Internat. J. Math. & Math. Sci. VOL. 21 NO. 4 (1998) 755-760
EFFICIENT ALGORITHMS FOR
SOLVINGNONLINEARVOLTERRA
INTEGRO-DIFFERENTIAL
EQUATIONSWITH
TWOPOINT BOUNDARY CONDITIONS
LE. GAREY andR.E. SHAW" Univrsxty ofNewBrunswick P.O.Box5050,SaintJohn,N.B.
Canada, E2L 4L5
755
(Received August19, 1996andinrevisedform December 2,1996)
ABSTRACT.
In
thispaperacollection of efficient algorithms are describedforsolvinganalgebraic systemwithasymmetric Toeplitz coecientmatrix. Systemsof this form arise when approximating thesolutionof boundary value Volterra integro-differential equationswith finite differencemethods.Inthenonlinearcase,
aniterativeprocedureisrequired andisincorporatedintothe algorithmspresented.Numericalexamples
illustrate theresults.
KEY WORDS AND PIIRASES. NonlinearVolterra integro-differential equation, symmetric Toeplitz matrix,algorithmefficiency.
AMS CLASSIFICATION CODE. 65R,45. 1. INTRODUCTION
The form ofthesystemtobe solvedin thispaperis
AY
F(Y)
(I I)whereF(Y)is a nonlinear function in anunknownvectorYandAis asymmetric Toeplitzmatrix.In [1,4]
circulant matricesand synunetric bandmatrices arediscussedin connection withsolvingalinearsystem. In
[7]
there is an application ofafast algorithmto asystem ofthe form(1.1). In [5]this algorithminvolves thesolutionofasecondorder Volterraintegro-differential equation. Thebackgroundfor that workandthework considered hereisbased on the Sherman-Morrison formula [2,p.113]. Considertwo
squarematricesAandBandtwocolumnvectors u andv relatedby A B nvT
If theinverseofBexistsandvTB’*u, 1, then
A"t
B
"
+B’muvTB
"
I/(I-vTB’u)
Inthispaperanapplication involving thenumerical solution ofasecond order boundary value problemof the Volterra type will be discussed.Inparticular, theproblemmodelwillbenonlinear inthe unknown and the computer implementationwillemploytheSherman-Morrison formula.Twonumerical
exampleswillbegiventocomparethismethodwith an efficientformof theLUmethod andavariantof the methoddescribedin
[5]
and[7].
2. THE DISCRETE SOLUTION
Considera nonlinearintegro-differentialequation of the form
y)(x)
f(x,y(x),z(x)),O<x a (2.1)Supported by NSERCof Canada
where
z(x)
"
K(x,t,y(t))dtsubject to the boundary conditions y(0)=b and y(a)=b Let
Rl={(x,t,y):0za,
lYl<(R)} andR2={(x,y,z):0xa,lyl<(R),lzl<(R)}. For equation
(2.1)
defined for points in R and R2, the followingconditions areassumed:
I) f and
K
areuniformlycontinuousineach variableii) for the function f and for all(x,y,z), (x,y,z)and(x,y,z)in
Rz,
]ffx,
y,z)-f(x,y,z)lly-
YI
lffx,
y,z) ffx,y,z)llz-
zl
iii) for the function
K
andforall(x,t,y)and(x,t,)
inR,
IK(x,t,y) K(x,t,y)l s
l=Jly-
YI
andiv) the functions
fr’
fz
and Kr are continuous and satisfy
I(x,t,y)<
0for all(x,t,y)6R and(x,y,z)6R2.f(x,y,z)>0,
fz(x,y,z)a0
andA proofof the uniqueness ofa solutionfor problems satisfying the abovecanbe foundinShaw
[6].
To develop a discrete solution for(2.1)
letIN={X.
x--ih, I=0(1)N,Nh=a}
be a partition of I=[0,a]. Ageneral k-step method ofsolution isgivenbywhere
giYa*i
h2
[i f(Xa*i’Yt*i’ga*i) nf0(l)lq-ki-O i-O
-.
h%x(..,j,y?,
>,,
"-o
0j-O
(2.2)
The coefficients
{wj}
denote the weights for thechoiceofaquadrature rule.I,naddition, there are special rules requiredin connection withstarting values. These rules haveweightsw,
msmax{k,s}
ands isrelatedto theorderofthe method.Notethat whenk=-2 thetwoboundaryconditionsprovide the required auxiliary
conditionsand the(N+1)x(N+1)system can thenbe solved.Inthelinearcaseonlyonesolveisrequiredbut inthe nonlinear caseaniterativeschememustbeemployed.
Forthe remainderofthissectionweconsideramodelof the form
ya)(x)
y(x) + g(x,y(x),z(x))withboundary values y(0)=b and y(a)fb. Applying atwo-step method to (2.3) gives rise to an
(IN+ l)x(N+ 1)system.Furthermore,if themethod for solvingthe differentialequationisdenotedby (p,o)
with
then a particular method, known as Numerov’s method, has
(ao,
al,a2)=(1,
-2,1) and[3o,[3 ,13
z)=(1/12,10/12,1/12).
ThequadratureruleQselectedaspartof thisparticular methodisthe fourthorder
Newton-Gregory
rule. Conditionsfor theconvergenceoftwopart methods ((p,o),Q)canbe foundNONLINEAR VOLTERRA INTEGRO-DIFFERENTIAL EQUATIONS 757
involving YoandYsto thefight handside togivean
(N-1)x(N-1)
system. This systemcanbe expressedas(j-h2B)V
h2(BG(Y)
+ S(Y)) R (2.4)where
,scr)--go]
h(.i-
1,
Denoting(J-h
2B)
byCwe have2 0
o
C C
C C
Of particularinterest isthatCisadiagonallydominantsymmetric Toeplitzmatrix.
3. ALGORITHM DEVELOPMENT
Tosolve system
(2.4)
aniterationschememustbeemployed.Using{i}todenote the=
iterategives(J-hZB)Y
i’* hZ(BG(Y
i)
+ s(yi)) + R (3.1)There areseveralwaystosolve system(3.1). For example,by efficiently usingLUdecompositiononthe Toeplitz tridiagonal symmetric system the operationcountisabout 9n,where n is the sizeof thesystem.
For repeatedapplicationswiththe same coefficientmatrixCand different fight hand sidesthecountis
5n.The formof theLUmethod usedcanbe described asfollows:
2)
3)
4)
5)
assign valuestoelements ofC:
co=l-hZ/12
c=-2.0-10h2/12
c=Co
compute superdiagonal ofU
for
j=2,N-uj=(u,c,-coS/Uj.,
performforward substitutiontosolve
Lx
(brepresents thefighthandsideof(3.1))and thesubdiagonal ofLisgeneratedfi’omthe superdiagonal ofU
Xl=bl
for
j=2,N-xj--bj-xjq
where=co/ujq
performbackward substitutiontosolveUy=x ys.--xs./us.
for
jfN-2,1,-j=(Xj’CoYj+
I)/Uj
if
IY
+ >tolerance,repeat from step3Workingwith
C,
notethat{c,I>{CO{ +lcxl,
%=5 and the divisionby -c givesa matrixwhich is tridiagonal symmetric and diagonallydominant.Thenonzero elementsineach roware given by(-1,.,
-1)with
.=2+(10h2/12).
Denotingthis matrixbyDgivesDyi.
_I 012(BGCY
i)
scxr
i))
R) (3.2)oo
Let u=(-g)e]and
v--e,
wheree=(1,O O)T.
PerturbDsuch that13
isgivenbyThenitiseasily seen that
Dffib-
uv’.
Method2uses anLUdecomposition of f) from[4]asfollows:Foreaseof notation, let
K
=h2(BG(yi)+s(Yi))+R
represent the siterateof the righthandsideofequation (3.2). Letaj=(’J-gJZ’2)d
whered=g’(1-’a’+))
",
m=N-1.The stepsinthe methodaregivenby1) factor
b
into itsLUdecomposition anddeterminetheaj,j= 1(1)N-2) solve the system Z’= KwhereZt=()
3) evaluate
Y’
(j)
where,+’
-
9
,
J=
(I)N-4) ifIY
/’-lil
>tolerance,repeat from step2Therearejustafew calculationsforstep and approximately4ncalculations for step2.The remaining steps require about 5noperationsforatotalof9n.For repeated applicationswith the same
13
the operationcount isabout 6n.Asa thirdapproach, Method3 utilizesthemethod givenin[2,p.113]. Toobtain anapproximate solutionthe stepsare asfollows:
1)
z) 3) 4) )factor15
imo itsLU decomposition solve13
W usolve
13
ZY
calculate theconstant
13--vZ/(1-vTW)
andsetYZ+I3W
if
IY
/’-1
>tolerance,repeatfromstep3As before, step essentially requires just afew operations and step 2 requires4n operations. Since
1,0 0)
3,
step4requires2noperations. The remaining steps giveatotal of9noperationsthatreducesto6nwithrepeatediterations.
To close this section, we willlookattheconvergenceof thethreemethods. Forthe nonlinear
problems, each methodisiterativein nature.Forthe directLU method,theapplicationof the methodis
equivalenttosolving
Y+’=
A"
(hX(BG(Y
)
+S(Y))
+ R).NONLINEAR VOLTERRA INTEGRO-DIFFERENTIAL EQUATIONS 759
fromwhich
IF-.,[I
<hI/[HIBNII
(3.3)whereLis aLipschitzconstantandconvergencetakesplaceprovided h2LI,IIB|<I.NotethatAisa continuousfunctionof hwhichtendstoJash-,0. The matrix-Aismonotone [3,p.360] and henceis invertible.Theinverseofa monotone matrixhas nonnegative elements.Since-Jand-Aaremonotoneand
(-A)-(-J)0we seethat
(-J")-(-Aa)0
[3,Theorem7.5,p.362].Itfollows,therefore,that0s(-A’l)s(-J).
From [3]itisalsoknownthat
IFI
N2/8 and thereforeIl
isbounded.Theothertwomethodsaresomewhat similar soonlythelast one will beanalyzed.Thereare two
partstotheiterationprocess.
In
thefirstpart, anapproximationisobtained.By
addinganappropriate correction, thenextiterateY
is givenbyY/
=b
"’
K +=
15
"v
b -’ K%.
(3.4)Again,exactvaluesfor thesolutionof the discreteproblemare substituted and the resultissubtracted from(3.4)togive
Ei+’
15
-’
hZB[G(Y)-G(Y)+S(Y)-S(Y)]
+I
-1v
x15
"h2B(CK’Y’)-G(Y)+S(Y)-S(Y))u.
(3.5)
Inparticular,v
Tb
"*h2B(G(Y)-G(Y)+S(Y)-S(Y))
isjusta constantbecauseof thedefinitionofv.Hence,
equation(3.5)
canbewrittenasE
i+’(I
-*+x
I
-1nv1
-1)h2B(G(yi).G(Y)+S(yi).S(y))
hZA-’B(G(y).G(Y)+S(y).S(y)).
By
takingnormsand assumingF
satisfiesaLipschitzconditionwehave thesameresultas(3.3).
4. NUMERICAL RESULTSTwo Volterra integro-differential equations oftheform(2.1)areusedtoillustratethemethods described.Method istheefficientform oftheLUmethod, Method2 is a variantofthe fast algorithm given
in[4]and
[6]
and Method3 isthe implementation of the Sherman-Morrison formula[see, 2].The iteration countfor all methods is givenalongwiththeaverage CPUtimeinseconds. AlLprogramswerenan on aSUN SPARC20 indoubleprecision arithmetic.Example1
y//
y -xxS
/ 2 x ((x-l)* + I) + at, y(0)=l, y(1)=l+e 4exactsolution:y= x +
stepsize(h)
10.01
n
"ll
1[
MethodII
Method
2][
Methodi,
"1
il
II#i, a,o
llavscP
avgCPUtime, .avgC.PUtime
llx
f
Example2: 3.//
er
+ 2 + + (x+t)y dt, y(0)=0, y(l)=l30
exactsolution: 3" x
Anotherapproachtosolving system
(3.1)
isgiveninthepaperbyYahandChung[7].
Their modificationreduces the operationcountforstep 3from5n to 4n+2twhere representsthenumber of operationsin the correction termapproximation(tsn).The actualvalue of dependsonthedegreeof the diagonaldominance
ofthe coefficientmatrixC. Todescribetheiralgorithm,one canreplacethe approximationtothecorrection terminstep3of Method 2by
yi.1 Z
Let
b=t
"]. Thevectorpisgivenby pfCu,b b’,0 0)Twherethenumber ofcomponents inpisdependenton
[.1
andatolerance%.Inparticular,% ischosen such thattog(ll 2)
t() >
togClbl)
Forthe matrixCas given in
(3.1),
as the step size h used in thek-stepmethod andquadraturerulein(2.2)
decreasesI.]
approaches2and bapproaches1. Thisgivesavalue of >Nand the method ofYanand Chung [7],as pointedoutin theirpaper,will notwork.CHEN,
M.,
Onthe solution ofcirculant linearsystems,SIAMJ.
Numer.Anal. 24(668-683),987.HAGER,
W.,
Appfied NumericalLinearAlgebra,PrenticeHall, 1988.HENRICI,
P.,
DiscreteVariable MethodsinOrdinaryDifferential
Equations, JohnW’dey,1962.ROJO, O., Anewmethod for solving symmetriccirculanttfidiagonalsystemsoflinearequations,
Computers
Math. Applic.20(1990),61-67.SHAW,
R.E.,
GAREY,
L.E. andGLADWIN,C.J.,
Afast algorithm for solving second order Volterraintegro-diffcrential equationswithtwo-point boundaryconditions,Congressus
Numerantium 106(1995), 193-203.SHAW,
R.E.,
Numerical Solutions for Two-Point Boundary Value Problems, Ph.D. Thesis, University ofNewBrunswick,1994.YAN,
W.M.andCHUNG,K.L.,
Afast algorithm for solving special tfidiagonal systems, Computing