Taylor & Francis, LTD., Mathematical Association of America The American Mathematical Monthly
Taylor & Francis, LTD., Mathematical Association of America The American Mathematical Monthly
Taylor & Francis, LTD., Mathematical Association of America The American Mathematical Monthly
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
Taylor & Francis, Ltd., Mathematical Association of America are collaborating with
JSTOR to digitize, preserve and extend access to The American Mathematical Monthly
This content downloaded from 128.32.10.230 on Thu, 11 Jun 2020 02:13:22 UTC
All use subject to https://about.jstor.org/terms
An Elementary View of Euler's
Summation Formula
Tom M. Apostol
11-. LII-I
f
0 1 2 3 4 . . . n-l n
This content downloaded from 128.32.10.230 on Thu, 11 Jun 2020 02:13:22 UTC
All use subject to https://about.jstor.org/terms
It is clear that dn + 1 > dn and that all the shaded pieces can be translated to the
left to occupy a portion of the rectangle of altitude f(1) above the interval [0, 1], as
shown in Figure 1. Because f is decreasing there is no overlapping of the
translated shaded pieces. Comparison of areas gives us the inequali
dn+ 1 < f(1). Therefore {dn} is increasing and bounded above, so it ha
C(f) = limn d(n). We refer to C(f) as the generalized Euler's constant associ-
ated with the function f. Geometrically, C(f) represents the sum of the areas of
all the curvilinear triangular pieces over the interval [1, co). These pieces can be
translated to fit inside the rectangle of area f(1) shown in Figure 1 (without
overlapping), so we have the inequalities 0 < C(f ) < f(1). Moreover, C(f ) - dn
represents the sum of the areas of the triangular pieces over the interval [n, oo).
These pieces can be translated to the left to occupy (without overlapping) a
portion of the rectangle of height f(n) above the interval [n, n + 1]. Comparing
areas we find
Theorem 1. Iff is positive and strictly decreasing on [1, oo) there is a positive constan
C(f) < f(1) and a sequence {Ef(n)}, with 0 < Ef(n) < f(n), such that
n n
Ef (k)= f(x) dx + C(f) + Ef(n), n = 2,3 .. (3)
k=1 1
Note. Eq. (3) tells us that the difference between the sum and the integral is equal
to a constant (depending on f) plus a positive quantity Ef(n) smaller than the last
term in the sum. Hence, if f(n) tends to 0 as n -> oo, then Ef(n) also tends to 0.
Proof: If we define Ef(n) = f(n) + d,] - C(f), then (3) follows from the definition
(1), and the inequality 0 < Ef(n) < f(n) follows from (2). 2
Example. When f(x) = 1/x, C(f) is the classical Euler's constant, often denoted
by C (or by -y), and (4) states that C = lim n ('k1(1/k) - log n). It is not
known (to date) whether Euler's constant is rational or irrational. Its numerical
value, correct to 20 decimals, is C = 0.57721566490153286060. In this case, Theo-
rem 1 says that
's 1 1
' k-= logn + C +E(n), where <E(n) <-.
k=1 k n
dn = I(k), (S)
k=1
This content downloaded from 128.32.10.230 on Thu, 11 Jun 2020 02:13:22 UTC
All use subject to https://about.jstor.org/terms
where
rk+1
I(k) = {f(k) -f (x)} dx. (6)
k
When f is positive and decreasing, as in Figure 2, 1(k) is the area of the shaded
curvilinear triangular piece over the interval [k, k + 1]. However, (5) and (6) are
meaningful for any integrable f.
k k+ 1
The integrand in (6) has the form u dv, where u = f(k) - f(x) and v = x + c,
where c is any constant. If we choose c = - (k + 1) and integrate by parts
(assuming that f has a continuous derivative), the integrated part vanishes and the
integral 1(k) reduces to
The last two terms on the right represent the error made when the sum on the
left is approximated by the integral fnl f(x) dx. The formula is useful because f
need not be positive or decreasing. In fact, f can be increasing or oscillating.
Variants of this formula will be obtained as we attempt to deduce more precise
information about the error.
The factor x - [x] is a nonnegative function with period 1. If f' has a fixed sign
(as it has when f is monotonic), the integral term in the error has the same sign as
f'. To decrease the error it is preferable to multiply f'(x) by a factor that changes
This content downloaded from 128.32.10.230 on Thu, 11 Jun 2020 02:13:22 UTC
All use subject to https://about.jstor.org/terms
sign so that some cancellation takes place in the integration. To introduce sign
changes, we translate the function x - [x] down by 4 and consider the new
function x - [xl - 4 whose graph is shown in Figure 3. The integral term in the
J 1 2 3 4
(X - [
The last term is equal to 4{f(n) - f(1)1. Using this in (7) we obtain the following
variant of the first-derivative form of Euler's summation formula:
The error integral does not change if the factor x - [x]I - 4 is replaced by P1(x)
because the two factors differ only at the integers. Therefore (8) can be written as
n n
Note the contrast between (10) and (3), which explicitly displays the generalized
Euler's constant C(f). To make (10) resemble (3) more closely, we assume that the
improper integral JlPl(x)f'(x) dx converges. Then we can write
n 00 00
jPl(x)
1 ~~ ~~1 n
This content downloaded from 128.32.10.230 on Thu, 11 Jun 2020 02:13:22 UTC
All use subject to https://about.jstor.org/terms
where
100
Ef(n)
2
= 2f
Eq. (11) has exactly the same form as (3), but (11) is more general because f is not
required to be positive or monotonic. The only restrictions on f are continuity of
f' and convergence of the improper integral
00
To see this, note that the Bernoulli function P1(x) is bounded; in fact, Figure 3
shows that IP1(x)l < 1 for all x, so (14) follows from (15).
JIf'(x)IdxJ nx
2 d n
Therefore (15) is satisfied and (12) expresses Euler's constant as an integral:
C = - f 2 dx.
This content downloaded from 128.32.10.230 on Thu, 11 Jun 2020 02:13:22 UTC
All use subject to https://about.jstor.org/terms
P2(x)
0 1 2 3 4 5
Figure 4
n 1 1 n
provided that f" is continuous. Repeated integration by parts leads to the general
form of Euler's summation formula, which involves higher order derivatives of f
and higher order periodic Bernoulli functions that represent polynomials on the
unit interval [0, 1]. To see exactly how the Bernoulli functions evolve in the process
we follow the method of the foregoing section and integrate the periodic function
3P2(t) from 0 to x to obtain another periodic function P3(x) whose derivative is
3P2(x). To guarantee that the integrated function P3(x) is periodic with period 1
we need JoP2(t) dt = 0. This property governs the choice of the constant c in
The integral of the quadratic polynomial x2 - x + c from 0 to 1 is equal to c - 6
so we choose c = 6 and take
() Pl(t)
P2(X) = 2f o 6 dt +-
Euler's summation formula can now be restated as follows:
ii n n
Moreover,
E f (k
k=11
where
1 1 1,.oo
C(f) = _f(1) - _P2(0)f'(1) - 2j P2(x)f"(x) dx, (18)
and
1 1 1 0o
Ef(n) = 2f (n) + fP2(0)f(n) + - 2P2(x)f" (x)dx. (19)
This content downloaded from 128.32.10.230 on Thu, 11 Jun 2020 02:13:22 UTC
All use subject to https://about.jstor.org/terms
x
To improve the error estimate we integrate P2(t) from 0 to x and define the
Bernoulli function P3(x) = 3foP2(t) dt so that P3(x) = 3P2(x). There is no need
to add a constant in this case because, on the unit interval [0, 1], P3(x) = x3 -2
+ 2x, and JJP3(t) dt = 0. The function P3 has period 1 because P2 has period 1
and f1P2(t) dt = 0. The graph of P3 is a bounded piecewise cubic curve, shown
Figure 5. Note that P3(x) vanishes at the integers. Integration by parts over [1, n]
gives us
n 1n
fP2(x)
I I3
provided f(3) is continuous. This equation, together with Theorem 3, gives a
third-derivative form of Euler's summation formula in which the second integral on
the right of (17) is replaced by fnP3(x)f (3)(x) dx. The corresponding changes in
(18) and (19) are replacement of the integrals by I yflP3(x)f (3)(X) dx and
3!n dxP3(x)f(3)(X) , respectively.
f1Pk(t) dt = 0. (21)
Periodicity implies that Pk(O) = Pk(l), and (21) shows that each of these values is
Bk. As already noted, on the closed interval [0, 1] each function Pk(x) is a
polynomial of degree k when k = 2 or 3. [The case k = 1 is special; P1(x) is a
linear polynomial x - 2 only on the open interval (0, 1) and is discontinuous at the
endpoints.] It is clear (and easily proved by induction) that on the closed interval
[0, 1] the function defined by (20) is a polynomial of degree k if k ? 2. We denote
this polynomial by Bk(x), the usual notation for Bernoulli polynomials. The first few
are
1 ~ ~ =2_X+1 3 1
Bl(x)=x-- 2 B2(x)=x2--x+ 6 B3(x)=x3 x2+ 2x,
4 X3+ X2 1 1 1 1
B4(X) =X-X2x x 0 B5(x) =x5 --X4 + x3 3 0x
B6(x) =x x5
= 6 3 1 1 1
5+-x4
2 10_ x2 42
+-.
This content downloaded from 128.32.10.230 on Thu, 11 Jun 2020 02:13:22 UTC
All use subject to https://about.jstor.org/terms
The Bernoulli periodic functions are periodic extensions of these polynomials given
by Pk(x) = Bk(X - [x]). The constants Bk = Pk() = Pk(l) are called Bernoulli
numbers. The first few are
1 1 1
B1 = - B2 B3 = B B4
2'2 6' 330B
1 1
B5 = O, B6 =- B7 = - , B8= -_
42 ' 7 ,B= 30
Next we show that our definitions of Bernoulli numbers and polynomials are
consistent with the usual definitions, provided we take BO(x) = 1 and Bo = 1. O
definition in (20) shows that the successive derivatives of these polynomials are
and hence
On the other hand, the Taylor expansion of any polynomial Bk(x) of degree k is
given by Bk(x) = = B(')(0)x"/r!, so (22) implies
This is the usual recursion formula for defining Bernoulli numbers (starting with
Bo = 1), and (23) is one of the standard ways of defining Bernoulli polynomials in
terms of Bernoulli numbers. Consequently, the numbers and polynomials that
appear in our treatment are the usual Bernoulli numbers and Bernoulli polynomi-
als that appear in the literature; see [1, p. 265], [2, p. 251], or [5, pp. 160-163].
It is well known that the Bernoulli numbers Bk with odd index k ? 3 are zero,
so only Bernoulli numbers with even index appear in the general form of Euler's
summation formula. It is also known [8, p. 533] that on the interval [0,1] the
Bernoulli polynomials satisfy the following inequalities for k ? 1:
Theorem 4. (General form of Euler's summation formula). For any function f with
a continuous derivative of order 2m + 1 on the interval [1, n] we have
This content downloaded from 128.32.10.230 on Thu, 11 Jun 2020 02:13:22 UTC
All use subject to https://about.jstor.org/terms
Moreover, if the improper integral f If (2m7 ?1) (X) dx converges then we also have
where
1 n B2
1 B2r
Ef (n) = j(n) + (2)!(2- 1) (n)
12dx2r(27
Example. When f(x) = l/x we have f(2n +1)(x) - (2m + 1)!/x2m1+2, and (26)
gives the following expression for the classical Euler's constant:
1 B2 B4 B2 m OOP2?17 (X)
C + +- -+ ++ ;;- 2? dx. (28)
2 2 4 2m 1 X2mt+2 28
The corresponding error term (27) becomes
1 B2 B4 B21 + 12 P2tn+2(I)
2 n 2 42 f 2m2dx. 29
Ef(n) = -~~ 2n2 4 n 2rnn2~ , x'' . (9
One is tempted to let m -> o in (28) and obtain an infinite series for Euler's
constant. However, the integral in (28) does not tend to 0 as m -> oo and, in fact, it
can be shown that the infinite series YB2k/(2k) diverges rapidly [see 6, p. 529], so
(28) is not very useful for calculating C. Nevertheless, as we show in the next
section, (25) and (27) can be used to calculate C very accurately.
1 B2 B4 B6 + c0 rP7(X)
Ef
Ef ( n) = 2n - (n) - 6n6
2n2 4n4 2 Jn
+1dx
x8
1 1 1 1 GO P7(X)
00 P7( X) 1 0oci 1
dx < ~-dx=
L 7(8 -42 71
This content downloaded from 128.32.10.230 on Thu, 11 Jun 2020 02:13:22 UTC
All use subject to https://about.jstor.org/terms
and (30) can be written in the form
n 1 1 1 1 1
C ,- - log + ~~- + + E(n), (31)
k-i k g 2n 12n2 120n4 252n6
where 0 < IE(n) I < 1/42n7. Using a hand calculator that displays 12 digits we
find Sl_1k - = 2.92896825381 and log 10 = 2.30258509299. If n = 10 the sum of
the error term E(n) plus the term with 252n6 in the denominator in (31) is too
small to influence the seventh digit. Neglecting these terms and retaining 8 digits
in the calculation we find
1 1 1
C - 2.92896825 - 2.30258509 - - + 1200 - 1200000
REFERENCES
1. Tomn M. Apostol, Introduction to Analytic Number Theory. Springer-Verlag, New York, 1976.
2. Tom M. Apostol, MathematicalAnalysis, Second edition. Addison-Wesley, Reading, Mass., 1974.
3. L. Euler, Methodus universalis serierum convergentium summas quam proxime inveniendi, Commen-
tarii academie scientiarum Petropolitanae, Vol. 8 (1736), pp. 3-9; Opera Omnia, Vol. XIV,
pp. 101-107.
4. L. Euler, Methodus universalis series summandi ulterius promota, Commentarii academie scientiarum
Petropolitanae, Vol. 8 (1736), pp. 147-158; Opera Omnia, Vol. XIV, pp. 124-137.
5. E. Hairer and G. Wanner, Analysis by Its History. Springer-Verlag, New York, 1996.
6. K. Knopp, Theory and Application of Infinite Series, R. C. Young, translator. Hafner, New York,
1951.
7. D. E. Knuth, Euler's constant to 1271 places, Math. of Computation, v. 16 (1962), pp. 275-281.
8. D. H. Lehmer, On the maxima and minima of Bernoulli polynomials, Amer. Math. Monthly, v. 47
(1940), pp. 533-538.
9. Colin Maclaurin, A Treatise of Fluxions. Edinburgh, 1742.
TOM M. APOSTOL received his Ph.D. in 1948 with a thesis in analytic number theory written under
the direction of D. H. Lehmer at UC Berkeley. He joined the Caltech faculty in 1950 and became
professor emeritus in 1992. His list of publications contains 58 research papers and several books,
including his pathbreaking Calculus in two volumes, first published in 1961, Mathematical Analysis
(1957), and Introduction to Analytic Number Theory (1976), all of which are still in print. He is director
of Project MATHEMATICS!, a prize winning series of videos and other educational activities he
initiated twelve years ago. His 50-year career in mathematics is described in an engaging article by
Don Albers, An Interview with Tom Apostol, published in the September 1997 issue of The College
Mathematics Journal.
Project MATHEMATICS!, 1-70 Caltech, Pasadena, CA 91125
[email protected]
This content downloaded from 128.32.10.230 on Thu, 11 Jun 2020 02:13:22 UTC
All use subject to https://about.jstor.org/terms