A PPENDIX C

G AUSSIAN QUADRATURE

appendix gives a derivation of the fundamental formulas for Gaussian

T HIS
quadrature, which were discussed but not derived in Section 5.6.2.
Gaussian quadrature, defined over the standard domain from −1 to 1,
makes use of an integration rule of the form
Z 1 N

−1
f ( x ) dx ≃ ∑ w k f ( x k ). (C.1)
k =1

The derivation of the positions xk of the sample points and the weights wk is
based on the mathematics of Legendre polynomials. The Legendre polyno-
mial PN ( x ) is an Nth-order polynomial in x that has the property
Z 1
x k PN ( x ) dx = 0 for all integer k in the range 0 ≤ k < N (C.2)
−1
and satisfies the normalization condition
Z 1 2 2
PN ( x ) dx = . (C.3)
−1 2N + 1
Thus, for instance, P0 ( x ) = constant, and the constant is fixed by (C.3) to give
P0 ( x ) = 1. Similarly, P1 ( x ) is a first-order polynomial ax + b satisfying
Z 1
( ax + b) dx = 0. (C.4)
−1
Carrying out the integral, we find that b = 0 and a is fixed by (C.3) to be 1,
giving P1 ( x ) = x. The next two polynomials are P2 ( x ) = 21 (3x2 − 1) and
P3 ( x ) = 12 (5x3 − 3x ), and you can find tables on-line or elsewhere that list
them to higher order.
Now suppose that q( x ) is a polynomial of degree less than N, so that it can
be written q( x ) = ∑kN=−01 ck x k for some set of coefficients ck . Then
Z 1 N −1 Z 1
q( x ) PN ( x ) dx = ∑ ck x k PN ( x ) dx = 0, (C.5)
−1 k =0 −1

514
 A PPENDIX C | G AUSSIAN QUADRATURE

by Eq. (C.2). Thus, for any N, PN ( x ) is orthogonal to every polynomial of

lower degree. A further property of the Legendre polynomials, which we will
use shortly, is that for all N the polynomial PN ( x ) has N real roots that all lie
in the interval from −1 to 1. That is, there are N values of x in this interval for
which PN ( x ) = 0.
Returning now to our integral, Eq. (C.1), suppose that the integrand f ( x ) is
a polynomial in x of degree 2N − 1 or less. If we divide f ( x ) by the Legendre
polynomial PN ( x ), then we get

f ( x ) = q( x ) PN ( x ) + r ( x ), (C.6)

where q( x ) and r ( x ) are both polynomials of degree N − 1 or less. Thus our

integral can be written
Z 1 Z 1 Z 1 Z 1
f ( x ) dx = q( x ) PN ( x ) dx + r ( x ) dx = r ( x ) dx, (C.7)
−1 −1 −1 −1

where we have used (C.5). This means that to find the integral of the poly-
nomial f ( x ) we have only to find the integral of the polynomial r ( x ), which
always has degree N − 1 or less.
But we already know how to solve this problem. As we saw in Section 5.6.1,
for any choice of sample points xk a polynomial of degree N − 1 or less can be
fitted exactly using the interpolating polynomials φk ( x ), Eq. (5.53), and then
the fit can be integrated to give a formula of the form
Z 1 Z 1 N

−1
f ( x ) dx =
−1
r ( x ) dx = ∑ w k r ( x k ), (C.8)
k =1

where Z 1
wk = φk ( x ) dx. (C.9)
−1

(See Eq. (5.60) on page 167.) Note that, unlike Eq. (C.1), the equality in Eq. (C.8)
is now an exact one (because the fit is exact).
Thus we have a method for integrating any polynomial of order 2N − 1 or
less exactly over the interval from −1 to 1: we divide by the Legendre polyno-
mial PN ( x ) and then integrate the remainder polynomial r ( x ) using any set of
N sample points we choose plus the corresponding weights.
This, however, is not a very satisfactory method. In particular the polyno-
mial division is rather complicated to perform. However, we can simplify the
procedure by noting that, so far, the positions of our sample points are uncon-
strained and we can pick them in any way we please. So consider again an

515
A PPENDIX C | G AUSSIAN QUADRATURE

integration rule of the form (C.1) and make the substitution (C.6), to get
N N N
∑ wk f (xk ) = ∑ wk q(xk ) PN (xk ) + ∑ wk r(xk ). (C.10)
k =1 k =1 k =1

But we know that PN ( x ) has N zeros between −1 and 1, so let us choose our N
sample points xk to be exactly the positions of these zeros. That is, let xk be the
kth root of the Legendre polynomial PN ( x ). In that case, PN ( xk ) = 0 for all k
and Eq. (C.10) becomes simply
N N
∑ w k f ( x k ) = ∑ w k r ( x k ). (C.11)
k =1 k =1

Combining with Eq. (C.8), we then have

Z 1 N

−1
f ( x ) dx = ∑ w k f ( x k ), (C.12)
k =1

where the equality is an exact one.

Thus we have a integration rule of the standard form that allows us to in-
tegrate any polynomial function f ( x ) of order 2N − 1 or less from −1 to 1 and
get an exact answer (except for rounding error). It will give the exact value for
the integral, even though we only measure the function at N different points.
We have not derived the closed-form expression for the weights wk given in
Eq. (5.64). The derivation of this expression is lengthy and tedious, so we omit
it here, but the enthusiastic reader can find it in Hildebrand, F. B., Introduction
to Numerical Analysis, McGraw-Hill, New York (1956).

G AUSS –K RONROD QUADRATURE

A widely used variant of Gaussian quadrature is Gauss–Kronrod quadrature,

which was mentioned briefly, but not defined, in Section 5.6.3. Gauss–Kronrod
quadrature provides an additional set of sample points interlaced between
those of ordinary Gaussian quadrature. By computing an estimate of an in-
tegral using just the ordinary Gaussian points, and then recomputing it using
the two sets of points combined, one gets two values whose difference gives
an estimate of the error on the result. Thus Gauss–Kronrod quadrature gives
results of accuracy comparable with Gaussian quadrature plus an estimate of
the error on the result (which Gaussian quadrature alone does not provide),
but does so at the expense of some addition computational effort, since one
must evaluate the integrand at all of the additional sample points.

516
 A PPENDIX C | G AUSSIAN QUADRATURE

The derivation of the Gauss–Kronrod formula is similar to that for ordinary

Gaussian quadrature. Suppose we choose N sample points to be the roots of
the Nth Legendre polynomial PN ( x ), as in standard Gaussian quadrature, and
an additional N + 1 other points, which we are free to place anywhere we like,
for a total of 2N + 1 points. For any integrand f ( x ) we can create a polynomial
approximation of degree 2N that matches the integrand exactly at these 2N + 1
points, for instance using the method of interpolating polynomials from Sec-
tion 5.6.1, then integrate that approximation to get an approximation to the
integral of f ( x ). If f ( x ) itself happens to be a polynomial of degree 2N or less,
then the calculation will be exact, apart from rounding error.
But now we note that we have N + 1 degrees of freedom in the positions
of our additional N + 1 sample points, which as we have said we can choose
in any way we like, and this suggests that, if we choose those points correctly,
we should be able to create an integration rule that is exact for polynomials
of degree N + 1 higher, i.e., polynomials of degree 3N + 1. Gauss–Kronrod
quadrature tells us how to pick the additional N + 1 points to achieve this.
The result is an integration rule with 2N + 1 sample points that is accurate for
polynomials up to degree 3N + 1, which is not as good as Gaussian quadrature
(which would be accurate up to degree 4N + 1 on 2N + 1 points), but it’s the
best we can do if we restrict our first N points to fall at the roots of PN ( x ).
To describe this another way, the N initial points at the roots of PN ( x )
are nested within (i.e., a subset of) the 2N + 1 points for the Gauss–Kronrod
quadrature, which is the crucial property that makes Gauss–Kronrod quadra-
ture attractive. It means we can evaluate our integral using standard Gaussian
quadrature on N points, and then again using Gauss–Kronrod quadrature on
2N + 1, as described above, and the second calculation requires us to evaluate
the integrand f ( x ) only at the newly added sample points. For the rest of the
points we can reuse the values from the first step. (Notice, however, that the
weights wk for the quadrature rule are different on the two steps, so one must
recompute the sum, Eq. (C.1). One can reuse values of f ( x ) on the second step,
which can save a lot of time, but one cannot reuse the value of the complete
sum.)
How then do we choose the additional N + 1 sample points for Gauss–
Kronrod quadrature? Let us define a new polynomial EN +1 ( x ) of degree N + 1
by
Z 1
x k PN ( x ) EN +1 ( x ) dx = 0, for integer k in the range 0 ≤ k ≤ N. (C.13)
−1

This formula gives us N + 1 conditions on EN +1 ( x ). If we also fix the nor-

517
A PPENDIX C | G AUSSIAN QUADRATURE

malization of EN +1 ( x ) (using any method we like), then we have N + 2 condi-

tions, which is enough to fix all N + 2 coefficients of the polynomial and hence
uniquely specify it. The polynomials EN ( x ) are known as Stieltjes polynomials.
In Gauss–Kronrod quadrature, we choose the additional N + 1 sample points
to be the roots of the polynomial EN +1 ( x ). Thus the complete set of 2N + 1
sample points is the set of roots of PN ( x ) plus the roots of EN +1 ( x ). The cor-
responding integration weights wk can then be calculated using Eq. (5.60) on
page 167.
Now suppose our integrand f ( x ) is a polynomial of degree 3N + 1 or less.
If we divide f ( x ) by PN ( x ) EN +1 ( x )—which is a polynomial of degree 2N + 1—
we get
f ( x ) = q( x ) PN ( x ) EN +1 ( x ) + r ( x ), (C.14)
where q( x ) and r ( x ) are polynomials of degree N or less. Then the integral of
f ( x ) over the standard interval from −1 to 1 is
Z 1 Z 1 Z 1 Z 1
f ( x ) dx = q( x ) PN ( x ) EN +1 ( x ) dx + r ( x ) dx = r ( x ) dx, (C.15)
−1 −1 −1 −1

where we have used (C.13) to eliminate the first term. This integral can now
be evaluated in the standard fashion
Z 1 2N +1

−1
r ( x ) dx = ∑ w k r ( x k ), (C.16)
k =1

where, as we have said, the sample points xk are the roots of the Legendre and
Stieltjes polynomials. Since r ( x ) is a polynomial of order N or less and there
are 2N + 1 sample points, Eq. (C.16) will always give an exact answer, to the
limits set by rounding error.
But now, using Eq. (C.14), we can also write
2N +1 2N +1 2N +1
∑ wk f ( x k ) = ∑ wk q( xk ) PN ( xk ) EN +1 ( xk ) + ∑ wk r ( x k )
k =1 k =1 k =1
2N +1
= ∑ w k r ( x k ), (C.17)
k =1

where the first sum has vanished because every sample point xk falls at a
zero of either PN ( x ) or EN +1 ( x ), so every term in the sum is zero. Combin-
ing Eqs. (C.15) to (C.17), we have
Z 1 2N +1

−1
f ( x ) dx = ∑ w k f ( x k ), (C.18)
k =1

518
 A PPENDIX C | G AUSSIAN QUADRATURE

where the equality is an exact one. Thus this particular choice of sample points
does indeed give us an integration rule that is exact for all polynomial inte-
grands of order 3N + 1 or less.
The Gauss–Kronrod integration method now involves the following steps:
1. We evaluate the integral of f ( x ) first using standard Gaussian quadrature
on N points.
2. We evaluate it again using Gauss–Kronrod quadrature on 2N + 1 points.
N of those points are the same as those for Gaussian quadrature—the
roots of PN ( x )—so we do not have to recalculate f ( x ) at these points. We
can reuse the values from step 1. Only the values at the N + 1 new points
have to be calculated, and this can save us a lot of time.
3. The second estimate of the integral (since it is the more accurate of the
two) gives us our final result. And the difference between the two es-
timates gives us an estimate of the error, by analogy with Eq. (5.66). In
fact, the difference only gives us an upper bound on the error. The ac-
tual error is probably significantly smaller, but unfortunately no precise
expression for the error is known, so in practice one usually just uses the
difference, bearing in mind that the true error may in fact be smaller than
this.

519