You are on page 1of 10

Mathematics and Computers in Simulation 54 (2000) 403412

GaussRadau formulae for Jacobi and Laguerre weight functions


Walter Gautschi
Department of Computer Sciences, Purdue University, West Lafayette, IN 47907-1398, USA
Dedicated to John R. Rice, on the occasion of his 65th birthday

Abstract
Explicit expressions are obtained for the weights of the GaussRadau quadrature formula for integration over
the interval [1, 1] relative to the Jacobi weight function (1t) (1+t) , >1, >1. The nodes are known to
be the eigenvalues of a symmetric tridiagonal matrix, which is also obtained explicitly. Similar results hold for
GaussRadau quadrature over the interval [0, ) relative to the Laguerre weight t et , >1. 2000 IMACS.
Published by Elsevier Science B.V. All rights reserved.

Keywords: GaussRadau formula; Jacobi weight function; Laguerre weight function

1. Introduction

For any positive measure d supported on the interval [a, b], with a a finite real number, and d having
moments of all orders, there exists a quadrature rule of the form
Z b n
X
f (t) d(t) = 0 f (a) + k f (tk ) + Rn (f ) (1.1)
a k=1

which is exact for polynomials of degree 2n,

Rn (f ) = 0, f P2n . (1.2)

It is called the ((n+1)-point) GaussRadau rule for the measure d. Its interior nodes tk are known
to be the zeros of n (; da ), the polynomial of degree n orthogonal with respect to the modified
measure da (t)=(ta) d(t). The weights are obtainable by interpolation at the nodes a, t1 ,
t2 , . . . , tn .
Golub [5] in 1973 observed that the formula can be obtained, more elegantly, via eigenvalues and
eigenvectors of a modified Jacobi matrix of order n+1,

0378-4754/00/$20.00 2000 IMACS. Published by Elsevier Science B.V. All rights reserved.
PII: S 0 3 7 8 - 4 7 5 4 ( 0 0 ) 0 0 1 7 9 - 8
404 W. Gautschi / Mathematics and Computers in Simulation 54 (2000) 403412

0 1 0
1 1

2


.. ..
Jn+1 (d) = 2 . . . (1.3)

..
. n1 n

0 n n
Here, k and k are the coefficients in the recurrence relation
k+1 (t) = (t k )k (t) k k1 (t), k = 0, 1, 2, . . . , 1 (t) = 0, 0 (t) = 1 (1.4)
satisfied by the (monic) orthogonal polynomials k ()= k (; d), and n is given by
n1 (a)
n = a n . (1.5)
n (a)

The nodes of Eq. (1.1) (including a) are then precisely the eigenvalues of Jn+1 (d), whereas the weights
j are expressible in terms of the first components vj,1 of the associated normalized eigenvectors vj ,
j = 0 vj,1
2
, j = 0, 1, 2, . . . , n, (1.6)
Rb
where 0 = a d(t) (cf. also [2,3]).
We show here that in the case of the Jacobi measure on [1, 1],
d(,) (t) = (1 t) (1 + t) dt, > 1, > 1, (1.7)
(,)
the quantity n in Eq. (1.5) as well as all the weights j = j in Eq. (1.1) can be computed explicitly in
terms of n, , and , thus obviating the need of computing eigenvectors. Our results generalize well-known
formulae for the Legendre measure d(0,0) (cf., e.g. [1], p. 103),
2 1 1 1 tk
(0,0) = , (0,0)
k = 0
= , k = 1, 2, . . . , n, (1.8)
0
(n + 1)2 (1 tk )[Pn (tk )]2 (n + 1) [Pn (tk )]2
2

where Pn is the Legendre polynomial of degree n. We obtain, in fact, an additional formula, which in the
case of the Legendre measure becomes
 2
(0,0) 2n + 3 1 tk
k = (0,1)
, (1.9)
(n + 1)(n + 2) [Pn+1 (tk )]2
(0,1)
with Pn+1 the Jacobi polynomial of degree n+1 relative to the Jacobi parameters =0 and =1. Similar
techniques can be used to derive explicit GaussLobatto formulae for general Jacobi weight functions
(cf. [4]).
Analogous results hold for the Laguerre measure on [0, ),
d() (t) = t et dt, > 1. (1.10)
Sections 24 are dealing with the Jacobi measure (1.7). After the explicit formula for the modified element
(,)
n in Eq. (1.3) is obtained in Section 2, the boundary weight 0 will be developed in Section 3, and
(,)
the interior weights k in Section 4. Section 5 deals analogously with the Laguerre measure (1.10).
W. Gautschi / Mathematics and Computers in Simulation 54 (2000) 403412 405

2. The modified Jacobi matrix


The monic polynomials orthogonal with respect to the Jacobi measure (1.7) will be denoted by
(,) (,)
k = k , and the Jacobi polynomials, as conventionally defined (cf., e.g., [6]), by Pk = Pk .
They are related by
 
1 2n + +
Pn (t) = kn n (t), kn = n . (2.1)
2 n
It is well known, furthermore, that
 
n n+
Pn (1) = (1) . (2.2)
n
By Eq. (1.5), we have
n1 (1) kn Pn1 (1)
n = 1 n = 1 n ,
n (1) kn1 Pn (1)
which, in view of Eqs. (2.1) and (2.2), gives
1 (2n + + )(2n + + 1)
n = 1 + n , n 1. (2.3)
2 (n + )(n + + )
On the other hand,
kn k2 kn1
2
hn
n = = ,
kn1 k 2 kn hn1
2

where
Z
2 1
2++1 (n + + 1) (n + + 1)
hn = Pn(,) = [Pn(,) (t)]2 d(,) (t) = .
1 2n + + + 1 (n + 1) (n + + + 1)
On substitution in Eq. (2.3), this yields
2n(n + )
n = 1 + , n 1. (2.4)
(2n + + )(2n + + + 1)

The modified Jacobi matrix Jn+1 (d(,) ) in Eq. (1.3) is now readily computable, the recursion coefficients
k and k for the Jacobi measure d(,) being explicitly known. This yields the GaussRadau formula in
terms of the eigenvalues and (first components of) the eigenvectors of the matrix (1.3). In the next two sec-
(,)
tions we develop explicit formulae for the weights j = j , which allow us to bypass the formula (1.6).

3. The boundary weight

Our concern, in this and the next two sections, is with the GaussJacobiRadau formula
Z 1 n
X
f (t) d(,) (t) = 0 f (1) + k f (tk ) + Rn (f ), (3.1)
1 k=1
(,)
where d(,) is the Jacobi measure (1.7). We first deal with the boundary weight 0 = 0 .
406 W. Gautschi / Mathematics and Computers in Simulation 54 (2000) 403412

For the Jacobi measure we have d1 (t)=(1+t) d(,) (t)=d(,+1) (t), so that the interior nodes
(,) (,+1)
tk = tk of Eq. (3.1) are the zeros of Pn ,

Pn(,+1) (tk ) = 0, k = 1, 2, . . . , n. (3.2)


(,+1)
Putting f(t)=Pn (t) in Eq. (3.1) then yields
Z 1
(,+1)
0 Pn (1) = Pn(,+1) (t) d(,) (t). (3.3)
1

After replacing by +1 in Eq. (2.2), the coefficient on the left is given by


 
(,+1) n n+ +1
Pn (1) = (1) . (3.4)
n
In order to compute the integral in Eq. (3.3), we begin writing, using the second relation in [6] (Eq. (4.5.4)),
(,) (,)
2 (n + 1)Pn+1 (t) + (n + + 1)Pn (t)
Pn(,+1) (t) = . (3.5)
2n + + + 2 1+t
Observing from (2.2) that
(,)
(1)n Pn (1)
n + 1 = (n + 1)   ,
n+
n
and similarly
(n + 1)
n+ +1=   (1)n+1 Pn(,) (1),
n+
n
we can write Eq. (3.5) as
(,) (,) (,) (,)
2(1)n (n + 1) Pn+1 (t)Pn (1) Pn (t)Pn+1 (1)
Pn(,+1) (t) =   . (3.6)
n+ 1+t
(2n + + + 2)
n
On the other hand, by the ChristoffelDarboux formula ([6], Eq. (4.5.2)),
(,) (,) (,) (,) X n
Pn+1 (t)Pn (1) Pn (t)Pn+1 (1) 1
= cn (, ) P (,) (t)P(,) (1),
(,)
(3.7)
1+t =0 h

where
(2n + + + 2) (n + + 1) (n + + 1)
cn (, ) = 2+ (3.8)
(n + 2) (n + + + 2)
W. Gautschi / Mathematics and Computers in Simulation 54 (2000) 403412 407
R1
and h(,)
= 1 [P(,) (t)]2 d(,) (t). Now integrating Eq. (3.6) with the measure d(,) gives, by
Eq. (3.7) and the orthogonality of the Jacobi polynomials,
Z 1
2(1)n (n + 1)cn (, )
Pn(,+1) (t) d(,) (t) =  ,
1 n+
(2n + + + 2)
n

or by an elementary computation, using Eq. (3.8),


Z 1
(n + + 1)
Pn(,+1) (t) d(,) (t) = (1)n 2++1 ( + 1) . (3.9)
1 (n + + + 2)
Combining Eqs. (3.3), (3.4), and (3.9) finally gives the desired result

(,) 2++1 ( + 1) (n + + 1)
0 =   . (3.10)
n+ +1 (n + + + 2)
n

In the case ==0, we recover the first equation of (1.8).


In almost all cases computed, the boundary weight (3.10) turned out to be more accurate than the
boundary weight computed by Eq. (1.6) (for j=0), often significantly so.

4. The interior weights


(,+1)
We now put f (t) = (1 + t)Pn (t)/(t tk ) in Eq. (3.1) and obtain, by virtue of Eq. (3.2),
Z (,+1)
1
Pn (t)
(1 + tk )k Pn(,+1)0 (tk ) = d(,+1) (t).
1 t tk
Applying once more the ChristoffelDarboux formula ([6], Eq. (4.5.2) with n replaced by n1, and
replaced by +1), we get similarly as in Section 3, using Eq. (3.2),
Z (,+1)
1
Pn (t) dn (, )
d(,+1) (t) = (,+1)
,
1 t tk Pn1 (tk )

where
2++1 (2n + + + 1) (n + ) (n + + 1)
dn (, ) = . (4.1)
(n + 1) (n + + + 2)
Thus,

(,) dn (, )
k = (,+1)0 (,+1)
. (4.2)
(1 + tk )Pn (tk )Pn1 (tk )
408 W. Gautschi / Mathematics and Computers in Simulation 54 (2000) 403412
(,+1)
4.1. The interior weights in terms of Pn+1 (tk )

Here, we use the second relation in ([6], Eq. (4.5.7) with replaced by +1) in combination with
Eq. (3.2) to obtain
(,+1)
1 tk2 (,+1)0 2(n + 1)(n + + + 2) Pn+1 (tk )
(1 + tk )Pn(,+1)0 (tk ) = Pn (tk ) = . (4.3)
1 tk 2n + + + 3 1 tk
(,+1)
The recurrence relation for the Jacobi polynomials Pk (cf. ([6], Eq. (4.5.1) with n replaced by n+1)),
on the other hand, yields, again using Eq. (3.2),
(,+1)
2(n + 1)(n + + + 2)(2n + + + 1)Pn+1 (tk )
(,+1)
= 2(n + )(n + + 1)(2n + + + 3)Pn1 (tk ),
that is
(,+1) (n + 1)(n + + + 2)(2n + + + 1) (,+1)
Pn1 (tk ) = Pn+1 (tk ). (4.4)
(n + )(n + + 1)(2n + + + 3)
It suffices now to insert Eqs. (4.1), (4.3), and (4.4) into Eq. (4.2) to obtain, after some computation,

(,) (2n + + + 3)2 (n++1) (n+ +2) 1 tk


k = 2+ (,+1)
, n 1, (4.5)
(n+1)(n++ +2) (n+2) (n + + + 3) [Pn+1 (tk )]2
(,) (,+1)
where tk = tk are the zeros of Pn . This is Eq. (1.9) when ==0.

(,)0
4.2. The interior weights in terms of Pn (tk )

We rewrite Eq. (4.2) by first noting from the first relation in ([6], Eq. (4.5.4) with n replaced by n1,
and replaced by +1), and also recalling Eq. (3.2), that
(,+1)
(+1,+1) 2(n + ) Pn1 (tk )
Pn1 (tk ) = .
2n + + + 1 1 tk
Furthermore, by ([6], Eq. (4.21.7)),
(+1,+1) 2
Pn1 (tk ) = P (,)0 (tk ).
n++ +1 n
Therefore,
(,+1) 2n + + + 1
Pn1 (tk ) = (1 tk ) P (,)0 (tk ). (4.6)
(n + )(n + + + 1) n
On the other hand, using Eq. (4.3) and the relation
(,+1) (n + )(n + + 1)(2n + + + 3) (,+1)
Pn+1 (tk ) = Pn1 (tk ),
(n + 1)(n + + + 2)(2n + + + 1)
W. Gautschi / Mathematics and Computers in Simulation 54 (2000) 403412 409
(,+1)
which follows from the recurrence relation for Pk (cf. ([6], Eq. (4.5.1) with n replaced by n+1))
and from Eq. (3.2), we obtain
(,+1)
2(n + )(n + + 1) Pn1 (tk )
(1 + tk )Pn(,+1)0 (tk ) = .
2n + + + 1 1 tk
This, together with Eq. (4.6), allows us to write Eq. (4.2) in the form

(,) (n + )(n + + + 1)2 dn (, ) 1


k = .
2(n + + 1)(2n + + + 1) (1 tk )[Pn(,)0 (tk )]2
Substituting the expression (4.1) for dn (, ) then gives

(,) n + + + 1 (n + + 1) (n + + 1) 1
k = 2+ . (4.7)
n + + 1 (n + 1) (n + + + 1) (1 tk )[Pn(,)0 (tk )]2
In the case of ==0, we recover the second relation in Eq. (1.8). For computational purposes, one will
(,)0 (+1,+1) (+1,+1)
probably want to replace Pn (tk ) by (1/2)(n + + + 1)Pn1 (tk ) and compute Pn1 (tk )
from its recurrence relation.

(,)
4.3. The interior weights in terms of Pn (tk )

We now rewrite Eq. (4.7) by using the second relation in ([6], Eq. (4.5.7)),
1
(1 tk2 )Pn(,)0 (tk ) = {(n + + + 1)[(2n + + + 2)tk
2n + + + 2
(,)
+ ]Pn(,) (tk ) 2(n + 1)(n + + + 1)Pn+1 (tk )},

and combining it with the relation


(,) (,)
2 (n + + 1)Pn (tk ) + (n + 1)Pn+1 (tk )
0= Pn(,+1) (tk ) = ,
2n + + + 2 1 + tk
which follows from Eq. (3.2) and the second relation in ([6], Eq. (4.5.4)), that is, with

(,) n + + 1 (,)
Pn+1 (tk ) = Pn (tk ).
n+1
The result is
(,) 2+ (n + + 1) (n + + 1) 1 tk
k = . (4.8)
n + + 1 (n + 1) (n + + + 2) [Pn(,) (tk )]2
In the Legendre case ==0, this reduces to the last relation in Eq. (1.8).
Numerically, the three formulae (4.5), (4.7), and (4.8) behave similarly. They produce less accurate
results than the formula (1.6) in about two-thirds of the cases computed, and more accurate results
otherwise.
410 W. Gautschi / Mathematics and Computers in Simulation 54 (2000) 403412

5. The GaussRadau formula for the Laguerre measure

Techniques similar to those in Sections 24, but much simpler, apply also to the Laguerre measure (1.10).
The final results (Eqs. (5.5) and (5.9)) are in fact known (cf. [1], pp. 223224)), but for completeness we
re-derive them from scratch.
The monic and conventional Laguerre polynomials, n() and L() n , are related by

(1)n ()
L()
n (t) = (t),
n! n
and we have
 
n+
L()
n (0) = , n = n(n + ). (5.1)
n
From Eq. (1.5), the modified element n in the matrix (1.3) is now
()
n1 (0) L()
n1 (0)
n = n = n ,
n() (0) nL()
n (0)

which, by Eq. (5.1), becomes, surprisingly simply,


n = n. (5.2)
The GaussLaguerreRadau formula to be considered is
Z n
X
f (t) d() (t) = 0 f (0) + k f (tk ) + Rn (f ), (5.3)
0 k=1

where d is the Laguerre measure (1.10), and, as before, Rn (P2n )=0. Since d0 (t)=t d() (t)=d(+1) (t),
()

the interior nodes tk = tk() are the zeros of L(+1)


n ,
L(+1)
n (tk ) = 0. (5.4)
The boundary weight is obtained by putting f(t)=L(+1)
n (t) in Eq. (5.3),
Z
0 L(+1)
n (0) = L(+1)
n (t) d() (t).
0

The coefficient on the left can be computed by Eq. (5.1), and the integral on the right by noting from ([6],
Eq. (5.1.13)) that
n
X
L(+1)
n (t) = L()
(t),
=0

and hence, by orthogonality,


Z n Z
X Z
(+1) ()
Ln (t) d (t) = L() ()
(t) d (t) = d() (t) = ( + 1).
0 =0 0 0
W. Gautschi / Mathematics and Computers in Simulation 54 (2000) 403412 411

There results
( + 1)
()
0 =  . (5.5)
n++1
n

For the interior weights k = ()k , we put f (t) = t Ln


(+1)
(t)/(t tk ) in Eq. (5.3) to obtain
Z (+1) Z (+1)
t Ln (t) () Ln (t) (+1)
k tk L(+1)0
n (tk ) = d (t) = d (t). (5.6)
0 t tk 0 t tk
The ChristoffelDarboux formula ([6], Eq. (5.1.11) with replaced by +1), similarly as in the beginning
of Section 4, yields
 
() ( + 2) n + + 1 1
k = (+1)0
. (5.7)
n+1 n tk Ln (tk )L()
n+1 (tk )

By ([6], Eq. (5.1.14) with replaced by +1), and Eq. (5.4),

tk L(+1)0
n (tk ) = (n + + 1)L(+1)
n1 (tk ),

where, by ([6], Eq. (5.1.13)) and Eq. (5.4),

L(+1) ()
n1 (tk ) = Ln (tk ). (5.8)

The recurrence relation for the Laguerre polynomials L(+1)


k , (cf. [6], Eq. (5.1.10)), in combination with
Eqs. (5.4) and (5.8), on the other hand, gives

L(+1) (+1) ()
n+1 (tk ) = (n + + 1)Ln1 (tk ) = (n + + 1)Ln (tk ),

so that finally
 
( + 1) n+ 1
()
k = . (5.9)
n++1 n [L()
n (tk )]
2

This is the analogue of Eq. (4.8). The analogue of Eq. (4.7) coincides with Eq. (5.9). This is because of
the identities ([6], Eqs. (5.1.13) and (5.1.14))
(+1)
L()0 () (+1)
n (t) = Ln1 (t) = Ln (t) Ln (t),

which, for t=tk , in view of Eq. (5.4), gives

L()0 ()
n (tk ) = Ln (tk ).

The numerical experience with these explicit formulae is much like in the case of the Jacobi measure, i.e.,
the boundary weight (5.5) is almost always considerably more accurate than the boundary weight obtained
via eigenvectors, whereas for interior weights, the formula (1.6) involving eigenvectors is generally more
accurate than the explicit formula (5.9).
412 W. Gautschi / Mathematics and Computers in Simulation 54 (2000) 403412

References

[1] P.J. Davis, P. Rabinowitz, Methods of Numerical Integration, 2nd Edition, Academic Press, Orlando, 1984.
[2] W. Gautschi, Algorithm 726: ORTHPOL a package of routines for generating orthogonal polynomials and Gauss-type
quadrature rules, ACM Trans. Math. Software 20 (1994) 2162.
[3] W. Gautschi, Orthogonal polynomials: applications and computation, in: A. Iserles (Ed.), Acta Numerica 1996, Cambridge
University Press, Cambridge, 1996, pp. 45119.
[4] W. Gautschi, High-order GaussLobatto Formulae, Electr. Trans. Numer. Analysis, accepted for publication.
[5] G.H. Golub, Some modified matrix eigenvalue problems, SIAM Rev. 15 (1973) 318334.
[6] G. Szeg, Orthogonal Polynomials, 4th Edition, Vol. 23, American Mathematical Society Colloquium Publications,
American Mathematical Society, Providence, RI, 1978.

You might also like