1 1
= · v − v · − 2 du
u u
v v
= + du
u u2
as before.
Secondly, there is the potential only for slight technical advantage in choosing for-
mula (2) over formula (1). An identical integral will need to be computed whether
we use (1) or (2). The only difference in the required differentiation and integration
occurs in the computation of du versus dU. In our example, for instance, we differen-
tiated u = x 1/2 rather than U = x −1/2 .
Acknowledgment. The author wishes to thank Cal Poly Pomona student Victor Reynolds for
his insightful idea.
Reference
1. J. Stewart, Calculus: Early Transcendentals, 4th ed., Brooks/Cole, 1999.
Estimating Definite Integrals
Norton Starr (nstarr@[Link]), Amherst College, Amherst, MA 01002
Many definite integrals arising in practice can be difficult or impossible to evaluate
in finite terms. Series expansions and numerical integration are two standard ways to
deal with the situation. Another approach, primitive but often very effective, yields
cruder estimates by replacing a nasty integrand with nice functions that majorize or
are majorized by it. With luck and skill, the bounds achieved suffice for the task at
hand. I was introduced to this method as a grad student instructor over forty years ago,
when I had the good fortune to learn some innovative teaching methods from Arthur
Mattuck. His supplementary notes for MIT’s calculus course included a section on
the estimation of definite integrals by an approach barely covered in texts back then.
Many first year calculus texts of that era touched on the method in connection with
comparison tests for improper integrals, but they seldom did anything with proper
integrals.
The situation has improved somewhat in recent years, with prominent texts at least
mentioning the basic idea within the chapter introducing the definite integral. Some-
times this is labeled the “Domination Rule” or “Comparison Property”. An informal
survey shows that most such books offer very few, if any, exercises in the method,
usually relatively trivial ones. The texts by Edwards & Penney [1] and Stewart [4] are
exceptional in providing more than a token selection of such problems. Unfortunately,
their exercises are duplicated in the early transcendental versions of these two texts,
thus making no use of the broader array of available functions.
Here is a sketch of the way I develop this form of estimation in my intermediate cal-
culus course. All integrals are understood to be over a closed, bounded interval [a, b]
60
c THE MATHEMATICAL ASSOCIATION OF AMERICA
and all functions assumed to be (Riemann) integrable. I start with the primitive obser-
b
vation that if f is nonnegative and integrable on [a, b] then a f ≥ 0. (This affords an
b
opportunity to remind students of a definition of a f and some of its implications.)
Next, using the linearity of the integral and the fact that sums and differences of inte-
b are integrable, I infer that for integrable f and g, if f ≤ g on [a, b]
grable functions
b
then a f ≤ a g. This last inequality is the key tool. As with any tool, its effective-
ness depends on the skill with which it’s wielded. For a given function f or g, the trick
is to dream up an appropriate comparison function that leads to a useful bound.
Several fundamental inequalities facilitate this effort:
1) sin x < x for all positive x. (This is easily illustrated with graphics devices, but
asking students why it’s true can lead to a review of basic methods from a first
course in calculus.)
√
2) ln x < x and ln x < x for all x > 0. (Again, easily illustrated and an occasion
for pointing to the relevance of earlier methods.)
3) x c < x d if c > d and 0 < x < 1. (As shown below, this simple observation can
turn messy integrals into trivial ones.)
4) If f and g are b continuous
b on [a, b] and f ≤ g yet f (c) < g(c) for some c in
[a, b], then a f < a g. (This can serve as an occasion to review continuity.)
b b
5) If f is integrable on [a, b] then so is | f |, and | a f | ≤ a | f |. (This continuous
analog of the triangle inequality is an easy way to eliminate pesky terms of the
form (−1)n .)
Examples.
6 6 √ 6
i) Which of 1 arctan(sin x) d x, 1 arctan x d x, 1 arctan x d x has the greatest
value? Which has the least value? Why? This simply relies on the monotone
increasing nature of the
√ inverse tangent function, together with the obvious re-
lations among sin x, x, and x over the interval [1, 6].
1/2 1/2 √
ii) Similarly, comparing 0 cos x 2 d x with 0 cos x d x amounts to combining
√
the inequality x 2 < x on (0, 1) with√ the observation that cosine is decreasing
on this interval, so cos(x 2 ) > cos( x). Thus the first integral is larger.
iii) Verify that
15
t3
dt < .003.
10 t6 + t2 − 1
On [10, 15], t 2 − 1 > 0, so deleting the t 2 − 1 in the denominator increases the
integrand to t −3 . Integrating this from 10 to 15 gives as an upper bound:
1 1 1 1
− 2
− 2 = ,
2 15 10 360
which is < .003.
iv) Verify that
1
3 1 9
< dt < .
4 0 1 + t4 10
VOL. 36, NO. 1, JANUARY 2005 THE COLLEGE MATHEMATICS JOURNAL 61
This can be delicate. The lower bound is readily obtained by observing that
1 + t 4 < 1 + t 2 on (0, 1) and using the inverse tangent. The second inequal-
ity appeared on a final examination at our college over thirty years ago, and
caused considerable consternation. One route would have been via partial frac-
tions, but our aim was to avoid such extensive computation. Various upper
bounds for 1/1 + t 4 were unsuccessfully proposed among those teaching the
course (they would have to grade the test and explain things to curious or un-
happy students). Finally my colleague David Armacost suggested showing, in-
stead, that one minus the integral is greater than one tenth. This is relatively
simple:
1 1 1
1 t4 t4 1
1− dt = dt > dt = .
0 1 + t4 0 1 + t4 0 1+1 10
v) One of the standard examples in calculus involves computing the volume and
surface area of the solid of revolution obtained by rotating the graph of y = 1/x
about the x-axis for x ≥ 1. Known as Gabriel’s Horn, this paradoxical object
has an easily computed finite volume, while the surface area is infinite. The area
computation,
A
1 1
lim 2π 1 + 4 d x,
A→∞ 1 x x
is a good exercise in techniques of integration. However, to show the unbound-
edness of the surface area it’s sufficient simply to observe that the integrand is
larger than 2π x1 and that the resulting 2π ln A blows up when A → ∞. (Cf.
“The Paradox of Gabriel’s Horn” in the Varberg, Purcell, Rigdon text [5, pp.
418, 419].)
vi) As a final application, we note that without appeal to the theorem allow-
ing termwise integrability of series, it’s possible to represent arctan(x) and
ln(1 + x) as power series. Starting with the fundamental finite geometric sum
formula:
1 x n+1 n
= 1 + x + x2 + x3 + · · · + xn + = x k + error term,
1−x 1−x k=0
substitute x = −t 2 and −t, respectively. Integrating from 0 to x yields arctan(x)
and ln(1 + x) as finite power series plus integrals of the error term, respectively.
It’s not difficult to show that as n → ∞, the integral of the error term disappears
for −1 ≤ x ≤ 1 in the case of arctan(x) and for −1 < x ≤ 1 in the case of
ln(1 + x), yielding the familiar power series for these functions.
For instance, in the case of ln(1 + x), let ε ∈ (0, 1) and suppose −1 + ε ≤
x ≤ 0. The error term integral is
x
(−t)n+1
dt.
0 1+t
Because x ≤ 0, it helps to change variables via t = −v, so 0 ≤ v ≤ −x = |x|.
This facilitates using the continuous analog of the triangle inequality. Thus the
magnitude of the error term integral satisfies, for fixed x ∈ [−1 + ε, 0],
62
c THE MATHEMATICAL ASSOCIATION OF AMERICA
|x| n+1 |x| n+1
x
(−t)n+1 |x| v n+1 v v
dt =
(−1) dv ≤
1+t 1 − v 1 − v dv = 1 −v
dv
0 0 0 0
|x| n+1
v |x|n+2
≤ dv = .
0 1 − |x| (n + 2)(1 − |x|)
This shrinks to zero as n → ∞, indeed uniformly for x ∈ [−1 + ε, 0].
The corresponding result when x ∈ [0, 1] is similar but easier, because there
is no need to avoid an endpoint at which the integrand blows up. For the arctan
function it’s also easy to show that, for −1 ≤ x ≤ 1, the error integral goes to
zero as n → ∞. This latter analysis has appeared in George Thomas’s calculus
texts at least since 1960 (see [2, pp. 689–690].)
The basic approach used in this sort of work may remind readers of the comparison
test for infinite series. However, the series method tends to be more simple in class-
room practice, for two reasons. First, the traditional repertory of comparison series is
relatively limited ( p-series and geometric series, with perhaps also factorial series).
Second, in the study of a series the primary goal is to find out whether it converges,
and for this the limit comparison test obviates the need to confirm an inequality. When
the task turns to bounding or estimating the sum of a convergent series or to assessing
the rate of growth of a divergent series, the approaches become more similar.
The above examples are elementary, yet they suggest the flexibility needed in ap-
proximations as well as the significant simplification provided by changing our point
of view. Students find these problems difficult precisely because they cannot be solved
by rote-learned algorithms. As Richard Feynman remarked when discussing skill in
making approximations, “This is very difficult to teach because it’s an art.” [3, p. 16]
This topic reinforces the view that calculus is the study of the behavior of functions.
More importantly, it shows mathematics as an art, as an endeavor probing relationships
among elements in an infinite universe.
Acknowledgment. I am grateful to the referee for alerting me to the presence of this method
in contemporary calculus texts.
References
1. C. H. Edwards and D. E. Penney, Calculus, 6th ed., Prentice Hall, 2002.
2. R. L. Finney, M. D. Weir, and F. R. Giordano, Thomas’ Calculus, 10th ed., Addison-Wesley Longman, 2003.
3. D. and J. Goodstein, “ ‘The Uncreative Scientist’—Feynman’s Other Lost Lecture”, Caltech News, 31:4 (1997)
12–17.
4. J. Stewart, Calculus, 5th ed., Thomson Learning, 2003.
5. D. Varberg, E. J. Purcell, and S. E. Rigdon, Calculus, 8th ed., Prentice Hall, 2000.
Added in proof: It has come to the author’s attention that there was an earlier Capsule
in this Journal illustrating the repeated use of the basic comparison property to ob-
tain improved estimates of definite integrals. See W. V. Underhill, Finding bounds for
definite integrals, The College Mathematics Journal 15 (1984) 426–429.
VOL. 36, NO. 1, JANUARY 2005 THE COLLEGE MATHEMATICS JOURNAL 63
