Contents lists available atScienceDirect
Journal of Computational Physics
www.elsevier.com/locate/jcp
A finite element based fast eigensolver for three dimensional anisotropic photonic crystals
So-Hsiang Chou
a,∗, Tsung-Ming Huang
b,∗, Tiexiang Li
c, Jia-Wei Lin
d, Wen-Wei Lin
daDepartmentofMathematicsandStatistics,BowlingGreenStateUniversity,OH,USA bDepartmentofMathematics,NationalTaiwanNormalUniversity,Taipei116,Taiwan cSchoolofMathematics,SoutheastUniversity,Nanjing211189,People’sRepublicofChina dDepartmentofAppliedMathematics,NationalChiaoTungUniversity,Hsinchu300,Taiwan
a r t i c l e i n f o a b s t r a c t
Articlehistory:
Received17December2017
Receivedinrevisedform3January2019 Accepted5February2019
Availableonline4March2019
Keywords:
Maxwell’sequations
Three-dimensionalanisotropicphotonic crystals
Face-centeredcubiclattice Finiteelementmethod
Null-spacefreeeigenvalueproblemandfast Fouriertransforms
ThestandardYee’sschemefortheMaxwelleigenvalueproblemplacesthediscreteelectric fieldvariable atthe midpointsoftheedges ofthegridcells.Itperformswell whenthe permittivityis ascalar field.However,when the permittivityis aHermitianfull tensor fielditwouldgenerateun-physicalcomplexeigenvaluesorfrequencies.Inthispaper,we proposea finiteelementmethod whichcanbe interpreted as amodifiedYee’sscheme toovercomethisdifficulty. ThisinterpretationenablesustocreateafastFFTeigensolver that cancomputeveryeffectivelythe bandstructureof theanisotropic photoniccrystal withSCandFCClattices.Furthermore,weovercometheusuallargenullspaceassociated with the Maxwelleigenvalue problem by derivinga null-space free discrete eigenvalue problemwhichinvolvesacrucialHermitianpositivedefinitelinearsystemtobesolvedin eachoftheiterationsteps.ItisdemonstratedthattheCGmethodwithoutpreconditioning convergesin37iterationsevenwhenthedimensionofamatrixisaslargeas5,184,000.
©2019ElsevierInc.Allrightsreserved.
1. Introduction
InthispaperweareconcernedwithafastnumericaleigensolverfortheMaxwelleigenvalueproblem
∇ × ∇ × E = λ ε E , λ = μω
2,
(1a)∇ · ε E =
0 (1b)posedon,whichisa primitivecellwithlattice vectorsa,=1,2,3,ofa periodicstructure inR3 such asa photonic crystal(PC). Here in (1), E=E(x) isthe electric field,
ε
is theperiodic permittivity complextensor field, andμ
isthe constantpermeabilityscalarfield.Mathematically,thisproblemarisesnaturallyfromthesource-freeMaxwellequationsby assumingthat thetimedependentelectricandmagneticfieldsaretime-harmonic withafactorofe−ιωt (ι
=√−1).Inan anisotropicmedium,thepermittivity
ε
maytakeaHermitianform[27–29]*
Correspondingauthors.E-mailaddresses:chou@bgsu.edu(S.-H. Chou),min@ntnu.edu.tw(T.-M. Huang),txli@seu.edu.cn(T. Li),jiawei.am05g@g2.nctu.edu.tw(J.-W. Lin), wwlin@math.nctu.edu.tw(W.-W. Lin).
https://doi.org/10.1016/j.jcp.2019.02.029
0021-9991/©2019ElsevierInc.Allrightsreserved.
ε =
⎡
⎣ ε
11ιε
12ιε
13− ιε
12ε
22ιε
23− ιε
13− ιε
23ε
33⎤
⎦ .
(2)Anisotropic PC material has many interesting and novel applications in practice, such as the hollow semiconductor nanorod [22], femtosecond laser pulses [38], photonic crystalfibers [2,6,37], among others. In 1993, it was shownthat the anisotropy ofthe material could split the degenerate band andnarrow the band gap ofa PC[41]. Laterin 1998, a largebandgapopeningbyusingananisotropicPCwas shownin[23,24].Thereafter,toexhibitlargebandgaptenability,a three-dimensional(3D)inverseopalPCstructureinfiltratedwithliquidcrystalswasproposedin[7].
Computational methods playan equallyimportantrole in the studyofthe anisotropic material.The spectral element method [30], the finite-element method (FEM) [3,4,10,11,35], the mixed finite element method [21,25,26], the finite- difference frequency-domain method(FDFD) [13,36], andtheplane wave expansion method[20] have been proposed to compute thebandstructureoftheanisotropicPC.Inthispaper,weareinterestedincreatingafiniteelementmethodthat resemblesthefinitedifferenceYee’sscheme[40] duetoitsrobustnessinsolvingtheMaxwell eigenvalueproblemsassoci- ated withisotropicmaterials. Yee’sschemeonlydiscretizes(1a) andleavesthedivergence-freecondition(1b).Thereason forthat isformerlythe exactsolutionof(1a) satisfies(1b). Indeed,thisstrategy workswell forisotropicmaterials. Thus, when itcomestothediscretizationof(1a)-(1b) orits manypossibleweak formulationsthroughfinitedifferenceorfinite element schemes, the traditionalwisdom is to usespatial discretizations that admit discrete analogueof the divergence freecondition.ThisclassofspatialdiscretizationsincludesYee’sscheme[14,16,40],theWhitneyform[5,39],thecovolume discretization[34],themimeticdiscretization[19] andtheedgeelement[32,33].Itshouldbenotedthatinallthesecases thepermittivity
ε
isascalarfield.Inthispaper,wewillfollowthisapproach,i.e.,wewilldiscretize(16a) insuchawaythat enablesustorigorouslyshowinTheorem5ofSec.5.2thatthediscretizedversion oftheweakdivergencefreecondition (16b) isautomaticallysatisfied,whichisbynomeansatrivialtask.Ontheotherhand,therearecertainlyotherapproaches, e.g., [4] thatkeep(16b). Theyhavethe advantageoftheavailabilityofageneralconvergencetheory of[3] foreigenvalue problemsposedongeneraldomains.However,forcuboiddomains ourapproachcan generatefasteigensolvers.Theavail- ability ofsuch solversis essential whenthesame largeMaxwell eigenvalue problemhas tobe solved alarge numberof timeswhenthematerialparametersintheconstitutiverelationsmustbevariedindesignproblems[17].On a 3D PC,due to the divergence-free condition (1b), a large nullspace of the underlyingeigenvalue problem will be produced. The presenceof the huge nullspace is a numericallychallenging problem in solving the eigenvalueprob- lem.Moreover, howtoefficientlysolvetheassociatedlinearsystemineachstepoftheeigensolverisanotherchallenging problem.
Inthe3Dnon-anisotropiccases(isotropicorbi-isotropicmedium),wehavehadsuccesses[8,14] intacklingtheseobsta- clesthatwerepresentwhenusingYee’sscheme[40] inwhichEisassignedtotheedgecentersandthemagneticfieldHto thefacecenters.Afterthediscretizationphase,ageneralizedeigenvalueproblem(GEP)withalargenullspaceisproduced.
Itisratherremarkablethatanull-spacefreemethodwithFFT(fastFouriertransform)-basedmatrix-vectormultiplications hasbeenfound[8,14] toexplicitlydeflatethishugenullspacesothattheGEPistransformedintoanull-spacefreestandard eigenvalueproblem(NFSEP).ThecoefficientmatrixoftheNFSEPisHermitianpositivedefinite(HPD)andwell-conditioned.
Therefore,theassociatedlinearsystemineachiterationoftheinvertLanczosmethodcanbeefficientlysolvedbythecon- jugategradient(CG)methodwithoutanypreconditioner.InthedispersivePCwithisotropicmedium(i.e.,
ε
isapiecewise nonlinear function of λ), a Newton-typeiterative method[18] withthe null-space free methodis proposed to solve the associatednonlineareigenvalueproblem.The successes of the null-space free method in isotropic and bi-isotropic PCs are partially due to the fact that the permittivity
ε
isapiecewiseconstantscalarfield,andthereisnoconflictintheplacementofdiscretepermittivityvariable.Inotherwords,itmesheswellwiththestandarddiscretizationbyYee’sscheme.However,fortheanisotropicmaterial,there is aconflictinthe placementofdiscretetensor permittivityvariable.Howtomodify Yee’sschemeso thatthe null-space freemethodcanbeappliedtoatensorfield
ε
isoneoftheissuesaddressedinthispaper.In an anisotropic medium,the permittivity
ε
isgiveninthe tensorform (2) whereε
i j, for1≤i,j≤3, are piecewise constants or piecewisenonlinear functions. Moreover,in [27,28],the tensorε
isHPD. AFDTDmethod [36] for modeling electromagnetic scatteringis proposed todiscretizethe generalanisotropicobjects.The importantstructure andproperty of the tensorε
in (2) were not considered in this discretization.In this paper, we will exclusively consider the casein whichtheHermitianmatrixin(2) satisfiesthatε
i j arepiecewiseconstant,andproposeanewfiniteelementschemewhich preservestheHermitianandpositivedefiniteness (ifε
has)inthediscretization.Themaincontributionsinthispaperare summarizedasfollows:•Our new finite element method can be interpreted asa modified finite difference Yee’s scheme. As a consequence, the discretizationofthedouble curloperator inourproposed methodisequivalent tothe corresponding operatorin Yee’s scheme so that the null-space free method[8,14] can be applied to solve the resultinggeneralized eigenvalue problem.Furthermore,awell-knownfeatureoftheclassicalfinitedifferenceYee’sschemeforisotropicmaterialsisthat thedivergencefreeconditionofthefieldisautomaticallysatisfiedbythefinitedifferencesolution.Inotherwords,one needs not todiscretize (1b) explicitly.We show inTheorem5 thatour methodpreservesthisfeatureforanisotropic materialsaswell.
• Fortheright-handsideof(1a),ourmethodprovidesanHPD-preservingcoefficientmatrixBε,i.e.,if
ε
isHPD,sois Bε. Furthermore,thematrixBεcanbefactorizedbyFFT-matrices,diagonalmatricesanda3×3 blockdiagonalmatrixwith asuitablepermutation.Thespecialfactorizationof Bε resultsinminimaleffort ininverting Bε whilestillkeepingthe correctphysicsintact(realfrequencies).• Basedon theabove importantfactorizationof Bε,the null-spacefree methodcanbe appliedto reduce theGEPinto a NFSEP with HPD coefficient matrix. The NFSEP can be solved by the invert Lanczos method and the CG method withoutpreconditioningcanbeusedtosolvethelinearsystemefficientlyateachiteration.Thecomputationalcostof thematrix-vector multiplicationin theCGmethodis almostequal tothat ofthe null-spacefree methodinisotropic PCs.
This paper is organized as follows. We introduce our proposed finite element method in Section 2. The equivalence betweenourmethodandYee’sschemeforthediscretizationofdouble-curloperatorisshowninSection3.Thediscretization of
ε
E isintroduced inSection 4. In Section 5, we show that the coefficient matrix corresponding toε
is HPD, andthe divergencefreeconditionisautomaticallysatisfiedfortheapproximateelectricfield.Finally,thenumericalefficiencyofthe developedschemesisstudiedinSection6.2. AfiniteelementmethodforMaxwelleigenvalueproblem
Inthissection, wepropose a weakformulationfortheMaxwell eigenvalueproblem(1) anda Galerkinfiniteelement methodtosolveit.
2.1. Weakformulationofproblem(1)
Denoteby(f,g)thecomplexinnerproductofCN-valuedfunctionsinL2(),i.e.,
(
f,
g) =
N i=1
fi
(
x)
gi(
x)
dx=
g∗
(
x)
f(
x)
dx.
(3)Here“∗”denotestheconjugatetranspose.Let beaprimitivecellwithlattice vectorsa,=1,2,3.Itisknownthatthe electricfieldsatisfiesthequasi-periodiccondition
E(
x+
a) =
eι2πk·aE(
x), =
1,
2,
3,
(4) forallx∈R3,wheretheBlochwavevectorkliesinthefirstBrillouinzoneB.CorrespondingtoE,wecandefineaperiodicfunctionu,itsassociatedmodulatingamplitude,as
u
(
x) =
e−ι2πk·xE(
x),
(5)whichsatisfiestheperiodiccondition
u
(
x+
a) =
u(
x), =
1,
2,
3.
(6)ThusonecanalsostudytheMaxwelleigenvalueproblem(1) bytheequivalentsystemintermsofu:
∇
k× ∇
k×
u= λ ε
u,
(7a)∇
k· ε
u=
0,
(7b)wheretheshiftedgradientoperator∇k:= ∇ +
ι
2π
k.Remark1.This u-system(7) has the advantageof havingFourier analysisat ourdisposal when analyzingthe existence, stability,andconvergenceissues,see[11,35].However,inthispaperwearemoreconcernedwithcreatingfasteigensolvers and the E-system (1) will be used. We will have occasions of interpreting some nice spectral-theory results from [35, Chapter4] backtotheE-system.
Weneedtointroducesomecomplex-valuedfunctionspaces.Let H
(
curl,) := {
v: → C
3|
v, ∇ ×
v∈
L2()}
withthenorm
||
v||
2H(curl,):= ||
v||
2L2()+ ||∇ ×
v||
2L2(),
andlet
C∞per
() := { φ |
: φ : R
3→ C , φ ∈
C∞( R
3), φ (
x+
a) = φ (
x), =
1,
2,
3} ,
C∞per() := {
v|
:
v: R
3→ C
3,
v∈
C∞(R
3),
v(
x+
a) =
v(
x), =
1,
2,
3}.
Definetheperiodicfunctionspaces
Hper
(
curl,) :=
the closure ofC∞per()
inH(
curl, ),
(8a)Hper1
() :=
the closure ofC∞per()
inH1(),
(8b)andtheirassociatedquasi-periodicfunctionspaces
Hqper
(
curl, ) := {
v:
e−ι2πk·xv∈
Hper(
curl, )},
(8c)Hqper1
() := {φ :
e−ι2πk·xφ ∈
H1per()}.
(8d)Note that inview of(5), (6) and the Fourierseries representation,functions inthe above two spaces satisfy the quasi- periodicconditions(4).
Theorem1.ForanyE,W∈C∞(R3)∩Hqper(curl,)andφ∈C∞(R3)∩Hqper1 (),wehave
(∇ × ∇ × E,W) = (∇ × E, ∇ × W),
(9a)(∇( ε E), φ) = ( ε E, ∇φ).
(9b)Proof. SinceEandWaresmooth,Green’sformulaimplies
(∇ × ∇ × E,W) = (∇ × E, ∇ × W) +
∂
ν × ∇ ×
E· W
d A,
(10)where
ν
isthe outwardunit normalvector totheboundary. Forconvenience, wenamethe sixboundary surfacesofthe primitivecellas∂0a1,∂1a1,∂0a2,∂1a2,∂0a3 and∂1a3,where∂
sa1= {
sa1+
aa2+
ba3|
a,
b∈
[0,
1]},
∂
sa2= {
sa2+
aa3+
ba1|
a,
b∈
[0,
1]},
∂
sa3= {
sa3+
aa1+
ba2|
a,
b∈
[0,
1]},
fors=0,1.Thenthesurfaceintegraltermin(10) canbedividedas
3=1
⎛
⎜ ⎜
⎝
∂0a
ν
,0× ∇ × E · W
d A+
∂1a
ν
,1× ∇ × E · W
d A⎞
⎟ ⎟
⎠
where
ν
,0= −ν
,1 are the outward unit normal vectors on ∂0a and ∂1a, respectively. In particular, from the quasi- periodicityofEandW,wehave∂0a
ν
,0× ∇ × E (
x) · W (
x)
d A=
∂0a
ν
,0× ∇ × E (
x+
a) · W (
x+
a)
d A=
∂1a
− ν
,1× ∇ × E (
x) · W (
x)
d A,
for=1,2,3.Therefore,theresultin(9a) isproved.
Toshow(9b),weobservethat
(∇ · ε E, φ) = ( ε E, −∇φ) +
∂
ε E · ν φ
d A.
(11)Again,thesurfaceintegraltermin(11) canbedividedas
=1
⎛
⎜ ⎜
⎝
∂0a
ε E · ν
,0φ
d A+
∂1a
ε E · ν
,1φ
d A⎞
⎟ ⎟
⎠ .
Fromtheperiodicityof
ε
,andthequasi-periodicityofEandφ,wehave∂0a
ε E (
x) · ν
,0φ (
x)
d A=
∂0a
ε (
x+
a) E (
x+
a) · ν
,0φ (
x+
a)
d A=
∂1a
ε (
x) E (
x) · ( − ν
,1)φ (
x)
d Afor=1,2,3.Theresultin(9b) isobtained. 2
Inview ofTheorem1,asimpledensityargumentsuggeststhatweconsiderthefollowingweakformulationofproblem (1):Findapair(λ,E)withλ∈CandE∈Hqper(curl,)suchthat
(∇ × E,∇ × W) = λ( ε E,W), ∀ W ∈
Hqper(
curl, ),
(12a)( ε E,∇φ) =
0, ∀ φ ∈
Hqper().
(12b) However, froma computationalelectromagnetics viewpoint,itis moreconvenienttowork withthefollowing twospaces whosedescriptionismoreaccessible:Forfunctionsu:R3→C3andφ:R3→C,wedefineHcurl
() = {
u|
:
u∈
H(
curl, ),
u(
x+
a) =
eι2πk·au(
x), =
1,
2,
3},
(13) Hgrad() = {φ|
: φ ∈
H1(), φ (
x+
a) =
eι2πk·aφ (
x), =
1,
2,
3}.
(14) Remark2.Alternatively,the quasi-periodic condition in (13) can be viewed asa wayto extendthe functionsoutsideof ,whiletheregularityconditionu∈H(curl,),togetherwiththequasi-periodiccondition, impliesatangentialboundary (trace)conditionon∂:uT
(
x+
a) =
eι2πk·auT(
x), =
1,
2,
3,
(15)whereuT=
ν
×uisthetangentialtraceofu.Infact, Hgrad()=H1qper() andHcurl()=Hqper(curl,).It isnothard toseethatthefirst relationholds.The proof ofthesecondrelationand(15) usessimilarargumentsinprovingtwocharacterizationsofH0(curl)areequivalent(cf.[31, Theorem3.33] or[12,Theorem2.12]).Seealso[1,Lemma2.1].Theanalogyisnotsurprising, ifwe viewthedifferenceof (15) asazero-tracecondition.
Insummary,weshallconsidertheweakformulationofproblem(1):Findapair(λ,E)withλ∈CandE∈Hcurl()such that
(∇ × E,∇ × W) = λ( ε E,W), ∀ W ∈
Hcurl(),
(16a)( ε E,∇φ) =
0, ∀ φ ∈
Hgrad(),
(16b) whereε
:R3→C3×3 isatensorfieldofthegeneralform(notnecessarilyHermitian)ε =
⎡
⎣ ε
11ε
12ε
13ε
21ε
22ε
23ε
31ε
32ε
33⎤
⎦ .
(17)Remark3.DuetotheexistenceofanexactdeRhamsequence(cf.[35,p.58])andRemark1,itimpliesthat
∇
Hgrad
()
⊂
Hcurl().
(18)ThuswecansetW= ∇φin(16a) andseethat[9] whenλ=0,thedivergencefreeconditionisautomaticallysatisfied.Our discreteweakformulationbelowwillrespectthisproperty(seesubsection5.2):
∇
Hhgrad
()
⊂
Hhcurl().
(19)Remark4.ItmaybecuriousthatwhythereisnoneedinintroducingaLagrangemultiplierintheaboveweakformulation.
See[35,Proposition4.20] forsuchacase.Usingthetechniquesin[35] via(5),wecanshowaresultsimilartotheProp.4.20:
Suppose that
ε
isHPD. Thenforeach k∈B,firstBrillouinzone,theeigenvalue problem(16) hasan increasing sequence {λj}j∈N ofnon-negativerealnumberstendingtoinfinity. Eacheigenvaluehasfinitemultiplicity andzeroisan eigenvalue ifandonlyifk=0.Thedetailswillbereportedelsewhere.Asaconsequence,zeroeigenvalueisnotonlynonphysical,but mathematicallyspuriousforthek=0 case,whenadiscreteweakformulationoftheMaxwelleigenvalueproblemproduces zeroeigenvalue.Anull-spacefreemethodwouldbecalledforinsuchacase.2.2. Discretizationsofweakformulationsof(16)
Inthissubsection,weuseaGalerkinfiniteelementmethodbasedontheedgeelementstoapproximatetheexactfield Ein(16a).Tothisend,let Hhcurl()andHhgrad()be3N-dimensionallinearsubspacesofHcurl()andN-dimensionallinear subspaceofHgrad(),respectively(tobedefinedin(23) and(24),respectively).ThenastraightforwardGalerkinmethodfor (16a)-(16b) is:Findapair(λ,E)withλ∈CandE∈Hcurlh ()suchthat
(∇ ×
E,∇ ×
W)
h= λ( ε
E,
W)
h, ∀
W∈
Hhcurl(),
(20a)( ε
E, ∇ φ)
h=
0, ∀ φ ∈
Hhgrad().
(20b)Here(f,g)hisadiscreteinnerproducttobespecifiedlater.
Tosimplifythediscussion,weonlygivethedetailswhenisasimplecubic(SC)primitivecellwiththelatticetransla- tionvectorsa=axˆ,=1,2,3,wherexˆisthe-thunitvectorinR3.Alongtheway,wealsogive sufficientremarksfor theface-centeredcubic(FCC)latticesothatitcanbehandledsimilarly.Lethx,hy,andhzbetheequalgridspacingsalong thex,y,andzdirections,respectively.Gridpointsareindexedfrom0 toni(i=1,2,3)withn1=a/hx,n2=a/hy,n3=a/hz. Thusthenumberoftheedgesparalleltothex-, y- andz-axes,respectively,excludingthoseontheright,rear,andtopsur- faces,isjustN=n1n2n3.
From nowon,weusexm=mhx andxmˆ = ˆmhx,wheremˆ =m+12.Furthermore,form= −1,0,. . . ,n1,let ˆcmˆ(x)denote thecharacteristicfunctionwhosevalueisoneover[xm,xm+1]andzeroelsewhere,ˆli denotethegenericglobalhatfunction supportedover[xi−1,xi+1],i.e.,ˆli(x)=δi,fori=1,. . . ,n1−1.Hereandhereafter,δidenotestheKroneckerdelta.Similar notationsareusedforthe yandzdirections.
Wenowdefinepiecewiseconstantfunctions
cˆi
(
x) = ˆ
cˆi(
x),
cˆj(
y) = ˆ
cˆj(
y),
ckˆ(
z) = ˆ
ckˆ(
z),
(21a) fori=0,n1−1, j=0,n2−1,k=0,n3−1,andc0ˆ
(
x) = ˆ
c0ˆ(
x) +
eι2πk·a1cˆ
nˆ1(
x),
cnˆ1−1(
x) = ˆ
cnˆ1−1(
x) +
e−ι2πk·a1cˆ
-1(
x),
(21b) c0ˆ(
y) = ˆ
c0ˆ(
y) +
eι2πk·a2cˆ
nˆ2(
y),
cnˆ2−1(
y) = ˆ
cnˆ2−1(
y) +
e−ι2πk·a2cˆ
-1(
y),
(21c) c0ˆ(
z) = ˆ
c0ˆ(
z) +
eι2πk·a3cˆ
nˆ3(
z),
cnˆ3−1(
z) = ˆ
cnˆ3−1(
z) +
e−ι2πk·a3cˆ
-1(
z),
(21d) aswellaspiecewiselinearfunctions,li
(
x) = ˆ
li(
x),
lj(
y) = ˆ
lj(
y),
lk(
z) = ˆ
lk(
z),
(22a) fori,j,k=0,andl0
(
x) =
⎧ ⎪
⎨
⎪ ⎩
−
h−x1(
x−
x1),
ifx∈ [
x0,
x1],
eι2πk·a1h−x1(
x−
xn1−1),
ifx∈ [
xn1−1,
xn1],
0
,
otherwise,
(22b)
l0
(
y) =
⎧ ⎪
⎨
⎪ ⎩
−
h−y1(
y−
y1),
ify∈ [
y0,
y1] ,
eι2πk·a2h−y1(
y−
yn2−1),
ify∈ [
yn2−1,
yn2] ,
0
,
otherwise,
(22c)
l0
(
z) =
⎧ ⎪
⎨
⎪ ⎩
−
h−z1(
z−
z1),
ifz∈ [
z0,
z1],
eι2πk·a3h−z1(
z−
zn3−1),
ifz∈ [
zn3−1,
zn3],
0
,
otherwise.
(22d)
Based on the piecewise constant andlinear functionsin (21) and (22), we now define bases forthe 3N-dimensional subspace Hcurlh ()andN-dimensionalsubspaceHhgrad()in(20),respectively,as
Hhcurl
() =
span{ φ
i1,j,ki, φ
2r,s,tj, φ
,3m,nk} ⊂
Hcurl()
(23) andHhgrad
() =
span{φ
i,j,k} ⊂
Hgrad(),
(24)wherei,j,karethestandardunitvectorsinR3,and
φ
i1,j,k(
x,
y,
z) :=
cˆi(
x)
lj(
y)
lk(
z),
(25a)φ
2r,s,t(
x,
y,
z) :=
lr(
x)
cˆs(
y)
lt(
z),
(25b)φ
3,m,n(
x,
y,
z) :=
l(
x)
lm(
y)
cnˆ(
z),
(25c)φ
i,j,k(
x,
y,
z) :=
li(
x)
lj(
y)
lk(
z)
(25d)withthesup-indicesvaryinginthesetS describedas
S
= {(
i,
j,
k) :
i=
0, . . . ,
n1−
1,
j=
0, . . . ,
n2−
1,
k=
0, . . . ,
n3−
1}.
(26)The components of an approximate electricfield E∈Hcurlh () are piecewise constant in one direction andpiecewise linearintheothertworemainingdirections,i.e.,
E1
(
x,
y,
z) =
(i,j,k)∈S
E1
|
ˆi,j,kcˆi(
x)
lj(
y)
lk(
z) ≡
(i,j,k)∈S
E1
|
ˆi,j,kφ
i1,j,k,
(27a) E2(
x,
y,
z) =
(i,j,k)∈S
E2
|
i,ˆj,kli(
x)
cˆj(
y)
lk(
z) ≡
(i,j,k)∈S
E2
|
i,ˆj,kφ
i2,j,k,
(27b) E3(
x,
y,
z) =
(i,j,k)∈S
E3
|
i,j,ˆkli(
x)
lj(
y)
ckˆ(
z) ≡
(i,j,k)∈S
E3
|
i,j,ˆkφ
i3,j,k.
(27c) Here in (27a), a subindex (ˆi,j,k) indicates that a quantity is located atthe midpoint ofthat edge. Bydirect evaluation, we seethat E1|ˆi,j,k=E1(xi+12,yj,zk).Similarinterpretations holdfor E2 and E3.The Ei’s arethefamiliaredge elements closelyrelated tothe finitedifference Yee’s scheme[9].In consistent withthe quasi-periodicconditions (15), it iseasily checkedthat
Left-Right
E2|
n1,ˆj,k=
eι2πk·a1E2|
0,ˆj,k,
j=
0:
n2−
1,
k=
0:
n3−
1,
E3|
n1,j,ˆk
=
eι2πk·a1E3|
0,j,ˆk,
j=
0:
n2−
1,
k=
0:
n3−
1,
(28a) Front-Rear E1|
ˆi,n2,k=
eι2πk·a2E1|
ˆi,0,k,
i=
0:
n1−
1,
k=
0:
n3−
1,
E3
|
i,n2,ˆk=
eι2πk·a2E3|
i,0,ˆk,
i=
0:
n1−
1,
k=
0:
n3−
1,
(28b) Top-Bottom E1|
ˆi,j,n3=
eι2πk·a3E1|
ˆi,j,0,
i=
0:
n1−
1,
j=
0:
n2−
1,
E2
|
i,ˆj,n3=
eι2πk·a3E2|
i,ˆj,0,
i=
0:
n1−
1,
j=
0:
n2−
1.
(28c)Forthederivationofthegeneralizedeigenvalueproblemfor(20a),thevectorizedformofthediscreteelectricfieldE in (27) canbewrittenas
e
=
e1 e2 e3
(29) with
e1
=
vec{
E1|
ˆi,j,k},
e2=
vec{
E2|
i,ˆj,k},
e3=
vec{
E3|
i,j,kˆ},
for(i,j,k)∈S.Here
vec
{
F} :=
⎡
⎢ ⎢
⎢ ⎣
vec
(
F(
0:
n1−
1,
0:
n2−
1,
0))
vec(
F(
0:
n1−
1,
0:
n2−
1,
1))
.. .
vec
(
F(
0:
n1−
1,
0:
n2−
1,
n3−
1))
⎤
⎥ ⎥
⎥ ⎦
forgivenF∈Cn1×n2×n3.
IntermsofthebasisfunctionsofHcurlh (),weexpectthediscreteweakformulation(20a) tohaveamatrixform
Ae
= λ
Bεe,
(30)whereeisasin(29),AandBεarematrixrepresentationsoftheleftandrighthandsidesof(20a),respectively.InSections3 and4,wewillusethetrapezoidalquadrature
(
f(
x),
g(
x)) =
xn
xm
f
(
x)
g(
x)
dx≈
hx 2⎡
⎣
f(
xm)
g(
xm) +
2n−1
i=m+1
f
(
xi)
g(
xi) +
f(
xn)
g(
xn)
⎤
⎦
≡ (
f(
x),
g(
x))
h (31)to approximatethe discretizationin(20a) with Hhcurl()in (23) sothat thematrix A in (30) isidenticalto thatof Yee’s schemeuptoafactorhxhyhz.
3. Trapezoidalquadratureapproximationfor(∇×E,∇×W)
ForagivenE∈Hhcurl(),from(20a) and(23) itsufficestoconsidertheexpressions
(∇ ×
E,∇ × φ
1i,j,ki), (∇ ×
E,∇ × φ
2i,j,kj), (∇ ×
E, ∇ × φ
3i,j,kk)
(32) with (i,j,k)∈S in(26) and φi,k,j, =1,2,3, in (25). In fact, the inner product (·,·) in(32) involves the evaluationof typicaltermslike(l,c),(c,l),(l,l)and(c,c).Fortechnicalderivation,insteadoftheexactintegrals(lr(x),li(x)),(ls(y),lj(y)) and(lt(z),lk(z))in(32),weusethetrapezoidalquadrature(31) toobtain(
lr(
x),
li(
x))
h= δ
rihx, (
ls(
y),
lj(
y))
h= δ
sjhy, (
lt(
z),
lk(
z))
h= δ
tkhz (33) fortheirapproximations.Let
c-1
(
x) =
eι2πk·a1cnˆ1−1(
x),
c-1(
y) =
eι2πk·a2cnˆ2−1(
y),
c-1(
z) =
eι2πk·a3cnˆ3−1(
z).
(34) Thenthefollowingrelationshold:∂
xli(
x) =
h−x1(
cˆi−1(
x) −
cˆi(
x)),
(35a)∂
ylj(
y) =
h−y1(
cˆj−1(
y) −
cˆj(
y)),
(35b)∂
zlk(
z) =
h−z1(
ckˆ−1(
z) −
ckˆ(
z)),
(35c)and
∂
yφ
1i,j,k=
h−y1cˆi(
cˆj−1−
cˆj)
lk, ∂
zφ
1i,j,k=
h−z1cˆilj(
ckˆ−1−
ckˆ),
(36a)∂
xφ
2i,j,k=
h−x1(
cˆi−1−
cˆi)
cˆjlk, ∂
zφ
2i,j,k=
h−z1licˆj(
ckˆ−1−
ckˆ),
(36b)∂
xφ
3i,j,k=
h−x1(
cˆi−1−
cˆi)
ljckˆ, ∂
yφ
3i,j,k=
h−y1li(
cˆj−1−
cˆj)
ckˆ,
(36c) for(i,j,k)∈S.Beforeprovingthefollowingtheoremconcerningacertainconsistencybetweenouredgeelementschemeandthefinite differenceYee’sscheme,wedenotethedifferencematrixwiththequasi-periodicconditionas
Kn,a
=
0 In−1 eι2πk·a 0−
In∈ C
n×n,
(37)for=1,2,3.
Theorem2.ForanSClattice,thevectorizedformsoftheinnerproductsdescribedin(32)takeonthematrixrepresentation
⎡
⎢ ⎣
vec
{ ( ∇ ×
E, ∇ × φ
1i,j,ki)
h}
(i,j,k)∈Svec
{(∇ ×
E, ∇ × φ
2i,j,kj)
h}
(i,j,k)∈Svec