ∈ R d
whih isdened as anoptimal ex-ponential parameter inaspetralell problem (seeLemma 2.1). Morepreisely,
it is proved in [11 ℄, [12℄ that the rst eigenfuntion onentrates as a boundary
layeron
∂Ω
inthediretion ofΘ
. WeshallprovethatthesamevetorparameterΘ
is alsoruial intheasymptoti analysisof(1.1).Notie that for large time and after a proper resaling the solution of (1.1)
should behave like therst eigenfuntion of theorresponding ellipti operator,
and thusonentratesinasmall neighbourhood of
∂Ω
inthediretionofΘ
. Weprove that this guess is orret, not only for large time but also for any time
t = O(1)
, namely thatu ε (t, x)
onentrates in the neighbourhood of the hot spot or onentration pointx c ∈ ∂Ω
whih depends onΘ
. The value ofΘ
anbe determined intermsof some optimality propertyof therst eigenvalue of an
auxiliary periodi spetral problem (see Setion 2). It should be stressed that,
in general,
Θ
does not oinide with¯ b
. As a onsequene, it may happen thattheonentrationpoint
x c
doesnot evenbelongtothesubset of∂Ω
onsistingofpointswhihareattainedbytranslationoftheinitial datasupportalong
¯ b
. Thisphenomenon is illustratedbynumerial examplesinSetion 7.
problems inthe unitell
Y
andimposeadditional onditions onthegeometryoftheompatsupportof
u 0
. Wethenstate ourmainresult(seeTheorem2.1)andgive its geometri interpretation. In Setion 3, in order to simplify theoriginal
problem (1.1), we use a fatorization priniple, as in [24℄, [18℄, [26 ℄, [11 ℄, based
on the rst eigenfuntions of the auxiliary spetral problems. As a result, we
obtain a redued problem, where thenew onvetion is divergene-free and has
zero mean-value. Studying the asymptoti behaviour of the Green funtion of
the redued problem, performed inSetion 4, isan important part of theproof.
It isbasedontheresultobtained in[1℄ forafundamentalsolution ofaparaboli
operator with lower order terms. Asymptotis of
u ε
is derived in Setion 5. InSetion 6 we study the ase when the boundary of the support of
u 0
has a atpart. To illustrate the main result of the paper, in Setion 7 we present diret
omputationsof
u ε
usingthesoftwareFreeFEM++[15 ℄. Anumber ofbasifatsfrom the theoryofalmost periodifuntionsis giveninSetion 8.
2 Auxiliary spetral problems and main result
Wedene an operator
A
and its adjointA ∗
byAu = − div(a ∇ u) + b · ∇ u, A ∗ v = − div(a T ∇ v) − div(b v),
where
a T
isthetransposedmatrixofa
. Following[8℄,forθ ∈ R d
,weintroduetwoparameterizedfamiliesofspetralproblems(diretandadjoint)intheperiodiity
ell
Y = [0, 1) d
.
e −θ·y A e θ·y p θ (y) = λ(θ) p θ (y), Y, y → p θ (y)
Y-periodi.
(2.1)
e θ·y A ∗ e −θ·y p ∗ θ (y) = λ(θ) p ∗ θ (y), Y, y → p ∗ θ (y)
Y-periodi.
(2.2)
The next result,based onthe Krein-Rutman theorem, wasproved in[11℄,[12℄.
Lemma 2.1. For eah
θ ∈ R d
, the rst eigenvalueλ 1 (θ)
of problem (2.1) isreal, simple, and the orresponding eigenfuntions
p θ
andp ∗ θ
an be hosen pos-itive. Moreover,
θ → λ 1 (θ)
is twie dierentiable, stritly onave and admits a maximum whihis obtained for a uniqueθ = Θ
.Theeigenfuntions
p θ
andp ∗ θ
dened byLemma 2.1, an be normalized byZ
Y
| p θ (y) | 2 dy = 1
andZ
Y
p θ (y) p ∗ θ (y) dy = 1.
Dierentiatingequation (2.1)withrespetto
θ i
,integrating againstp ∗ θ
andwrit-ingdownthe ompatibility onditionfor theobtained equation yield
∂λ 1
∂θ i = Z
Y
b i p θ p ∗ θ + a ij (p θ ∂ y j p ∗ θ − p ∗ θ ∂ y j p θ ) − 2 θ j a ij p θ p ∗ θ
dy.
(2.3)Obviously,
p θ=0 = 1
,and, thus,∂λ 1
∂θ i (θ = 0) = Z
Y
b i p ∗ θ=0 + a ij ∂ y j p ∗ θ=0
dy := ¯ b i ,
(2.4)whih denes the omponents
¯ b i
of the so-alled eetive drift. In the presentpaper we assume that
¯ b 6 = 0
(or, equivalently,Θ 6 = 0
). The ase¯ b = 0
an bestudied by lassial methods (see, for example, [25℄). The equivalene of
¯ b = 0
and
Θ = 0
is obvioussineλ 1 (θ)
isstritly onave witha unique maximum.We need to make some assumptions on the geometry of the support
ω
(alosed setasusual) ofthe initialdata
u 0
withrespetto thediretion ofΘ
. Onepossible setof onditionsis thefollowing.
(H3) Theinitialdata
u 0 (x)
isaontinuousfuntioninΩ
,hasaompat supportω ⋐ Ω
and belongs toC 2 (ω)
. Moreover,ω
isaC 2
-lassdomain.(H4) Thesourepoint
x ¯ ∈ ∂ω
,atwhihtheminimuminmin x∈ω Θ · x
isahieved,isunique (see Figure1(a)). Inother words
Θ · (x − x) ¯ > 0, x ∈ ω \ { x ¯ } .
(2.5)(H5) The point
x ¯
is ellipti and∂ω
is loally onvex atx ¯
, that is the prinipalurvatures at
x ¯
havethesamesign. Morepreisely,inloaloordinatestheboundary of
ω
insome neighborhoodU δ (¯ x)
of thepointx ¯
an be denedby
z d = (Sz ′ , z ′ ) + o( | z ′ | 2 )
forsomepositivedenite
(d − 1) × (d − 1)
matrixS
. Herez ′ = (z 1 , · · · z d−1 )
are the orthonormal oordinates inthe tangential hyperplane at
x ¯
, andz d
isthe oordinateinthenormal diretion.
(H6)
∇ u 0 (¯ x) · Θ 6 = 0
.Remark 2.1. In assumption
( H3 )
it is essential that the supportω
is a stritsubsetof
Ω
,i.e.,doesnottouhtheboundary∂Ω
(seeRemark5.3forfurther om-ments on this issue). However, the ontinuity assumption on the initialfuntion
u 0
is not neessary. It will be relaxed in Theorem 5.2 whereu 0 (x)
still belongsto
C 2 (ω)
but isdisontinuousthrough∂ω
. Of ourse, assumingontinuityor notwill hange the order of onvergene and the multipliative onstant in front of
the asymptoti solution.
Notethat assumption
( H4 )
impliesthatΘ 6 = 0
isanormal vetorto∂ω
atx ¯
.Eventually, assumption
( H6 )
is required beause,u 0
being ontinuous inΩ
,we have
u 0 (¯ x) = 0
.To avoidexessivetehnialities for the moment, we state our main resultin
a looseway(see Theorem 5.1for apreise statement).
ω Ω
Θ
¯ x
x c
b b
(a)
Θ ω
Ω
b
¯ x
b
x c
(b)
Figure 1: Denition of the soure point
x ¯
and of the onentration pointx c
.Theorem 2.1. Suppose onditions
( H1 ) − ( H6 )
are satised andΘ 6 = 0
. Ifu ε
is a solutionof problem (1.1) , then, for any
t 0 > 0
andt ≥ t 0
u ε (t, x) ≈ ε 2 ε d−1 2 e − λ 1(Θ) ε2 t e Θ·(x ε −¯ x) M ε p Θ x ε
u(t, x), ε → 0,
where
(λ 1 (Θ), p Θ )
is the rst eigenpair dened by Lemma 2.1 andu(t, x)
solvesthe homogenized problem
∂ t u = div(a eff ∇ u), (t, x) ∈ (0, T ) × Ω, u(t, x) = 0, (t, x) ∈ (0, T ) × ∂Ω, u(0, x) = ∇ u 0 (¯ x) · Θ
| Θ | δ(x − x), ¯ x ∈ Ω.
(2.6)
Here
a eff
isa positive denitematrix, dened by (4.7),M ε
is a onstant, denedin Theorem 5.1,depending on
p Θ
,on thegeometry of∂ω
atx ¯
andon therelativeposition of
x ¯
inεY
(see Remark 5.1 and Figure 2), andδ(x − x) ¯
is the Diradelta-funtion atthe point
x ¯
.TheinterpretationofTheorem2.1intermsofonentrationorndingthehot
spot isthe following. Up toamultipliative onstant
ε 2 ε d −1 2 M ε
,thesolutionu ε
is asymptotially equal to the produt of two exponential terms, a periodially
osillating funtion
p Θ x ε
(whih is uniformly positive and bounded) and the
homogenized funtion
u(t, x)
(whih is independent ofε
). The rst exponential terme − λ 1(Θ) ε2 t
indiates a fast deay in time, uniform in spae. The seond ex-ponential term
e Θ·(x ε −¯ x)
is the root of a loalization phenomenon. Indeed, it is maximumatthosepointsontheboundary,x c ∈ ∂Ω
,whihhaveamaximaloor-dinate
Θ · x
,independently ofthepositionofx ¯
(seeFigure1(b)). These(possiblymultiple) points
x c
are thehot spots. Everywhereelse inΩ
the solution isex-ponentiallysmaller,foranypositive time. Thisbehaviouranlearly beheked
on thenumerial examplesofSetion 7. Itis ofoursesimilar tothebehaviorof
theorrespondingrsteigenfuntion asstudied in[12 ℄.
priniple (see, for example, [24℄, [18℄, [26℄, [11℄) inSetion 3 we make a hange
of unknown funtion in suh a way that the resulting equation is amenable to
homogenization. After that,the new unknownfuntion
v ε (t, x)
isrepresentedin terms of the orresponding Green funtionK ε (t, x, ξ)
. Studying theasymptotibehaviourof
K ε
isperformedinSetion 4. Finally, we turnbakto theoriginalproblem and write down the asymptotis for
u ε
in Setion 5 whih nishes theproof ofTheorem 2.1.
Remark 2.2. Theorem 2.1holds true even if we add a singular zero-order term
of thetype
ε −2 c( x ε )u ε
in theequation (1.1). Thiszero-orderterm will be removedby the fatorization priniple and the rest of the proof is idential. With some
additionalworkTheorem2.1anbegeneralizedtotheaseofso-alled ooperative
systems for whih a maximum priniple holds. Suh systems of diusion equa-
tions arise in nulear reator physis and their homogenization (for the spetral
problem) wasstudied in [12℄.
3 Fatorization
Werepresent asolution
u ε
of theoriginalproblem (1.1) inthe formu ε (t, x) = e −
λ 1(Θ) t
ε 2 e Θ·(x ε −¯ x) p Θ x ε
v ε (t, x),
(3.1)where
Θ
andp Θ
are dened inLemma 2.1. Notiethat thehange ofunknownsis well-dened sine
p Θ
ispositive andontinuous. Substituting (3.1) into (1.1) , multiplying the resulting equation byp ∗ Θ x ε
and using (2.2) , one obtains the
following problemfor
v ε
:
̺ Θ x ε
∂ t v ε + A ε Θ v ε = 0, (t, x) ∈ (0, T ) × Ω, v ε (t, x) = 0, (t, x) ∈ (0, T ) × ∂Ω, v ε (0, x) = u 0 (x)
p Θ x ε e − Θ·(x ε −¯ x) , x ∈ Ω,
(3.2)
where
̺ Θ (y) = p Θ (y) p ∗ Θ (y)
andA ε Θ v = − ∂
∂x i a Θ ij x ε
∂v
∂x j + 1
ε b Θ i x ε
∂v
∂x i ,
and theoeientsof theoperator
A ε Θ
aregivenbya Θ ij (y) = ̺ Θ (y) a ij (y);
b Θ i (y) = ̺ Θ (y) b j (y) − 2 ̺ Θ (y) a ij (y) Θ j +a ij (y)
p Θ (y) ∂ y j p ∗ Θ (y) − p ∗ Θ (y) ∂ y j p Θ (y) .
(3.3)