See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/241101118

Simpson's Rule Is Exact for Quintics

Article  in  The American Mathematical Monthly · February 2006

DOI: 10.2307/27641865

CITATIONS READS
8 2,547

1 author:

Lou Talman
Metropolitan State University of Denver
6 PUBLICATIONS   70 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Lou Talman on 01 August 2017.

The user has requested enhancement of the downloaded file.

 Simpson’s Rule Is Exact for Quintics
Louis A. Talman

1. INTRODUCTION. Everyone who has taught the error estimate for Simpson’s
rule in a few freshman calculus courses knows that the chances are very good that
someone who hasn’t quite thought things through will look at the error estimate and
then ask “Yes, but how can we find the error exactly?” The questioner is usually thun-
derstruck by the observation that if we knew how to do so in general, then we wouldn’t
speak of “Simpson’s rule for approximating integrals,” but of “Simpson’s rule for eval-
uating integrals.” Nevertheless, under certain circumstances we can find exact error. It
is well known, for instance, that Simpson’s rule error is zero whenever the integrand is
a polynomial of degree three or less. What is less well known is the fact that we have
adduced in our title: as we show in this article, we can easily find the exact error (see
equation (5)) for Simpson’s rule approximations to integrals of polynomial functions
of degree four or five. We accomplish some other things as well, including an exten-
sion of the theorem generally known as the “First Mean Value Theorem for Integrals.”
We begin by reviewing some background.
We all learned in our elementary calculus courses that

if K 2 is a bound for | f (2) (x)|
b
on the interval [a, b], then the error E nT in replacing a f (t) dt with its n-subdivision
trapezoidal rule approximation

b−a n
Tn = [ f (xk ) + f (xk−1 )],
2n k=1

with xk = a + k(b − a)/n for k = 1, . . . , n, satisfies the inequality

 T
 E  ≤ K 2 (b − a) .
3
n (1)
12n 2
We also learned that if K 4 is a bound for | f (4) (x)| on [a, b], then the error E 2n
S
in

b
replacing a f (t) dt with its 2n-subdivision Simpson’s rule approximation

b−a n
S2n = [ f (x2k−2 ) + 4 f (x2k−1 ) + f (x2k )] , (2)
6n k=1

where xk = a + k(b − a)/(2n) for k = 0, 1, . . . , 2n, satisfies the inequality

(b − a)5
|E 2n
S
| ≤ K4 . (3)
180(2n)4
Most of us didn’t learn then, however, why these things are so. If our calculus books
said anything at all about why these inequalities work, they probably just referred
us to a numerical analysis book. Those of us who pursued the matter found that the
numerical analysis books couched the proofs in terms of Lagrange interpolation in the
context of general Newton-Cotes quadratures (see, for example, [6] or [9]). Most who
had gotten that far threw up their hands at that point: the arguments are not accessible
to freshmen.

144 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 113
 In October 2003, D. Cruz-Uribe and C. J. Neugebauer [2] gave an elementary ar-
gument that establishes (1). They based their note in Mathematics Magazine upon a
more advanced—and more thorough—treatment that they had given in [3], and in both
papers they also discussed error estimates for the trapezoidal rule applied to functions
that might not possess bounded second derivatives. While their methods extend to the
cases of Simpson’s rule in which the integrand possesses just one or two derivatives,
Cruz-Uribe and Neugebauer were not able to apply them to obtain higher order esti-
mates for smoother functions. In particular, they couldn’t use their methods to derive
(3) for Simpson’s rule. Moreover, they explicitly asked whether it was possible to give
an elementary derivation of error estimates for application of Simpson’s rule to func-
tions possessing either four or, even better, just three derivatives. Other authors (see
[1, pp. 607–609], [5, p. 571, ex. 8], [7, p. 146, ex. 23], [8, p. 119, ex. 17], or [10,
pp. 501–2]) give elementary arguments that establish (3). But while the argument in
[1] avoids the general context of Newton-Cotes, it still relies on Lagrange interpola-
tion. Furthermore, the arguments given in [1], [5], [7], [8], and [10] break down when
we try to apply them to functions that do not possess fourth derivatives.
In this article, we show first how to exploit Taylor’s theorem with integral remainder
to represent the error in an elementary numerical quadrature as a certain integral. This
part of our argument does not depart substantively from an argument given in [2],
though the authors do not explicitly mention Taylor’s theorem. We then show how
to evaluate such an integral by using a symmetric derivative. Taylor’s theorem with
Lagrange remainder also plays a role in some of these latter calculations. Our methods,
which are accessible to freshmen, allow us to achieve two kinds of results. For starters,
we demonstrate how to establish the results that the numerical analysis books give for
the numerical quadratures of elementary calculus:

Theorem. If f is a continuously differentiable function on [a, b] for which f (2) (u)

b point u of (a, b) and if Tn is the n-subdivision trapezoidal rule approxi-
exists at each
mation to a f (t) dt, then there exists ξ in (a, b) such that
 b
(b − a)3
f (t) dt = Tn − f (2) (ξ ) .
a 12n 2

Theorem. If f is a continuously differentiable function on [a, b] for which f (2) (u)

b point u of (a, b) and if Mn is the n-subdivision midpoint rule approxi-
exists at each
mation to a f (t) dt, then there exists ξ in (a, b) such that
 b
(b − a)3
f (t) dt = Mn + f (2) (ξ ) .
a 24n 2

Theorem. If f is a thrice continuously differentiable function on [a, b] for which

f (4) (u) exists at each

b point u of (a, b) and if S2n is the 2n-subdivision Simpson’s rule
approximation to a f (t) dt, then there exists ξ in (a, b) such that
 b
(b − a)5
f (t) dt = S2n − f (4) (ξ ) . (4)
a 180(2n)4
5
From (4), we will see in section 4 that if q(x) = k=0 ak x k is a fifth degree polyno-
mial, then

February 2006] SIMPSON’S RULE IS EXACT FOR QUINTICS 145

  b
{q(a) + 4q[(a + b)/2] + q(b)}(b − a)
q(t) dt =
a 6
  
1 a+b
− 5a5 + a4 (b − a)5 . (5)
120 2

This is the sense in which Simpson’s rule is exact for quintic polynomials.
We also illustrate how one can use our technique to establish error estimates for the
integral approximation schemes of elementary calculus, including not just the trape-
zoidal rule, but the midpoint rule and Simpson’s rule as well, even when the integrands
are not as smooth as the standard results require. We do not pursue these latter esti-
mates in any depth, though we identify an underlying principle and give an example
to point the reader in what we think is a fruitful direction.

2. SOME PRELIMINARY FACTS. We will make use of some facts that, while
generally known and elementary in nature, do not ordinarily appear in the standard
calculus sequence. Because they do not ordinarily appear, we provide proofs.
First, we note that the mean value theorem takes on a special form for quadratic
functions.

Theorem 1 (Quadratic Mean Value Theorem). If q is a polynomial of degree at

most two, then

q(b) − q(a) = q  (ξ )(b − a)

for all real numbers a and b, where ξ = (a + b)/2.

Proof. Write q(x) = αx 2 + βx + γ . Then, with ξ as in the statement of the theorem,

we have


q  (ξ )(b − a) = 2α · 12 (a + b) + β (b − a) = α(b2 − a 2 ) + β(b − a)
= (αb2 + βb + γ ) − (αa 2 + βa + γ ) = q(b) − q(a).

Everybody knows that continuous functions have the intermediate value property.
It may surprise some readers to learn that derivatives—which need not be continuous
functions—also have this property (see [4]).

Theorem 2 (Darboux). Suppose that a function f is differentiable on (a, b) and that

f  (x1 ) = α < λ < β = f  (x2 ) for points x1 and x2 of (a, b). There is a number ξ
between x1 and x2 such that f  (ξ ) = λ.

Proof. In order to be definite, we assume that x1 < x2 . Consider the function g defined
on the interval (a, b) by g(x) = f (x) − λx. Then g  (x) = f  (x) − λ, and g  (x1 ) <
0 < g  (x2 ). Because of this latter pair of inequalities, neither g(x1 ) nor g(x2 ) can be
the minimum value taken on by g in the interval [x1 , x2 ]. On the other hand, as a
continuous function on [x1 , x2 ], g must assume a minimum value somewhere in the
interval, so must achieve it at some ξ in (x1 , x2 ), at which point g  (ξ ) = 0.

Theorem 2 has an important consequence that is not as well known as it should be.

146 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 113
Corollary (Positive Linear Combination Property of Derivatives). Suppose that f
is a function differentiable on (a, b) and that n is a positive integer. For k = 1, 2, . . . , n
let θk belong to (a, b), and let αk be a positive real number. Then there exists ξ in (a, b)
such that

n 
n
αk f  (θk ) = f  (ξ ) αk .
k=1 k=1

In particular, we can replace any convex combination (i.e., all αk > 0 and αk = 1)
of values that f  assumes in (a, b) with its value at some particular (albeit unspecified)
point in the interval.

Proof. For k = 1, 2, . . . , n put

αk
λk = n .
j =1 αj

Take M to be the largest of the numbers f  (θ1 ), f  (θ2 ), . . . , f  (xn ) and m to be

the smallest. Then M = f  (θk1 ) and m = f  (θk2 ) for certain k1 and k2 . The sum
 n    
k=1 λk f (θk ) is a weighted average of the numbers f (θ1 ), f (θ2 ), . . . , f (θn ), and
as such must lie between M and m. Darboux’s theorem now yields a number ξ in
(a, b) with the desired property.

We will also have need of Taylor’s theorem with remainder, including two standard
forms of the remainder sometimes found in calculus books.

Theorem 3 (Taylor’s Theorem with Integral Remainder). If f is a function whose

derivative of order n + 1 is continuous throughout some open interval I centered at
the origin and if h is a number in I , then
 
n
f (k) (0) k 1 h (n+1)
f (h) = h + f (t) (h − t)n dt. (6)
k=0
k! n! 0

Proof. By the fundamental theorem of calculus,

 h
f (h) = f (0) + f  (t) dt,
0

which is (6) for n = 0. Expand the integral using integration by parts, taking u = f  (t),
du = f  (t) dt, dv = dt, v = −(h − t). Repeat inductively.

Theorem 4 (Taylor’s Theorem with Lagrange Remainder). If f is a function whose

derivative of order n + 1 is defined throughout some open interval I centered at the
origin and if h is a number in I , then there exists a point ξ lying between 0 and h such
that
n
f (k) (0) k f (n+1) (ξ ) n+1
f (h) = h + h .
k=0
k! (n + 1)!

We could derive the latter theorem as a corollary to Theorem 3 if we made the

stronger assumption that f (n+1) were continuous on I . The proof relies on the first

February 2006] SIMPSON’S RULE IS EXACT FOR QUINTICS 147

mean value theorem for integrals, which we state presently. However, we require the
theorem in the stronger form that we have given. We omit the proof (the usual one that
appears in calculus texts is based upon Rolle’s theorem).
Finally, we appeal to (and eventually extend) the first mean value theorem for inte-
grals:

Theorem 5 (First Mean Value Theorem for Integrals). If f and ϕ are both contin-
uous on the interval [a, b] and if ϕ(t) ≥ 0 for all t in [a, b], then there exists ξ in (a, b)
such that
 b  b
f (t)ϕ(t) dt = f (ξ ) ϕ(t) dt.
a a

Proof. If ϕ(t) vanishes for all t in [a, b], the conclusion is trivially true, so we may as-
sume
that ϕ takes on a positive value somewhere in the interval, from which it follows
b
that a ϕ(t) dt > 0. Let m be the minimum value taken on by the continuous function
f in the interval [a, b], and let M be its maximum value. Then m ϕ(t) ≤ f (t)ϕ(t) ≤
M ϕ(t) holds for all t in [a, b], implying that
 b  b  b
m ϕ(t) dt ≤ f (t)ϕ(t) dt ≤ M ϕ(t) dt.
a a a

Consequently,

b
f (t)ϕ(t) dt
m≤ a

b ≤ M.
a
ϕ(t) dt

Because the quotient of the integrals is a value that lies between the minimum and
maximum values taken on by f in the interval [a, b], the intermediate value theorem
implies that there must be ξ in (a, b) such that

b
f (t)ϕ(t) dt
f (ξ ) = a
b .
a
ϕ(t) dt

This is the conclusion we sought in the nontrivial case.

It should be clear that the conclusion of the first mean value theorem for integrals
still follows if we replace the nonnegativity hypothesis on ϕ with the more general
condition that ϕ not change sign in [a, b].

3. ERROR IN SIMPSON’S RULE. We now turn to the error in a Simpson’s rule

approximation for the integral of a general (smooth) function. In what follows, we

b
always think of the 2n-subdivision Simpson’s rule approximation S2n for a f (t) dt in
the form that we gave in (2). This form is not optimal for numerical computation, but
it has the merit that each of its summands corresponds to the (signed) area over the
interval [x2k−2 , x2k ] between the x-axis and a certain parabola.
In this theorem and its proof, as well as elsewhere, we will have to consider deriva-
tives of functions when the domains of those functions are closed intervals. Our argu-
ments require continuity of those derivatives throughout those closed intervals. In this
situation, we follow the convention that the derivative of such a function at an endpoint
of its domain is the appropriate one-sided derivative.

148 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 113
Theorem 6 (Error in Simpson’s Rule). If f is a thrice continuously differentiable
function on [a, b] for which f (4) (u) exists at each u in (a, b), then there is a point ξ in
(a, b) such that
 b
(b − a)5
f (t) dt = S2n − f (4) (ξ ) . (7)
a 180(2n)4

Proof. We first consider the error in a single one of the n Simpson’s rule summands.
We can simplify matters appreciably by assuming that the interval associated with that
summand is centered at the origin. We introduce the error function E given on [0, h],
where h = (b − a)/(2n), by
 u
u
E(u) = f (t) dt − [ f (−u) + 4 f (0) + f (u)] ,
−u 3

and we examine E  , E  , and E (3) :

1 u
 
E  (u) = [2 f (−u) − 4 f (0) + 2 f (u)] + f (−u) − f  (u) ;
3 3

u  
E  (u) = − f (−u) − f  (u) − f (−u) + f  (u) ;
3 3
u

E (3) (u) = f (3) (−u) − f (3) (u) .
3
The function f and its derivatives are all continuous functions (at least in short in-
tervals centered at the origin), so E and its derivatives are continuous on [0, h] if h
is sufficiently small. Moreover, E(0) = E  (0) = E  (0) = 0. Taylor’s formula with
integral remainder thus gives
 h
1  1
E(h) = E(0) + E  (0)h + E (0)h 2 + E (3) (t)(h − t)2 dt
2 2 0

1 h

= f (3) (−t) − f (3) (t) t (h − t)2 dt. (8)
6 0

Next, we define a function F on [0, h] by

 (3) (3)
 f (−t) − f (t)
if t  = 0,
F(t) = t

−2 f (4) (0) if t = 0.

We observe that
f (3) (−t) − f (3) (t) f (3) (−t) − f (3) (0) f (3) (t) − f (3) (0)
=− − −→ −2 f (4) (0)
t −t t
(9)

as t → 0+ . Combining (9) with the continuity of f (3) , we find that F is continuous on

[0, h]. This means that we can write
 
1 h
(3) (3)
 1 h
f (−t) − f (t) t (h − t) dt =
2
F(t)t 2 (h − t)2 dt,
6 0 6 0

February 2006] SIMPSON’S RULE IS EXACT FOR QUINTICS 149

and we can invoke the first mean value theorem for integrals to assert the existence of
η in (0, h) such that

F(η) h 2 f (3) (−η) − f (3) (η) 5
E(h) = t (h − t)2 dt = h . (10)
6 0 180η
The existence of f (4) then allows us to apply the mean value theorem to the numerator
of the right-most quotient in (10). This produces a number θ in (−η, η) such that
f (4) (θ)(−2η) 5 h5
E(h) = h = − f (4) (θ) .
180η 90
Thus, the kth interval [x2k−2 , x2k ] associated with Simpson’s rule contains a point
θk such that
 x2k
b−a (b − a)5
f (t) dt − [ f (x2k−2 ) + 4 f (x2k−1 ) + f (x2k )] = − f (4) (θk ) .
x 2k−2 6n 90(2n)5
(11)
We can therefore obtain the error in the 2n-subdivision Simpson’s rule approximation
by summing the quantities on the right-hand side of (11).
According to Darboux’s theorem, we can find ξ in (a, b) such that

n
f (4) (θk ) = n f (4) (ξ ).
k=1

Consequently,

n
(b − a)5 (b − a)5 f (4) (ξ )(b − a)5
− f (4) (θk ) = − · n f (4)
(ξ ) = −
k=1
90(2n)5 90(2n)5 180(2n)4

for a certain ξ in (a, b). This establishes (7).

4. SIMPSON’S RULE FOR QUARTIC AND QUINTIC POLYNOMIALS. In

this section we turn our attention to the behavior of the error term when the inte-
grand is a polynomial of degree four or five. We demonstrate how to evaluate the error
exactly in these cases.
Consider first that if p(x) = a4 x 4 + a3 x 3 + a2 x 2 + a1 x + a0 is a fourth-degree
polynomial, then p (4) (x) = 24a4 is constant. Thus, according to Theorem 6, the er-

b
ror in the 2n-subdivision Simpson’s rule approximation to a p(t) dt is

p (4) (ξ )(b − a)5 a4 (b − a)5

− = − .
180(2n)4 120n 4
It follows that for any positive integer n
 b
a4 (b − a)5
p(t) dt = S2n − .
a 120n 4
There is no need to take n to be anything other than 1, so
 b
a4 (b − a)5
p(t) dt = S2 − . (12)
a 120

150 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 113
 If q(x) = a5 x 5 + a4 x 4 + a3 x 3 + a2 x 2 + a1 x + a0 is a fifth-degree polynomial, we
must be just a little bit more clever. Recall from (10) that
q (3) (−η) − q (3) (η) 5
E(h) = h
180η
for a certain η lying in (0, h). If q is quintic, then q (3) is quadratic and Theorem 1
allows us to write
q (4) (0)(−2η) 5 q (4) (0) 5
E(h) = h =− h .
180η 90
Translating this from the interval [−h, h] to the interval [a, b] and again noting that
there is no reason to choose n different from 1, we conclude that
 b
q (4) [(a + b)/2]
q(t) dt = S2 − (b − a)5 .
a 2880
Because q (4) (x) = 120a5 x + 24a4 , it follows that
 b   
1 a+b
q(t) dt = S2 − 5a5 + a4 (b − a)5 . (13)
a 120 2
Of course, we could have obtained equation (12) from equation (13) by simply
putting a5 = 0. However, we thought it more interesting to derive (12) independently.

5. SOME EXTENSIONS. Before proceeding with this final section of the paper, the
reader may want to try to use the ideas that underlie the proof of Theorem 6 to derive
the classical error expressions for the trapezoidal rule and midpoint rule approxima-
tions that we gave at the beginning of this paper. We begin the section with the core of
just such an argument for each.
The central issue in the proof of Theorem 6 is the evaluation of the integral that
appears on the right-hand side of equation (8). If we write a zero-degree Taylor poly-
nomial with integral remainder for the error in a single summand of a trapezoidal
approximation,
 h
E(h) = f (t) dt − h[ f (−h) + f (h)],
−h

h
we learn that we must evaluate the integral 0 [ f  (−t) − f  (t)] t dt. (The latter is es-
sentially the integral on the right-hand side of equation (3) of [2], though our route to
it has been a little different from the one taken in that work.) Now we note that we can
extend the quotient function F(t) = [ f  (−t) − f  (t)]/t continuously to the origin by
defining F(0) = −2 f  (0). We can then see in much the same way as in the proof of
Theorem 6 that there exist η in (0, h) and ξ in (−h, h) such that
 h 

  f  (−η) − f  (η) h 2 2 f  (ξ )h 3
f (−t) − f  (t) t dt = t dt = − .
0 η 0 3
A somewhat similar problem arises in calculating error in the midpoint rule, the
integral in question in this case being
 h
[ f (−t) − 2 f (0) + f (t)] dt.
0

February 2006] SIMPSON’S RULE IS EXACT FOR QUINTICS 151

Here we introduce the function

 f (−t) − 2 f (0) + f (t) if t  = 0,
F(t) = t2
 
f (0) if t = 0,

and—after establishing the continuity of F at the origin by using L’Hôpital’s rule—we

discover that
 h 
f (−η) − 2 f (0) + f (η) h 2
[ f (−t) − 2 f (0) + f (t)] dt = t dt
0 η2 0

h3
= f  (ξ ) (14)
3
for certain η in (0, h) and ξ in (−h, h).1
Our results for Simpson’s rule, the trapezoidal rule, and the midpoint rule differ
in form, but there is a common thread in the arguments that establish them. We can
isolate this thread by considering the functions that we designated F in each of the
three arguments. In each case, we were able to define F(0) as the limit of a certain
quotient. The limiting values of these quotients are, except for a constant factor, what
many call the first

f (x0 + h) − f (x0 − h)
D1 f (x0 ) = lim
h→0 2h
and the second
f (x0 + h) − 2 f (x0 ) + f (x0 − h)
D2 f (x0 ) = lim
h→0 h2
symmetric derivatives of f at the point x0 . Some authors also call the second symmet-
ric derivative the “Schwartz derivative” or the “Riemann derivative.” It is a well-known
fact, which we have reestablished for the sake of completeness, that if f (k) (x0 ) exists
for k = 1 or 2, then Dk f (x0 ) also exists and is equal to f (k) (x0 ). The converse is not
true (see [11, pp. 22–23]).
We could have attempted to apply the first mean value theorem for integrals (Theo-
rem 5) to the integral that appears in (8). If we had done so, we would have obtained η
in (0, h) for which

f (3) (−η) − f (3) (η) h
E(h) = t (h − t)2 dt.
6 0

As in the argument we gave earlier, the form of the numerator that appears in the
fraction here suggests that we apply the mean value theorem at the same time that we
1 Equation (14) comes from the observation that, by Taylor’s formula with Lagrange remainder, there exist

ξ1 and ξ2 in (0, h) for which

f (−η) = f (0) − f  (0)η + 1
2 f  (−ξ1 )η2 , f (η) = f (0) + f  (0)η + 1
2 f  (ξ2 )η2 ,
so that


f (−η) − 2 f (0) + f (η) = 1
2 f  (−ξ1 ) + f  (ξ2 ) η2 = f  (ξ )η2 ,
where ξ in (−h, h) arises from the corollary to Theorem 2.

152 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 113
evaluate the integral. Doing these things results in a value θ in (−η, η) for which

f (4) (θ)(−2η) h 4 ηh 4
E(h) = · = − f (4) (θ) .
6 12 36

At this point we realize that we have no way of eliminating the number η from the
formula except by passing to magnitudes. Moreover, the denominator is only two-fifths
of what we know it should be. This line of reasoning is therefore unsatisfactory—not
just because it results in an estimate rather than an evaluation, but also because that
estimate is not as tight as it ought to be.
We were able to obtain a stronger result because the difference [ f (3) (−t) − f (3) (t)]
has a zero at t = 0. Introduction of the symmetric derivative allowed us to transfer this
zero, in the form of a power of t, to the part of the integrand that we wanted to leave
inside the integral when we applied Theorem 5. Similar considerations came into play
in our other arguments. In point of fact, however, it is not the symmetric derivative
as such that is crucial to our reasoning, but the observation that the correct portion of
the integrand vanishes in a helpful way. We capture this thread in the arguments as a
theorem:

Theorem 7 (Extended Mean Value Theorem for Integrals). Let n be a positive in-
teger. Suppose that g is (n − 1)-times continuously differentiable on [a, b], that t0
lies in [a, b], that g (n) (t) exists for every t in {t0 } ∪ (a, b), and that g (k) (t0 ) = 0 for
k = 0, 1, . . . , n − 1. If a function ϕ is continuous on [a, b] and if the function t  →
ϕ(t)(t − t0 )n does not change sign at any point of (a, b), then there exists η in (a, b)
such that
 
b
g (n) (η) b
g(t)ϕ(t) dt = ϕ(t) (t − t0 )n dt.
a n! a

Proof. By our hypotheses and n − 1 applications of L’Hôpital’s rule,

g(t) g (n−1) (t) 1 g (n−1) (t) − g (n−1) (t0 ) g (n) (t0 )

lim = lim = lim = .
t→t0 (t − t0 )n t→t0 n! (t − t0 ) n! t→t0 t − t0 n!

This means that the function t  → g(t)/(t − t0 )n has a removable singularity at t0 , so

the function G : [a, b] → R given by


 g(t)
 if t  = t0 ,
(t − t0 )n
G(t) =

 1
 g (n) (t0 ) if t = t0 ,
n!

is continuous. Consequently, we can appeal to Theorem 5 to find a number ξ in (a, b)

such that
 b  b  b
g(t)ϕ(t) dt = G(t)(t − t0 ) ϕ(t) dt = G(ξ )
n
ϕ(t) (t − t0 )n dt.
a a a

If ξ = t0 , we simply take η = t0 = ξ , and we are done. If ξ  = t0 , we apply Taylor’s

theorem with Lagrange remainder to g(ξ ) to find an η interior to the interval deter-

February 2006] SIMPSON’S RULE IS EXACT FOR QUINTICS 153

mined by t0 and ξ (and a fortiori interior to [a, b]) such that
 
1 1 n−1
0 g (n) (η)
G(ξ ) = g(ξ ) = (ξ − t0 ) +
k
(ξ − t0 ) n
(ξ − t0 )n (ξ − t0 )n k=0 k! n!
1 (n)
= g (η).
n!

We invite the reader to apply Theorem 7 to derive the classical error expressions
for the trapezoidal rule and the midpoint rule, as well as error expressions for the
trapezoidal rule, midpoint rule, and Simpson’s rule approximations to integrals whose
integrands are not as smooth as the classical error expressions require. In each case,
Taylor’s formula with integral remainder provides an expression that we can use the
extended mean value theorem for integrals to evaluate.
For example, in the case of the midpoint rule and a twice-differentiable function f ,
Theorem 7 immediately gives an η in (0, h) for which
 
h
1
  h
[ f (−t) − 2 f (0) + f (t)] dt = f (−η) + f  (η) t 2 dt,
0 2 0

whence by the corollary to Theorem 2 there is a ξ in (−h, h) satisfying (14).

By way of a further example, if we know that f is a twice continuously differen-
tiable function on [a, b], we can write the Simpson’s rule error function defined in
the proof of Theorem 6 as a first-degree Taylor polynomial with integral remainder.
Equivalently,
 h 

 
−3E(h) = f  (−u) − f  (u) + f  (−u) + f  (u) u (h − u) du
0
 
h
 h

= f  (−u) − f  (u) (h − u) du + f  (−u) + f  (u) u(h − u) du.
0 0

We can then invoke Theorem 7 in conjunction with the corollary to Darboux’s theorem
to reduce the first of these two integrals to − f  (θ)h 3 /3 for some θ in (−h, h). We
are then able to combine Theorem 5 with the aforementioned corollary to reduce the
second integral to f  (η)h 3 /3 for a certain η in (−h, h). Dividing by −3, translating
to each of the Simpson’s rule subintervals of [a, b], and summing, we infer that there
exist ξ1 and ξ2 in (a, b) for which

 (b − a)3
S
E 2n = f  (ξ1 ) − f  (ξ2 ) . (15)
18(2n)2
Equation (15), for twice continuously differentiable functions, is similar to the estimate
(1.7) given for Lipschitz functions by Cruz-Uribe and Neugebauer in [3]. That (15) is
an order two estimate, whereas (1.7) of [3] is order one, we attribute to the additional
smoothness that we have imposed upon the integrand.
If, in addition, we know that f (3) (t) exists and satisfies | f (3) (t)| ≤ K 3 for all t in
(a, b), we can apply the mean value inequality immediately before summing and then
conclude after the summation that
(b − a)4
|E 2n
S
| ≤ K3 .
9(2n)3

154 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 113
 This estimate is not the best available: its denominator is no more than one-eighth of
what it should be (see Exercises 51 and 52 in [6, chap. 5]). Nevertheless, the estimate
has the right order of decay, and it doesn’t depend on the mysteries of an “influence
function.” We therefore have a partial answer to the question raised in [2] about using
elementary techniques to estimate the errors that may arise when we apply Simpson’s
rule to functions whose fourth-order derivatives are not available.

ACKNOWLEDGMENT The author presented, under the same title, a preliminary report on this paper at
the Annual Meeting of the Rocky Mountain Section of the Mathematical Association of America in Colorado
Springs in April 2004.

REFERENCES

1. T. M. Apostol, Calculus, vol. 2, 2nd ed., Blaisdell, Waltham, MA, 1969.

2. D. Cruz-Uribe and C. J. Neugebauer, An elementary proof of error estimates for the trapezoidal rule,
Math. Mag. 76 (2003) 303–306.
3. , Sharp error bounds for the trapezoidal rule and Simpson’s rule, J. Ineq. Pure Appl. Math. 3(4)
(2002), article 49; available at http://jipam-old.vu.edu.au/v3n4/031_02.html.
4. J. G. Darboux, Mémoire sur les fonctions discontinues, Ann. Sci. École Norm. Sup. Sér. 2 4 (1875) 57–
112.
5. M. P. Fobes and R. B. Smyth, Calculus and Analytic Geometry, vol. 1, Prentice-Hall, Englewood Cliffs,
NJ, 1963.
6. F. B. Hildebrand, Introduction to Numerical Analysis, 2nd ed., McGraw-Hill, New York, 1974.
7. J. M. H. Olmstead, Real Variables, Appleton-Century-Crofts, New York, 1959.
8. , Advanced Calculus, Appleton-Century-Crofts, New York, 1961.
9. A. Ralston, A First Course in Numerical Analysis, McGraw-Hill, New York, 1965.
10. R. E. Williamson, R. H. Crowell, and H. F. Trotter, Calculus of Vector Functions, 3rd ed., Prentice-Hall,
Englewood Cliffs, NJ, 1972.
11. A. Zygmund, Trigonometric Series, 2 vols., Cambridge University Press, Cambridge, 1977.

LOU TALMAN earned his B.A. from the College of Wooster in 1966 and his Ph.D. from the University of
Kansas in 1975. Since then he has taught at Kansas State College at Pittsburg (now Pittsburg State University),
Southern Illinois University at Carbondale, the University of Central Arkansas, and the College of Ganado.
He recently completed his second decade at Metropolitan State College of Denver, from whence he can spend
much of the brief mountain summer in Colorado’s high country.
Department of Mathematical and Computer Sciences, Metropolitan State College of Denver, Denver, CO
80217-3362
[email protected]

February 2006] SIMPSON’S RULE IS EXACT FOR QUINTICS 155

View publication stats