Contents lists available atScienceDirect
Journal
of
Computational
Physics
www.elsevier.com/locate/jcp
Short note
Convergence
of
energy
stable
finite-difference
schemes
with
interfaces
Magnus Svärd
a,
∗
,
Jan Nordström
b,
caDept.ofMathematics,UniversityofBergen,P.O.Box7803,5020Bergen,Norway
bComputationalMathematics,DepartmentofMathematics,LinköpingUniversity,SE-58183Linköping,Sweden
cDepartmentofMathematicsandAppliedMathematics,UniversityofJohannesburg,P.O.Box524,AucklandPark2006,SouthAfrica
a
r
t
i
c
l
e
i
n
f
o
a
b
s
t
r
a
c
t
Articlehistory: Received6May2020
Receivedinrevisedform19November2020 Accepted20November2020
Availableonline1December2020
Keywords: Convergencerate Finitedifference Stability Well-posedness Interface
We extendtheconvergenceresultsinSvärdandNordström(2019)[7] forsingle-domain energy-stable high-orderfinite differenceschemes,to includedomainssplitintoseveral grid blocks. The analysis also demonstrates that reflective boundary conditions enjoy the sameconvergenceproperties.Finally, webrieflyindicatethattheseresults (andthe previousonesin[7])alsoholdinmultipledimensions.
©2020TheAuthor(s).PublishedbyElsevierInc.Thisisanopenaccessarticleunderthe CCBYlicense(http://creativecommons.org/licenses/by/4.0/).
1. Background
The convergencerateoffinite-differenceschemes thatare closedattheboundarieswithstencils oflower order accu-racy1 thantheinteriorstencilsisalong-standingproblemthathasbeentreatedextensivelyintheliteratureine.g.[3,4,1,5]. Recentlytheproblemwas addressed in[7],whereitwasshownthatenergy-stable schemes,satisfyingsome natural con-straints,automatically“gain”asmanyordersontheboundaryasthehighestspatialderivative.
Thecasewhenthedomainissplitinto(atleast)twocomputationaldomainsthatareconjoinedataninterface,wasnot treatedin[7].Theinterfacecaseisfundamentallydifferentfromtheboundarycase.Inthetheoryforboundaryconditions, the semi-discretisation ofthe initial-boundary value problemis Laplacetransformed. This allowsa decomposition ofthe modesintothose that decayintothedomain fromtheboundary,andshould be suppliedwithboundary conditions,and those thatdonot. Theboundednessofthe boundarydataandtheenergystability ofthescheme,enableda proofofthe convergenceresult.
Thisproof,however,doesnotimmediatelyencompassinterfaces,sincetheytakedatafromtheconjoiningdomain,which mayormaynotbesufficientlybounded.Moreover,thetreatmentataninterfaceneednotabidetotheminimalnumberof boundaryconditionsrequirement,sincedataataninterfacecanbeconsideredexact.Likewise,theprevious theorydidnot consider reflectiveboundaries,i.e., those thatfeedout-goingwaves backinto thedomainvia thein-going characteristics, sincesufficientboundednessofthesolutionthatisfedbackhastobeestablished.
*
Correspondingauthor.E-mailaddresses:Magnus.Svard@uib.no(M. Svärd),jan.nordstrom@liu.no(J. Nordström).
1 Convergencerateistheratebywhichthesolutionerrordecreases withgridrefinements.Theorderofaccuracyisgivenbythetruncationerrorand canbedifferentatdifferentpoints.Itrepresentshowaccuratelytheequation(notthesolution)isapproximated.
https://doi.org/10.1016/j.jcp.2020.110020
0021-9991/©2020TheAuthor(s).PublishedbyElsevierInc.ThisisanopenaccessarticleundertheCCBYlicense (http://creativecommons.org/licenses/by/4.0/).
Here,weaddressbothinterfacesandreflectiveboundaryconditionsandwediscusspossiblegeneralisationstomultiple dimensions.Specifically,werequire
(1) thatthesemi-discreteschemeisenergystable,hasapolynomialnullspace(possiblyempty),isnullspaceconsistent,and nullspaceinvariant.(Alltheassumptionsusedin[7].)
(2) that thenumber ofinterface conditions,as seenfrom oneside, is consistent withtheminimal number ofboundary conditions.Thatis,datafromacrosstheinterface canbereplaced bysmooth andboundeddata,andresultina well-posedproblem.
Remark.Thesecondrequirementimmediatelyridsoneoftheproblemsdiscussedabove.Itprohibitsover-specificationattheinterface. Furthermore,itisthemostrobustchoicethatismostoftenusedinpractice.
Theprogrammeforanalysinginterfaceaccuracyistoviewtheinterfaceasanexternalboundaryand,withtheaidofthe requirementsabove,theaprioriestimatesandprevioustheory,infertheoptimalconvergencerates.
Wewilldemonstratethisforafewexamples.Thegeneralisationsareobviousandwillonlybediscussedbriefly. 2. Thehyperboliccase
Weconsidertheadvectionequationontherealline,
ut
+
aux=
0,
a>
0.
(1)We splitthedomainintotwo: x
<
0 and x>
0 withaninterface atx=
0 anddiscretise eachwithaSummation-by-Parts (SBP) scheme (see [6]). (We assume appropriate trail-off conditions to the left andright and denote the left and right domain asL
,R.) The gridis equidistantwith indexingxi=
ih, i=
0,
±
1,
±
2,
...
,such that x0 representsthe interface in bothdomains.vtL
+
a P−1Q vL=
0 (2)vtR
+
a P−1Q vR= −
P−1a(
v0R−
v0L)
e0 where vL= (...
vL−1
,
vL0)
T and vR= (
v0R,
vR1,
...)
T are the unknowns to the left and right and e0= (
1,
0,
...
0)
T. The SBP operator has the following properties: P=
PT>
0, Q+
QT=
B=
diag(
−
1,
0,
...,
0,
1)
and P−1Q v is a first-derivative approximation.(See[7] formoreinformation.)Toadherewiththesecondrequirementabove,thereisonlyapenaltyterm totherightinagreementwiththecharacteristicdirection.Remark.Generally,onecanpenalisebothequationsatx0and,withaproperscaling,stabilitycanbeproven.Althoughstableand
convergent,sucha“two-way”couplingrenderstheleftproblemill-posedviewedasastand-aloneproblem.Hence,Requirement(2)is notsatisfied.
Weapplytheenergymethodinasomewhatunconventionalway.Multiplythefirstequationin(2) byvLP toobtain
vL2 t+
a(
v L 0)
2=
0.
The left problemis well-posed asa stand aloneproblemand we obtainthe bound vL0
∈
L2(
0,
T)
. Next,we multiply the secondequationin(2) by(
vR)
TP , vR2 t−
a(
v R 0)
2= −
2av0R(
v0R−
vL0).
Byrecastingtheboundaryterms,weobtain vR2 t+
a((
v R 0−
v0L)
2=
a(
v0L)
2.
This problemis bounded, andwell-posed,since vL0
∈
L2(
0,
T)
. Hence,the theory of[7] appliesto both theleft andright problemsindividuallyandweconcludethat(2) willconvergewithoptimalrates.Thisprocedurecanbegeneralisedtosymmetricsystems,ut
+
Aux=
0 discretisedbyvtL
+ (
A⊗
P−1Q)
vL= +(
A−⊗
P−1)((
vL0−
vR0)
⊗
e0L)
(3)vtR
+ (
A⊗
P−1Q)
vR= −(
A+⊗
P−1)((
vR0−
v0L)
⊗
e0R)
whereeL,R arebothoneatx
=
0 and zeroelsewhere.(FormoreinformationonKroneckerproducts, see[6].)ThedivisionA
=
A++
A− refers to thesigns ofthe eigenvalues such thatonly in-going characteristicsare penalisedon each side of the interface. As in the scalar case, the outgoing characteristics on each side serve as data for the other side and are independentlyboundedinL2(
0,
T)
suchthatoptimalratesareobtained.2.1. Reflectiveboundaryconditions
Another wayof looking at (1) is obtained by transforming the left domain, x
<
0, toξ >
0. Thatis, by x= −ξ
and∂x
= −∂
ξ,suchthat,utL
−
auξL=
0,
ξ >
0.
(4)Relabelling
ξ
asx,uL=
u1,uR=
u2,weobtainthesystem, u1 u2 t+
−
a 0 0 au1 u2 x
=
0,
x>
0withtheboundarycondition, u1
(
0,
t)
=
u2(
0,
t)
.(Inparticularwe note thattheproblemiswell-posed ifu2(
0,
t)
=
g(
t)
∈
L2
(
0,
T)
.) The previously derived apriori estimate guarantees that u1(
0,
t)
∈
L2(
0,
T)
andwe canview theproblemasa well-posedproblemwithareflective externalboundarythatconsequentlyenjoyoptimalconvergencerates.3. Theparaboliccase
Weconsiderthesamesetupfortheadvection-diffusionequation,ut
+
aux=
uxx,withaninterfaceatx
=
0 andtrail-off conditionstotheleftandright.TheSBP second-derivativeoperatorisgivenas:D2=
P−1(
−
A+
B S)
where A is symmet-ric positive semi-definite and S holds afirst-derivative stencilsat theboundary points. Following [2], theapproximation becomes,P vL
+
a Q vL=
(
−
A+
B S)
vL+
σ
1e0(
v0L−
v0R)
+
σ
2e
0((
SvL)
0− (
SvR)
0)
(5)P vR
+
a Q vR=
(
−
A+
B S)
vR+
σ
3e0(
vL0−
v00)
+
σ
4e
0((
SvR)
0− (
SvL)
0).
In[2] thestablechoicesofthe
σ
’sarederived.Furthermore,theseσ
’shavethefollowingdependenciesofh:σ
1,3∼
O(
h−1)
andσ
2,4∼
O(
1)
.Considertheleftproblem.Itisanapproximationoftheadvection-diffusionequationwithboundarycondition
α
LuL+
hβ
LuLx=
α
LuR+
hβ
LuxR=
f(
uR,
uRx).
(6)Asufficientconditionforthisstand-aloneproblemtobewell-posedisuR
∈
L2(
0,
T)
anduxR∈
L2(
0,
T)
.(Ormoreprecisely,huRx
∈
L2(
0,
T)
.)Furthermore,wesimultaneouslydemandthemirroredconditionsontherightside.Remark.NotethattheinterfaceconditionisanapproximationofuL
(
0,
t)
=
uR(
0,
t)
andnotaRobin-typeconditionsinceσ
1,3areO(
h−1)
andσ
2,4are
O(
1)
.Furthermore,parabolicityrequiresaboundaryconditiononallboundariesimplyingthataninterfaceconditiononbothsidesoftheinterfacedoesnotimplyover-specificationofthestand-aloneproblems,inaccordancewithRequirement (2).
Next, we investigatethe stability of the stand-aloneproblems. Thatis, we want thenumerical approximationsof (6) to be bounded in L2
(
0,
T)
.By perusing the stability proof in[2], itis clearthat(
SvL)
0 and(
SvR)
0 are indeed bounded in L2(
0,
T)
. (See p. 358-9. The eigenvalues associated with(
SvL)
0
+ (
SvR)
0 and(
SvL)
0− (
SvR)
0 are non-zero and non-vanishing.) However, a bound on vR0 or vL0,which isneeded forstand-alonewell-posedness of thetwo problems,is not obtainedfromtheinterface stability.(Thecorresponding eigenvalueiszeroonp.359.)Torecoverthenecessaryestimates on vR0 orvL0,weproceedviatheenergyestimateof(5) whichgivesthebounds
sup t∈{0,T}
v L≤
C
,
T 0(
vL)
AvLdt≤
C
,
(7) sup t∈{0,T}v R≤
C
,
T 0(
vR)
AvRdt≤
C
.
Sincetheestimatesobtainedwiththe A-matrixabovecorrespondtoL2-estimatesofthegradients,weintendtouseSobolev embeddingtoboundvL andvR inL2
(
0,
T;
L∞(L
D
=
h−1⎛
⎜
⎜
⎜
⎜
⎜
⎝
−
1 1 0. . .
0−
1 1 0. . .
. .
.
−
1 1 0 0⎞
⎟
⎟
⎟
⎟
⎟
⎠
,
such that Dv is an approximation of the derivative in the domain and
Dv2h=
hvTDTDv bounds the L2-norm of the derivativeofv (interpretedasapiecewiselinearfunction).Lemma3.1.Foranygridfunctionv,vTAv
>
cDv
2
hforsomec
>
0.Proof. By assumptionofnullspaceconsistency, A has onezeroeigenvalue associatedwithits rowssummingtozero.(By symmetryitscolumnsthereforealsosumtozero.)Furthermore,itscomponentsare
O(
h−1)
andweassumethat A isN×
N. ACholeskyfactorisationisgivenasUTU=
A.Since A issemi-definite,itwillresultinonezeroonthediagonalofthe upper-triangularmatrixU .Furthermore,thebandedstructureofA ispreservedinU .Since A1=
0,wemusthaveU 1=
0. Hence,theelementinthelowerrightcornerhastobe0.SincetherowsofU sumtozero,wecanwriteU
=
R D,whereR isan N×
N matrixwiththelastrowbeingzero.(The lastcolumnisalsotakentobezerosinceitdoesnotcontributeanywayduetothebandstructure.)Furthermore,noother rowof R sumto zero,since (most) rowsof A=
DTRTR D aresecond derivative approximationsandthe two derivatives areaccountedforby DT andD.Hence, RTR issymmetricpositivesemi-definitewithonezeroeigenvalueassociatedwith the last rowandcolumn. Moreover, we denotethe upper-left(
N−
1)
× (
N−
1)
submatrix of RTR as M, which isthen symmetricpositivedefiniteandhaselementsofO(
h)
.Furthermore,since R isbanded, M≥
chI forsome c>
0 whereI isthe
(
N−
1)
2-identitymatrix.Consequently,(
Dv)
TRTR(
Dv)
≥
cDv
2
h,sincethezerointhelast rowof RTR coincidewith thezeroinDv andM definesanorm.
By(7) andLemma3.1,wehaveDvL,R
∈
L2(
0,
T;
L2(L
,R
))
which,togetherwiththeL2boundsin(7) onvL,R andSobolev embedding,impliesvLi,R∈
L2(
0,
T;
L∞(L
,R
))
foralli.Withtheseaprioriestimates,andsincewealreadyknowfromthe stabilityproofthat(
SvL,R)
0
∈
L2(
0,
T)
,therightproblemboundsthe“data” f in(6) fortheleftproblem(andviceversafor therightproblem).Remark.Above, we used thebound on
(
SvL,R)
0 obtained from the stabilityproof. However, since onlyh
(
SvL,R)
0∈
L2(
0,
T)
wasneeded, we couldalsohaveobtainedthatdirectlyfrom theL2
(
0,
T;
L2(L
,R))
bounds on DvL,R,whichgives√
hDvL,R
∈
L2(
0,
T;
L∞(L
,R
))
.Seeingthat(
SvL,R)
0arelinearcombinationsof(
DvL,R)i
,theL2(
0,
T)
boundson√
h(
SvL)
0and√
h(
SvR)
0 follow.Weendthissectionwithafewremarks.ThesamereasoningappliestoparabolicsystemssincetheytooyieldL2
(
0,
T)
boundsonboththevariablesandtheirgradients,whichistheonlyrequirementneededtoapplythetheoryin[7]. Further-more,by mirroringone domainandrewritinga parabolicequation(scalar orsystem)such thattheinterface turnsintoa reflectiveboundary,theconvergencerateofthelatterfollowsimmediately.4. Multi-dimensions
Apossible problemwiththeprevious procedureis that L∞ isnot embedded in H1 inmultipledimensions. However, in multipledimensionsthe boundednessrequirements are suitablymilder.Consider the 2D advection-diffusion equation,
ut
+
aux+
buy=
(
uxx+
uy y),withperiodicityinthey-direction(andN points).Using von Neumann analysisin the y-direction reduces the problemto k
=
1...
N 1-D problems,one foreach mode,ˆ
vLk,R inthey-direction.They-derivativeisreducedtoalow-ordertermandcanbeignored.The H1 embeddingimpliesan
L2
(
L∞)
boundonvˆ
k,forallk=
1...
N.EachmodeconvergeswithoptimalrateandbyParseval’srelation,sodoesvL,R. Thisreasoningshowsthataboundon0T
∂
L,R
(
uL,R
)
2dsdt issufficientintheperiodic(intheextradimensions)case, andgivenbythe H1 embedding.Theoryforthenon-periodiccaseislacking,butitappearsreasonablethatthesamebound wouldbesufficient.CRediTauthorshipcontributionstatement
MagnusSvärd:Writingoriginaldraft.JanNordström:Reviewingandediting. Declarationofcompetinginterest
The authors declare that they haveno known competingfinancial interests or personal relationships that could have appearedtoinfluencetheworkreportedinthispaper.
References
[1]S.Abarbanel,A.Ditkowski,B.Gustafsson,Onerrorboundsoffinitedifferenceapproximationstopartialdifferentialequations-temporalbehaviorand rateofconvergence,J.Sci.Comput.15(2000)79–116.
[2]M.H.Carpenter,J.Nordström,D.Gottlieb,Astableandconservativeinterfacetreatmentofarbitraryspatialaccuracy,J.Comput.Phys.148(1999).
[3]B.Gustafsson,Theconvergencerate fordifferenceapproximationstomixedinitialboundary valueproblems,Math.Comput.29 (130)(Apr.1975) 396–406.
[4]B.Gustafsson,Theconvergenceratefordifferenceapproximationstogeneralmixedinitialboundaryvalueproblems,SIAMJ.Numer.Anal.18 (2)(Apr. 1981)179–190.
[5]M.Svärd,J.Nordström,Ontheorderofaccuracyfordifferenceapproximationsofinitial-boundaryvalueproblems,J.Comput.Phys.218 (1)(October 2006)333–352.
[6]M.Svärd,J.Nordström,Reviewofsummation-by-partsschemesforinitial–boundary-valueproblems,J.Comput.Phys.268(2014)17–38.