Generalized Gaussian Quadrature Rules for Systems of Arbitrary Functions

Author(s): J. Ma, V. Rokhlin and S. Wandzura

Source: SIAM Journal on Numerical Analysis , Jun., 1996, Vol. 33, No. 3 (Jun., 1996),
pp. 971-996
Published by: Society for Industrial and Applied Mathematics

Stable URL: https://www.jstor.org/stable/2158491

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact [email protected].

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

Society for Industrial and Applied Mathematics is collaborating with JSTOR to digitize,
preserve and extend access to SIAM Journal on Numerical Analysis

This content downloaded from

187.19.243.229 on Sun, 24 Mar 2024 12:26:12 +00:00
All use subject to https://about.jstor.org/terms
 SIAM J. NUMER. ANAL. ( 1996 Society for Industrial and Applied Mathematics
Vol. 33, No. 3, pp. 971-996, June 1996 007

GENERALIZED GAUSSIAN QUADRATURE RULES FOR SYSTEMS

OF ARBITRARY FUNCTIONS*

J. MAt, V. ROKHLINt, AND S. WANDZURAt

Abstract. A numerical algorithm is presented for the construction of generalized Gaussian quadrature rules,
originally introduced by S. Karlin and W. Studden over three decades ago. The quadrature rules to be discussed
possess most of the desirable properties of the classical Gaussian integration formulae, such as positivity of the
weights, rapid convergence, mathematical elegance, etc. The algorithm is applicable to a wide class of functions,
including smooth functions (not necessarily polynomials), as well as functions with end-point singularities, such as
those encountered in the solution of integral equations, complex analysis, potential theory, and several other areas.
The perfonnance of the algorithm is illustrated with several numerical examples.

Key words. numerical integration, quadrature rule, Gaussian quadrature

AMS subject classifications. 65D30, 65D32

1. Introduction. Classical Gaussian quadrature rules are extremely efficient when the
functions to be integrated are well approximated by polynomials. When the functions to be
integrated are very different from polynomials, Gaussian quadratures do not perform well;
many particularly difficult and important problems involve the integration of functions of the
form
n

(1) f (x) =L i * (i,

i=l

where each of the functions qi has its own singularity at one of the ends of th
function f can only be evaluated in toto, the coefficients oai being unavail
is encountered in the solution of integral equations with singular kernels, in the numerical
complex analysis, in the numerical solution of elliptic partial differential equations on regions
with corners, and in many other situations. While such problems are normally dealt with by
means of various ad hoc procedures (see, for example, [1], [8]), these schemes lack the rapid
convergence, stability, and elegance of the Gaussian rules.
In fact, in [6], a far-reaching generalization of the classical Gaussian quadratures is intro-
duced, replacing the polynomials with functions from an extremely wide class. The quadra-
ture rules of [6] possess most of the desirable properties of the classical Gaussian integration
formulae, such as positivity of the weights, rapid convergence, mathematical elegance, etc.
Unfortunately, it is not clear from [6] how such quadrature rules can be obtained numerically.
In this paper, we present a numerical scheme for the construction of such generalized
Gaussian quadratures. The algorithm is applicable to a variety of functions, including smooth
functions (not necessarily polynomials), as well as functions with end-point singularities.
The paper is organized as follows. In ?2, we restate the relevant results from [6], and
in ?3, we summarize the numerical techniques to be used in this paper. In ?4, we develop
the analytical apparatus to be used in the numerical construction of the generalized Gaussian
quadrature rules. We extend those analytical tools to functions with end-point singularities in
?5. The actual numerical algorithm is presented in ?6, and the performance of the algorithm
is demonstrated with numerical examples in ?7.

*Received by the editors February 16, 1994; accepted for publication July 22, 1994.
t Department of Computer Science, Yale University, New Haven, CT 06520 (majh @ math.ucla.edu). The research
of these authors was supported in part by Defense Advanced Research Project Agency contract F49620/91/C/0084
and in part by Office of Naval Research grant N00014-89-J-1527.
tHughes Research Laboratories, 3011 Malibu Canyon Road, Malibu, CA 90265. The research of this author
was supported in part by Defense Advanced Research Projects Agency contract F49620/91/C/0064.

971

This content downloaded from

187.19.243.229 on Sun, 24 Mar 2024 12:26:12 +00:00
All use subject to https://about.jstor.org/terms
972 J. MA, V. ROKHLIN, AND S. WANDZURA

2. Mathematical preliminaries. In this section, we summarize several classical results

from [6] to be used in ??4 and 5.

2.1. Chebyshev systems.

DEFINITION 2.1 (Chebyshev system). A finite sequence offunctions {ipl, 'P2, ... q }pm I
be referred to as a Chebyshev system if and only if each of them is continuous on [a, b], and
the determinants

( 1 (X1) y?1 (X2) * P * (Xm)

(P2 (X1l) (P2 (X2) ... (P2 (X. )
(2) det I

(Pm(Xi) Ym(X2) Ym(Xm)

are nonzero for any set of m points Xi, X2, ... ., XE E [a, b] such that xi 7& x
Following are several important cases of Chebyshev systems (for more examples, see
[6]).
EXAMPLE 2.1. For any natural m, the functions 1, x, x2, ... ., xm constitute a Chebyshev
system. Moreover, if Oc1, cr2, ... , Olm is a sequence of distinct real numbers, then the system
{x(Xi } is a Chebyshev system on any interval [a, b] c (0, oo).
EXAMPLE 2.2. For any n distinct real numbers oti, l2 , ... ., On, the functions ealx, xealx,
ea2X, xea2x, ... , eanx, xeanxconstitute a Chebyshev system on any interval [a, b] c (-oc, oc).
DEFINITION 2.2 (Hermite system). A finite sequence of functions {fPl, 'P2w ... , Y2n } will
be referred to as a Hermite system on the interval [a, b] if and only if P E C1 [a, b] for all
i = 1, 2, ... , 2n, and the determinants

( p(Xi) P'P(Xl) PlP(x2) P'1(x2) ... Pi1(Xn) yP(xn)

(P2(X1) (Pq(Xl) (P2 (X2) (X2) ... (P2 (Xn) (P2 (X) 1
(3) det 2

YP2n(Xl) (P'n(Xl) YP2n(X2) y?2n(X2) ... Y2n(Xn) q42(xn)

are nonzero for any set of n points Xi, X2, . .. , Xn E [a, b] such that xi 7& xj for any i 7& j
DEFINITION 2.3 (extended Hermite system). Afinite sequence offunctions {y1, 'P2, * ... *
will be referred to as an extended Hermite system if it is both Chebyshev and Hermite.
REMARK 2.1. The extended Hermite systems are a slight generalization of the extended
Chebyshev systems of [6].
Following are several important cases of extended Hermite systems (for more examples,
see [6]).
EXAMPLE 2.3. For any natural n, the functions 1, x, x2, ... ., 2n-l constitute an extended
Hermite system. Moreover, if oi, Ol2, ... , O2n is a sequence of distinct real numbers, then the
system

(4) ~~~~~~~~XC 1Js X(X2 9 .. 9x C2n

is an extended Hermite system on any interval [a, b] C (0, oo).

EXAMPLE 2.4. An important special case of the proceeding example is thefinite sequence
offunctions

(w) 1, an atxrr n xn+ X2 X2+er rxn-ea nXn-u+

with ot an arbitrary noninteger real number.

This content downloaded from

187.19.243.229 on Sun, 24 Mar 2024 12:26:12 +00:00
All use subject to https://about.jstor.org/terms
 GENERALIZED GAUSSIAN QUADRATURE RULES 973

2.2. Gaussian quadratures. We will be considering integrals of the form

rb

(6) f q(x)p(x)dx,

where q: [a, b] -+ R is a nonnegative function to be referred to as the weight function, and

yp: [a, b] -* R is a function from a suitably chosen class. A quadrature rule is an expression
of the form

(7) Tn (P) = Ewi * ((xi),

i-l1

with xi E [a, b] and wi E R for all i = 1, 2, . .., n. The points xi and coefficients wi are
referred to as the nodes and weights of the quadrature formula (7), respectively, while the
expression (7) itself is viewed as an approximation to the integral (6). Normally, quadrature
formulae are chosen to be exact on certain chosen sets of functions, most frequently polyno-
mials up to some fixed order m. An n-point quadrature formula is referred to as a Gaussian
quadrature if and only if it integrates exactly all polynomials of orders up to 2n - 1.
We will generalize the notion of the classical Gaussian quadrature somewhat by intro-
ducing the following definition.
DEFINITION 2.4 (Gaussian quadrature). Suppose that

(8) {[?P, (P2,9 . . . Y2nI

is a set of integrable functions [a, b] -+ R. We w

Gaussian with respect to the system (8) if and on
(8). In other words, a Gaussian rule is an n-point ru
We will refer to the nodes and weights of a Gaussian quadrature as the Gaussian nodes and
weights, respectively.
REMARK 2.2. Obviously, a classical Gaussian quadrature rule is a Gaussian quadrature
rule for which

1 )= 1,

02(X) =x,

(9) ......

'P2n (X) =

The principal r
THEoREM 2.1 (
a Chebyshev system on the interval [a, b]. Then there exists a unique n-point quadrature
rule (7) that is Gaussian with respect to the functions {(Pl, (P2w . .., Y2n }. Furthermore, all the
weights wi, W2 . w . , wnof the quadrature are positive.
As for smooth functions, Gaussian quadrature rules also exist for a variety of functions
with end-point singularities. The following theorem is an immediate consequence of Theorem
2.1.
THEOREM 2.2. Suppose that functions (Pi : (a, b] -* R are continuous and integrable on
[a, b]for all i = 1, 2, ... , 2n. Suppose also that the function r(x) > 0 is continuous on (a, b]
and integrable on [a, b]. Suppose further that functions *i are defined by the formula

(10) Vri(x) = yj (x)

This content downloaded from

187.19.243.229 on Sun, 24 Mar 2024 12:26:12 +00:00
All use subject to https://about.jstor.org/terms
 974 J. MA, V. ROKHLIN, AND S. WANDZURA

and that

(11) ~~~~~~~lim *jx--a(x) < oo

for all i = 1, 2, .. ., 2n. Supposefinally that the functions lit, /2, ...
constitute a Chebyshev system on the closed interval [a, b].
Then there exists a unique n-point quadrature rule (7) that is Gauss
functions {f1, 02, ..0. , Y2n}. Furthermore, all the weights w , W2, ..
are positive.
Proof. The theorem is proved by applying Theorem 2.1 to the new

(12) 4(x) = q(x) - r(x),

and the new set of functions {M', *2, * * * ,
EXAMPLE 2.5. For any natural n, and real number 0 < ct < 1, the unique n-point
quadrature on the interval [0, 1] with respect to the functions

(13) 1, x- X, x, 1a x 2 x2-a x n-1 xn-l-a

can be obtained via the following Chebyshev system (see Example 2.3):

(14) Xa, 19 X1+a, X, X2+a x2 xn-l+a xn-1

on the interval [0, 1] with the weight function 4(x) = q(x) * x

3. Numerical preliminaries. In this section, we collect the r

be used in ??4 and 5. They can be found, for example, in [3],

3.1. Nested Chebyshev approximation. For any nonnegativ

polynomial Tn of order n is defined by the formula

(15) Tn(cosO) = cos(nO).

Clearly, 1Tn(x)l < 1 forx E [-1, 1].

The Chebyshev polynomials constitute an
the inner product

(16) (f,g)=f f(x).g(x)dx.

Therefore, any function f E C0 [-1, 1] c

00

(17) f(x) = ai Ti(x),

i=o

with the coefficients ai given by

(18) as = (f, T).

Lemma 3.1 states that the Ch

functions. Its proof can be fo
LEMMA 3.1. Suppose thatn a
further that the coefficients
x E [-1, 1],

(19) f(x)-Eai.Ti(x) =0
i=o n
In particular, if f E C00, t

This content downloaded from

187.19.243.229 on Sun, 24 Mar 2024 12:26:12 +00:00
All use subject to https://about.jstor.org/terms
 GENERALIZED GAUSSIAN QUADRATURE RULES 975

I IlI I . I I I I I I I

FIG. 1. Subintervals clustering near the endpoints.

OBSERVATION 3.1. For functions with end-point singularities, such as f (x) - mx,
we can build a structure on the given interval, consisting of subintervals clustering near
the end points (see Figure 1), and then use the Chebyshev expansion (17) to approximate
the functions on each subinterval. On each of the subintervals, the Chebyshev expansion
converges superalgebraically.

3.2. Nonlinear equations. We will be considering systems of nonlinear equations of the

form

(20) F(x) = 0,

where x = (xl, . .X T E

fi (X1, . I ., Xn)
f2(Xlg .. * * Xn)
(21) F(x)= ...

fn (Xl . ,Xn)

DEFINITION 3.1. The Jacobian matrix of mapping F in (21) is defined by the formul

8/i ... ah
axI ax"
(22) DF(x) = . .
_f ... /S
ax1 axn

The following two lemmas about the solution of systems of nonlinear equations (20) are
well known (see [3], [9], for example).
LEMMA 3.2 (Newton's method). Suppose that F : Rn Rn is continuously differentiable
in an open convex set D C Rn, and the mapping G: Rn Rn is defined by the formula

(23) y = x - (DF(x))-1F(x).

Suppose also that x* E Rn is the zero of F, and

(24) IIDF(x*) II <f.

Suppose further that there exist two positiv

thatIIx-x*II < r, and

(25) 1 IDF(x) - DF(y) I 1 y Y I Ix - yl I

for any x, y such that Ix - x*I I < r, I - x*I I < r. Suppose finally that xo is an a
point in Rn, and the sequence x1, X2 ..., ofpoints in Rn is defined by the formul

(26) xk+l = G(xk)

for all k= 0, 1, 2, ....

This content downloaded from

187.19.243.229 on Sun, 24 Mar 2024 12:26:12 +00:00
All use subject to https://about.jstor.org/terms
976 J. MA, V. ROKHLIN, AND S. WANDZURA

Then there exist E > 0 and a > 0 such that the sequence generated by (26) converges to
x*, and

(27) IIXk+1 - X*I < 0IIXk - XI2

for any xo such that IIxo-x*II < E.
LEMMA 3.3 (modified Newton's method). Suppose that under the assumptions of Lemma
3.2, x* is the zero of the mapping F: D -- Rn. Suppose also that

(28) Ao, A1, A2, . .

is a sequence of n x n nonsingular m
formula

(29) y = x-A A1F(x).

Suppose further that there exists a

(30) IlAk - DF(xk) I1 < M * IIF(xk) I I.

Suppose finally that xo is an arbitrary point in Rn, and the sequence xl, X
Rn is defined by the formula

(31) Xk+l = G(Xk)

for allk = 0, 1,2, ....

Then there exist E > 0 and a > 0 such that the sequence generated by (31) converges to
x*, and

(32) I IXk+1-X* - I < 1 IXk - X*lI

for any xO such that IIxo - x*I < E.
3.3. Continuation method. The Newton algorithm for the solution of systems of non-
linear equations is an extremely powerful technique, provided that a satisfactory initial point
is available. In many cases, a starting point is not available directly, but can be obtained by the
process known as the continuation method (otherwise referred to as the homotopy method).
Following is a brief description of the technique.
Suppose that we are trying to solve a system of nonlinear equations

(33) F(x) = 0,

with F: Rn -- Rn satisfyin
is not available. Suppose
satisfying the following c
1. The mapping Fo: Rn

(34) Fo(x) = G(0, x)

has a simple structure, so that the s

(35) Fo(x) = 0

is unique and known.

2. ForallxeR ,

(36) G(1, x) = F(x).

This content downloaded from

187.19.243.229 on Sun, 24 Mar 2024 12:26:12 +00:00
All use subject to https://about.jstor.org/terms
 GENERALIZED GAUSSIAN QUADRATURE RULES 977

3. For all t E [0, 1], the equation

(37) G(t,x) =0
has a unique solution x(t) E R
of x(t).
4. x is a continuous (or better, Lipschitz) function of t.
Under the above conditions, the following procedure yields the solution of the equation
(33).
1. For a sufficiently large m, construct the points t = (i -1)/(m - 1) on the interval
[0, 1], with i = 1, 2, ... , m, and consider the solutions of the equation (33) for t = t1 with
i = 1,2,...,m.
2. Clearly, we know the solution x(O) of the equation,

(38) G(ti, x) = 0
when i = 1, and for all i = 2, 3, ... , m, we solve the equation (37) by means of Lemma 3.3,
using x(ti-1) as the initial approximation.
3. Since tm = 1, the result of the final step of this process is the solution of the equation
(33).
A detailed discussion of the continuation techniques can be found in [4], where the
convergence of the above scheme is proven (in a much more general environment) for all
sufficiently large m.

4. Analytical apparatus. In this section, we develop analytical tools to be used in the

numerical construction of the Gaussian quadratures whose existence follows from Theorem
2.1.

4.1. Construction of Gaussian quadratures. The following lemma is an immediate

consequence of Definition 2.3 of extended Hermite systems.
LEMMA 4. 1. Suppose that the functions {Cp1, 5?2, . . . , 92n} constitute an Hermite system on
the interval [a, b], andxl, X2, . v. xn are n points on the interval [a, b] such thatxi $& xj for any
i =# j. Then there exist such unique coefficients tij, ,ij, i = 1, 2, ... , n, j = 1, 2, . . ., 2n
that
JT Q(Xk) = 0,
(39) 1 (x)= 3k
( li (Xk) = 5ik,

(40) { i (Xk) = 0ik,

for all i = 1, 2, ... , n, k = 1, 2, ... , n, with

functions ai, rii defined by the formulae
2n

(41) ci (x) = E (xi j * f (x),

j=1
and
2n

(42) /iE (x) = , ,Bi, i (Pi (x).

j=1

Furthermore, there exist unique coefficients aij, bij with i = 1, 2, .. ., 2n, j = 1, 2, .. ., n,

such that
n

(43) Pi (x) = Lai, j j (x) + bi, j * i.j (x)

j=l

for all i = 1, 2, ..., 2n.

This content downloaded from

187.19.243.229 on Sun, 24 Mar 2024 12:26:12 +00:00
All use subject to https://about.jstor.org/terms
978 J. MA, V. ROKHLIN, AND S. WANDZURA

In other words, there exist unique linear combinations o,, o2, . . n.

functions 9p1, 902, ... -, 92n satisfying the conditions (39), (40), (43
991, 992, . * *, 992n are linear combinations of the functions or1, o2, . .
Theorem 4.1 below is the principal analytical tool of this paper. It
and sufficient conditions under which a quadrature is Gaussian with
system.
THEOREM 4.1. Suppose thatfunctions

(44) 901, 902,. ***2n

constitute a Hermite system on the int

Gaussian with respect to the function

rb

(45) fbq(x)ai(x)dx = 0

for all i = 1, 2, .. ., n. In this case, the

formula

(46) wi = Jbq(x) 7j(x)dx,

with functions or, a2, . . . c,1 t, r72, ...,

Proof First, we show that for any Gau
are satisfied. Indeed, since it integrates
integrates exactly all their linear combin
follow immediately from (39), (40).
Suppose now that the nodes xi, x2, .. ., x n are such that the conditions (45) are satisfied,
and the coefficients Wi, W2, .. ., wn are defined by the formula (46). We will show that the
n-point quadrature (7) is Gaussian with respect to the system (44).
Due to Lemma 4.1, there exist coefficients Ci,, jBi,j, i = 1, 2, . ..., n, j = 1,2,..., 2n
such that
n

(47) 9j= T(1o

(47)~~~~~~~~~~P , (aij b1j,
orj + bij i7j)
j=1

for any i = 1, 2, . .. , 2n. Thus

b n b b

(48) q(x
a j=l a aj=

Combining (48) with (45), (46), we have

b n

(49) fbq(x)(ij(x)dx = bijwj

j=1

On the other hand, combining (39), (40), and (48), we obtain

n n n

E Wpj i(Xj) = L Wj E (aikck(Xj) + bik77k(Xj))

j=1 j=j k=1
n

(50) = bijwj.
j=l

This content downloaded from

187.19.243.229 on Sun, 24 Mar 2024 12:26:12 +00:00
All use subject to https://about.jstor.org/terms
 GENERALIZED GAUSSIAN QUADRATURE RULES 979

Combining (49) and (50), we finally get

b n

(51) f q(x)pi (x)dx =E w(pi (xj).

j=1
for all i = 1, 2, . .. ,2n. []
Theorem 4.2 below follows immediately from Theorem 4.1. It describes the Gaussian
nodes as the solution of a system of nonlinear equations.
THEOREM 4.2. Suppose that the functions {qpj, qp2, .. ., pn I constitute a Hermite system,
and functions {fr,, 72, . .. , rn} and {r i, ?72, . . ., z1n} are defined by the formulae (39), (40),
(41), and (42). Suppose further that S is a subset of Rn consisting of all finite sequences
{x1, X2, ... , xn } such that xi ? xj whenever i 5 j . Suppose finally that the mappin
F: S -* Rn is defined by the formula

XI xl'fa q (x)al (x)dx

(52) X2 fa Iq(x)r2(x)dx J

Xn f(a q (x)orn (x)dx

Then {xi, X2, ...,x I are the Gaussian nodes with respect to the system of functions
{qp1, q02, * pn. , yn} if and only if

(53) ~~~~F(x , . . .,xn) =

4.2. Computation of Gaussian quadratures. In this subsection, we observe that the
modified Newton method in ?3.2 assumes a particularly simple form when it is applied to
the system of equations (53), and establish Theorem 4.5, the principal numerical tool of this
paper. Theorem 4.5 shows that an extremely simple iterative scheme converges quadratically
for the system of equations (53).
Theorem 4.4 below provides an analytical expression for the Jacobian matrix of the
mapping F defined by the formula (52). Theorem 4.3 is the consequence of Lemmas 4.2 and
4.3, and will be used in the proof of Theorem 4.4. Theorem 4.5 follows immediately from
Lemma 3.3, Corollary 4.2, and Theorem 4.4.
LEMMA 4.2. Suppose that functions {q'1, 02 .. , 2n } form an extended Hermite system
with qPi E C3 [a, b] for all i = 1, 2,..., 2n. Suppose also that

(54) Xl, x2,.. ., Xn

are n distinct points on the interv

(55) r, (O2 * * * 7n ) 77 1 * * * s 77n

are determined by the set ofpoints (54) via formulae (39) and (40).
an integer such that 1 < 1 < n, and a is a real number such that

(56) xi,... I xl_1, xI + 3, x1+19. * I Xn

are n distinct points on interval [a, b]. Suppose finally that the functio

(57) 61, (2, * * 6n 51, )2 * * * &

are determined by the set ofpoints (56) via formulae (39) and (40).
Then there exist coefficients ail and Pi, with i = 1, 2, ... , n, such that

(58) &i (x) = ri (X) + ceii - Cl (X) + fiij * r1 (x)

This content downloaded from

187.19.243.229 on Sun, 24 Mar 2024 12:26:12 +00:00
All use subject to https://about.jstor.org/terms
980 J. MA, V. ROKHLIN, AND S. WANDZURA

for alli 1,i = 1,2,...,n, and

(59) al(x) = Call * ai(x) + ,ii * (x)

Proof. Due to Lemma 4.1, there exist unique coefficie

j = 1, 2, ...,n, suchthatforall i = 1, 2,..., 2n,
n

(60) ai(x) = E (ctij * orj(x) + ,iij * W)

j=1

for any x E [a, b]. Differentiating (60), we have

n

(61) &' (x) = (ij * Oc'(x) + ,8ij * 7j(x)).

j=1

For the functions (57), the conditions (39), (40) assume the form

(62) &f &(xi+3) = 0,

(2 & (xi+3) = 3i1,
and

&i (xk) = 0,
(63) &?(Xk) = Sik

for all i = 1, 2, ... , n, k = 1, 2, ... , n, and k $ 1.

For any k $ 1, evaluating (60) at Xk and applying conditions (39), (40), (63), we obtain

(64) Pik = 0

for alli = 1,2,...,n.

Similarly, for any k $ 1, evaluating (61) at Xk and applying conditions (3
we have

(65) Oxik = 3ik

for alli = 1, 2, ...,n.

Now, (58), (59) immediately follow from (60), (64), (65). [
LEMMA 4.3. Suppose that under the assumptions of Lemma 4.2, the coeffcients ail and
Pil are defined via formulae (58) and (59) for all i = 1, 2, .. ., n, Then

(66) ail = - 3 yi (xl) + Q(82)

(67) 8i1 = Q(82)

for alli $ 1,i = 1,2,...,n, and

(68) ol, = 1- * or, (xl) + )

(69) pl8 = 3 + O(32).

Proof. Expanding the functions

(Ti, 77i, (fi, 77i

into the Taylor series at x1, we have

(70) i vi(xi+3) = a,(Xl)+8 (Xl)+ (32),

(70) ~ ~~~~ ori (xi +A= n r fi .n., f {

This content downloaded from

187.19.243.229 on Sun, 24 Mar 2024 12:26:12 +00:00
All use subject to https://about.jstor.org/terms
 GENERALIZED GAUSSIAN QUADRATURE RULES 981

(71 (xi + 8) = oi'(xi) + a <'(xl) + Q(o2)

(71) 7 (Xl + 8) = 17(X1) + a 17 (Xl) + o(2

for alli = 1, 2, ...,n.
Evaluating (58) and (59) at x = (xl + 8) and using the conditions (39), (40), (62), (70),
we obtain

(72) 8i1 = -an 3 + Q(32)

for alli = 1, 2, ..., n.

Differentiating (58) at x = (xl + 8) and using conditions (39), (40), (62), (71), we have

0a1 = (1 + a * T(xi)) (1 + 0(82))

(73) = 1 -3 * al(xl) + O(32).

Similarly, differentiating (59) at x = (xl + 8) and ap

and (71), we get

(74) cx,, = -3. "(Xl) + Q(82)

foralli = 1,2,d,...,n,andi $ 1.
Finally, combining (72) with (73), (74), we have

(75) fil = 3 + Q(32),

(76) 8i1 8 = Q(32)

foralli = 1,2,...,n,andi $ 1. [
Combining Lemmas 4.2 and 4.3, we now obtain the following the
THEOREM 4.3. Under the assumptions of Lemma 4.2,

(77) ali(x) = ali(x) - * or, (xi) *r C(x)-_8. * l(X) + o(82)9

(78) &i(x) = cri(x) - ai"(xi) aCl(x) + 0 ( )2)
for anyi = 1,2,...,n, andi $l1.
The following theorem is an immediate consequence of Theorem 4.3. It provides a
expression for the Jacobian of the mapping (52), showing that the latter is nearly dia
the vicinity of the solution of the equation (53).
THEOREM 4.4. Suppose thatfunctions {p1, p2 . p. ., ) }form an extended Hermite sy
and qPi E C3[a, b]for i = 1, 2, ... , 2n. Suppose further that xl, X2, ... , x, are n dist
points on the interval [a, b], andfunctions

(71,9 (72 * .. * (Tns t)l 771 2 * .. * 77tn

are determined byformulae (39) and (40). Then the Jacobian DF(x) of the mapping F d
by (52) is given by the formula

fqa q (x) (x) dx 0 ... 0

b~~~~~~

(79) DF(x) =f- f'q(x)7)2(x)dx f.. O |-E(x),

0 ? .. f a q (x) in(x) dx

This content downloaded from

187.19.243.229 on Sun, 24 Mar 2024 12:26:12 +00:00
All use subject to https://about.jstor.org/terms
 982 J. MA, V. ROKHLIN, AND S. WANDZURA

where x = (Xi, X2, ... Xn)T, and matrix E(x) is given by the for

01(X1) .. * 0"(Xn) b 1( (X (x)dx 0 . .. 0

(80) E(x) = 0 0
n (X)) E ( = n (Xn) ) (f ?x?*)*0 . a' q(x)omn(x)dx)

Proof Suppose that 1 is an integer such that 1 < 1 < n, and 3 is a real numb

(81) x1, . .. , xlhi, xl + 8, xl+1, . ..., Xn

are n distinct points on the interval [a, b]. Suppose further that the

(82) 1, 62, * * *n, t19 /29 * * *

are determined by the set of points (81) via formulae (39) and
Combining (77), (78) with Definition 3.1 of the Jacobian matrix and the definition (52)
of the mapping F, we immediately obtain

a lb
(DF(x))il = ax 1b q(x)oai (x) dx

= li fa q(x)&l (x)dx - fq(x)OaI(x)dx

3--*O~~~~
b b

(83) =-1 (xi) q (x)a,(x)dx - q(x) i(x)dx

for any i 1, i 1, 2, ... , n, and

ab
(DF(x))11 = - q(x)ol (x)dx
ax,
= qli (x)61i (x)dx-bq (x)cui (x)dx

(84) = qJ(xl) q(x)al(x)dx. [

Corollary 4.1 follows immediately from Theorem 2.1 and Theorem 4.4, and Corollary
4.2 is the consequence of Corollary 4.1 and Theorem 4.4.
COROLLARY 4.1. Suppose that under the assumptions of Theorem 4.4, the function F:
Rn Rn is defined by (52). Then there exists a unique x* = (x*, x*, ... , xn*)T E Rn such
that

(85) F(x*) = 0,

and the Jacobian matrix

fa qb~~~~~
(x) i7 (x) dx O .
(86) DF(x*) = -_0 q. fX)q7)2(x)dx ... O

? ? .. f a q (X) 7n (x) dx
is nonsingular, where the functions rn7, r72, ..., zj,n are determin
X*, X*, ... ., x* via the formula (40).

This content downloaded from

187.19.243.229 on Sun, 24 Mar 2024 12:26:12 +00:00
All use subject to https://about.jstor.org/terms
 GENERALIZED GAUSSIAN QUADRATURE RULES 983

COROLLARY 4.2. Suppose that under the assumptions of Theore

Rn- Rn is defined by (52), and x* is the unique zero of F. T
differentiable, and there exist three positive real numbers r, f, an

(87) I IDF(x*)-' 11 < p

and

(88) IIDF(x) - DF(y)II yIIx -y

for any x and y such that IIx-x*II < r, IlY - x*II < r.
The following theorem is the principal numerical tool
extremely simple iterative scheme converges quadratical
and is an immediate consequence of Lemma 3.3, Corolla
THEOREM 4.5. Suppose thatfunctions {CPi, CP2, ... 2n }f
and the mapping G: Rn -* Rn is defined by the formula

(89) Yj=Xj+ Jf q(x)c1i(x)dx

fa q(x) i(x)dx
with i = 1, 2, ... , n, and the functions

O'l ,(72 ,..* s(yn977 1 9772 9.. 77)n

defined by the points xi, X2, .. ., x n via the formulae (39) and (40). Suppose further
qOj E C3[a, b] for all i = 1, 2, ... , 2n, and the function F: Rn -* Rn is defined b
Suppose finally that x* is the unique zero of F, that xo is an arbitrary point in Rn, an
sequence xi, x2, .., ofpoints in Rn is defined by the formula

(90) Xk+1 = G (Xk)

for allk = 0, 1,2....

Then there exists E > 0 and a > 0 such that the sequence xl, x2, .. ., gener
converges to x*, and

(91) IIXk+1 - X*l < llXk - X*112

for any initial point xo such that I Ixo-x* I I < e.

REMARK 4.1. In Theorem 4.5, we impose the condition

(92) 'Pi E C3[a, b]

for all i = 1, 2, ... , 2n. However, it can be
cessively restrictive, and a somewhat more
tinuously differentiable, and qp4 satisfy th
modified Newton method (90) will still con

5. Integration of singular functions. In t

Gaussian quadrature rules established in ?4
end-point singularities. We will first intr
Hermite systems in the case of singular fu
Theorem 5.2 for the construction of Gaussi
construction for Gaussian quadratures for f
DEFINITION 5.1 (Chebyshev system). Sup
(a, b]. Then afinite sequence of functions {qp

This content downloaded from

187.19.243.229 on Sun, 24 Mar 2024 12:26:12 +00:00
All use subject to https://about.jstor.org/terms
984 J. MA, V. ROKHLIN, AND S. WANDZURA

system on X if and only if it constitutes a Chebyshev system -on every closed subinterval
[c, d] c X (see Definition 2.1).
DEFINITION 5.2 (extended Hermite system). Suppose that X c R is either (a, b), or
[a, b), or (a, b]. Then a finite sequence offunctions {'1p, q'2, .. . -, (Pn will be referred to a
an extended Hermite system on X if and only if it constitutes an extended Hermite system on
every closed subinterval [c, d] c X (see Definition 2.3).
The following is an important example of extended Hermite systems.
EXAMPLE 5.1. For any natural n, the functions

(93) 1, xa, x, x1+a, x2 ,x2+ ... xn-1 xn-+

constitute an extended Hermite system on the interval (0, 1] with x

real number (see Example 2.3 and Definition 5.2).
THEOREM 5.1. Suppose thatfunctions {(O1, (02, ... 0 '02nn} are all
constitute a Chebyshev system on (a, b]. Suppose further that on th
n-point Gaussian quadrature (7) is given by the nodes

(94) xl3 9 X239. 9 * * ) X

and the weights

(95) W(, W2 *** Wn

for any E (O,b-a).

Then for any E > 0, there exists 80 > 0 such thatfor all a < 80,
b n

(96) q (x) oi (x)dx- w(8)(oi(x/)) <E

j=1

for all i = 1, 2. 2n.

Proof. Due to Definition 2.4 of the Gaussian quadratures, we have

b n

(97) q (x) (p o(x)dx = L uj)(oj(x))

+ yj=1

for all i = 1, 2, ... , n. Subtracting

j q (x)(po (x)dx

from both sides of (97), we obtain

,b n ,b {b

f q (x) o
j=a 1+
ra+3

(98) jac q (x) p1 (x)dx.

Due to the assumption that fp

for any E > 0, there exists 80 >
ra+3

(99) q(x)(po (x)dx < E

for all i = 1, 2, ... , 2n.

This content downloaded from

187.19.243.229 on Sun, 24 Mar 2024 12:26:12 +00:00
All use subject to https://about.jstor.org/terms
 GENERALIZED GAUSSIAN QUADRATURE RULES 985

Now, (96) follows immediately from (98), (99). [

THEOREM 5.2. Suppose that under the assumptions of Theorem 2.2 and Theorem 5.1

(100) P1 -1.

Then there exists a uniqu

(101) (01, 902, . * * .(02n

such that all the nodes X1, X2, . X. , Xn lie

Wl, W2, ..., Wnare positive. Furthermore, f

(102) limx1 = Xi

and

(103) lm w wig

where the nodes xi and the weights wi are defined in Theorem 5.1.
Proof Due to Theorem 2.2, there exists a unique Gaussian quadrature (7) with respect to
the functions (101) on the interval [a, b] such that all the nodes xl, X2, ... , xn lie in the open
interval (a, b), and all the weights w 1, W2, ..., wn are positive.
On the other hand, due to Theorem 2.1, for any a E (0, b - a), there exists a unique
Gaussian quadrature (7) with respect to the functions (101) on the interval [a + 8, b] such that
the nodes xi E (a + 8, b), and the weights

(104) 0) > 0
for all i = 1, 2, ... , n. Comb
obtain
b n

(105) / q(x)(pj(x)dx = Lw W
J+ i=l

Now, for any a E (0

(106) XS = (x(), x2 ),... ,x x)) T

(107) wS = ( w( ), .. .,w)) T
where xi and w( are the Gaussian nodes and weights with
interval [a + 8, b].
Clearly, for any a E (0, b - a), we have

(108) llxsll < c /

with c = max(lal, Ibi).
Combining (105) and (104), we obtain

(109) Ilwsl < d VH

for any a E (0, b - a), with d given by the formul
b

(110) d= jbq(x)dx.

We will show that for all a E (0, b - a), ther

vectors xs, and only one limit point for the set

This content downloaded from

187.19.243.229 on Sun, 24 Mar 2024 12:26:12 +00:00
All use subject to https://about.jstor.org/terms
 986 J. MA, V. ROKHLIN, AND S. WANDZURA

Suppose that there exists a sequence of positive real numbers 81, 82, such that
limk,, 'k = 0,

k *oo

(112) M WSk =V'

k->ooo

with y = (YI, Y2, ...,Yn) and v = (vl, v2, ... , vn)T.

Due to Theorem 5.1, Definition 2.4 of the Gaussian quadratures, and conditions (111),
(112), we have

rb rb
bq (x)cp; (x)dx = lim f q (x)(p (x)dx
Ja ~ ~~~~~ ke.oo J+3k
n

= Lm k) E (i(xy )

(113) = LVjoi(yj)
j=1

for alli= 1, 2, .. ., 2n.

Due to (113) and the uniqueness of the Gaussian quadrature (7) with respect to the
functions (101) on the interval [a, b], we have

(114) yi = x,
(115) Vi = Wi
for all j = 1, 2, .. ., n, where xl,, x
of the Gaussian quadrature (7) with respect to the functions (101) on [a, b]. In other words,
the set of vectors xs for all a E (0, b - a) has a unique limit point

XO= (xi,x2, ... X

and the set of vectors ws for all a E (0, b - a) has a unique limit point

Wo = (Wi, W2, ..., Wn) T.

Now, the formulae (102) and (103) follow from (108), (109), and the fact that each of the
two sets of the vectors xs and ws possesses only one limit point in Rn. [1
REMARK 5.1. Clearly, the condition pl =1 in Theorem 5.2 can be relaxed. To ensure the
boundedness of the Gaussian weights, we need only to impose the condition that there exists
a real e > 0, and 2n real numbers al, a2, ... a2n such that
2n

(116) Lci * p (x) >E

i=l

for all x E (a, b).

6. The numerical algorithm. We can now compute Gaussian quadratures for both
smooth functions and functions with end-point singularities using the numerical apparatus
developed in ??4 and 5. The Gaussian quadrature rules for an extended Hermite system can
be obtained by solving a system of nonlinear equations (53). Due to Theorems 4.5 and 5.2,
the modified Newton's method defined by the formula (90) converges quadratically when it is
applied to the system of equations (53).

This content downloaded from

187.19.243.229 on Sun, 24 Mar 2024 12:26:12 +00:00
All use subject to https://about.jstor.org/terms
 GENERALIZED GAUSSLAN QUADRATURE RULES 987

As is well known, the Newton method ig sensitive to the c

xo, and we use the continuation method (see ?3.3 above) to
given an extended Hermite system

(117) O1, , ..., (02n,

on the interval [a, b], we construct a family of extended Hermite systems

(118) t t
(118)~~~~~~~~~ 1 ( p2 * s(2ns
with t E [0, 1], and such that

(1 19) W(X) =x

(120) f-O (x)-fi (x)=

for all x E [a, b], and i = 1, 29 .... 2n

equations (53) corresponding to the extended Hermite system (118) via the formulae (39),
(40). Clearly, for the extended Hermite system (119), the solution of the system of equations
(53) is known (see Remark 2.2), and we use the continuation method (see ?3.3) to obtain the
solution of (53) for the Hermite system (117).
In the numerical examples of the following section, the one-parameter families of Hermite
systems are constructed as follows.
1. For Hermite systems of the form

(121) JO, J1 ..., * 2n-19

the one-parameter family of systems is

(122) t(x) = (1 - t) * xi-' + t * Oi(x)

for i = 0, 1, ..., 2n-1.
2. For Hermite systems of the form

(123) 1, nx, x, x nx, x2, x2 ix, ... , xn-1 xn-1 x,

the one-parameter family of systems is given by (122).

3. For Hermite systems of the form

(124) 1, s(x), x, xs(x), x2, x2s(x) ... , xn-1 xn-1s(x),

with

(125) s(x) = x',

and ox an arbitrary noninte

(126) 1, s(t, x), x, x * s(t, x) x 2, x2 s(t, x), ... x1 x1 s(t, x),

with

(127) s(t, x) = xt'.

REMARK 6.1. The necessary number m of steps in the continuation process (see ?3.3) is
significantly reduced if, prior to the application of the above procedure, the original system
(117) is orthonormalied (for example, via the Gram-Schmidtprocess). To do that, we discretize
the originalfunctions (Pi at nested Chebyshev nodes (see ?3.1), andperform the Gram-Schmidt
procedure on the obtainedfinite-dimensional representations.
The following is the formal description of the numerical algorithm (excluding the con-
tinuation process).

This content downloaded from

187.19.243.229 on Sun, 24 Mar 2024 12:26:12 +00:00
All use subject to https://about.jstor.org/terms
988 J. MA, V. ROKHLIN, AND S. WANDZURA

Initialization
Comment [ Build the structure for integration and interpolation. ]
Step 1
do
Subdivide [a, b] into subintervals clustering near end-points (see Figure 1).
enddo
Orthogonalization (optional)
Comment [ Perform Gram-Schmidt orthogonalization on the given set of functions.]
Step 2
doi = 1,2,...,2n
do j = 1,2, ...,i-1

Orthogonalize the ith function ypi with respect to jth funct

endo
enddo
Nested Chebyshev Approximation (optional)
Comment [ Generate the nested Chebyshev expansions for the orthogonalized functions
(see ?3.1). ]
Step 3
doi = 1,2,...,2n
Construct the local Chebyshev expansion of the ith function qpi
Observation 3. 1.
enddo
Newton's Iteration
Comment [ Conduct Newton's iteration to find Gaussian nodes and weights.]
Step 4
do
Construct functions o-i and 77i for i = 1, 2, ... , n via formulae (39) and (40).
enddo
Step 5
do
Adjust Gaussian nodes {xi } via formulae (90).
enddo
Step 6
do
Compute error = IIxk+I - xkII
If error > e, Go to Step 4.
enddo

REMARK 6.2. The procedure described above requires the construction of the functions
oi, ?7i, i = 1, 2, ... , n, given the functions Cpi, i = 1, 2, ... , 2n. The latter is possible for any
extended Hermite systems and is equivalent to inverting the matrix (3). Obviously, for many
choices offunctions qp, 5?2 . , (02n, including the numerical examples given in the following
section, the matrix (3) will be ill conditioned.
Thus, in order to obtain the double precision results presented in this paper, the authors
have performed the computations in extendedprecision (REAL * 16)for generating the Gauss-
ian quadratures up to order 20, and in Mathematica (120 digits operations) for generating
the Gaussian quadratures of higher orders.

7. Numerical results. We have implemented the numerical algorithm described in ?6

for the computation of Gaussian quadrature formulae, and tested it on various examples.

This content downloaded from

187.19.243.229 on Sun, 24 Mar 2024 12:26:12 +00:00
All use subject to https://about.jstor.org/terms
 GENERALIZED GAUSSIAN QUADRATURE RULES 989

EXAMPLE 7.1. Gaussian quadratures with respect to the system offunctions

(128) 1,Inx,x,xlnx,x2,x2lnx, ... n-1,xn-1 nX

on [0, 1] are given in Table 1, and tested on selected functions in Table 5.

EXAMPLE 7.2. Gaussian quadratures with respect to the systems offunctions

(129) 1 x xxl+a x2 x2+a. xn-1 xn-l+ac

on [0, 1] are given respectively in Tables 2-4 for

1 1 2
39 3' 3'
and tested on selected functions in Tables 6-8, respectively.
REMARK 7.1. The algorithm of this paper has been applied to a large selection offunctions
{Jj }. Space limitations prevent us from presenting more examples here. A detaiWed report on
our numerical experiments can be found in [7].

8. Conclusions. A numerical algorithm has been presented for the construction of the
generalized Gaussian quadrature rules, introduced in [6]. The quadrature rules of this paper
possess most of the desirable properties of the classical Gaussian integration formulae, such
as positivity of the weights, rapid convergence, mathematical elegance, etc. The algorithm
is applicable to a wide class of functions, including smooth functions (not necessarily poly-
nomials), as well as functions with end-point singularities, such as those encountered in the
solution of integral equations, complex analysis, potential theory, and several other areas.

TABLE 1
Gaussian quadrature for products of polynomials and logarithmic finction.
I l N
Ok(x)dx = WiOk(xi) for k = 1, 2,..., 2N
O ~~~i=l

where {p} = {I, hix, x, x hix, ..., xN 1

[WN Nodes xi I Weights w

5 0.565222820508010E-02 0.210469457918546E-01
0.734303717426523E-01 0.130705540744447E+O0
0.284957404462558E+O0 0.289702301671314E+O0
0.619482264084778E+O0 0.350220370120399E+O0
0.915758083004698E+O0 0.208324841671986E+O0
10 0.482961710689630E-03 0.183340007378985E-02
0.698862921431577E-02 0.134531223459918E-01
0.326113965946776E-01 0.404971943169583E-01
0.928257573891660E-01 0.818223696589036E-01
0.198327256895404E+O0 0.129192342770138E+O0
0.348880142979353E+O0 0. 169545319547259E+O0
0.530440555787956E+O0 0.189100216532996E+O0
0.716764648511655E+O0 0.177965753961471E+O0
0.875234557506234E+O0 0.133724770615462E+OO
0.975245698684393E+OO 0.628655101770325E-01

15 0.105784548458629E-03 0.403217724648460E-03
0.156624383616782E-02 0.306297843478700E-02
0.759521890320709E-02 0.978421211876615E-02
0.228310673939862E-01 0.215587522255813E-01
0.523886301568200E-01 0.383230673708892E-01
0.100758685201213E+O0 0.588981990263004E-01
0.170740768849943E+O0 0.811170299392595E-01
0.262591206118993E+O0 0.102122101972069E+O0
0.373536505184558E+O0 0.1 18789059030401E+O0
0.497746358414533E+O0 0. 128210316446694E+O0
0.626789031392373E+O0 0.128163327417093E+O0
0.750516103461408E+O0 0.1 17489465888492E+O0
0.858255335207861E+O0 0.963230185695904E-01
0.940141291212346E+O0 0.661345398318934E-01
0.988401595986342E+O0 0.296207140035355E-01

This content downloaded from

187.19.243.229 on Sun, 24 Mar 2024 12:26:12 +00:00
All use subject to https://about.jstor.org/terms
990 J. MA, V ROKHLIN, AND S. WANDZURA

TABLE 1

(cont.)

N . Nodes xi Weights w= [II Nodes xi ; Weights wi j

20 0.352330453033401E-04 0. 134499676467758E-03 35 0.401098093701052E-05 0.153323972344389E-04
0.5260939825174lOE-03 0.103477692295062E-02 0.603496493961345E-04 0.1 19314925936177E-03
0.258751954058141 E-02 0.337726367723322E-02 0.300616431782387E-03 0.396956727600450E-03
0.79344719483804l E-02 0.767355619359468E-02 0.938413243142584E-03 0.926719742189694E-03
0.186828881374457E-01 0.142054962855420E-0l 0.2261 10954001356E-02 0.177722277468937E-02
0.370976733697505E-01 0.229844384632086E-0l
0.461861900484087E-02 0.300450660226082E-02
0.653124886740214E-01 0.337363605577136E-01 0.840960506728443E-02 0.464967479805987E-02
0.10504850471 1551 E+00 0.459147630734522E-01
0.140655627785534E-01 0.673701850949898E-02
0. 157359691819002E+00 0.587404799428040E-01
0.220332949959123E-01 0.927268937587387E-02
0.222430062767455E+00 0.71265013161 1020E-01
0.327563552513284E-01 0.122439660322322E-01
0.299443765654 IOOE+00 0.824518089775832E-01
0.466560658935003E-01 0. 156191410212273E-01
0.386542446943882E+00 0.912682015163873E-01
0.641127372512586E-01 0.193480355175793E-01
0.480876453826790E+00 0.967797159091613E-01
0.8544771 11324605E-01 0.233631296006668E-01
0.578747932205507E+00 0.982381433400897E-01
0.1109068299941 1SE+00 0.275812764915292E-01
0.675835475840038E+00 0.951553030540297E-01
0.140645892246187E+00 0.319059507976849E-01
0.767482460872564E+00 0.873556504104574E-01
0. 174718595617135E+00 0.362299639512362E-01
0.849025253970320E+00 0.750027772122717E-01
0.213067396094360E+00 0.404385652051107E-01
0.916133703241664E+00 0.585972958082337E-01
0.255517621930679E+00 0.444128342291967E-01
0.965135427900256E+00 0.3894725054961 14E-01
0.301775083274460E+00 0.480332619050301E-01
0.993303536456954E+00 0. 171372052681059E-01
0.35142731 1172301E+00 0.511834096518771E-01
25 0.148805205646725E-04 0.568460660251700E-04 0.403948448291627E+00 0.5375353472157 IOE-01
0.223091159576968E-03 0.439997585769338E-03 0.458707701 144855E+00 0.556440694728935E-01
0.1 10464364905582E-02 0. 145071890475996E-02 0.514981153335511 E+00 0.567688466612984E-01
0.341946946888592E-02 0.334401873817378E-02 0.571966634792308E+00 0.570579701395159E-01
0.815052929503389E-02 0.630809954735919E-02 0.628801246341 147E+00 0.564602408366056E-01
0. 164289374947977E-0 1 0. 104488723103534E-01 0.684581055249250E+00 0.54945061 1514993E-01
0.294459835598650E-01 0.157795036631362E-0 1 0.738382408163506E+00 0.52503756563515 1 E-01
0.483575697079336E-01 0.222157908473762E-01 0.789284255328327E+00 0.491502708662106E-01
0.741870939197324E-01 0.295777024141012E-01 0.836390845802486E+00 0.449212104911785E-01
0.107732955883526E+00 0.375970456071838E-01 0.878854138847408E+00 0.398752335254109E-01
0.149486638258087E+00 0.459308515949819E-01 0.915895282821554E+00 0.340918004630718E-01
0. 199566730959288E+00 0.5417972366570(0E-01 0.946824542300002E+00 0.276693300400168E-01
0.257673355831325E+00 0.619100915223073E-01 0.971059120835061 E+00 0.207228620420109E-01
0.323066266406720E+00 0.686790748928476E-01 0.988138551329524E+00 0.133817082903205E-01
0.394568512069187E+00 0.740604961651023E-01 0.997739722313237E+00 0.579513447814728E-02
0.470596049553408E+00 0.776705045127631 E-01
40 0.237684143879143E-05 0.908716648479552E-05
0.549212146433180E+00 0.791912877674548E-01 0.357941 198659536E-04 0.708096570276615E-04
0.628203942243430E+00 0.783914525088150E-01
0. 178564191524854E-03 0.236107924249676E-03
0.705177201069316E+00 0.751418416138532E-01 0.558572352357276E-03 0.552941324982586E-03
0.777664184415814E+00 0.694258212157633E-01 0. 134949361265519E-02 0. 106471907669186E-02
0.843238762138573E+00 0.613433919048206E-01
0.276558618788719E-02 0.180898827559081 E-02
0.899632416106180E+00 0.511088512980029E-01
0.505522381417743E-02 0.281624906827513E-02
0.944844733405715E+00 0.390421895640092E-01
0.849332816182978E-02 0.410894500639753E-02
0.977242575226693E+00 0.255554713626328E-01
0. 133728905726761 E-01 0.570065525023859E-02
0.99564721545644 1E+00 0.11 1503547267078E-0 1 0. 199957958342846E-01 0.759550947534350E-02
30 0.732379744272606E-05 0.279892154309547E-04 0.286631788855304E-01 0.978784093517871 E-02
0.1 10044700457775E-03 0.217365526502542E-03 0.396655579573064E-01 0. 122620873954733E-01
0.546918326183967E-03 0.720703586534390E-03 0.532729943956627E-01 0. 149929436836712E-01
0.170185751910164E-02 0. 167446096505499E-02 0.697255301257928E-01 0.179457635281876E-01
0.408386360971437E-02 0.319128240641 147E-02 0.892241483693333E-01 0.210772023407832E-01
0.8300041 17688234E-02 0.535378831352934E-02 0.11 1922491943615E+00 0.243360867646222E-01
0. 150229781560800E-0 1 0.820962136808196E-02 0. 137919556510407E+00 0.276644913094057E-01
0.249539236157546E-01 0.1 17680292130848E-0l 0.167253553899293E+00 0.309989973545750E-01
0.387833861710629E-01 0. 159981435048914E-01 0.199897113629064E+00 0.342721053423017E-01
0.571508984811764E-01 0.208290410936243E-0 1 0.235753959625130E+00 0.374137672104434E-01
0.806057414726552E-01 0.261515976613093E-01 0.274657164616264E+00 0.403530031221956E-01
0. 109570394234518E+00 0.318220682694564E-01 0.316369047605705E+00 0.430195644057483E-01
0. 144308373001501E+00 0.376672559998067E-01 0.360582741027397E+00 0.453456033756603E-01
0. 184897949427532E+00 0.434910622632234E-01 0.406925414636825E+00 0.472673103983455E-01
0.231213001548255E+00 0.490821532284030E-01 0.454963103785795E+00 0.487264791953468E-01
0.28291 1960197208E+00 0.542224286612626E-01 0.504207051427098E+00 0.496719629360194E-01
0.339435481190853E+00 0.586959442909769E-01 0.554121436898801 E+00 0.50060986 1199254E-01
0.400013113109450E+00 0.622979181149365E-01 0.604132331 112845E+00 0.498602805316812E-01
0.463678856939306E+00 0.648434457072330E-01 0.653637688015750E+00 0.490470175905769E-01
0.529295142741036E+00 0.661755598512544E-01 0.702018156811904E+00 0.476095141198307E-01
0.595584395325645E+00 0.661722952772005E-01 0.748648479051785E+00 0.455476938170290E-01
0.661 167040396915E+00 0.647524589229209E-01 0.792909219783803E+00 0.428732923992053E-01
0.724604528176033E+00 0.618798583499546E-01 0.834198572923854E+00 0.396098004030230E-01
0.784445734734942E+00 0.575658036634420E-01 0.871943978073835E+00 0.357921438452470E-01
0.839274951478663E+00 0.518697691924657E-01 0.905613289425416E+00 0.314661094009314E-01
0.887759597617343E+00 0.448981784955672E-01 0.934725247515507E+00 0.266875279291532E-01
0.928695795963228E+00 0.368013625003695E-01 0.958859023272801E+00 0.21521241 1445673E-01
0.961050059187410E+00 0.277688703774136E-01 0.977662642064842E+00 0. 160399123812805E-01
0.983995703521289E+00 0. 180238737841607E-01 0.990860246620079E+00 0.103230260805101 E-01
0.996945958679763E+00 0.782767019549676E-02 0.998259972471242E+00 0.446223271379884E-02

This content downloaded from

187.19.243.229 on Sun, 24 Mar 2024 12:26:12 +00:00
All use subject to https://about.jstor.org/terms
 GENERALIZED GAUSSIAN QUADRATURE RULES 991

TABLE 2
Gaussian quadrature for products of polynomials andfractional powers.
N

]Y k (x)dx wi pk (xi) for k = 1,2,... 2N

N-I N-1
where {pi} = {1, s(x), x, xs(x), .x x s(x)} with s(x) = x 3

EN Nodes xi ] Weights wi
5 0.831174531456776E-02 0.279778782123048E-01
0.863966362795308E-01 0.142398935990482E+00
0.305943516943443E+00 0.291943668689807E+00
0.635656558720652E+00 0.339447240627363E+00
0.920004207024857E+00 0.198232276480043E+00

10 0.751897878625465E-03 0.259000938099737E-02
0.87466609426837 lE-02 0.157085175605557E-01
0.372684363855664E-01 0.438972012175236E-01
0.100986759928202E+00 0.851979518885884E-01
0.209274931334610E+00 0.131185462315937E+00
0.360730157291068E+00 0.16927942889149 1E+00
0.540868549812902E+00 0.18661224712397 1E+00
0.723966527445448E+00 0.174225330076043E+00
0.878692655294032E+00 0.130242598410538E+00
0.975963746745065E+00 0.610612531343552E-01

15 0.168121349224655E-03 0.582014429388243E-03
0.200427399422478E-02 0.366362217710688E-02
0.888445820257933E-02 0.108842247988913E-01
0.254380748331252E-01 0.230661756216959E-01
0.56616336647950 IE-01 0.400108265603416E-01
0. 106639742234206E+00 0.604643908138577E-01
0.177998880856086E+00 0.822581407444590E-01
0.270675783247184E+00 0.102603879506980E+00
0.381720959785030E+00 0.118500791181315E+00
0.505267785653412E+00 0.127191992481750E+00
0.632999186646096E+00 0.126600784497532E+00
0.755011893747991E+00 0.1 15679059680403E+00
0.860961671393287E+00 0.946133646757434E-01
0.941329076190074E+00 0.648572778769515E-01
0.98863660805779 IE+00 0.290234549535849E-01

20 0.565910606963981 E-04 0.196275471368437E-03

0.680854395380959E-03 0.125263851254803E-02
0.306270110956997E-02 0.380561564887333E-02
0.894917200513315E-02 0.832217213031223E-02
0.204442383726108E-01 0.150414196754775E-01
0.397611592475325E-01 0.239388575134562E-01
0.689512869857525E-01 0.347155654274889E-01
0.109635234283660E+00 0.468137478429227E-01
0.162762389006638E+00 0.594574979240510E-01
0.228422459407150E+00 0.717153419416687E-01
0.305728258549570E+00 0.825791227715473E-01
0.39278268978053 IE+00 0.910522631915573E-01
0.486735430814252E+00 0.962395518442346E-01
0.583926798874029E+00 0.974303994336976E-01
0.680108431067843E+00 0.941680397096348E-01
0.770723438587922E+00 0.862983560861 IOOE-01
0.851223225892569E+00 0.739938065555806E-01
0.917394745239490E+00 0.577502214287376E-01
0.965671257181204E+00 0.383577114718547E-01
0.993407695194012E+00 0.168713954188783E-01

This content downloaded from

187.19.243.229 on Sun, 24 Mar 2024 12:26:12 +00:00
All use subject to https://about.jstor.org/terms
992 J. MA, V. ROKHLIN, AND S. WANIDZURA

TABLE 3
Gaussian quadrature for products of polynomials andfractional pow
I NN
(Pk(x)dx = W Owik(xi) for k = 1, 2,..., 2N
O ~~~i=l

where {(pi} = {1, s(x), x, xs(x), ..., xN, xN

[N Nodes xi Weights wi
5 0.325694885051994E-02 0.143436825038591E-01
0.600861836255002E-01 0.11763546204798 1E+00
0.262200689766926E+00 0.286500581347155E+00
0.601527262010120E+00 0.361906424437754E+00
0.910998040648555E+00 0.219613849663250E+00

10 0.261010670878138E-03 0.1 16634009622830E-02

0.533570406879994E-02 0.1 12056554482295E-01
0.279923598777842E-01 0.369538395786576E-01
0.845001067179312E-01 0.782012224225048E-01
0.186967540598809E+00 0.126971878207205E+00
0.336449186747933E+00 0.169715037099809E+00
0.519422386872603E+00 0.191665171997203E+00
0.709119391752190E+00 0.181908136234217E+00
0.871552940481438E+00 0.137424134774237E+00
0.974480099088773E+00 0.647885841417101 E-01

15 0.559212686135403E-04 0.250636999997100E-03
0.1 16764741757783E-02 0.248629682453753E-02
0.635876913141328E-02 0.868418049215750E-02
0.202601345786029E-01 0.200174031540757E-01
0.481469972090806E-01 0.365707842279019E-01
0.947893055727041E-01 0.572506624316614E-01
0.163312504831184E+00 0.798969454587593E-01
0.254266678317515E+00 0.101582504459800E+00
0.365070714120304E+00 0.1 19053915578673E+00
0.489939651507906E+00 0.129244090759107E+00
0.620326692664507E+00 0.129774994114533E+00
0.745828824368843E+00 0.1 19369417740098E+00
0.855429899278146E+00 0.981047682668204E-01
0.938900129707038E+00 0.674683579975495E-01
0.988155905432853E+00 0.302450414943277E-01

20 0.184194907598179E-04 0.826457102725699E-04
0.387587458269692E-03 0.829421040741860E-03
0.213955259577947E-02 0.295732155968210E-02
0.695120073310895E-02 0.702434283841401E-02
0.169467664404234E-01 0.133575962599994E-01
0.344417452000176E-01 0.220068985356573E-01
0.616535209242469E-01 0.327254003368527E-01
0.100407215362434E+00 0.449795404557805E-01
0.151866019233245E+00 0.579877214786053E-01
0.216313281828314E+00 0.707843838119479E-01
0.293009074582624E+00 0.823037488978054E-01
0.380137222473773E+00 0.914756213036679E-01
0.474850532149502E+00 0.973244581548574E-01
0.573412898781505E+00 0.990625442531322E-01
0.671427980742799E+00 0.961685890414821E-01
0.764136026657879E+00 0.884443478050775E-01
0.846753983774620E+00 0.760438602249420E-01
0.914829857705509E+00 0.594724986507429E-01
0.964581196697193E+00 0.395568714306716E-01
0.993195777887590E+00 0.174121882096675E-01

This content downloaded from

187.19.243.229 on Sun, 24 Mar 2024 12:26:12 +00:00
All use subject to https://about.jstor.org/terms
 GENERALIZED GAUSSIAN QUADRATURE RULES 993

TABLE 4
Gaussian quadrature for products of polynomials andfractional powers.
I N

]Yk (x)dx = wi (k (xi for k=1,2.... ,2N

Jo~ ~~~~~~- N-1

where {i} = {1, s(x), x, xs(x), .. N1, XN s(x)l with s(x) =

N Nodes xi Weights w,
5 0.127054140407448E-02 0.798095300782248E-02
0.462681413226638E-01 0.102702832369157E+00
0.237083298080667E+00 0.282009089112635E+00
0.581208702007414E+00 0.374825379887486E+00
0.905561736397846E+00 0.232481745622901E+00

10 0.945849186124229E-04 0.599207947722056E-03
0.379603281250484E-02 0.895661938070323E-02
0.233952614888275E-01 0.332290268550819E-01
0.759462672091217E-01 0.742813877203913E-01
0.175082164491838E+00 0.124481407759041E+00
0.32329556501452 1E+00 0.169777375189087E+00
0.507679537823808E+00 0.194331900792034E+00
0.700934090961697E+00 0.186099994751647E+00
0.867600470065010E+00 0.141387747529539E+00
0.973657038209960E+00 0.668553320747537E-01
15 0.197855962731515E-04 0.125556576654582E-03
0.809581023980278E-03 0.193257660045135E-02
0.517321860359321 E-02 0.757902277249522E-02
0.177160194789544E-01 0.184325783933937E-01
0.438708712542835E-01 0.347412411366044E-01
0.886977166657579E-01 0.555083582856597E-01
0.155667671624679E+00 0.785867325255755E-01
0.245646859906903E+00 0. 100978745423403E+00
0.356264972533447E+00 0.1 19295279382622E+00
0.481792161982642E+00 0.130299495385019E+00
0.613565444840732E+00 0.131446824433812E+00
0.740915815910203E+00 0.121332382739473E+00
0.852464602157936E+00 0.999716176822775E-01
0.937596443363357E+00 0.688685445138014E-01
0.987897722986291E+00 0.309010441487584E-01

20 0.644080669471020E-05 0.408973476604909E-04
0.265385497209493E-03 0.636150271129477E-03
0.171795135681289E-02 0.254437715881698E-02
0.599650565966228E-02 0.637191713688041E-02
0.152297986688364E-01 0.124939147170132E-01
0.317830016969808E-01 0.210015095348042E-01
0.579590115876318E-01 0.316775070055125E-01
0.956908382982334E-01 0.440030511092655E-01
0.146256150019543E+00 0.571949651553830E-01
0.210043256963358E+00 0.702705143585106E-01
0.286393214270049E+00 0.821337066503373E-01
0.373535741616497E+00 0.916751595069523E-01
0.468627945387644E+00 0.978762191044640E-01
0.567895206083029E+00 0.999075512797006E-01
0.666863963449413E+00 0.972128770925840E-01
0.760667402163021E+00 0.895698602397595E-01
0.844398042013510E+00 0.771222579836698E-01
0.913476675531666E+00 0.603802128097188E-01
0.964005785910473E+00 0.401896290968344E-01
0.993083879404685E+00 0.176977224410030E-01

This content downloaded from

187.19.243.229 on Sun, 24 Mar 2024 12:26:12 +00:00
All use subject to https://about.jstor.org/terms
994 J. MA, V. ROKHLIN, AND S. WANDZURA

TABLE 5
Integration by the Gaussian quadrature for the products of polynomia
p1 ~~N

j f(x)dx wi f (xi)

f(x) = sin(15x)

f (x)dx = 0.1 17312527523921 E+00

N Computed integral Absolute error Relative error

5 0.941018184543160E-01 0.2321 1 E-0 l 0.19785E+00
10 0.1 17457260225742E+00 0.14473E-03 0.12337E-02
15 0.1 17312524081768E+00 0.34422E-08 0.29342E-07
20 0.1 17312527523928E+00 0.70222E-14 0.59859E-13

f (x) = X 25

fo f (x)dx = 0.384615384615385E-01
N Computed integral Absolute error Relative error
5 0.230835034135673E-01 0.15378E-01 0.39983E+00
10 0.384161409736514E-01 0.45397E-04 0.1 1803E-02
15 0.384615375071555E-01 0.95438E-09 0.24814E-07
20 0.384615384615384E-01 0.55511E-16 0.14433E-14

f (x) = -x25 in x

fo f (x)dx = 0.147928994082840E-02
N Computed integral Absolute error Relative error
5 0.203228387592046E-02 0.55299E-03 0.37382E+00
10 0. 149328817343850E-02 0.13998E-04 0.94628E-02
15 0.147929073676245E-02 0.79593E-09 0.53805E-06
20 0.147928994082847E-02 0.70256E-16 0.47493E-13

f (x) = sin(55x)

f f (x)dx = 0.177795135225099E-0 1
N Computed integral Absolute error Relative error
25 0.174229968967879E-01 _ 0.35652E-03 0.20052E-01
30 0.17779799853459 1 E-0 1 0.28633E-06 0.16105E-04
35 0.177795134768564E-01 0.45653E-10 0.25678E-08
40 0. 177795135225109E-01 0.10374E- 14 0.58346E-13

f (x) = X 85

fo f (x)dx = 0. 116279069767442E-01
N Computed integral Absolute error Relative error
25 0. 116278805943580E-01 0.26382E-07 0.22689E-05
30 0. 116279069416079E-01 0.35136E-10 0.30217E-08
35 0.1 162790697673 18E-0 1 0.12405E- 1 3 0.10668E-1 1
40 -0.1 16279069767444E-01 0.20470E-15 0.17604E-13

f(x) = -x85 lnx

fo f (x)dx = 0. 135208220659816E-03
N Computed integral Absolute error Relative efror
25 0.135212681693139E-03 0.4461OE-08 0.32994E-04
30 0.135208229595502E-03 0.89357E- 11 0.66088E-07
35 0. 135208220664142E-03 0.43263E-14 0.31998E-10
40 - 0. 0135208220659815E-03 0.84026E-18 0.62145E1

This content downloaded from

187.19.243.229 on Sun, 24 Mar 2024 12:26:12 +00:00
All use subject to https://about.jstor.org/terms
 GENERALIZED GAUSSIAN QUADRATURE RULES 995

TABLE 6
Integration by the Gaussian quadrature for the products of polynomials and fractional power s(x) = x 3 in
Table 2.
fl N
f (x)dx = wi f (xi)
Oi=1

f(x) = sin(lOx)

fo f (x)dx = 0.183907152907645E+00
N Computed integral Absolute error Relative error
C 0.203627012649713E+00 0. 19720E-0I 0.10723E+00
10 0. 183906319056362E+00 0.83385E-06 0.45341 E-05
15 0. 183907152908719E+00 0. 10739E- 1 0.58396E- 11
20 0. 183907152907644E+00 0.86042E-15 0.46786E- 14

f(x) =x25

f0 f (x)dx = 0.384615384615385E-01
N Computed integral Absolute error Relative error
5 0.246599205561149E-01 0.13802E-01 0.35884E+00
10 0.384268064253340E-01 0.34732E-04 0.90303E-03
15 0.384615378508424E-01 0.61070E-09 0.15878E-07
20 0.384615384615384E-01 0. 11102E-15 0.28866E-14

f (x) x25s(x)

fo f (x)dx = 0.379746835443038E-01
N Computed integral Absolute error Relative error
5 0.239835389273907E-01 0.13991E-01 0.36843E+00
10 0.379360956044503E-01 0.38588E-04 0.101 61 E-02
15 0.379746827352660E-01 0.80904E-09 0.21305E-07
20 0.379746835443037E-01 0. 12490E- 15 0.32890E-14

TABLE 7
Integration by the Gaussian quadrature for the products of polynomials andfractional power s(x) = x 3 in
Table 3.
rl N

f (x)dx wi f (xi)
Oi=l

f(x) = sin(lOx)

fo f (x)dx = 0.183907152907645E+00
N Computed integral Absolute error Relative error
5 0. 181420301850517E+00 0.24869E-02 0.13522E-0 I
I10 0. 183906605077266E+00 0.54783E-06 0.29788E-05
15 0.183907152908166E+00 0.52114E-12 0.28337E- 1 I
20 0.183907152907645E+00 0.30531E-15 0.16601E-14

f(x) =x25

f,_f f (x)dx = 0.384615384615385E-01

N Computed integral Absolute error Relative error
5 0.213608283352790E-01 0.17101E-01 0.44462E+00
10 0.384022552100793E-01 0.59283E-04 0.15414E-02
15 0.384615369753817E-01 0.14862E-08 0.38640E-07
20 0.384615384615384E-01 - 0.69389E-16 - 0.18041E-14

f(x) = x 25s(x)

fO f (x)dx = 0.389610389610390E-01
N Computed integral Absolute error ] Relative error
5 0.220351282880372E-01 0.16926E-01 0.43443E+00
10 0.389074113404340E-01 0.53628E-04 0.13764E-02
1i5 0.389610378319899E-01 0.11290E-08 0.28979E-07
20] 0.389610389610389E-01 0.27756E-16 0.71239E-15

This content downloaded from

187.19.243.229 on Sun, 24 Mar 2024 12:26:12 +00:00
All use subject to https://about.jstor.org/terms
996 J. MA, V. ROKHLIN, AND S. WANDZURA

TABLE 8
Integration by the Gaussian quadrature for the products of polyn
Table 4.
1N

j f(x)dx = w f (xi)

f (x) = sin(lOx)

fo f (x)dx = 0183907152907645E+00
N Computed integral Absolute error Relative error
5 0.156179751994470E+00 0.27727E-01 0.15077E+00
10 0.183907466630775E+00 0.31372E-06 0.17059E-05
15 0.183907152906471 E+00 0. 11742E-11 0.63847E-11
20 0.183907152907646E+00 0.10825E-14 0.58859E-14

f(x) x25

fo f (x)dx = 0.384615384615379E-01
N Computed integral Absolute error Relative error
5 0. 194693376078146E-01 0.18992E-01 0.49380E+00
10 0.383840696841997E-01 0.77469E-04 0.20142E-02
15 0.384615361512299E-01 0.23103E-08 0.60068E-07
20 0.384615384615379E-01 0.52042E-15 0.13531E-13

f (x) =x 25s(x)

fo f (x)dx = 0.394736842105263E-01
N Computed integral Absolute error Relative error
5 0.208006141309651E-01 0.18673E-01 0.47305E+00
10 0.394099645206005E-01 0.63720E-04 0.16142E-02
15 0.394736828687551E-01 0.13418E-08 0.33992E-07
20 0.394736842105259E-01 0.40939E-15 0.10371E-13

REFERENCES

[1] B. ALPERT, Rapidly-Convergent Quadrature for Integral Operators with Singular Kernels, Lawre
Laboratory, University of California at Berkeley, December 1990.
[2] P. DAvis AND P. RABINowITz, Methods of Numerical Integration, Academic Press, Inc., Orlando, FL, 1984.
[3] J. E. DENNIS and R. B. SCHNABEL, Numerical Methods for Unconstrained Optimization and Nonlinear Equa-
tions, Society for Industrial and Applied Mathemathics, Philadelphia, PA, 1996.
[4] F. A. FICKEN, The continuation method for functional equations, Comm. Pure Appl. Math., 4 (1951),
pp. 435-456.
[5] D. G0rrLEB AND S. A. ORSZAG, Numerical Analysis of Spectral Methods: Theory and Applications, Regional
Conference Series in Applied Mathematics, Society for Industrial and Applied Mathematics, Philadelphia,
PA, 1977.
[6] S. KARLIN AND W. STuDDEN, Tchebycheff Systems with Applications in Analysis and Statistics, John Wiley
(Interscience), New York, 1966.
[7] J. MA, V. ROKHLIN, AND S. WANDZURA, Generalized Gaussian Quadrature Rules For Systems of Arbitrary
Functions, Research Report YALEU/DCS/RR-990, Department of Computer Science, Yale University,
1993.
[8] V. RoKHIuN, End-point corrected trapezoidal quadrature rules for singular functions, Comput. Math. Appl.,
20 (1990), pp. 51-62.
[9] J. STOER AND R. BULIRSCH, Introduction to Numerical Analysis, Springer-Verlag, New York, 1983.
[10] A. H. STROUD AND D. SEcREsT, Gaussian Quadrature Formulas, Prentice-Hall, Inc., Englewood Cliffs, NJ,
1966.

This content downloaded from

187.19.243.229 on Sun, 24 Mar 2024 12:26:12 +00:00
All use subject to https://about.jstor.org/terms