HAL Id: hal-00844174
https://hal.archives-ouvertes.fr/hal-00844174
Submitted on 13 Jul 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
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
Numerical simulation of a kinetic model for chemotaxis
Nicolas Vauchelet
To cite this version:
Nicolas Vauchelet. Numerical simulation of a kinetic model for chemotaxis. Kinetic and Related
Models , AIMS, 2010, 3 (3), pp.501-528. �10.3934/krm.2010.3.501�. �hal-00844174�
N. Vau helet
UPMC,UnivParis06,UMR7598LJLL,ParisF-75005Fran e;
CNRS, UMR7598LJLL,Paris,F-75005Fran e,
andINRIAProjetBANG,
Tel.: (+33)(0)144273772 Fax: (+33)(0)144277200
E-mailaddresses: vau heletann.jussieu.fr
Abstra t
This paperis devoted to numeri al simulations of a kineti model des ribing
hemo-taxis. This kineti framework has been investigated sin e the 80's when experimental
observationshaveshownthatthemotion ofba teriaisdueto thealternan eof'runs and
tumbles'. Sin eparaboli andhyperboli modelsdonottakeintoa ountthemi ros opi
movement of individual ells, kineti models have be omeof a great interest. Dolak and
S hmeiser (2005) have then proposed a kineti model des ribing themotion of ba teria
responding to temporal gradients of hemoattra tants along their paths. An existen e
result for this system is provided and a numeri al s heme relying on a semi-Lagrangian
method is presented and analyzed. An implementation of this s heme allows to obtain
numeri al simulationsofthe modelandobserve blow-uppatternsthatdiergreatly from
the aseof Keller-Segeltype of models.
Keywords. Chemotaxis; Kineti equations;semi-Lagrangianmethod; onvergen e analysis.
AMS subje t lassi ations: 92C17;92B05;65M12; 82C80.
1 Introdu tion and modelling
Chemotaxis is the phenomenon in whi h ells dire t their movements a ording to ertain
hemi als in their environment. A possible issue of a positive hemota ti al movement is
the aggregation of organisms involved to form a more omplex organism or body. Many
attempts for des ribing hemotaxis from a Partial Dierential Equations viewpoint, i.e. for
a large population, have been onsidered. At the ma ros opi level the most famous is the
Patlak, Keller and Segel model[28, 34℄. Although this models have been su essfully used to
des ribeaggregation ofthe population(see [25, 26, 41℄ forsurveys), these ma ros opi models
have several short omings, for instan e they do not take into a ount the detailed individual
movement of ells.
Therefore another approa h involving kineti equations to des ribe hemotaxis has been
distribution fun tion
f
of ba teria at timet
, positionx ∈ ω
and velo ityv ∈ V
and of theon entration of hemoattra tant
S
. The system writes inthe following way :
∂
t
f + v · ∇
x
f =
Z
v
′
∈V
(T [S](v
′
→ v)f(v
′
) − T [S](v → v
′
)f (v)) dv
′
,
−∆S + S = ρ(t, x) :=
Z
v∈V
f (t, x, v) dv,
(1.1)ompleted with the initial ondition
f (0, x, v) = f
0
(x, v).
(1.2)The turningkernel
T [S](v
′
→ v)
denotes the rate of ells hanging their velo ity fromv
′
to
v
.This system models the evolution of agellated ba teria su h as E. Coli. It has been
observed that aba teriummovesalong straightlines, suddenlystop to hoose a newdire tion
andthen ontinuemovinginthenewdire tionuntilthe ellsre eptorssaturate. Themovement
of the ba teriumis thendue to the alternan eof these 'run'and 'tumble' phases[1,40℄. Cells
areableto omparethepresent hemi al on entrationstopreviousonesandthustorespondto
temporalgradientsalongtheirpaths. The de isionto hangedire tionand turnorto ontinue
moving depends then on the on entration prole of the hemi al
S
along the traje tories ofells and detailed models have been proposed in [18, 19, 16℄. In [16℄ the authors propose to
onsider simply aturning kernel of the form :
T [S](v
′
→ v) = φ(∂
t
S + v
′
· ∇
x
S).
(1.3)Therateof turningisgreaterif thegradient on entrationalongthetraje tory
∂
t
S + v · ∇
x
S
is negativethan whenitispositive. Experimentally,intheabsen e ofgradientsof on entration,an individual ell of E. Coli performs a random walk with a mean duration of run times of
1s
(see [30℄). Due to the inuen e of the hemoattra tant a ellsensing a positive gradientofon entration has a run 4 times longer. Then, in this simpliedmodel, we onsider that
φ
isa positive nonin reasingsmooth fun tion; morepre isely,
φ ∈ C
∞
(R),
φ
′
< 0,
φ(z) =
1
ifz < −α,
1/4
ifz > α,
(1.4)
for agiven positive
α
small.Inthis work,weare interestedintheevolutionoftheba teria on entrationinaPetri box,
whi his approximated by a bounded domain
ω ⊂ R
2
. The velo ity of ba teriahas a onstant
modulus
V
,thereforewetakeV = S
V
:= {v ∈ R
2
with
kvk = V }
. We denoteΩ = ω × V
. Thesystem an bethen rewritten in the followingway :
∂
t
f + v · ∇
x
f =
Z
v
′
∈V
φ(∂
t
S + v
′
∇
x
S)f (v
′
) dv
′
− 2πφ(∂
t
S + v∇
x
S)f (v),
f (0, x, v) = f
0
(x, v),
−∆S + S = ρ(t, x) :=
Z
v∈V
f (t, x, v) dv,
(1.5)This system is ompleted with the spe ular ree tion onditions
where
ν(x)
istheoutwardunitve toratthepointx
oftheboundary∂ω
. Andforthehemoat-tra tant on entration,we set Neumannboundary onditions :
∂
ν
S(t, x) = 0,
∀ x ∈ ∂ω.
(1.7) Thissystem is omposedofakineti equation oupledtoanellipti equation. Therefore weproposeinthisworktouse te hniques whi hhaveproven theire ien y forthenumeri al
res-olutionofthe Vlasov-Poissonsystem inplasmaphysi stodealwiththe numeri alapproa hof
system(1.5). LagrangianmethodslikeParti le-In-Cellmethodswhi h onsistofapproximating
theplasmabyanitenumberofma ro-parti lesareusuallyperformedforthe Vlasovequation
(see [4℄). However, thesemethodsareknown tobevery useful forlarges ale problemsbut are
very noisy and do poor job on the tailof the distribution fun tion. To remedy this problem,
Eulerian methods have been proposed. They onsist of dis retizing the Vlasov equation on a
mesh of phase spa e. Among them, nite volume s hemes are known to be robust and
om-putationally heap [9, 12, 14, 22℄ but very onstrained by a CFL ondition. Semi-Lagrangian
methodsare otherkindsof Eulerianmethodallowingtoobtaina uratedes riptionof the
dis-tributionfun tion [39, 2, 3, 15℄. They onsist of dire tly omputing the distribution fun tion
atea htimestep onaxed Cartesianmesh ofthe phasespa e byfollowingthe hara teristi s
urvesba kward andinterpolatingthevalueatthe baseofthe hara teristi s. Wereferto[23℄
for a review on Eulerianmethods. Then, we will use in this work a semi-Lagrangian method
for the numeri alresolution of model(1.5).
The paper is organized as follows. In the next se tion we state and prove an existen e
and uniqueness result for the system (1.5)(1.7). For the sake of simpli ity this study is
onsidered in the whole domain
R
2
, whi h allows to have an expli it expression of
S
thanksto the Bessel potential. Se tion 3 is devoted to the numeri al resolution of this system. We
rst re all the semi-Lagrangian method used for the dis retization of the kineti transport
equation. Then we present the algorithm of resolution of the whole system (1.5). Finally, an
analysis of this s heme under additional assumptions furnishes a onvergen e result in
L
2
ofthe dis reteapproximationtowards the solution ofmodel(1.5). Numeri alsimulations,whi h
shows the aggregation phenomenon observed forba teria E.Coli, are presented and dis ussed
in omparison to those for Keller-Segelin se tion4.
2 Existen e result
Forthe sake of simpli ity,we onsider in this se tion that
ω = R
2
and
Ω = R
2
× V
.
The existen e of solutionstokineti models of hemotaxishas been investigated inseveral
papers. In[11,27℄,globalexisten efortheinitialvalueproblem(1.1)in
R
3
andin
R
2
hasbeen
obtainedundertheassumptionthattheturningkernelis ontrolledby termsinvolving
S(t, x+
v)
andS(t, x − v
′
)
. Dispersive methods are used to obtain a priori estimates. These resultshas been extended in [8℄formore generalassumptions onthe turningkernel. All thesepapers
took intoa ount the ee t ofthe gradient ofthe hemi alsignaland showed global existen e
of solutions. However, these rigorous global existen e results have not in luded the temporal
derivative of the signal in the growth ondition of the turning frequen y. In ref. [17℄, the
authorsinvestigate global existen e of solutions(not ne essary unique) for general hyperboli
hemotaxis models where the turning kernel takes into a ount the temporal derivative of
the hemoattra tant through the evolution of internal states but only in the one-dimensional
(1.5) onsidered in the whole domain
Ω = R
2
× V
and to establish existen e and uniqueness
of global-in-time solution. The main novelty is due to the dependan e of the ross-se tion of
the turningoperatoron the timederivative ofthe hemoattra tant on entration. Weuse the
followingexpression for the hemoattra tant on entration :
S(t, x) = (G ∗ ρ(t))(x),
whereG(x) =
1
4π
Z
∞
0
e
−
π
|x|
4s
2
−
s
4π
ds
s
.
The fun tion
G
isknown asthe Besselpotential. The idea toover ome thedi ulty raisedbythe term
∂
t
S
inthe turning kernel expression (1.3) isto use the onservation of the density :∂
t
ρ + ∇
x
· J = 0,
J(t, x) =
Z
V
vf (t, x, v) dv.
(2.1) Wehave then∂
t
S = G ∗ ∂
t
ρ = −G ∗ ∇
x
· J = −∇
x
G ∗ J,
(2.2)where the onvolutionbetween two ve tors is dened by
∇
x
G ∗ J = ∂
x
1
G ∗ J
1
+ ∂
x
2
G ∗ J
2
. Itimplies a ontrol onthe partial derivatives of
S
with respe t to time thanks to ontrols onJ
(see [7℄). We an then rewritethe problem (1.5) as
∂
t
f + v · ∇
x
f =
Z
v
′
∈V
φ(∇
x
G ∗ (v
′
ρ − J))f(v
′
) dv
′
− 2πφ(∇
x
G ∗ (vρ − J))f(v),
f (0, x, v) = f
0
(x, v),
(2.3)Werst denethe notionof weak solutions for(2.3) on
(0, T ) × Ω
.Denition 2.1 We say that
f
is a weak solution of (2.3) onL
q
(0, t
0
; L
p
(Ω))
fort
0
> 0
andp, q ≥ 1
, if forany test fun tionψ ∈ D([0, t
0
) × Ω)
, we haveZ
(0,t
0
)×Ω
(∂
t
ψ + v · ∇
x
ψ)f dxdvdt = −
Z
(0,t
0
)×Ω
Z
V
φ(∇
x
G ∗ (v
′
ρ − J))f(v
′
)ψ dtdxdvdv
′
+2π
Z
(0,t
0
)×Ω
φ(∇
x
G ∗ (vρ − J))fψ dtdxdv +
Z
Ω
f
0
(x, v)ψ(0, x, v) dxdv,
where the ells density
ρ
and the urrent densityJ
are dened byρ(t, x) =
Z
V
f (t, x, v) dv,
J(t, x) =
Z
V
vf (t, x, v) dv.
(2.4)Westate the following existen e and uniqueness result :
Theorem 2.2 Assume
f
0
∈ L
1
+
∩ L
∞
(Ω)
and that the turning kernelT
is dened by (1.3)(1.4). Then the initial value problem (2.3) admits a unique global weak solution
f
satisfyingf ∈ L
∞
((0, ∞); L
1
+
∩ L
∞
(Ω))
. Moreover, iff
0
∈ W
2,2
(Ω) ∩ W
1,∞
(Ω)
then for all
t
0
> 0
, there exists a onstantC
0
depending ont
0
and on the data su h thatthe weak solution of (2.3) satiseskfk
L
∞
((0,t
0
);W
2,2
(Ω)∩W
1,∞
(Ω))
≤ C
0
;
andk∂
t
f k
L
∞
((0,t
of a unique solution. The smoothness assumption (1.4) on
φ
provides thanks to a xed pointpro edure the uniqueness of solution. Then, thanks to a priori estimates, we extend this
solution up to
t
0
for allt
0
> 0
and then re over global existen e. The main tool in the proofof Theorem 2.2is ana priori estimategiven in the following Lemma:
Lemma 2.3 Let
t
0
> 0
andf
0
∈ L
∞
(Ω)
. Letf
be a weak solution of (2.3), su h thatf ∈
L
1
((0, t
0
), L
1
+
∩ L
∞
(Ω))
. Then we have for a.e.t ∈ (0, t
0
)
,kf(t, ·, ·)k
L
1
(Ω)
= kf
0
k
L
1
(Ω)
andkfk
L
∞
((0,t
0
);L
∞
(Ω))
≤ C(kf
0
k
L
∞
(Ω)
) e
2πt
,
where the onstant
C(kf
0
k
L
∞
(Ω)
)
depends only on the initial data.Proof. First the onservation of the mass shows that
kf(t, ·, ·)k
L
1
(Ω)
= kf
0
k
L
1
(Ω)
.
From the bound
1/4 ≤ φ ≤ 1
ofφ
(1.4) we dedu e∂
t
f (t, x, v) + v · ∇
x
f (t, x, v) ≤
Z
V
f (t, x, v
′
) dv
′
.
Integrating alongthe traje tories, we nd
f (t, x, v) ≤ f
0
(x − tv, v) +
Z
t
0
ρ(s, x + (s − t)v) ds.
(2.5)
We an boundtherighthandside termbyits
L
∞
x
norm. Integratingwithrespe ttov
provideskρ(t, ·)k
L
∞
(R
2
)
≤ 2πkf
0
k
L
∞
(Ω)
+ 2π
Z
t
0
kρ(s, ·)k
L
∞
(R
2
)
ds.
We obtain a bound on the
L
∞
norm on
ρ
thanks to Gronwall's inequality and on lude theproof with (2.5).
Proofof Theorem2.2. Thelo al-in-timeexisten eisobtainedbyaxedpointargument.
Let
t
0
> 0
, the mapF
onL
1
((0, t
0
); L
∞
(Ω))
is dened for allf
by :F(f)
is a weak solutionof the problem
∂
t
F + v · ∇
x
F =
Z
v
′
∈V
φ(∇
x
G ∗ (v
′
ρ − J))F(v
′
) dv
′
− 2πφ(∇
x
G ∗ (vρ − J))F(v),
F(0, ·, ·) = f
0
,
whereρ =
R
V
f (t, x, v) dv
andJ =
R
V
vf (t, x, v) dv
. We willshow that this map denes a on-tra tiononL
1
((0, τ ), L
∞
(Ω))
forτ
smallenough. Letf
1
andf
2
begiven inL
1
((0, t
0
), L
∞
(Ω))
and denotingF
12
= F(f
1
) − F(f
2
)
we have∂
t
F
12
+ v · ∇
x
F
12
=
Z
V
φ(∇
x
· G ∗ (vρ
1
− J
1
))F
12
(v
′
) dv
′
− 2πφ(∇
x
· G ∗ (vρ
1
− J
1
))F
12
−2πF(f
2
)(φ(∇
x
· G ∗ (vρ
1
− J
1
)) − φ(∇
x
· G ∗ (vρ
2
− J
2
))
+
Z
V
F(f
2
)(v
′
)(φ(∇
x
· G ∗ (v
′
ρ
1
− J
1
)) − φ(∇
x
· G ∗ (v
′
ρ
2
− J
2
)) dv
′
,
with thenotations
ρ
i
=
R
V
f
i
(t, x, v) dv
andJ
i
=
R
V
vf
i
(t, x, v) dv
. We an rewritethis identityinthe following way
∂
t
F
12
+ v · ∇
x
F
12
+ 2πφ(∇
x
· G ∗ (vρ
1
− J
1
))F
12
= G,
(2.6) whereG(t, x, v) =
Z
V
φ(∇
x
· G ∗ (v
′
ρ
1
− J
1
))F
12
(v
′
) dv
′
− 2πF(f
2
)(φ(∇
x
· G ∗ (vρ
1
− J
1
))−
φ(∇
x
· G ∗ (vρ
2
− J
2
)) +
Z
V
F(f
2
)(v
′
)(φ(∇
x
· G ∗ (v
′
ρ
1
− J
1
)) − φ(∇
x
· G ∗ (v
′
ρ
2
− J
2
)) dv
′
.
(2.7)Usingthe hara teristi sof the system, we an rewriteequation (2.6) as
d
ds
e2πφ(∇
x
· G ∗ (vρ
1
− J
1
))(τ, x + v(τ − t))F
12
(s, x + v(s − t), v)
=
e2πφ(∇
x
· G ∗ (vρ
1
− J
1
))(τ, x + v(τ − t))G(s, x + v(s − t), v).
Atthe initialtime,
F
12
(0, ·, ·) = 0
. Integrating the latterequality between0
andt
, we haveF
12
(t, x, v) =
Z
t
0
exp
2π
Z
s
t
φ(∇
x
· G ∗ (vρ
1
− J
1
))(τ, x + v(τ − t)) dτ
G(s, x+v(s−t), v) ds.
Sin e
φ
(1.4) is bounded from below by1/4
we dedu e that for all0 < t < t
0
,|F
12
(t, x, v)| ≤
Z
t
0
|G(t − s, x − vs, v)| ds.
(2.8)
Moreover, from(2.7) and the assumptionson
φ
(1.4), we dedu e|G| ≤
Z
V
|F
12
(v
′
)| dv
′
+ kφ
′
k
∞
2πF(f
2
) +
Z
V
F(f
2
)(v
′
) dv
′
×
× (|∇
x
G ∗ (J
1
− J
2
)| + V |∇
x
G ∗ (ρ
1
− ρ
2
)|) ,
where
V = max
v∈V
kvk
. Noti ingthat|J
1
− J
2
| ≤ V |ρ
1
− ρ
2
|
, we have moreoverk|∇
x
G ∗ (J
1
− J
2
)(t, ·)| + V |∇
x
G ∗ (ρ
1
− ρ
2
)(t, ·)|k
L
∞
(R
2
)
≤ V k∇
x
Gk
L
1
(R
2
)
k(ρ
1
− ρ
2
)(t, ·)k
L
∞
(R
2
)
.
Finally,from (2.8)wededu e the bound
kF
12
(t, ·, ·)k
L
∞
(Ω)
≤ C
1
Z
t
0
kF
12
(t − s, ·, ·)k
L
∞
(Ω)
ds+
C
2
Z
t
0
kF(f
2
)(t − s, ·, ·)k
L
∞
(Ω)
k∇
x
Gk
L
1
(R
2
)
k(ρ
1
− ρ
2
)(t − s, ·)k
L
∞
(R
2
)
ds.
Therefore, using a Gronwall Lemma, we on lude that for
τ > 0
small enough,F
denes aontra tion on
L
1
((0, τ ), L
∞
(Ω))
. It allows to onstru t a unique solution as the xed pointof the map
F
on the interval(0, τ )
. Using the a priori estimates established in Lemma 2.3,we an extend this solution on
(0, t
0
)
for allt
0
> 0
and we have a bound on this solution inL
∞
For the proof of the se ond point of Theorem 2.2, let us assume that
f
0
∈ W
2,∞
(Ω)
andthat
t
0
> 0
is xed. By dierentiating with respe t tox
1
the kineti equation (2.3) satised byf
, we obtain∂
t
∂
x
1
f + v · ∇
x
∂
x
1
f + 2πφ(∇
x
· G ∗ (vρ − J))∂
x
1
f =
R
V
φ(∇
x
· G ∗ (v
′
ρ − J))∂
x
1
f (v
′
) dv
′
+
R
V
∇
x
G ∗ (v
′
∂
x
1
ρ − ∂
x
1
J)φ
′
(∇
x
G ∗ (v
′
ρ − J))f(v
′
) dv
′
−2π∇
x
G ∗ (v∂
x
1
ρ − ∂
x
1
J)φ
′
(∇
x
G ∗ (vρ − J))f(v).
The righthand side isbounded by
Z
V
|∂
x
1
f (v
′
)| dv
′
+ 8πV kφ
′
k
L
∞
kfk
L
∞
(Ω)
k∇
x
Gk
L
1
(Ω)
Z
V
|∂
x
1
f (v
′
) dv
′
|.
Integrating alongthe hara teristi sandpro eedingasabove,weobtainthatforall
0 ≤ t ≤ t
0
,|∂
x
1
f (t, x, v)| ≤ C
1
|∂
x
1
f
0
(x − tv, v)| + C
2
Z
t
0
Z
V
|∂
x
1
f (t − s, x − sv, v)| ds,
where
C
1
andC
2
stand for nonnegative onstants depending only ont
0
and on the data. Bythe same token as in proof of Lemma 2.3 using Gronwall's inequality, we obtain a bound on
∂
x
1
f
inL
∞
((0, t
0
); L
∞
(Ω))
. Dierentiating (2.3) with respe t tox
2
andv
, we dedu e by the same token that for allt ∈ (0, t
0
)
, we havef (t, ·, ·) ∈ W
1,∞
(Ω)
. With a similar argument we
dedu e after straightforward al ulationsthat
f
isbounded inL
∞
((0, t
0
); W
2,2
(Ω))
. Moreover from (2.3) we have an expression of∂
t
f
with respe t tof
,ρ
,J
,∇
x
G
and∇
x
f
allowing to obtain aboundon∂
t
f
inL
∞
((0, t
0
); L
∞
(Ω))
.Remark 2.4 Forthesakeof simpli ity,thisexisten eresulthasbeenestablishedinthedomain
ω = R
2
. However, this existen e and uniqueness result is still available in a bounded domain
ω ⊂ R
2
providedthattheboundary onditionsallows tousetheellipti regularityfortheellipti
equationsatised by
S
. For numeri al analysis,we will onstraint the domainω
to be bounded and make use of this existen e result in this framework.3 Numeri al approa h
Inthisse tion,wepresentthenumeri alapproa hforsolving(1.5). The omputationaldomain
isdenedby
(x, v) ∈ Ω = ω×S
V
whereω
isare tangulardomainofR
2
,
ω = [0, L
x
]×[0, ℓ
y
]
,andS
V
is thesphereS
V
= {v ∈ R
2
,
su hthat
kvk
2
= V }
, foragiven onstant velo ityV > 0
. Weredu e this 4-dimensional problem to a 3-dimensional problem by onsidering the ylindri al
oordinate
θ
.At the boundary of the domain, we assume to have spe ular ree tion at the boundaries
on entration
S
we set Neumann onditions onthe entire boundary :f (x, 0, θ) = f (x, 0, 2π − θ),
forθ ∈ [0, π], x ∈ [0, L
x
],
f (x, ℓ
y
, θ) = f (x, ℓ
y
, 2π − θ),
forθ ∈ [π, 2π], x ∈ [0, L
x
],
f (0, y, θ) = f (0, y, π − θ),
forθ ∈ [0, π/2] ∪ [3π/2, 2π], y ∈ [0, ℓ
y
],
f (L
x
, y, θ) = f (L
x
, y, π − θ),
forθ ∈ [π/2, 3π/2], y ∈ [0, ℓ
y
],
∂
x
S(0, y) = ∂
x
S(L
x
, y) = 0,
fory ∈ [0, ℓ
y
],
∂
y
S(x, 0) = ∂
y
S(x, ℓ
y
) = 0,
forx ∈ [0, L
x
].
(3.1)Obviously, inthe
θ
dire tionf is2π
-periodi . Weintrodu ethenodes(x
i
= i h
x
)
i=0,··· ,N
x
−
1
,(y
j
= j h
y
)
j=0,··· ,N
y
−
1
and(θ
k
= k h
θ
)
θ=0,··· ,N
θ
−
1
where
h
x
= L
x
/(N
x
− 1)
,h
y
= ℓ
y
/(N
y
− 1)
andh
θ
= 2π/N
θ
. We denotex
i
= (x
i
1
, y
i
2
)
withi = i
1
+ i
2
N
x
and wemeshthe domainω
withre tangulartrianglesusingthenodesx
i
. There-forethetriangulationisregularandallniteelementsareaneequivalenttoasinglereferen eelement. We denote the time step
∆t
and sett
n
= n ∆t
for
n = 0, . . . , N
t
.3.1 Semi-Lagrangian methods
In this se tion we des ribe the semi-Lagrangian method used for the numeri al resolution in
ω ⊂ R
2
of the kineti equation:
∂
t
f + v · ∇
x
f =
Z
S
V
φ(∂
t
S + v
′
· ∇
x
S)f (v
′
) dv
′
− 2πφ(∂
t
S + v · ∇
x
S)f (v),
(3.2)ompleted with the initial datum (1.2). We assume in this se tion that the hemoattra tant
on entration is known and we will therefore onsider the turningkernel as a fun tion of
t
,x
andv
:T (t, x, v) = φ(∂
t
S(t, x) + v · ∇
x
S(t, x))
. Its numeri al approximation will be denotedT
h
.Semi-Lagrangian methods onsist in al ulating the distribution fun tion at time
t
n+1
=
t
n
+ ∆t
thanks to the one whi h has been obtained at the time
t
n
by using the onservation
relationalongthe hara teristi s urves. Werstdenethe hara teristi s
(X, Θ)
ofthesystem for all0 ≤ s ≤ t
by :
d
ds
X(s; x, θ, t) = v
Θ
,
withv
Θ
= (V cos Θ, V sin Θ) ;
X(t; x, θ, t) = x,
d
ds
Θ(s; x, θ, t) = 0 ;
Θ(t; x, θ, t) = θ,
(3.3)
if
X(s; x, θ, t) ∈ (0, L
x
) × (0, ℓ
y
)
. Therefore the velo ity remains onstant, ex ept when the traje torymeetsaboundaryofthedomain: ifthereexistsatimes > 0
su hthatX(s; x, θ, t) ∈
(0, L
x
)×{0, ℓ
y
}
thentheangleΘ(s; x, θ, t)
is hangedinto2π−Θ(s; x, θ, t)
;ifthereexistsatimes > 0
su hthatX(s; x, θ, t) ∈ {0, L
x
}×(0, ℓ
y
)
thenΘ(s; x, θ, t)
issubstitutedbyπ−Θ(s; x, θ, t)
.More pre isely, wehave
X(s; x, θ, t) = x + v
θ
(s − t),
withv
θ
= (V cos θ, V sin θ),
Θ(s; x, θ, t) = θ,
for alls ∈ R
+
for whi h the traje tory does not ross the boundaries. By ree tion, if the
t < t
0
< s
,X(s; x, θ, t) = x + v
θ
(t
0
− t) + v
2π−θ
(s − t
0
)
andΘ(s; x, θ, t) = 2π − θ.
Ifthe traje tory rosses the boundary
x = 0
orx = L
x
attimet
1
, thenfort < t
1
< s
, wehaveX(s; x, θ, t) = x + v
θ
(t
1
− t) + v
π−θ
(s − t
1
)
andΘ(s; x, θ, t) = π − θ.
Obviously, if the traje tory meets several times the boundaries, we make others ree tions.
Usingthe hara teristi s,we an rewrite the kineti equation (3.2) inthe following way :
d
ds
f (s, X(s; x, θ, t), Θ(s; x, θ, t)) =
Z
2π
0
(T (s, X(s; x, θ, t), θ
′
)f (s, X(s; x, θ, t), θ
′
) dθ
′
−2π T (s, X(s; x, θ, t), Θ(s; x, θ, t)) f(s, X(s; x, θ, t), Θ(s; x, θ, t))).
(3.4)The semi-Lagrangian method relies on a dis rete approximation of relation (3.4). We
assumetoknow thedistribution fun tionattime
t
n
. Weuse anexpli itintime Eulers heme
to ompute this quantity attime
t
n+1
= t
n
+ ∆t
on ea h point
(x
i
, θ
j
)
of the grid. It leads to the followingsystem :f
h
(t
n+1
, x
i
, θ
j
) = f
h
(t
n+1
, X(t
n
; x
i
, θ
j
, t
n+1
), Θ(t
n
; x
i
, θ
j
, t
n+1
))
+∆t
Z
2π
0
(T
h
(t
n
, X(t
n
; x
i
, θ
j
, t
n+1
), θ
′
)f
h
(X(t
n
; x
i
, θ
j
, t
n+1
), θ
′
) dθ
′
−2π∆t T
h
(t
n
, X(t
n
; x
i
, θ
j
, t
n+1
), Θ(t
n
; x
i
, θ
j
, t
n+1
)) f
h
(X(t
n
; x
i
, θ
j
, t
n+1
), Θ(t
n
; x
i
, θ
j
, t
n+1
))),
(3.5)where
f
h
andT
h
stand forapproximationsoff
andT
. Forthe sakeof larity,we willuse fromnow on the notations
X
n
i,j
(s)
instead ofX(s; x
i
, θ
j
, t
n
)
andΘ
n
i,j
(s)
instead ofΘ(s; x
i
, θ
j
, t
n
))
.The dis retizationrelies on two main steps :
1. Findthepoint
(X(t
n
; x
i
, θ
j
, t
n
+∆t), Θ(t
n
; x
i
, θ
j
, t
n
+∆t)) := (X
i,j
n+1
(t
n
), Θ
n+1
i,j
(t
n
))
. Start-ingfrom(x
i
, θ
j
)
,itsu es tofollowba kward the hara teristi s urvesduring the time step∆t
. Tothis end, wehave tosolve (3.3). Sin e the resolution ofthe se ond equation in(3.3) is lear, this step is simple (ina generalframeworksee e.g. [39, 2,15℄)X
i,j
n+1
(t
n
) = x
i
− ∆t v
Θ
n+1
i,j
,
Θ
n+1
i,j
(t
n
) = θ
j
.
Obviously, if the traje tory rea hes a boundary of the domain, then we use spe ular
ree tion :
X
n+1
i,j
(t
n
)
is repla ed by its symmetri with respe t to this boundary andΘ
n+1
i,j
(t
n
)
by
2π − Θ
n+1
i,j
(t
n
)
at the verti al boundaryy = 0
ory = ℓ
y
, orbyπ − Θ
n+1
i,j
(t
n
)
atthe boundaryx = 0
orx = L
x
.Therefore the omputationof the footof the hara teristi sis exa t.
2. Sin e the fun tions
f
andT
at timet
n
are only known on the nodes of the mesh, we
interpolate these fun tions at the points
(X
n+1
i,j
(t
n
), Θ
n+1
i,j
(t
n
))
. A tually, sin e we haveΘ
n+1
i,j
(t
n
) = θ
j
orπ − θ
j
or2π − θ
j
, by takingN
θ
odd we only have a 2D interpolationmethod to implement. In fa t, we have that
2π − θ
j
= θ
N
θ
−
j
andπ − θ
j
= θ
N
θ
/2−j
ifan use a linear interpolation : we dene the linear interpolation operator
Π
onto thespa e of Lagrangian polynomialsof degree lesser than orequal to1 by :
Πf (x, y) =
N
x
N
y
−
1
X
i=0
f (x
i
)l
i
(x, y),
withl
i
∈ P
1
(ω)
su h thatl
i
(x
j
) = δ
ij
,
(3.6) whereP
1
(ω)
is the set of pie ewise polynomialswhose restri tion onea htriangleof the
meshholdsinpolynomialsfun tionof degreelesserthanorequalto1. Theinterpolation
error estimations inSobolev spa esgive that for all
f ∈ W
2,2
(ω)
(see [13℄ Theorem 16.2
p. 128)
kf − Πfk
L
2
(ω)
≤ C max{h
2
x
, h
2
y
} kfk
W
2,2
(ω)
.
(3.7)Moreover, this linear interpolationmethodpreserves the
L
∞
bound :
kΠfk
L
∞
(ω)
≤ kfk
L
∞
(ω)
.
(3.8)For more pre ision on solutions, a Hermite spline interpolation method whi h is a well
established high order interpolation method an be used. We refer to [39, 15℄ for more
details about this interpolation step. However, spurious os illations (e.g. Runge
phe-nomena) an appear with high order interpolationmethods.
Finally,usingadis retizationofthe integralin(3.5),theapproximationofthe distribution
fun tion attime
t
n+1
is obtained by
f
h
(t
n+1
, x
i
, θ
j
) = Πf
h
(t
n
, X
i,j
n+1
(t
n
), Θ
i,j
n+1
(t
n
)) − 2π∆t Π(T
h
f
h
)(t
n
, X
i,j
n+1
(t
n
), Θ
n+1
i,j
(t
n
))+
+∆t
N
X
θ
−
1
k=0
h
θ
Π(T
h
f
h
)(t
n
, X
i,j
n+1
(t
n
), θ
k
).
(3.9)
3.2 Numeri al resolution of the oupled system
We re all the following notations :
X
n+1
i,j
(t
n
)
instead ofX(t
n
; x
i
, θ
j
, t
n+1
)
,Θ
n+1
i,j
(t
n
)
instead ofΘ(t
n
; x
i
, θ
j
, t
n+1
))
andΠ
istheinterpolationoperator. Thenumeri alresolutionofthe oupledproblem (1.5) is then ta kled in the following way. We assume that the approximation of
the distribution fun tion
f
h
and of the hemoattra tant on entrationS
h
are known at timet
n
at ea h nodes of the mesh. We des ribe the pro ess to ompute
f
h
andS
h
at timet
n+1
.As des ribed previously, the distribution fun tion at time
t
n+1
is approximated thanks to the
relation:
f
h
(t
n+1
, x
i
, θ
j
) = Πf
h
(t
n
, X
i,j
n+1
(t
n
), Θ
i,j
n+1
(t
n
)) − 2π∆t Π(T
h
f
h
)(t
n
, X
i,j
n+1
(t
n
), Θ
n+1
i,j
(t
n
))+
+∆t
N
X
θ
−
1
k=0
h
θ
Π(T
h
f
h
)(t
n
, X
i,j
n+1
(t
n
), θ
k
),
(3.10)
where the turning kernel
T
h
is omputed by adis retization of (1.3). As we have seen for theto get an estimation of the term involving temporal derivativein the denition of
T
. In fa t, the quantity∂
t
S
satises−∆∂
t
S + ∂
t
S = −∇ · J,
(3.11)ompletedwith Neumannboundary onditionsdedu ed from(3.1). We denethe
approxima-tion of the urrent by
J
h
(t
n
, x
i
) =
N
X
θ
−
1
k=0
h
θ
v
θ
k
f
h
(t
n
, x
i
, θ
k
).
(3.12)Then the approximation of
∂
t
S
, denotedS
th
, is omputed by solving (3.11) withJ
h
usingonforming
P
1
nite elements :∀ V
h
∈ X
h
,
Z
ω
(∇S
th
· ∇V
h
+ S
th
V
h
) dx =
Z
ω
J
h
· ∇V
h
dx,
(3.13)where
X
h
isthe set of fun tion ofC
0
(ω)
whose restri tion toea h triangleon the mesh holds
in
P
1
. Wedene the approximation
T
h
of the turningkernel byT
h
(t
n
, x
i
, θ
j
) = φ
S
th
(t, x) +
1
∆t
(S
h
(t
n
, x
i
) − ΠS
h
(t
n
, X
i,j
n
(t
n−1
))
.
(3.14)On e
f
h
is known at timet
n+1
, the hemoattra tant on entration is updated by solving
the followingellipti equation with onforming
P
1
nite elements:−∆S
h
(t
n+1
, x
i
) + S
h
(t
n+1
, x
i
) = ρ
h
(t
n+1
, x
i
) :=
N
X
θ
−
1
k=0
f
h
(t
n+1
, x
i
, θ
k
) h
θ
.
(3.15)This system is ompleted with boundary onditions (3.1).
From basi error estimates on ellipti problem,wehave (see [13℄ Theorem 18.1 p. 138) :
Proposition 3.1 Let
u
be a solution of the variational problema(u, v) = l(v)
wherea
isbilinear ontinuous symmetri oer ive on
H
1
(ω)
and
l
is linear ontinuous onH
1
(ω)
. Then,
if
u
h
is the dis rete approximation omputed by onformingP
1
nite elements, there exists a
nonnegative onstant
C
su h thatku − u
h
k
H
1
(ω)
≤ C max{h
x
, h
y
}kuk
H
2
(ω)
.
3.3 Convergen e analysis
We are interested in this se tion in the onvergen e of the s heme (3.10)(3.15) towards
so-lutions of model (1.5). Convergen e analysis of semi-Lagrangianmethodin the framework of
Vlasov-Poissonsystem havebeenobtained in[3℄;thoseresultsarepresented in
L
∞
. Duetothe
la k ofregularity inour ase ausedby the presen e ofthe turningoperator,we presentinthe
following theorem a onvergen e result in
L
2
. The main result of this se tion is presented in
Theorem3.4 belowunderthefollowingadditionalassumptionallowingtosimplifyallintegrals
allparti lesareinitially onnedinthe enter ofthedevi eandthatthetime
t
0
issmallenough to avoid parti les to meet the boundary of the domain. Thus all traje tories are straight linesand
X(s; x, θ, t) = x + v
θ
(s − t);
Θ(s; x, θ, t) = θ,
for0 ≤ s ≤ t.
Moreover, sin e
f
h
vanishes near the boundary of the domainω
, we haveZ
ω
f
h
(t, X(s; x, θ, t), θ) dx =
Z
ω
f
h
(t, x, θ) dx,
for
0 ≤ s ≤ t
small enough.From now on, we x
t
0
> 0
small enough and assumef
0
∈ W
2,2
∩ W
1,∞
(Ω)
is hosen
su h that Assumption 3.2holds. Therefore, in this onvergen e analysis ells donot 'see' the
boundary. Thes heme (3.10)(3.15)allowstodenetheapproximatedfun tion
f
h
onlyonthenodes on the mesh. We extend this denition on all
(x, θ) ∈ ω × [0, 2π]
thanks to the linearinterpolation operator
Π
(3.6). We rst establish the positivity and a priori estimates on thedis reteapproximation
f
h
.Lemma 3.3 Let
t
0
> 0
and assume thatf
0
∈ W
2,2
∩ W
1,∞
(ω × [0, 2π])
is a nonnegative
fun tion su h that Assumption 3.2 holds. If
∆t ≤ 1/(2π)
, then the s heme dened by (3.10)(3.15) givesa nonnegative approximation
f
h
of thedistribution fun tion. Moreoverf
h
satises the followingestimate∀ t ∈ (0, t
0
),
kΠf
h
(t, ·, ·)k
L
∞
(ω×[0,2π])
≤ e
2πt
0
kf
0
k
L
∞
(ω×[0,2π])
,
kΠf
h
(t, ·, ·)k
H
1
(ω×[0,2π])
≤ C
0
,
where
C
0
is a nonnegative onstant depending ont
0
andkf
0
k
H
1
(ω×[0,2π])
.Proof. From the denition of (3.6) and the assumptions on
φ
(1.4) we have that for allnonnegative fun tion
f
1
4
Πf ≤ Π(T f) ≤ Πf.
Therefore assuming
f (t
n
, ·, ·)
nonnegative, we dedu e from(3.10) that
f
h
(t
n+1
, x
i
, θ
j
) ≥ (1 − 2π∆t) Πf
h
(t
n
, x
i
− v
θ
j
∆t, θ
j
) ≥ 0.
Moreover, from(3.10) and (3.8) we have
kΠf
h
(t
n+1
, ·, ·)k
L
∞
(ω×[0,2π])
≤ (1 + 2π∆t) kΠf
h
(t
n
, ·, ·)k
L
∞
(ω×[0,2π])
.
Applying adis reteGronwallinequalityallowsto on lude theproof ofthe
L
∞
bound.
Dier-entiating (3.10)with respe t to
x
1
gives∂
x
1
f
h
(t
n+1
, x
i
, θ
j
) = ∂
x
1
Πf
h
(t
n
, x
i
− v
θ
j
∆t, θ
j
)
−2π∆t ∂
x
1
Π(T
h
f
h
)(t
n
, x
i
− v
θ
j
∆t, θ
j
) + ∆t
P
N
θ
−
1
k=0
h
θ
∂
x
1
Π(T
h
f
h
)(t
n
, x
i
− v
θ
j
∆t, θ
k
).
(3.16)By linearity of the interpolationoperator,we have that
From the denition of
T
h
(3.14),we dedu e thatk∂
x
1
(ΠT
h
)(t
n
)k
L
2
(ω)
≤ k∂
x
1
S
th
(t
n
)k
L
2
(ω)
+
1
∆t
k∂
x
1
S
h
(t
n
, ·) − ∂
x
1
ΠS
h
(t
n
, X(t
n−1
; ·, θ, t
n
))k
L
2
(ω)
.
From the ellipti regularity onthe equation(3.13) and the bound of
f
h
inL
∞
(ω × [0, 2π])
, wehavethat
S
th
isbounded inH
1
(ω)
. The ellipti regularity on (3.15)gives
kS
h
(t
n
, ·) − ΠS
h
(t
n
, X(t
n−1
; ·, θ, t
n−1
))k
H
1
(ω)
≤ kρ
h
(t
n
, ·) − Πρ
h
(t
n
, X(t
n−1
; ·, θ, t
n
))k
L
2
(ω)
.
Usinga Taylor expansion gives
ρ
h
(t
n
, x) − Πρ
h
(t
n
, x − v
θ
∆t) =
Z
0
−
∆t
v
θ
· ∇Πρ
h
(t, x + v
θ
s) ds.
From Assumption 3.2wededu e that
kρ
h
(t
n
, x) − Πρ
h
(t
n
, x − v
θ
∆t)k
L
2
(ω)
≤ ∆tV k∇(Πρ
h
)(t
n
, ·)k
L
2
(ω)
.
Thus,
k∂
x
1
(ΠT
h
)k
L
2
(ω×[0,2π])
≤ C(1 + k∇(Πρ
h
)k
L
2
(ω×[0,2π])
).
Doingthe same with
x
2
,(3.16) leads to the following estimatek∇Πf
h
(t
n+1
)k
L
2
(ω×[0,2π])
= k∇Πf
h
(t
n
)k
L
2
(ω×[0,2π])
+ C∆t(1 + k∇Πf
h
(t
n
)k
L
2
(ω×[0,2π])
).
A Gronwallinequality allows to on lude the proof.
Theorem 3.4 Let
t
0
> 0
and assume thatf
0
∈ W
2,2
∩ W
1,∞
(ω × [0, 2π])
is a nonnegative
fun tion su h that Assumption 3.2 holds. Let
f
be the global weak solution of (1.5) on(0, t
0
)
and
f
h
be its approximation omputed atthe nodes of themesh thanksto the algorithm(3.10)(3.15) where
Π
is dened in (3.6). Then there exists a nonnegative onstantC
depending ont
0
,f
0
and the data su h that
kf − Πf
h
k
L
∞
(0,t
0
;L
2
(ω×[0,2π]))
≤ C(∆t + h
2
+
h
∆t
+ h),
where
h = max{h
x
, h
y
, h
θ
}
.Proof. From Theorem 2.2wehave
f (t, ·, ·) ∈ W
1,∞
∩ W
2,2
(ω × [0, 2π])
We denethe global
error attime
t
n+1
by
ǫ
n+1
= kf(t
n+1
, x, θ) − Πf
h
(t
n+1
, x, θ)k
L
2
(ω×[0,2π])
.
(3.17) A rst remark isthat sin e hara teristi s are straight lines the numeri al omputationof thehara teristi s
(X, Θ)
isexa t. From (1.5), wededu e that for0 ≤ s ≤ t
d
ds
f (s, x − v
θ
(t − s), θ) =
Z
2π
0
T (s, x − v
θ
(t − s), θ
′
)f (s, x − v
θ
(t − s), θ
′
) dθ
′
−2π T (s, x − v
θ
(t − s), Θ
n
(s))f (s, x − v
θ
(t − s), θ),
T (t, x, θ) = φ(∂
t
S + v
θ
· ∇
x
S),
and
S
being the solution of the ellipti problem−∆S + S =
Z
2π
0
f (t, x, θ) dθ.
We dedu e from the regularity of the fun tion
f
proved in Theorem 2.2 that the fun tions 7→ f(s, x − v
θ
(t − s), θ)
is bounded inW
2,2
(0, t
0
)
. Hen e a Taylor expansion gives, underAssumption 3.2:
f (t
n+1
, x, θ) = f (t
n
, x − v
θ
∆t, θ) + ∆t
Z
2π
0
T (t
n
, x − v
θ
∆t, θ
′
)f (t
n
, x − v
θ
∆t, θ
′
) dθ
′
−2π∆t T (t
n
, x − v
θ
∆t, θ)f (t
n
, x − v
θ
∆t, θ) + O
L
2
(∆t
2
),
whereO
L
2
(∆t)
means that there existsC > 0
su h thatkO
L
2
(∆t
2
)k
L
2
≤ C∆t
2
. Therefore, using the denition (3.10), we rewritethe dieren ef (t
n+1
, x, θ) − Πf
h
(t
n+1
, x, θ)
asf (t
n+1
, x, θ) − Πf
h
(t
n+1
, x, θ) = f (t
n
, x − v
θ
∆t, θ) − Πf
h
(t
n
, x − v
θ
∆t, θ)
+∆t
Z
2π
0
(T f )(t
n
, x − v
θ
∆t, θ
′
) dθ
′
−
N
X
θ
−
1
k=0
h
θ
Π(T
h
f
h
)(t
n
, x − v
θ
∆t, θ
k
)
!
−2π∆t ((T f)(t
n
, x − v
θ
∆t, θ) − Π(T
h
f
h
)(t
n
, x − v
θ
∆t, θ)) + O
L
2
(∆t
2
).
To evaluate the global error
ǫ
n+1
, we de omposef (t
n+1
, x, θ) − f
h
(t
n+1
, x, θ)
asf (t
n+1
, x, θ) − Πf
h
(t
n+1
, x, θ) = f (t
n
, x − v
θ
∆t, θ) − Πf
h
(t
n
, x − v
θ
∆t, θ)
+∆t((1 − Π)I
1
+ I
2
+ I
3
) + O
L
2
(∆t
2
),
(3.18) whereI
1
=
N
X
θ
−
1
k=0
h
θ
(T f )(t
n
, x − v
θ
∆t, θ
k
) − 2π(T f)(t
n
, x − v
θ
∆t, θ),
(3.19)I
2
=
N
X
θ
−
1
k=0
h
θ
(T f − T
h
f
h
)(t
n
, x − v
θ
∆t, θ
k
) − 2π(T f − T
h
f
h
)(t
n
, x − v
θ
∆t, θ),
(3.20)I
3
=
Z
2π
0
(T f )(t
n
, x − v
θ
∆t, θ
′
) dθ
′
−
N
X
θ
−
1
k=0
h
θ
(T f )(t
n
, x − v
θ
∆t, θ
k
).
(3.21) TakingtheL
2
normof (3.18) implies with Assumption 3.2
ǫ
n+1
≤ ǫ
n
+ ∆t(k(1 − Π)I
1
k
L
2
(ω×[0,2π])
+ kI
2
k
L
2
(ω×[0,2π])
+ kI
3
k
L
2
(ω×[0,2π])
) + C∆t
2
.
(3.22)Wewillestimate ea h term separately thanks tothe following Lemmata.
Lemma 3.5 Let
(f, S)
being solution of (1.5) forf
0
∈ W
1,∞
∩ W
2,2
(ω × [0, 2π])
. If
I
1
is dened by (3.19), then there exists a nonnegative onstantC
su h that for allt ∈ (0, t
0
)
,k(1 − Π)I
1
k
L
2
(ω×[0,2π])
≤ C max{h
2
Proof. Let
0 ≤ t ≤ t
0
. From Theorem 2.2, we have thatf (t, ·, ·) ∈ W
1,∞
∩ W
2,2
(ω × [0, 2π])
.Therefore, by ellipti regularity, we dedu e that
∇S(t, ·) ∈ W
3,2
(ω)
. Moreover, (3.11) implies
−∆∂
t
S(t, ·) + ∂
t
S(t, ·) = −∇ · J(t, ·) ∈ W
1,2
(ω).
Then theellipti regularityfurnishes∂
t
S(t, ·) ∈ W
3,2
(ω)
and sin ewith our denition on
φ
weget
φ
′
∈ C
∞
c
(R)
,wededu e thatT (t, ·, ·) = φ(∂
t
S + v
θ
· ∇S) ∈ W
3,2
(ω × [0, 2π])
. Then∂
x
2
i
x
j
(T f ) = (∂
x
i
x
j
T )f + ∂
x
i
T ∂
x
j
f + ∂
x
j
T ∂
x
i
f + T (∂
x
i
x
j
f )
∈ L
2
(ω × [0, 2π]).
Hen e(T f )(t, ·, ·) ∈ W
2,2
(ω × [0, 2π])
. Finally, the result of the Lemma is a straightforward
onsequen e of the interpolation error (3.7).
Lemma 3.6 Let
f
andT
beingdened previously,there existsa nonnegative onstantC
su hthat for all
0 ≤ t ≤ t
0
andx
∈ ω
,Z
2π
0
(T f )(t, x, θ
′
) dθ
′
−
N
X
θ
−
1
k=0
h
θ
(T f )(t, x, θ
k
)
L
2
(ω)
≤ Ch
2
θ
.
Proof. Let
0 ≤ t ≤ t
0
andx ∈ ω
. As noti ed in the proof of Lemma 3.5 we have that(T f )
belongsto
W
2,2
(ω ×[0, 2π])
. Hen etheresultofLemma3.6isa onsequen eofthewell-known
error estimatefor the trapezoidalrule : if
g ∈ W
2,2
(0, 2π)
, there exists
θ ∈ (0, 2π)
su h thatZ
2π
0
g(θ) dθ −
N
X
θ
−
1
k=0
h
θ
g(θ
k
)
= h
2
θ
π
6
∂
2
g(θ)
∂θ
2
.
(3.23)The two previous Lemmata allow us to estimate the terms involving
I
1
andI
3
in (3.22).For
I
2
, we need rst toestimate the errorT − T
h
with respe t tof − f
h
.Lemma 3.7 Let assume thatAssumption 3.2 holds. Let
T
be denedin (1.3)(1.4) andT
h
beits approximation omputed with (3.14). Then, there exists
C > 0
su h that forn = 1, . . . , N
t
,we have
k sup
α∈[0,2π]
|T (t
n
, ·, α) − ΠT
h
(t
n
, ·, α)|k
L
2
(ω)
≤ C(∆t + max{h
x
, h
y
} +
max{h
x
, h
y
}
∆t
+ h
2
θ
+
+kf(t
n
, ·, ·) − f
h
(t
n
, ·, ·)k
L
2
(ω×[0,2π])
).
Proof. Let
n ∈ {1, . . . , N
t
}
,x
∈ ω
andα ∈ [0, 2π]
. We have from(3.14) that|T (t
n
, x, α) − ΠT
h
(t
n
, x, α)| = |φ(∂
t
S(t
n
, x) + v
α
· ∇
x
S(t
n
, x)) − ΠT
h
(t
n
, x, α)|
≤ kφ
′
k
L
∞
(|∂
t
S(t
n
, x) − S
th
(t
n
, x)| + |v
α
· ∇
x
S(t
n
, x) −
1
∆t
(S
h
(t
n
, x) − ΠS
h
(t
n
, x − v
α
∆t))|),
(3.24)where
S
th
is dened in (3.13). We will estimateseparately ea h term of the sum of the righthand side. Letus introdu e
S
e
t
aweaksolution ofompleted with Neumann boundary onditions, where
J
h
is dened in (3.12) at the nodes ofthe mesh and extended on
ω
thanks tothe linear interpolationoperatorΠ
. From Proposition3.1, we dedu e that
k e
S
t
− S
th
k
L
2
(ω)
≤ C max{h
x
, h
y
}k e
S
t
k
W
2,2
(ω)
≤ C max{h
x
, h
y
}k∇ΠJ
h
k
L
2
(ω)
,
where the ellipti regularity on equation (3.25) is used. And Lemma 3.3 allows to bound the
term
k∇ΠJ
h
k
L
2
(ω)
. Moreover, from (3.11)and (3.25), we dedu e thatk∂
t
S − e
S
t
k
L
2
(ω)
≤ CkJ − ΠJ
h
k
L
2
(ω)
≤ C(h
2
θ
+ kf − Πf
h
k
L
2
(ω×[0,2π])
),
where wehave used the errorestimate given by the trapezoidalrule (3.23)to estimate
J − J
h
for
f ∈ W
2,2
(ω × [0, 2π])
. We on lude then that
k∂
t
S−S
th
k
L
2
(ω)
≤ k∂
t
S− e
S
t
k
L
2
(ω)
+k e
S
t
−S
th
k
L
2
(ω)
≤ C(h
2
θ
+kf −Πf
h
k
L
2
+max{h
x
, h
y
}).
(3.26) Weintrodu eS
e
weak solutionof the ellipti problem−∆ e
S + e
S = Πρ
h
,
(3.27)ompleted with Neumannboundary onditions. Therefore, wehave that
k sup
α∈[0,2π]
|v
α
· ∇
x
S − v
α
· ∇
x
e
S|k
L
2
(ω)
≤ V kS − e
Sk
W
1,2
(ω)
≤ Ckρ − Πρ
h
k
L
2
(ω)
≤ C(h
2
θ
+ kf − Πf
h
k
L
2
(ω×[0,2π])
),
(3.28)whereweusetheellipti regularityforequation(3.27)andtheerrorestimateforthetrapezoidal
rule (3.23). Moreover,
Πρ
h
belonging toL
∞
(ω) ֒→ L
2
(ω)
, we have by ellipti regularity on
(3.27) that
S ∈ W
e
2,2
(ω)
. A Taylorexpansion givesthat for all
x
∈ ω
and allα ∈ [0, 2π]
,e
S(t
n
, x) = e
S(t
n
, x − v
α
∆t) − v
α
· ∇ e
S(t
n
, x)∆t + O
L
2
(ω)
(∆t
2
).
Hen e, for all
x
∈ ω
k sup
α∈[0,2π]
|v
α
· ∇
x
S(t
e
n
, x) −
1
∆t
(S
h
(t
n
, x) − ΠS
h
(t
n
, x − v
α
∆t))|k
L
2
x
(ω)
≤ C∆t+
1
∆t
k e
S(t
n
, x) − S
h
(t
n
, x)k
L
2
x
(ω)
+
1
∆t
k sup
α∈[0,2π]
| e
S(t
n
, x − v
α
∆t) − ΠS
h
(t
n
, x − v
α
∆t)|k
L
2
x
(ω)
.
Assumption 3.2impliesthat the lasttwoterms of thesum are equals. Sin e
S
h
isobtained bysolving equation (3.27) with onforming
P
1
niteelements,Proposition 3.1impliesthat