Photocopying permitted bylicenseonly theGordon and BreachScience Publishers imprint.
Printed in Singapore.
On Optimal Quadrature Formulae
FLAVIA LANZARA*
Dipartimentodi Matematica,Universit& degfiStudi diRoma "LaSapienza", PiazzaleAldoMoro,2 00185Rome,Italy
(Received 15 April 1999;Revised4May1999)
A procedureto construct quadrature formulaewhich are exactforsolutionsoflinear differentialequationsand areoptimalinthesenseof Sardis discussed.Wegive necessary and sufficient conditions under whichsuchformulae do exist.Several formulaeobtained by applyingthismethodareconsideredandcomparedwithwell known formulae.
Keywords: Quadrature formulae;Errorestimates;Influencefunction MathematicsSubjectClassification: 65D32
1.
INTRODUCTION
Let [a,
b]
be abounded,closed interval of the realaxis and let n be a positive integer.Considerthe following integralbu(x)g(x)dx
(1.1)
whereuE
C’([a, b])
andg(x) Ll(a, b).
g(x)isthe weight functionand issupposedtobenon-zero on a setof positivemeasure.Denote byxl,...,Xmrndifferent points of the interval[a, b]suchthat a
xo <_ x < <
Xm Xm+ b(1.2)
* E-mail:lanzara@uniromal.it.
201
and denoteby Ealinear differential operator of ordern:
dn
dn-k
E
+ a(x)
dxn_(1.3)
k=l
where
ak(x)
Ecn-k([a, b]),
k 1,...,n.In
[4]
Ghizzetti and Ossicini consider the following general quad- ratureformula:/a u(x)g(x)
dxAhiu(h-1)(Xi)
At-R[u],
h=l i=1
(1.4)
relevantto the integral
(1.1),
to the nodes(1.2)
and to the differential operator(1.3),
with thefollowingcondition:e[u]
0 0(1.5)
thatis
(1.4)
is exactwhenuis solutionof thelinear differentialequation=0.
Fixedthe weightg(x),thenodes x,..., Xm and the operatorE,in
[4]
amethodtodetermineall thequadratureformulae of type
(1.4),
which satisfycondition(1.5),
isgiven. Inordertodothat,considerthe adjoint operator ofE:E*-(-1)
dnn
dn-k
+ (- 1)
-kdx,,_kak(x)
k=l
andthe reduced operators
d dr-k
Er --dx -+- ak(x)
dxr_k r 0,... n-k=l
with theiradjoint operators
Er* -(_1) -dxr
dr-+- (- )r-k dr-k
dxr-kak
(X),
k=l
Let
0(x)
andgm(X be the solutions of the equationE*[]
gwhichsatisfy,respectively, the initial condition
(1.6)
h) (a)
O,q(mh)(b) O,
h-- 0,...,n- and letl(X),..., m-l(X)
bern- arbitrarysolutionsof(1.6).
By
assumingin(1.4)
Ahi "-E;_h[i--i_l]x=xi
h 1,...,n; i= 1,...,m(1.7)
and
rn rxi+
R[U] E / i(x)E[u(x)]
dxi=0 xi
(1.8)
thequadrature formula
(1.4)
satisfiescondition(1.5).
Conversely,if(1.4)
and(1.5)
hold true, then thereareuniquely
determined 1,...,m-1,solutions of
(1.6),
suchthat, togetherwithqa0andqam,makevalid(1.7)
and(1.8)
[4,pp.27-32].
Appropriately choosing the weightfunctiong(x),thenodesx,...,Xm,
the differential operator E and the functions q,...,m_ many of the knownquadrature formulaecan be found as particularcases
(see
[4,pp.80-147]).
Since
(1.4)
depends on(m- 1)n
free parameters, it is naturalto try to determinethese parametersinsuchawayquadrature formula(1.4)
isoptimalin some sense.
This problem has been investigated by many authors, see e.g.
[1,2,5,6,8,11,12,19,20].The firstresult of the presentpaperisthat there is oneandonlyoneway ofmaking
(1.4)
optimalinSard’ssense 11, p. 38, 16,p.176]
by choosingin asuitableway all the free parameters.This is shownbyTheorem 2.I.Moreover
inSection2weconsiderthequadratureformulan-p
u(x)g(x)
dxE Ahiu(h-)(Xi) + R[u],
h=l i=1
(E[u]
0= R[u] 0), (1.9)
fora fixedp:
<
p<
n 1, andweinvestigate the existence and unique- ness ofthe bestquadratureformula in thesenseof Sard.Thisproblem has alreadybeen discussed by several authors, mainly concerning the operatorE-d"/dx" (see [11,13-15,18]
andtheirreferences).
We remark that formula
(1.9)
is interesting in the applications, especiallyif we require onlytheknowledge of thefunction’svaluesat given points.InTheorem2.111 we give necessary and sufficient conditions under which it ispossibletomake
(1.9)
optimalinSard’ssense.Thistheorem contains somepreviousresults" ifp n andm>
n, in[7]
ithas been proved that the optimal quadrature formula in Sard’s sense can be written,in oneand onlyoneway,ifthe differential operatorEhasthe propertyWinthesenseofPolya(see [5,6,10]).
In[17],
the author extends thisresulttothecaseofthe operatorEof type(1.3)
under the follow- ing hypothesis:the only solution of the equationEu-0
vanishing at the nodes isu 0.
(1.10) As
faras ourresultisconcerned,it mustberemarked thatifcondi- tions(2.12)
are not satisfied, it is not possible to write a quadrature formula of type(1.9).
Thereforethispaper providesacompletesolution totheproblemof constructingquadratureformulae of type(1.9),
opti- malinSard’ssense. Moreoverbymeansofourmethodit ispossibleto obtainseveralnewquadrature formulae.The proofs of Theorems 2.I and2.III lead to the explicitconstruc- tion of quadrature formulae useful in the applications. In Section 3 severalexamplesaregiven.InExamples(e)-(g)weconsiderequidistant nodesand the operatorE--
d"/dx
becausewewant tocompare thenew formulaewiththeclassical ones.Moreoverweconstructother formulae bychoosingdifferentoperators anddifferentnodes.2.
MAIN RESULTS
Let
{vl(x),...,vn(x)}
be a fundamental system of solutions of the homogeneousequationE*v-O
(2.1)
and let
Vp(X)
be a particular solution of(1.6).
The solutions 9i(X), 1,...,m of(1.6)
canbewritten asi(x) lp(X) + c i)lh(X),
1,...,m(2.2)
h=l
where
{c i)) (h
1,..., n,i-- 1,...,m-1)
denote(m- 1)n
arbitrary constants.Define thefunction
(x)
as(x)
qoi(x), x E(xi, xi+l],
i=0,...,m.iscalled the "influence function"orthe"Peanokernel".
From
(1.8)
wededuce thatb
dx.
Theremainder
R[u]
canbeestimated indifferent ways.By
applyingCauchy-Schwarzinequalitywededuce1/2
IR[u]l < IlE[u]l]2 [,I,(x)]
2dxXi+l
-IIE[u]ll= [/(X)]
2dxi=0 xi
(2.3/
where
Wesay that
(1.4), (1.5)
is"optimal" in thesense of Sard[11,16]
ifthe(m 1)n
constants{ e
i)}
arechoseninsuchawaytominimizeb
[I’ (X)dx.
Ofcourse wemay alsowrite
(/a )
IR[u]l _< Ilg[u]ll(b a)
1/2[(x)]
2dx(2.4)
where
Theestimate
(2.4)
permitsus tocompareourquadrature formulaewith the classical ones, where the following appraisal forR[u]
isused:Lb 0 /’Xi+l
IR[u]l < [IE[u][l Iff(x)[
dx-IIE[u]ll ]i(x)[
dx.(2.5)
Xi
Unfortunatelyinmanycasesitis noteasytostudythe sign of the func- tion ffthatistoapply theestimate
(2.5).
THEOREM 2.1 Thereexists auniquequadrature
formula of
type(1.4)- (1.5)
thatisoptimalinSard’ssense.Set,fori=1,...,rn- 1,
’i(Ii),,,,, ci))
xi+[i(X)]
2dx.X
(2.6)
Them terms
.f’i(cli),..., C(n
i) areindependent. Thereforeto minimizeJa [ (x)]
2dxit is sufficient to minimizeeachterm.If the function i(x) is zero a.e. in (xi,xi+l) (i.e. g(x) is zero a.e. in (Xi,
Xi+l)
then.i(cli), (n i))
--O. Otherwisei(cli), C(n i))
O. Inthiscase, from
(2.2),
wededuce that,n rXi+
ffSi(cli) ci)) Z
c.(i)hc,
(i)/ Vh(X)V,(x)
dxh,k xi
..jl__2ci)
h=lLxi+i Vp(X)Vh(X)
dx+ i
xixi+l[Vp(X)]
2dx.Itiswell known that suchapolynomial hasapositive minimum in
n
and, denoting by
t?
(i)(li),..., (i))
thesolutionof thefollowingsystemofn linearequations
Vh(X)V:(X)
dxVp(X)V.(x)
dx,h= xi xi
k-- 1,...,n, wehave
’i(Ii),..., (n i)) min.T’i(cli),..., C(ni)).
By
choosing theconstants{c
i)}
in(2.2)
equal to{i)}
weuniquelydeterminethe functions pi(x), 1,...,m 1,solutionsof
(1.6)
and, by meansof(1.7)
and(1.8),
weuniquelydetermine aquadratureformula which is exact for the solutions of the equationE[u]
=0. It is the"optimal" quadrature formulainthesenseof Sard.
Since
..i(li, (ni,)
h=i) I
Xixi+’Vp(X)Vh(X)
dx+ I
Xixi+’[Vp(X)]
2dxfrom
(2.3)
weobtain(2.7)
Inthiswayit ispossibletoconstruct alotofnewquadratureformulae (see,e.g.,Examples
(a)-(c)
in Section3).
Formulaeobtained in thisway havethe disadvantage that thederivativesupto the ordern- of the integrand function appear in the nodes. The choice of the(m- 1)n
arbitraryconstantsin theclassicalformulaeisbasedonthe requirement that the derivatives ofuhave not to appearin(1.4).
Inthe following wewill see how it is possible, byvirtue ofthe arbitrary choice of theSince {lh}h=
...
is asystem of linearly independentsolutionsof(2.1)in[a,b]then {1)h}h=l...
arelinearly independentfunctions in[xi, xi+l],Vi= m- 1.functions
{i}i---1
m--l,toavoid thepresenceof the derivatives ofu(x)
inthenodesand,atthesametime,to"optimize"thequadratureformula.
Suppose
thatwehavefixed a differentialoperator(1.3),
the nodes(1.2)
and the weight functiong(x) in the interval[a,b].
Consider the quad- ratureformula(1.4),
togetherwith(1.7)
and(1.8).
Letusfixaninteger<p <
n- 1. In[4]
the authors give necessary and sufficient condi- tions inorder to write aquadrature formula where the valuesu(h)(xi)
of the derivatives of order higher thann-p- aredropped,thatis a quadratureformula of the followingkind"
n
U(x)g(x)
dxAhiu(h-1)(Xi)
-4-R[u],
h-1 i=1
0
g[u] 0). (2.8)
Ifsomeconditions are satisfied, the functions 1,...,(/gm-1 (solutions of
(1.6))
canbedetermined insuchawayAhi En*-h[i i-1]x=xi
O, h=n-p+l,...,n; i-- 1,...,m(2.9)
that is tosaywecanwrite aquadratureformula of type
(2.8).
Let
(ul(x),..., u,,(x))
be n linearly independent functions, solutions ofEu 0.Assume
uus.,
(j 1,...,n)
in(2.8):
n-P
_
(h_l)Z
hiUj(Xi)-- uj(x)g(x)dx,
j-l,...,n.h=l i=1
(2.10)
Itis possibleto write aquadrature formula of type
(2.8)
ifandonlyif then linearsystem(2.10)
withm(n-
p)unknowns {Ahi}
has solutions.Inorderto discussthis system,consider the transposed homogeneous system
(h-l)
(2.11)
cju)
(xi)--O,
h--1,...,n-p; i--l,., mj=l
Ifthe rank of thematrix
(ul
h-l)(Xi))
(withn rowsandm(n-p) col-umns)
isequalto n(itmustbem(n
p)>_ n)
then(2.11)
hasno non-trivialsolutions and thelinearsystem
(2.10)
hasm(n-
p)-nlinearly indepen- dent solutions. Then a quadrature formula of type(2.8)
depends onm(n
-p) nfree parameters.(h-l)
If the rank of the matrix
(,j (xi))
isless than n, that isn q, q_>
(itmustbe
m(n
-p)>_
n q)then(2.11)
has q linearly independent solu- tionsCr=(Crl,..., Crn), (r=
1,...,q). Inthiscase(2.10)
hassolutionsifandonlyifthe followingcompatibilityconditionsare satisfied:
/a
bCj
uj(x)g(x)
dxO,
j--1
r 1,...,q.
(2.12)
Thenit is possibleto write a quadrature formula ofthe type
(2.8)
in0(3m(n-p)-(n-q)
differentways.
Consider the followinghomogeneous boundaryvalueproblem
o, (2.13)
U(h)
(Xi)
0, h 0,...,n p 1; 1,...,m.The general solution of the equation Eu-O is given by
u(x)-
-4 cjuj(x), (cl,... ,c,,)
denoting arbitrary parameters.By
imposing theboundaryconditions we obtainexactlythe system(2.11).
In[4]
the authorsprovedthe following.THEOREM 2.11
If
problem(2.13)
admitsonlythesolutionu(x)-0,
it ispossible to write a quadrature
formula of
the type(2.8)
in ocre(n-p)-"different
ways (itmust ben<
m(n-p)).If
problem(2.13)
has q (q>_ 1)
linearly independentsolutions
vr(x) jn= Cjuj(x), (r
1,...,q), for-
mula
(2.8)
canbewrittenonlyif
theconditions(2.12)
aresatisfied.
Thenwemaygetaquadrature
formula (2.8)
inO0m(n-p)-n+qdifferent
waysanditmustben p
<
n q<_ m(n
-p).Sets
m(n
p) n+
q, whereweassume q 0if(2.13)
has only the solutionu 0.Suppose
thatit ispossibletowritequadrature formula of type(2.8).
Ifq 0it issufficienttoassumethe numberof the nodesm
> n/(n
p).If q
>
0 thisispossible if andonlyif conditions(2.12)
aresatisfiedandit mustbem
> (n
q)/(n -p).FromTheorem2.11,aquadrature formula of type
(2.8)
canbewritteninec differentways. Then itispossibletofind1,. m-1,solutions of
(1.6),
whichsatisfy the system(2.9).
If s-0, ,...,m_ are uniquely determined and a quadrature formula of type
(2.8)
canbe writtenin auniqueway.Suppose
s>
0.Sincethequadratureformula(2.8)
canbewritteninocdifferentways, them- functionsq,...,
m-
depend altogetheron sarbitrary parameters: C1,...,C.
These constants C1,...,C
canbeuniquely determined such that thequadrature formula
(2.8)
is"optimal"thatis
fa
b[q(x)]
9dx hasaminimum.ThereforeTHEOREM 2.11I Supposethatone
of
the followingtwoconditionsholds true:(i)
(2.13)
hasonlythesolutionu=0;(ii)
(2.13)
has q linearly independent solutions and the compatibility conditions(2.12)
aresatisfied.
Then there exists a unique quadrature
formula of
type(2.8)
that isoptimalin the sense
of
Sard.If(2.13)
haseigensolutions andconditions(2.12)
are notsatisfied
thenquadrature
formulae of
type(2.8)
donotexist.Weknowalreadythatifs 0wehave onlyonequadrature formula of type
(2.8).
Let now s
>
0. Because of Theorem 2.11, there exist 1,...,qm-, defined as in(2.2),
satisfying system(2.9). (2.9)
is a linear systemwithn(m-1)
unknowns{c} i)}
and mp equations and it has s linearly independentsolutions (itmustben(m- 1)-
s_> 0).
Then itispossible to assumec)
i) aijO+ Chcti ", a
EN,
i- 1,...,m-1; j-1,...,n.h=l
The rank ofthematrix
{a}
of order(m 1)n
xsisequaltos(we
haves
< (m 1)n).
Itfollows that thefunctions{i}i= m--l,whichsatisfy conditions
(2.9),
canbewritten asi(X) i(X) -’
Chw
i)(x),
1,...,m(2.14)
h=l
wherewi)(x)=,?
-1ahvj(x),
ij h-1,...,s; i=1,...,m-1, are solu-tions of
(2.1)
and{ i)
aresolutionsof(1 6).
Setwh
(x) w
i)(x),
x E(xi,xi+l],i=11...,
m 1;h 1,...,s.{Wh(X)}h--1
sisasystemoflinearlyindependentfunctions in[a, b].
Otherwise thereexisth=l(dl,...,
dhWh(X) d) - (0,..., 0)
O,VX
Esuch that[a, b].
Then
VX (Xi, Xi+l)
i= 1,... ,m-thatis
VX
(Xi, Xi+l),
1,...,m 1.Since
{Vh(X)}h=
are linearly independent functions in (Xi, Xi+l)i-1,...,m- 1,itmust be
adh-O,
i= l,...,m- l, j= l,...,n.(2.15)
h=l
(2.15)
is a linear homogeneous systemwithn(m- 1)
equations and s unknownswhosematrixhasrank equaltos.Thendl ds--0.
Define
m-1
f
Xi+F(C) .=
axi[i(X)]
2dx,(2.16)
whereC
(C1,..., Cs).
WehaveF(C) >
0becauseg(x)isnotzero on a setof positivemeasure.From
(2.3)
itfollows that(fxX fXm+ )1/2
IR[u]l <_ IIE[u][I2
2dx+ F(C) + [m(X)]
2dxXm
Set
l
x’+lW(k
i)W
i)L
bAkr (x) (x)
dxWk(X)Wr(X
dx,i=1 xi
m-lixi+l(k
Ok .=
axi W)(X)i(X
dx, k 1,...,sk,r 1,... ,s;
and
i
xi+lD
[i(X)]
2dx.i= xi
Thereforewehave ,,s
F(C) Z AkrCkCr +
2BkCk +
D.k,r k=l
The function
F(C)
hasapositive minimumins.
Consider the systemOF/OCh--O,
h= 1,...,s which corresponds to the following linear system"-AhrCr -+- Bh
0, h 1,...,s.(2.17)
r=l
Sincethematrix
{Ahr}h,r=
ispositivedefinite then(2.17)
hasone and only one solution-{l,..., ’s}
which corresponds to the minimumofF:F
(7)
minF(C ).
Inthis case we assume in
(2.14)" Ci i,
i--1,...,S.3.
EXAMPLES OF OPTIMAL
QUADRATUREFORMULAE
Inthis Section weshall givesomeexamples of quadratureformulaeof type
(1.4) (see
Examples(a)-(d))
and(1.9) (see
Examples (e)-(i),(1))
obtainedbyapplying the methods ofSection2. Observe thatifXl a, from
(1.7)
and(1.8)
it follows that the function 0 must not be considered. Analogously ifXm-b
it is not necessaryto consider the functionm.
Example
(a)
Letusassumein(1.4) g(x) / x/x
a,n 2, Ed2/dx 2,
m 2,xl a,x2 b.
A
quadratureformula of type(1.4)
canbewritten in2
differentways. The "optimal"oneis:dx
3- v/b
a[24u(xl) + lu(x2)]
4
(b a)3/214u’(xl) u’(x2)]-+- R[u]
+-i-
By
applying(2.3)
wededuceIR[u]l E[u] 1[2 (b a) 2.
Example (b) Forthe sake of simplicity consider the interval [a,
b]
[0,
1].
Let us apply the general rule by assumingg(x)= 1/x/Y,
n 2,E---dZ/dx 2,
m--3 and the nodes:x
=0, x2 an arbitrary point of the interval(0,
1),x3 1.Then4
[ l+2v/-4x-2
+
4 u’(0) +
4(1
/v/-2S)
u3
+ 12V/ (1 -+- + V/Y): 16x:z + 42 u’(1) ] +R[u]
From
(2.3)
2
V/ T(Xfl
IR[u]l <_ IIE[u]ll: 1- (1 + x/)
where
7-()
9+ 27- 37
2183 + 128
4+ 448 448
6128
7+ 192
8+ 64 9.
Example
(c) Assume
n--2,E- d2/dx
x2-1, g(x)= in
[a,
b]--[0,1].
Then3(d/dx) +
2,m--2,xl 0,U(X)
dx3(e2 1)
(
3e2(e +
4e+ 1)(u(0) + u(1)) +
e2+
4e+ (u’(0)
/u’(1))/ R[u]; [R[u][ < IIE[u]ll=g
where
K-
/2i
+2e-+
4e+ ez
e)-) -<
0.03512.Example
(d)
Let E, [a,b] and g(x) be the one considered in the Example(c).
Assume
m-3,xl-0,x2-1/2,x3 1.Then3(e- 1) u(x)
dx2(e + 4v/ + 1) (u(0) + 2u(1/2) + u(1))
e
+ 4x/ 13x/- + ) (u’(0) u’(1))+ R[u]’, IR[u]l IlE[u]ll2K.
where
K-
4 + + 4x/- 4x/- +
5ee) -<
0.009184.Assume
n>_
2,m>_
2 andR
dxn
(3.1)
We investigateiftherecanexistquadratureformulae of the follow- ing type:
fab U(X)g(x)
dx i=1 Aliu(xi) -+- R[u], (E[u]
0= R[u] 0);
(3.2)
this isequivalenttoconsider
(2.8)
with p n 1.In the case we are considering, the homogeneous boundary value problem
(2.13)
becomesdn
Tx" u(x) o,
u(xi) O,
i-- 1, m.(3.3)
Ifm
>
n, problem(3.3)
has no non-trivial solutions thatis condition(1.10)
is satisfied. It follows that it is possible to write a quadrature formula of type(3.2)
inm-n
differentways.Ifn
_>
3,consider(2.8)
with p n- 2. Weinvestigateif it is possible towrite aquadrature formula of typeu(x)g(x) x
m
Z[Aliu(xi) + A2iut(xi)] -+- R[u]
i=1
(e[u] o n[u] o). (3.4)
Ifn
_<
2m thehomogeneousboundary valueproblem dnxn u(x)
O,U(Xi)
U(Xi)
0,has nonon-trivialsolutions.Thereforewemay getaquadrature formula of the type
(3.4)
inOQ2m-ndifferentways.In the following we give several examples ofquadrature formulae obtainedbyapplyingthemethod given inSection 2 totheseparticular
cases, for differentvalues ofn and m and for different choices of the weightfunctiong(x). IntheExamples
(h),
(i)and(1)
weconsidermore generalcases. The optimal quadrature formula canbewritten in one andonlyoneway.Of courseit ispossibletofindalot ofother formulae.Example
(e) Assume,
in(3.1),
n=2 and letg(x) in[a,b].
Consider the following equidistant nodesxiof theinterval[a,b]"
b-a
X a
+ (i 1)h,
1,..., m, h(m 1)" (3.5)
These particular formulae(withm
< 19)
werealreadyobtainedbySard in[10].
By
assuming,in(3.5),
m-- 2 wemay getaquadrature formula of the type(3.2)
inonlyoneway:weobtain theclassicaltrapezoidal rule.Ifm 3the optimalquadratureformula of the type
(3.2)
isu(x)
dx1----
2b) (b)t +
3u+ R[u].
From
(2.3)
and(2.4)
weget(b-a)
5/2(b-a)
IR[u][ < IIE[u]12
32/ <lE[u]ll
32---" (3.6)
Nowassume m 4 in
(3.5). By
applying thegeneralruleweobtainthe following optimalquadratureformula:fa
gu(x)
dxb-a[4u(a)+llu(2a+b)
30 3+1 lu(
a+2b)
3+4u(b) + R[u].
(2.3)
and(2.4)
give the followingestimatesforthe remainder:(b-a)
5/IR[u]I _< IIE[u]ll2 < IlE[u]ll (b-a)
54v (3.7)
Assume,
in(3.5),
m--5.Wemay get thequadratureformula(3.2)
in3differentways.Theoptimaloneis
fab u(x)
dxb-a[llu(a)+32u(3a+b) 11----
4+
26u(a+
24
+ lu(b) + R[u]
with
(b a)
5/2< IIE[u]ll (b a) IR[u]I IIE[u]ll2
32
lx/]-0
32lx/]-
Finally assume,in
(3.5),
m--6.Wehave the optimal formula:fab u(x)
dxb-a[15u(a)+43u(4a+b)
190 5+
378(3a +
52b) +378(2a+3b)
5 +438(a+54b) +15u(b) 1 +R[u]
(3.8)
with
[R[u]l [Ig[u]ll= (b a)
5/(b a)
50i-i IIE[u]ll
50lT (3.9)
Now we compare all the formulae obtained up to now with the classical compositetrapezoidal rule[3,pp.
40-42]:
z u(x)
dxu(x) +
2Z
i=2u(xi) + u(x) + R[u];
(b a)
(3 10) IR[u]l _< g[u]ll
12(m- 1) ’
(3.6), (3.7), (3.8)
and(3.9)
give better estimates forR[u]
than(3.10)
for m-3,4,5, 6, respectively. In fact, estimates(3.6), (3.7), (3.8)
and(3.9)
give:m=3:
m=4:
m--5:
m=6:
IR[u]l IIE[ulll(b- a)
0.01398;IR[u]l IIE[u]ll(b- a)
0.00586;IR[u]l <_ IlE[u]l[(b- a)
0.00305;IR[u]l _< IIg[u]ll(b- a)
0.00188,while
(3.10)
gives, respectively, m--3"m--4’
m--5"
mm6
[R[u][ <_ IIE[u]ll(b- a)
0.02084;IR[u]l IIg[u]ll(b- a)
0.00926;IR[u31 <_ IIg[u]ll(b- a)
0.00521;IR[u]l _< IIg[u311(b- a)
0.00334.Example
(f) Assume,
in(3.1), (3.2),
n 4,g(x)= in[a, b] andthe nodes(3.5).
Considerthe quadrature formula
(3.4).
Ifm>
2 thenwemay get a quadrature formulainO32m-4different ways. Ifm-3, among the z2 differentquadrature formulae the "optimal"oneisthe following:
b
u(x)
dx b624a[149u(a) + 326u(a +
2b) + 149u(b)
15
(b a)Z[u’(a) u’(b)] + R[u]
+
with
1 (b -a)
9/2< E[u]] 1 (b- a) (311)
IR[u]l _< IIE[u]ll
4608 4608
Compare
the last formula with the trapezoidal rule with "end correction"(see
[3,p. 105; 21, p.66])"
fa
gu(x)
dxb-alu(a)+2u(a+b)
2+ u(b)
+ (b -4_______ a)2 [u’(a) u’(b)] + R[u]’,
IR[u]l < IIg[u]ll (’ a)5
11520
(3.12)
(3.11)
provides betterestimateforR[u]
than(3.12)
because(3.11)
givesIR[u]l g[u]l (b- a)
7.5451 x 10-4while
(3.12)
givesIR[u]l [IE[u]ll (b- a)
8.6856 x 10-4.
By
assumingin(2.8)
m 3(m 4)
andp 3the functions{i}
i--1,2,3such that(2.9)
holds are uniquely determined and we obtain the classical Cavalieri-Simpson’s rule(3/8 rule).
Now assume in
(3.2)
m--5. It is possible to write this quadrature formulainc different ways.Theoptimaloneisfab u(x)dx=241--- b-a[_ U(Xl)%-
24833u(x2)%-
22156 783
2483
u(x4)
21-u(x5) + R[u]
/ 3
(x3)
By
applying(2.3)
(b a)
9/2,/6557
IR[u]l <_ IIE[u]ll=
24576
V15855"
Assume,
in(3.2),
m--6.By
applyingthegeneralmethoddescribedin Section2weobtainthe following optimal quadrature formula:fab u(x)
dx- 54 105b-aI 3674(u(x) + u(x6)) +
1102098(u(x2)+ u(xs))
76819
%- 8
(U(X3)
%-"(X4))
%-R[u].
From
(2.3)
(b a)
9/:z,/61
633<__ IIE[u]ll=
50000
W151494
50000
(3.13)
If we apply the
3/8
rule to the interval [Xl,X4] and the Cavalieri- Simpson’s rule to [X4,X6]
we obtain the following estimate for the remainder:< iiE[u]ll (b-aj < iiE[u]ll (b-a)51
5556 l0-464285
that isworsethan
(3.13)
because(3.13)
givesIR[u]l IIE[u]ll(b- a)51.2757
x 10-4.
Form-7 the optimal quadrature formulais
a
U(X)
dxb-a 645007 741681
+
4049
73499_____1 (u(x:z) + u(x6))
36 37
(u(xl) + u(x7)) +
2
(u(x3) + u(xs)) +
358707u(x4)l + R[u].
From
(2.3)
and(2.4)
(b a)
9/2,/210047
IR[u]l <
36288
V7050030
IIg[u]l (b a) /. _2_!
0_047
36288
V7
050030" (3.14)
Considerthe compound
3/8
rule:b (x)
dx--U-
b a[U(Xl) -+- 3U(X2) -+- 3U(X3)
nt-2U(X4)
+3u(xs)
q-3u(x6) -+- U(XT)]
q-R[u];
IR[u]I < IIE[u]l (b a) < tlE[u]l (b a)59.6451
10-6.103 680
(3.14)
gives betterestimatesthan(3.15)
because(3.14)
givesIR[u]l < IIE[u]ll(b- a)54.7567
x 10-6.
Assume
m 9and, for brevity,[a,b] [0, 1].Thequadrature formula(3.2)
dependsonfivearbitrary parameters. The "optimal"oneisu(x)
dx-0.041393[u(xl)+ u(x9)]-+-0.165878[u(x2) + u(x8)]
+ 0.0898962[u(x3) + u(x7)] + 0.151826[u(x4) + u(x6)
+ 0.102004u(xs) + R[u];
]R[u]l < IIE[u]1121.35792
x 10-6.
Ifm--11, among the
007
quadrature formulae of type(3.2),
the"optimal"oneis:
u(x)= 0.033182[u(xl)
dx+ u(x)] +
0.132281[u(x2) + u(xl0)]
+ 0.073287[u(x3) + u(x9)] +
0.1lgZ85[u(x4) + u(x8)]
+ 0.087894[u(xs) + u(x7)] + 0.110141u(x6) + R[u];
IR[u]l <_ IIE[u]125.04696
10-7< IIE[u]ll5.04696
10-7. (3.16) Compare
the lastquadrature formulawiththe composite Cavalieri- Simpson’s rule.Intheclassicalformulawehave the followingestimate forR[u]:
IR[u]l _< I[E[u]ll
(3.17) 180(m 1)4"
Form-11,
(3.17)
givesworst estimateR[u]
than(3.16)
because(3.17)
givesIR[u]l IIE[u]ll5.55556
x 0-7.
Example(g) Letus assumein
(3.1)
n 6and,for the sake ofsimplicity, [a, b] [0, ]. Let g(x) in[0,1]
andassumethe nodes(3.5).
Form 5(m
6),(3.2)
is the classical Boole’s rule(Newton-Cotes
6-pointrule)
[3,p.63]. By
assumingm 7in(3.5),
formula(3.2)
canbewritten inc differentways. The "optimalone"isthe following:u(x)
dx 522 593[u(x)+ u(x7)] +
6 574 999[u(x2) + u(x6)]
10482832 26207080
2504 563 3 969777
+
52414 160[u(x3) + u(xs)] +
13103 540u(x4) + R[u].
From
(2.3)
IR[u]l IIE[u]1124.7703
x 10-8.
Ifm 8, thequadratureformula
(3.2)
dependson twofreeparameters.The "optimal
one"
isu(x)=
dx0.0441851[u(xl)+ u(xs)] + 0.20338[u(x2)+ U(XT)]
+ 0.0830825[u(x3)+ u(x6)]
+ 0.169352[u(x4) + u(xs)] + R[u];
IR[u]] < IIE[u]l]2.6128
x 10-8.
By
assuming,in(3.2),
rn--9 we may geta quadrature formula in oc3 differentways. The"optimal"oneisu(x)
dx 0.0374541[u(xl) + b/(x9)] -+- 0.187974[u(x2) + u(x8)]
+
0.0341871[u(x3) + u(x7)] + 0.23889[u(x4) + u(x6)]
+ 0.0299175u(x5)+ R[u];
]R[u]] < [IE[u]119.7.8991
x 10-9_< E[u]] 7.8991
x 10-9. (3.18)
Compare (3.18)
withtheestimateforR[u]
in the compositeBoole’s rule withninenodes:IR[u]l _< IIE[u][I8,073
l0-9. (3.19) (3.18)
isbetter than(3.19).
Example
(h) By
assumingn 2,m 3and[a,b]
[0, 27r],consider the operatorEd2/dx
2+
1, thenodesxl 0,x2 7r,x3 27r and the weight functiong(x)= in [0,27r].
Weinvestigateifthere exists aquadrature formula of type(2.8)
with p=1. The homogeneous boundary value problem(2.13)
has the solution v(x)=sin(x). Then q= and the compatibilitycondition(2.12)
issatisfied. Thenitispossibleto write a quadrature formula(2.8)
inoc2differentways.By
assumingl(x) -cos(x) (4/7r)
sin(x)+
1;(x) -cos(x)
/(4/7r)
sin(x)+ weobtainthe "optimal"one:/0 u(x)
dx= 1-u(0)--8u(vr)-
1+u(27r)+R[u],
Example
(i)
Let be n 4, m 3, [a,b]
[0, 27r], E=(d2/dx
2/1)(d2/
dx2+ 9),
the nodes xl=0, x2=Tr,x3--27r
and the weight functiong(x) in
[0, 27r].
Consider thequadratureformula(2.8)
withp 3.The homogeneous boundaryvalue problem(2.13)
has three linearly inde- pendent solutions (q=3)
and the compatibility conditions(2.12)
are satisfied.A
quadratureformula(2.8)
with p- 3 can be writtenin oe differentways. The"optimal"onecanbewritten in oneandonlyin one way, by assumingcos(x) cos(3x) 16sin(x)
/ 16 p(x)
9 8 72 457r
1357r sin(3x);
cos(x) cos(3x)
16sin(x)-
16sin(3x)
2(x)
9 87--- + i357r
Then
2 128
u(x) dx--4- [u(0)+ 2u(Tr)+ u(27r)] + R[u];
IR[u]l _< IIE[u]ll2g
where
/3_57r
512K-
V864
36457r<
0.28732.Letusnote thatinExamples
(h)
and(i) thehypothesis(1.10)
is notsatisfied.
Example
(l) Assume
[a,b]
[0, 1], n 2, Ed2/dx
23(d/dx) +
2,g(x) 1, m 3, xl=0, x2 log(2e/3),x3 and p-- 1.
A
quadrature formula of type(2.8)
canbewritten incc different ways. The optimal one canbewritten in auniqueway, thatis:1 u(x)
dx4(e
12e2+ 4)(2e
35e+
243) u(0)
9
(4e
10e2/7e1)
8e(e + 4)(2e- 3) u(log(2e/3))
-2e2
+
12e 94e(e
/4) u(1) + R[u]; IR[u]l I[E[u]ll2g
where
K
/-63 +
483e768e2(e +
952e/4)
2 412e<
0.01543.Acknowledgments
The writer wishes tothank Prof.Dr.K.J. FoersterandDr. K.Diethelm fortheir commentsand forsomebibliographicalreferences.
References
[1] R.P. Agarwal and P.J.Y.Wong,ErrorInequalitiesinPolynomialInterpolation and their Applications,KluwerAcademicPublishers, Dordrecht,1993.
[2] B.D.Bojanov, Onthe existence ofoptimal quadrature formulae for smoothfunctions.
Calcolo,16, 1979,61-70.
[3] P.J.DavisandP. Rabinowitz, MethodsofNumerical Integration, AcademicPress, 1975.
[4] A. Ghizzetti andA. Ossicini, Quadrature Formulae,BirkhiuserVerlag Basel und Stuttgart,1970.
[5] G.R.Grozev,Optimal quadrature formulae fordifferentiable functions.Calcolo, 23, 1986,67-92.
[6] K. Jetter,Optimale Quadratureformelnmit semidefinitenPeano Kernen. Numer.
Math.,25, 1976,239-249.
[7] S.KarlinandZ.Ziegler, Chebyshevian splinefunctions.J.SiamNumer. Anal., 3,1966, 514-543.
[8] G.Lange,OptimaleDefiniteQuadratureformeln. Numer.Integr.,1979,187-197.
[9] G. Polya, Onthe mean valuetheorem correspondingto agivenlinearhomogeneous differential equation. Trans.Amer.Math.Soc.,24,1922,312-324.
[10] A.Sard,Linear Approximation.AMS,ProvidenceR.L,1963.
[11] G. Schmeisser, Optimale Quadratureformeln mit semidefiniten Kernen. Numer.
Math.,20, 1972,32-53.
[12] I.J. Schoenberg, Onthebest approximations oflinear operators. Akad. Wetensch.
Indag.Math.,26, 1964,155-163.
[13] I.J. Schoenberg, On monosplinesof least deviationand bestquadrature formulae.
J.SiamNumer.Anal.,2(1),1965,144-170.
[14] I.J. Schoenberg, On monosplines of least square deviationand best quadrature formulaeII.J.SiamNumer. Anal.,3(2),1966,321-328.
[15] I.J. Schoenberg, Monosplinesandquadratureformulae.In:TheoryandApplications
ofSplineFunctions.Ed.T.N.E.Greville,AcademicPress,1969, pp. 157-207.
[16] S.Seatzu,Sullacostruzionedella migliore approssimazionedifunzionali lineari nel senso diSard. Calcolo, 15, 1978, 171 179.
[17] S.Seatzu,Sulleformule diquadraturaottimalinelsenso di Sard.Rend.Sem. Fac.Sci.
Univ.Cagliari,XLIX,1979,349-358.
[18] S.Seatzu,Formulediquadratura migliorinel senso diSardconquasitutti coefficienti uguali. Rend.Mat.,s.VI1,1,1981, 159-176.
[19] H. Strauss, OptimaleQuadratureformeln und Perfekt Splines. J.Approx. Theory, 27(3),1979, 203-226.
[20] F.Stummel andK.Hainer,IntroductiontoNumericalAnalysis,Scottish Academic Press,1980.