Mathematical Association of America
Mathematical Association of America
Mathematical Association of America
Apostol Reviewed work(s): Source: The American Mathematical Monthly, Vol. 106, No. 5 (May, 1999), pp. 409-418 Published by: Mathematical Association of America Stable URL: http://www.jstor.org/stable/2589145 . Accessed: 31/05/2012 11:26
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp 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].
Mathematical Association of America is collaborating with JSTOR to digitize, preserve and extend access to The American Mathematical Monthly.
http://www.jstor.org
d,l= Ef(k)
k=l
-f(x)
1
dx,
n=2,3...
(1)
11-. LII-I
n-l
19]
409
It is clear that dn+ 1 > dn and thatall the shaded pieces can be translated the to of leftto occupya portionof the rectangle altitudef(1) above the interval 1], as [0, shown in Figure 1. Because f is decreasing there is no overlappingof the translatedshaded pieces. Comparisonof areas givesus the inequalities0 < d,, < and bounded above,so it has a finite limit {dn} is increasing dn+ 1 < f(1). Therefore C(f) = limn d(n). We referto C(f) as the generalized Euler's constant associated withthe function Geometrically, f. C(f) represents sum of the areas of the all the curvilinear triangular pieces over the interval[1, co). These pieces can be translatedto fit inside the rectangleof area f(1) shown in Figure 1 (without so overlapping), we have the inequalities0 < 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 translatedto the left to occupy (withoutoverlapping)a portionof the rectangleof heightf(n) above the interval n + 1]. Comparing [n, areas we find 0 < C(f) -dn <f(n), From these inequalitieswe can easilydeduce: n = 2,3,.. (2)
on Theorem1. Iff is positive and strictly decreasing [1,oo) there a positive is constant C(f) < f(1) and a sequence{Ef(n)}, with0 < Ef(n) < f(n), such that n n Ef (k)= f(x) dx + C(f) + Ef(n), n = 2,3 .. (3)
k=1 1
betweenthe sum and the integral equal Note. Eq. (3) tellsus thatthe difference is to a constant (dependingon f) plus a positivequantity Ef(n) smallerthanthe last termin the sum. Hence, if f(n) tends to 0 as n -> oo,then Ef(n) also tends to 0. from definition the Proof: If we defineEf(n) = f(n) + d,] - C(f), then(3) follows 0 from (1), and the inequality < Ef(n) < f(n) follows (2). 2 If f(n) -> 0 as n -> oo,then(3) implies n C(f) = lim Ef(k)
k-=l f1
n ff(x) dxX|.
/
(4)
oftendenoted Example. When f(x) = 1/x, C(f) is the classical Euler's constant, - log n). It is not by C (or by -y),and (4) states that C = lim ('k1(1/k) n known(to date) whetherEuler's constantis rationalor irrational. numerical Its In value, correctto 20 decimals,is C = 0.57721566490153286060. thiscase, Theorem 1 says that ' k-=
k=1
's
logn + C +E(n),
where
<E(n)
<-.
no longerassume that f is positiveor decreasing.At the outsetwe require only that the integralfn f(x) dx existsfor each integern ? 2. The key insightis to noticethatthe difference in (1) can be written d, as
,i-1
dn
k=1
I(k),
(S)
410
[May
where I(k)
=
rk+1
as When f is positiveand decreasing, in Figure 2, 1(k) is the area of the shaded [k, triangular piece over the interval k + 1]. However,(5) and (6) are curvilinear f. for meaningful any integrable
{f(k)
-f (x)} dx.
(6)
k+ 1
of Figure 2. Geometric interpretation the integral 1(k) as the area of the shaded region.
in The integrand (6) has the formu dv,where u = f(k) - f(x) and v = x + c, where c is any constant. If we choose c = - (k + 1) and integrateby parts partvanishesand the the derivative), integrated thatf has a continuous (assuming integral1(k) reduces to I(k)
=
k symbolx variesfromk to k + 1, so the quantity in In thisintegral dummy the the integrandcan be replaced by [x], the greatest integer < x. Make this in and substitute (6) to find replacement
n-1 dn n-i
=
n?I(k)
k=1
n
k=1
(+
[x]
n
1)f'(x) dx
=f(x
=
f
n
- [x])f'(x)
dx -
f'(x) dx
(x - [x])f'(x)
dx + f(1) - f(n).
termsto obtain: Now use the definition dn in (1) and rearrange of formula).For anyfunction form Euler's summation of Theorem2. (First-derivative a [1, with continuous derivative theinterval n] we have on f
k=111
Ef(k)
f(x) dx +
(x-
[x])f'(x)
dx +f(1).
(7)
the represent errormade when the sum on the The last two termson the right left is approximated the integralfnl f(x) dx. The formulais useful because f by need not be positive or decreasing.In fact, f can be increasingor oscillating. Variants of this formulawill be obtained as we attemptto deduce more precise about the error. information withperiod 1. If f' has a fixedsign function The factorx - [x] is a nonnegative the termin the errorhas the same signas (as it has when f is monotonic), integral f'. To decrease the errorit is preferable multiply f'(x) by a factorthatchanges to
1999]
AN ELEMENTARY VIEW OF EULER'S SUMMATION FORMULA
411
sign To takesplace in the integration. introduce signso thatsome cancellation the new the x changes, translate function - [x] downby 4 and consider we function - [xl - 4whosegraph shown Figure The integral x is in 3. term the in
4 x
[x]
- changessign.
(X -
[x])f'(x)
dx =
is the The lastterm equal to 4{f(n) - f(1)1.Usingthisin (7) we obtain following formula: form variant thefirst-derivative ofEuler'ssummation of
E f(k)
=
f
n
11n
dx.
f(x) dx + fx
[x]
2 )f(x)
dx + 2{ff(n) +f(1)}.
(8)
integration parts thesecond by in by Further variations be obtained repeated will of integral theright (8). on x The factor - [x] - 4has thevalue whenx is an integer. modify this We when a that at factor slightly makeitvanish theintegers, property is desirable to Bernoulli we integrate parts. do thiswe introduce To fiunction: by Pj(x), thefirst P1(x)
=x-
[x] -
if x + integer
by x if The error integral does notchange thefactor - [x]I- 4is replaced P1(x) as differ at theintegers. Therefore can be written (8) only becausethetwofactors
n
0O
ifx = integer.
k=1
, f(k)
J |f(x)
1
dx + |Pl(x)'(x)
1
dx +-{f(n) 2
+f(1)}.
(10)
Notethecontrast between and(3), which explicitly displays generalized the (10) that the we (3) Euler's constant C(f). To make(10) resemble moreclosely, assume Thenwe can write improper integral JlPl(x)f'(x)dx converges.
jPl(x)f'(x)
1 n
dx =
n
00
Pl(x)f'(x) ~~ ~~1
n
dx -
fPl(x)f'(x) n
00
dx,
E f(k)
(1)
412
[May
fPI(x)f'(x) fPl(x)f'(x)
00
dx
(12)
2f(n) 2
dx.
Eq. (11) has exactly same formas (3), but (11) is moregeneralbecause f is not the on of The onlyrestrictions f are continuity requiredto be positiveor monotonic. f' and convergence the improperintegral of
fPl(x)f'(x)
00
dx.
(13)
lim
fPl(x)f'(x)
00
dx = 0.
(14)
lim
-*oon
fIf'(x)I dx
= -
0.
(15)
To see this,note that the BernoullifunctionP1(x) is bounded; in fact,Figure 3 showsthat IP1(x)l < 1 forall x, so (14) followsfrom(15). Example. When f(x) = l/x we have f'(x)
oo G,,IIOol
1/X2
and
1
JIf'(x)IdxJ
nx
2d
dx.
4. FURTHER ANALYSIS OF THE ERROR TERM. Alternateformsof both the for can be obtainedby errortermand the formula the generalizedEuler's constant repeated integration parts. First we introducea new functionP2(x) whose by 2 derivative 2P1(x) at all noninteger is values of x. The factor is used so that P2(x) is the second Bernoulli periodic functionthat appears in Euler's summation we formula. Therefore requirethat P2(x)
=
2f Pl(t) dt + c,
(16)
where c is a constantto be specifiedlater. The functionP2 is quadratic on the interval 1]. In fact,P2(x) = X2 _ x + c if0 < x < 1. Its graphis a parabolic arc [0, the joiningthe points(0, c) and (1, c). Outside thisinterval graph(shownin Figure translations thisparabolicarc because P2 has period 1. of 4) consistsof horizontal To see this,we use the fact that P1 has period 1 and that JfP1(t) = 0, which dt [a, impliesthat fbP1(t) dt = 0 forany interval b] of length1. Therefore
P2(X + 1)
-
P2(X)
2f xPl(t)
dt = 0.
413
1999]
P2(x)
P2 Because of periodicity, has the constantvalue c = P2(0) at the integers. in by Integration partsshowsthatthe integral (10) is jPi(x)f'(x)
1
n
dx = -P2(0){f'(n)
~~~~22
-f'(1)}
- -
by Repeated integration partsleads to the general providedthatf" is continuous. of which involveshigherorder derivatives f formula, formof Euler's summation on that representpolynomials the and higherorder periodic Bernoullifunctions evolvein the process how the Bernoullifunctions [0, unitinterval 1]. To see exactly the sectionand integrate periodicfunction we followthe methodof the foregoing is P3(x) whose derivative 3P2(t) from0 to x to obtain anotherperiodicfunction P3(x) is periodicwithperiod 1 function 3P2(x). To guaranteethatthe integrated governs choice of the constantc in (16). the we need JoP2(t)dt = 0. This property x2 0 The integral the quadraticpolynomial - x + c from to 1 is equal to c - 6 of so we choose c = 6 and take
P2(X)
() =
+2fo Pl(t) dt 6
can now be restatedas follows: formula Euler's summation formula).For anyfuncof form Euler's summation Theorem3. (Second-derivative [1, on secondderivative theinterval n] we have a tionf with continuous
ii
k=1
Ef(k)=|f(x)
12
dx
n -P2(X)f"(x)
dx
1 + -{f(n) 2 +f(1)}. (17)
1 +-P2(0){f'(n) 2
-f'(1)}
E f (k)
1
ff(x)
dx + C(f) + Ef(n),
where 1
-
C(f)
and
_f(1)
_P2(0)f'(1)
1,.oo 2j P2(x)f"(x)
dx,
(18)
(x)dx.
(19)
414
[May
3fxP2(t)dt.
To improve error the the estimate integrate we P2(t) from to x and define 0 Bernoulli function P3(x) = 3foP2(t)dt so that P3(x) = 3P2(x). There is no need to add a constant this in case because, theunit on interval 1],P3(x) = x3 -2 [0, + 2x, and JJP3(t) = 0. The function has period1 becauseP2 has period1 dt P3 cubiccurve, shown in and f1P2(t) = 0. The graph P3 is a bounded of piecewise dt by over[1,n] Figure NotethatP3(x) vanishes theintegers. 5. at Integration parts
gives us
fP2(x)f"(x)
I
dx=
1n
f(3)is continuous. equation,together Theorem3, givesa provided This with integral on third-derivative ofEuler's form summation formula which second in the theright (17) is replaced changes in of by fnP3(x)f(3)(x) dx.The corresponding (18) and (19) are replacement the integrals of by IyflP3(x)f (3)(X) dx and
3!n
I3
of MATION FORMULA. The strategy obtaining generalversion Euler's for a in P1(x) in (9) we introduce, succession, periodic functions P2(x), P3(x),..., period1, and a sequence constants k suchthat of Bk
Pk(x)
=
5. BERNOULLI
FORM OF EULER'S
SUM-
(21)
Pk(x) is a Bk. As alreadynoted,on the closed interval 1] each function [0, polynomialof degree k when k = 2 or 3. [The case k = 1 is special; P1(x) is a at linear on polynomial - 2 only theopeninterval 1) andis discontinuous the x (0, It endpoints.] is clear(and easily proved induction) on theclosedinterval by that of [0,1] thefunction defined (20) is a polynomial degreek ifk ? 2. We denote by few for The this polynomial Bk(x), theusualnotation Bernoulli by polynomials. first are 1~ ~ 1 3 =2_X+1
Bl(x)=x-2 B2(x)=x2--x+
x
Periodicity impliesthat Pk(O) = Pk(l), and (21) showsthat each of these values is
B3(x)=x3
--X4
x2+
2x,
B4(X)
=X-X2x
X3+ X2
1
0
B5(x) =x5
x3
1 30x
1 3 = 6 x5 +-x4 B6(x) =x 2 5
_ x2 +-. 10 42
of The Bernoulli periodic functions periodicextensions these polynomials are given by Pk(x) = Bk(X - [x]). The constants = Pk() = Pk(l) are called Bernoulli Bk numbers. The first feware
B1 = B5 = O,
1
2'2
B2 1 42 '
1 6'
B B3 = 330B
B4 B8=
-_
1 1 30
B6 =-
= B7 7
,B=
of are Next we show that our definitions Bernoullinumbersand polynomials we consistent withthe usual definitions, provided take BO(x) = 1 and Bo = 1. Our in of are definition (20) showsthatthe successivederivatives these polynomials B'(x)
= kBk.l(x),
B"(x)
k(k
l)Bk
(x),.
..,
B( (x) = r!()Bk.(x),
and hence
r!(k)Bk .(0)=r!(k)Bk
(22)
On the otherhand, the Taylor expansionof any polynomialBk(x) of degree k is = = B(')(0)x"/r!, so (22) implies
Bk(X)
(k)BrX B
=
(23)
=
Pk(W)
Bk
for k ? 2, we findthat
Bk=
P E
k
)Bk-,
fork?2.
with This is the usual recursion formulafordefining Bernoullinumbers(starting in Bernoullipolynomials Bo = 1), and (23) is one of the standardwaysof defining terms of Bernoulli numbers.Consequently,the numbersand polynomialsthat and Bernoullipolynomiare appear in our treatment the usual Bernoullinumbers als thatappear in the literature; [1, p. 265], [2, p. 251], or [5, pp. 160-163]. see It is well knownthatthe BernoullinumbersBk withodd index k ? 3 are zero, so onlyBernoullinumbers witheven index appear in the generalformof Euler's summationformula.It is also known [8, p. 533] that on the interval[0,1] the Bernoullipolynomials satisfy following the inequalitiesfork ? 1:
IB2k(x)l < IB2kI and
IB2k?l(x)I
The method we have outlined leads to the followingodd-order derivative versionof Euler's summation A on formula. proofis easilygivenby induction the order2m + 1. Theorem4. (General formof Euler's summation formula).For anyfunction with f a continuous derivative order2m + 1 on theinterval n] we have of [1,
E f(k)
1
ff(x)
dx + (2m +
f
1i) l
+ I( xf
(2m?+l) (x)
dx (24)
+E
m B
A 1 (2rN
(f(2r-1)(n)
_f(2r-1)(1)}
1 -ff(1) +f(n)}.
f ?1) then integral If(2m7 (X) dxconverges wealso have if Moreover, theimproper
k=1
f(k)
ff(x)
1
dx + C(f) + Ef(n),
(25)
l))
dx
(26)
Ef(n)
1)(n)
1)!
P2,n?(X)f(2m+l)(X)
dc.
and (2m + 1)!/x2m1+2, (26) Example. When f(x) = l/x we have f(2n +1)(x) for expression the classical Euler's constant: givesthe following
C
1
2
1
=
+ +-
B2
-+
B4
++
B2 m
2m
;;-
dx.
(28)
28
4n Ef(n) -~~ 2n2 2rnn2~ , x'' (9 -> o in (28) and obtain an infinite series for Euler's One is temptedto let m in constant. However,the integral (28) does not tend to 0 as m -> ooand, in fact,it [see 6, p. 529], so rapidly series YB2k/(2k) diverges can be shownthatthe infinite as is not veryuseful for calculatingC. Nevertheless, we show in the next (28) section,(25) and (27) can be used to calculate C veryaccurately.
.
2n
B4
42
12 + B21 P2tn+2(I)
2m2dx.
29
6. CALCULATION OF EULER'S CONSTANT. We use Euler's summationforTake f(x) = l/x in (25) and 7 mula to calculatethe first digitsin Euler's constant. it rewrite as 'I1 (30) C= Ejk-logn-Ef(n),
k=l
2 2n2
B2
1
4n4
1
B4
6n6 +1dx
1
B6
Jn x8
+
c0 rP7(X)
2n
12n2
00 P7( X)
120n4
6,
252n6
we get
0oci
L
1
71
GO P7(X) x8
L
1999]
7(8
dx <
-42
~-dx=
417
k-i
,-
- log k g
2n
12n2
~~-
120n4
252n6
+ E(n),
(31)
where 0 < IE(n) I < 1/42n7. Using a hand calculator that displays 12 digitswe findSl_1k - = 2.92896825381and log 10 = 2.30258509299.If n = 10 the sum of in the errorterm E(n) plus the termwith252n6 in the denominator (31) is too small to influence the seventhdigit.Neglecting these termsand retaining digits 8 in the calculationwe find C - 2.92896825 - 2.30258509 =
1 1 1 - + 1200 - 1200000
0.62638316 - 0.05000000 + 0.00083333 - 0.00000083 = 0.57721566 This calculation,using m = 3 and n = 10 in (29) and (30), which guarantees 7 decimal places, actuallygivesthe first correctdigitsof C. Knuth[7] used (29) 8 and (30) with m = 250 and n = 10,000 to calculate the value of C to 1,271 decimal places. This note outlinesonlyone applicationof Euler's summation formula.Others can be found in Knopp's treatise[6]. One of them uses the increasingfunction formula the logarithm n!. Euler's for of asymptotic f(x) = log x to deriveStirling's summationformulaand its relationto Bernoulli numbersand polynomials provides a treasuretroveof interesting materialsuitable for elementary enrichment calculus courses.
REFERENCES 1. Tomn Apostol, Introduction Analytic M. to NumberTheory. New York, 1976. Springer-Verlag, 2. Tom M. Apostol, MathematicalAnalysis, Second edition.Addison-Wesley, Reading,Mass., 1974. 3. L. Euler, Methodusuniversalis serierum convergentium summasquam proxime Commeninveniendi, tarii academie scientiarumPetropolitanae,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 Its History. by Springer-Verlag, New York, 1996. 6. K. Knopp, Theory and Application Infinite of Series,R. C. Young, translator. Hafner,New York, 1951. to v. 7. D. E. Knuth,Euler'sconstant 1271places, Math. of Computation, 16 (1962), pp. 275-281. Amer. Math. Monthly, 47 v. 8. D. H. Lehmer,On themaximaand minimaof Bernoulli polynomials, (1940), pp. 533-538. 9. Colin Maclaurin,A Treatise Fluxions.Edinburgh, of 1742. TOM M. APOSTOL receivedhis Ph.D. in 1948 witha thesisin analyticnumbertheory written under the directionof D. H. Lehmer at UC Berkeley.He joined the Caltech faculty 1950 and became in professoremeritusin 1992. His list of publicationscontains 58 research papers and several books, includinghis pathbreakingCalculus in two volumes,firstpublished in 1961, Mathematical Analysis to (1957), and Introduction Analytic NumberTheory (1976), all of whichare stillin print.He is director of ProjectMATHEMATICS!, a prize winningseries of videos and other educational activitieshe is initiatedtwelveyears ago. His 50-yearcareer in mathematics described in an engagingarticle by withTom Apostol,published in the September 1997 issue of The College Don Albers, An Interview Mathematics Journal. Project MATHEMATICS!, 1-70 Caltech, Pasadena, CA 91125 [email protected]
418
[May