A Special Case of the Filon Quadrature
Formula*
By Lloyd D. Fosdick
Introduction. The Filon quadrature formula [1] is used for the numerical evalua-
tion of integrals of the form
= J fix) sin ikx)dx , C = J
(1)
£ = / fix) sin ikx)dx , C = / fix) cos ikx)dx .
J a
J a
The Filon formula is advantageous
over usual numerical integration formulas for
smooth fix), especially for large k, since the number of points which need be tabu-
lated depends on the behavior of fix) rather than on fix) sin kx or fix) cos kx. Under
certain circumstances
the Filon quadrature
formula reduces to a simple form,
namely
(2) S* = ii-l)m/k){fia) - fib)} , C* = ii-l)m/k)[fib) - fia)} ,
where the asterisk is used to denote the inexact result produced by the quadrature
formula; a and the integer m are related through the condition
(3)
a = mir/k for S , a = im + h)ir/k for C
and, finally, the integration interval (a, b) satisfies the condition
(4)
(6 - a) = 2iir/k ,
where i is any integer. It would seem that the error associated with the use of Eq.
(2) as approximations
for S and C would be intolerable. However, this suspicion is
unfounded when ky> 1. Let E, and Ec be the error terms, thus
(5)
S = S* + Es and C - C* + Ec,
then it is shown below that
(6)
\E,\ á M(b - a)/k3 and \E.\ ¿ M(6 - a)/kz
where
(7)
M =
max
dx
Since the error is proportional to k~3, it will be small for large k, provided that the
third derivative of fix) is not large relative to fc.
Thus, there is an interesting set of circumstances under which accurate estimates
of the integrals in Eq. (1) can be obtained by a trivial computation.
This work was
stimulated by some recent work of Clendenin [2], who indicated that formulas of the
type shown in Eq. (2) would not be very well suited for practical computations.
Since this was not supported by a computation
of the error bounds, we decided to
Received May 8, 1967.
* This work is supported in part by the AEC under contract US AEC AT( 11-1) 1469.
77
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
78
L. D. FOSDICK
determine them. Since analysis of the error in Filon quadrature is rare, Luke's work
[3] being one of the rare cases, many details are given here.
The Filon Formula. This quadrature
formula is derived as follows. In Eq. (1)
fix) is replaced by a polynomial approximation,
in particular a second-degree poly-
nomial which agrees with fix) at three points. Since the integrals
(8)
S = J xm sin ikx)dx , C = I xm cos ikx)dx ,
a
a
are obtainable in closed form this procedure leads to a quadrature
formula. Follow-
ing the usual pattern in constructing quadrature formulas, the interval of integration
(a, b) is subdivided into p panels, each of length 2h. The integration formula is ap-
plied to each panel; in this application the polynomial approximation
of fix) is re-
quired to agree with fix) at the endpoints and midpoint of the panel. Finally the
sum of the contributions
from each panel gives the desired quadrature
formula.
These formulas are
(9)
S = h[aifo cos kxo — ftp cos kx2p) + ßS2p + yS2p-x] + Es,
(10)
C = h[a(J2p sin kx2p — /0 sin kx0) + ßC2p + yC2p-x] + Ec,
where
p
(11)
S2p = 2/í»sm
kx2i — s[/o sin kxo + f2p sin kx2p],
i—0
p
(12)
Sip-x = ]C/2t-isin/b;2._i,
V
(13)
C2p — Ylrlii cos kx2i - M/o cos kxo + f2p cos kx2p],
•-0
p
(14)
C2p_i = S/2,-1 cos fccii-i,
»—1
(15)
a = I/o + (sin 20)/202 - (2 sin2 0)/03,
(16)
ß = 2((1 + cos2 e)/82 - (sin 2d)/d3) ,
(17)
7 = 4((sin 6)/93 - (cos 9)/d2) ,
(18)
6 = kh ,
(19)
fi = fix/) , Xi+x — Xi = h , xo = a , x2p = b ,
and Es and Ec are the error terms associated with using the first term on the right of
Eqs. (9) and (10) as an approximation
for S and C. Where 6 is small it is necessary
to replace the expressions
for a, ß, and y by power series in 0 to avoid the loss of
significant figures due to cancellation
in these expressions; this fact has been over-
looked in a recently published algorithm [4].
The Error Term. Peano's theorem [5] is used to put the error terms Es and Ec in
a useful form. Define the operator
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
A SPECIAL CASE OF THE FILON QUADRATURE FORMULA
79
(20)
Lif) = / fix)tix)dx - E bifiXi) ,
•o
¿-1
where Lif) vanishes when/ is a polynomial of degree n or less, the Xi are contained
in the closed interval [a, b], tix) is piecewise continuous in this interval, and the
in + l)th derivative oí fix) is continuous in this interval. Then, by Peano's theorem,
(21)
Lif)= I fn+l\t)Kit)dt,
J a
where K(t), the Peano kernel, is given by
(22)
Kit) = (l/n!)L,((*
- t)n+) ,
and
(23)
ix - t)\ = ix - t)n , x^t,
ix - t)n+ = 0 , x < t,
and the subscript x in Lx denotes that x, rather than t, is regarded as the variable. In
the present application Es is to be identified with Lif) to obtain a useful expression
for the error in the quadrature formula for the sine integral; similarly, Ec is identified
with Lif) for the error in the cosine integral.
Let us now direct our attention to applying the quadrature formula on one panel;
thus we use Eqs. (9) and (10) with p = 1. The contribution to S from one panel is
fxi+ih
(24)
Si = /
fix) sin ikx)dx,
J xi
and the contribution
to C from one panel is
/xi+ih
fix) cos ikx)dx.
xi
After a change of variables S, and C¿ can be expressed as follows:
/+h
r+h
giiz) sin (kz)dz + sin (kxf + 6) j g^z) cos ikz)dz ,
-h
J -h
/+h
r+h
giiz) cos ikz)dz — sin (fcc,- + 0) / 0,(z) sin ikz)dz ,
(26)
where
(27)
giiz) = fiz + Xi + h) .
Let us define the two integrals appearing in these expressions as
r+i>
/•+*
(28)
Si = I giiz) sin ikz)dz , c, = / <7,(z) cos ikz)dz,
J-h
J-h
and consider applying the Filon quadrature
formula to them. The task now is to
determine the Peano kernel for each case; first we consider s,-. Identifying the opera-
tor L with the error term associated with using the Filon quadrature
formula for s,-
we have
(29)
Lig/) = J giiz) sin ikz)dz + h[a cos 0 - | sin »j(?<(*) - 0,-(-A)) .
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
80
L. D. FOSDICK
It is clear that Lig,) is zero when g,iz) is a polynomial of second degree, since the
quadrature
formula is designed to be exact in this case. On the other hand it is not
exact for a polynomial of third degree, as may be verified by substituting
z3 for g¿z)
in Eq. (29) ; this is different from the situation for Simpson's rule which is also de-
signed to be exact for polynomials of second degree, but which is exact for poly-
nomials of third degree too. Consequently, the Peano kernel is
(30)
Ksit) = I ] J ix — t)Y sin ikx)dx + h[ a cos 0 — - sin i
x «a - t)+2 - i-h - t)Y)
Notice that
(31) i-h-t)+
= 0, -hgtgh,
ih-t)+
= h-t,
-h^t^h
hence the kernel can be written
(32)
Ksit) = %y ix - t)2sin ikx)dx + h\a cos 8 - | sin 0 j(A - t)2'
After executing the integration and some algebraic manipulations
this becomes
(33)
^-^(..(Xy)+!(„,_
„(A))}.
Similarly, one obtains for the kernel associated with the next integral in Eq. (28)
*.„_£{,(*._«..)
((i-!)•-<-
i);)
+»s5í(>-iy-T0-í)-?(-—(f))}-
It will be noticed that Kdt) contains a multiplier A4 instead of A3; this arises from the
fact that the Filon formula is exact for c,-, Eq. (28), when g^z) = z3. Thus in this case
the situation is analogous to Simpson's rule. Equation (34) differs from Luke's result
(Eq. (19) in [3]) which contains an error. A term — 2(— t/h)+3 in Eq. (34) was
omitted by Luke. This omission stems from the omission of the j = 0 term in Eq.
(12) of his paper. As a result the regions of definiteness for (?i(s, 0) and G2(s, 0) in his
paper are incorrect, but the error curves shown in Figs. 1-5 in his paper are correct.*
The Simple Formulae. It is apparent that under the conditions cited in the intro-
duction, Eqs. (3) and (4), the simple expressions in Eq. (2) result from Eqs. (9) and
(10) ; notice that Eq. (4) implies that 0 is an integral multiple of w,
(35)
0 = mr.
Recalling the definition of Sit Eq. (26), it is seen that
r+h
(36)
Si = cos ((m + ii + 1)«)*") / giiz) sin ikz)dz ,
* Private correspondence.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
A SPECIAL CASE OF THE FILON QUADRATURE FORMULA
81
hence only the kernel 2f ,(£) enters into a consideration of the error associated with
applying the Filon formula to this panel. Since we now have 0 = nir, substitution of
this value in K,it), Eq. (33), yields
(37)
K,it) = Qi3/nY) (cos nir — cos inwt/h)) .
The quadrature
error is given by Eq. (21), hence for this case
Hg¿) = cos Urn + (• + l)n)x) /+ g,i3\t)
(38)
xlw(cosn,r-cos(x))r-
Since the kernel does not change sign over the interval of integration, the mean-value
theorem can be applied to obtain
Hg/) = 9ii3)iï) cos Um + Ü + l)»)x) f+k (—Y
(39)
/
/ \\
X I cosnw — cos ( ~ J Jdt, — A<£
and, performing the integration,
(40)
Ligt) = gYYt) cos ((m + (i + l)n)ir) (A/n7r)3(2A) cos (nir) .
Summation of Ligi) over the p panels yields the error Es. Using the definition of 0,
Eq. (35), and the fact that 2hp = b — a, the inequality (6) results.
A parallel calculation yields the same inequality for the error in the cosine term.
The kernel K dt) does not enter in this computation,
since the coefficient of the
cosine integral in Eq. (26) vanishes.
Acknowledgment.
The author wishes to thank Yudell L. Luke for his helpful
comments on this work.
Department
of Computer Science
University of Illinois
Urbana, Illinois
1. P. J. Davis & Ivan Polonsky
in Handbook of Mathematical
Functions
with Formulas,
Graphs, and Mathematical Tables, National Bureau of Standards,
Applied Math. Series, No. 55,
U. S. Government Printing Office, Washington, D. C, 1964. MR 29 #4914.
2. W. W. Clendenin, Numer. Math., v. 8, 1966, pp. 422-436. MR 34 #982.
3. Y. L. Luke, Proc. Cambridge Philos. Soc, v. 50, 1954, pp. 269-277. MR 15, 992.
4. Linda Teijelo, "Algorithm 255," Comm. ACM, v. 8, 1965, p. 279.
5. P. J. Davis, Interpolation
and Approximation,
Blaisdell, New York, 1965.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
Dostları ilə paylaş: |