Sums of Powers of Integers

Author(s): B. L. Burrows and R. F. Talbot

Source: The American Mathematical Monthly , Aug. - Sep., 1984, Vol. 91, No. 7 (Aug. -
Sep., 1984), pp. 394-403
Published by: Taylor & Francis, Ltd. on behalf of the Mathematical Association of
America

Stable URL: http://www.jstor.com/stable/2322985

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. and Mathematical Association of America are collaborating with JSTOR
to digitize, preserve and extend access to The American Mathematical Monthly

This content downloaded from

194.117.40.104 on Sun, 14 Jun 2020 18:44:03 UTC
All use subject to https://about.jstor.org/terms
 SUMS OF POWERS OF INTEGERS

B. L. BURROWS AND R. F. TALBOT

Department of Mathematics, North Staffordshire Polytechnic, Beaconside, Stafford, England ST] 8 O

1. Introduction. The Bernoulli family produced several mathematicians, the first of whom was
Jacob (1654-1705). He was the son of Nicolaus, a merchant in Basle, and it is reported [1] that
Jacob had the motto 'Invito patre sidera verso' (against my father's will I study the stars). Despite
his father's opposition, Jacob devoted his life to the study of mathematics and astronomy. His
most famous work is 'Ars Conjectandi,' which was published after his death, in 1713, and which
reflected his interest in probability. It was in this book that the famous Bernoulli numbers were
introduced in connection with finding sums of powers of integers. Jacob was very enthusiastic
about the technique that he had developed and in comparing his calculations with those of
Bullialdus (1605-1694) he states [2] 'with the help of this table it took me less than half of a
quarter of an hour to find that the tenth powers of the first 1000 numbers being added together
will yield the sum

91,409,924,241,424,243,424,241,924,242,500.'
In this article we show that a simple formula will give highly accurate answers to such sums and
with the help of a very crude electronic calculator we can obtain, in 9 seconds,
1000

L r10 9.14104 X 1031

r=1

which is in error by 0.0005%!

The approximation used was

(1.1) E rk' (n?2)k+1

r=1

and the idea stems from another observation of Bernoulli's that the sums can be expressed in the
form p([n(n + 1)]) when k is odd and (2n + 1)p([n(n + 1)]) when k is even, where p(x) is a
polynomial [3]. We refine the approximation (1.1) and extend the results to the case where
any real number.

2. E Ir k (k is a Positive Integer). It is well known that this sum can be expressed as a

polynomial in n of degree k + 1. Bernoulli's observation mentioned earlier suggests that there is
some advantage to be gained by considering the equivalent polynomial in n + 4 which is also of
degree k + 1.
We shall use the method of undetermined coefficients to find this polynomial which we shall
denote by Sk+l(n + 4). Thus
n

(2.1) E rk= Sk+ n + -2-

r==12

Brian L. Burrows: I obtained my B.Sc. from the University of Surrey in 1968 and my doctorate from
Nottingham University in 1971 for research into quantum mechanics. Since then I have published work in numerical
analysis, variational methods, combinatorics and quantum theory. My nonmathematical interests include campanol-
ogy, avoiding administration, and being beaten at squash by Dr. Talbot.

Richard F. Talbot: I was awarded my B.Sc. by Nottingham University in 1966 and I obtained my doctorate from
the same university in 1969 for research in the area of fluid mechanics. Since that time my mathematical interests
have diversified into combinatorics and more recently numerical analysis. Apart from my family my main
nonmathematical interests are local history, sport in general including cricket (similar to baseball but more
intellectual!), and squash.

394

This content downloaded from

194.117.40.104 on Sun, 14 Jun 2020 18:44:03 UTC
All use subject to https://about.jstor.org/terms
 SUMS OF POWERS OF INTEGERS 395

where
/ 1\ ~k+1

(2.2) Sk+i(ln+ 2= Eat(n +

From (2.1) it follows that

n-I/

(2.3) E = Sk+( n

and so equations (2.1) and (2.3) give

(2.4) Sk+1( n + )Sk+1 n -)- nk.

Now, using (k) to denote the familiar binomial coefficien

(2.5) n( 2) ( _ 2) ,E (y n jl+-)
so that the right-hand side is either an even or an odd polynomi
rewritten
k-i1 t

(2.6) F, at (J)(2 n t (I +( 1+) = n k

t=i =

Notice that a0 will have to be determ

can now be found by equating in
following results:

ak+ +
(2.7) 1) 1
ak+(1 = 1,

a(k) = 0,

(2.7) ak*1( (2)+a- 1) ?

ak(3 )( )+ ak2( ) = 0,

and generally

ak+(k + 1)(1 )P+(1 +(-1)p+2) +

(a- k - ( P-1 (I (- 1 )) + *+
We first observe that the second of the
ak-2, ak-4, ... are all zero. From the first of
akl,ak-3,... in turn, we have

1
ak+i = k + 1'
ak+l =-

( (k
(2.8) ak-3=ona

This content downloaded from

194.117.40.104 on Sun, 14 Jun 2020 18:44:03 UTC
All use subject to https://about.jstor.org/terms
 396 B. L. BURROWS AND R. F. TALBOT [August-September

ak-5=-31 (k)
k=2~ 7-63 kY
127 k
ak7 2 7*240 \7
Thus, for example, when k = 3 we obtain

L r3 = s4=n + 2 n + - + n + ao.
r=1 2 S( 4) 2( 8[
When n = 1, S4(3) = 1 and so ao = 1
Therefore,
/ 1\ 1/I 1\~2 2

(2.9) S4( + ) 4{(n + -

= -n2(n + 1)2,
4

this final form being the one that is well known.

The results above can perhaps be better appreciated from a computational point of view.
Writing X for N + 2 and S(k) for y.r k, then we have shown that

(2.10) S(k) = ak+lXk+l + ak-lXkl1 + ak-3Xk-3 +

where the coefficients are given by (2.8) and the last term is ao if k is odd a
Thus when k > 4, we may rewrite (2.10) in the form

(2.11) S(k) = k+ 1k (1_(2 )+ (__ 4)_

+ 1 12 X2 240 X4

From (2.11) it is clear that good results may be obtained from the

(2.12) S(k) Xk k + i
+1.

We may illustrate these results by considering E35= r5 and, comparing the exact value of
333,263,700 with just two approximations that can be obtained from (2.10) or equivalently (2.11).
The first approximation, expressed in (2.12), gives

L r5 _3__ = 333,594,489,
r=1l

which is in error by about 0.1%.

The second approximation, using just th

r5 6 - 35 54 = 333,263,608

which is in error by 92 or about 0.0000003%.

The natural extension of this work is to the case where k is not

3. The General Case. In this section we generalize the previous re

all real k deriving a result that extends (2.11). In particular we sh
tion obtained in (2.12) provides surprisingly accurate estimates w
The formulae we will derive for k not a positive integer result
methods of asymptotic analysis to sums of powers of integers. A
appears in [4]; in particular equation 3.4.6 of [4] gives a result for

This content downloaded from

194.117.40.104 on Sun, 14 Jun 2020 18:44:03 UTC
All use subject to https://about.jstor.org/terms
 1984] SUMS OF POWERS OF INTEGERS 397

powers of n and the Riemann Zeta function.

Consider the Taylor series expansion of f(x) = xk about the integer x = r.

(3.1) f(x) = f(r) +(x - r)f'(r) + (-2! f"(r) +

+ (x _r) fq( ) + (x r 1(Zr)'

where Zr = r + Or(x - r) and 0 < Or < 1.

Using the notation

(3.2) (k) k(k- l) (k- r+ 1)

for non-integer k and integer r, we may rewrite (3.1) i
q to be even without loss of generality.

(3.3) xk = rk +( )rk1(X - r) +(2)rk2(X - r)2 +

1
+( kq r
Integrating from r - 2 to r + 2 then gives

(3.) fr (i kdX k(k)kr (1)3(k)k4r2(1)5

(3.4) xkdx 1/2+ 2 ) k 3 (2 ) k4 5 (2)+
Summing equation (3.4) from r = 1 to n yields

(3.5) g(k) = r+ k 2 ) Er + ( 4 E r k +
where

g(k) =k 1(n + 2) -2 (k 0 - 1)

and g(-1) = log(2n + 1). Write S(k) = En.4rk. Then equatio

(3.6) g(k)=S(k) + k S(k- 2)+ k S(k- 4)+

+( k(q I 1)2qS(k - q) + Rq+

where

R ( =q
Rq+1 ~k f+
) r+ (1/2) k--( )q+ldX
)f?(1/2)Zir1(x -r)?d
and Zr iS given in (3.1).
Thus for q > k - 1,
k n r+ p 1\2) k-q-1
JRq+1I < (qf ,i r -~ - jx - rjq?1 dx,
q + I r=1 - (1/2) 2,
that is,

(3.7) (q< k - r 1 2()

IR+i q + I)r=1( 2 q +2

This content downloaded from

194.117.40.104 on Sun, 14 Jun 2020 18:44:03 UTC
All use subject to https://about.jstor.org/terms
398 B. L. BURROWS AND R. F. TALBOT [August-September

and so

jRq+ij <

From this we can easily deduce that for fixed k and n, lirn q4 R q + 1=0 so that the series
(3.6) converges.
Putting k, k - 2, k - 4,... in turn in equation (3.6) gives the set of equations
(3.8)

g(k) = S(k)?( k S(k -2) +(k)IS(k -4) +(k SI 6

\213.-2 2415 .2 4 \617 26(

g(k -2) = S(k -2) + ( )S(k -4)?(k2) 1S(k -6) +..

k213 .22241 5 24

g(k -4) = S(k -4) +(k4)l2 S(k -6) s..

g(k -6) = S(k -6) +

These equations can be

(3.9) g =MS,
where

g(k) 1 (~3X-22 (~5-24 (6I7-2 6

g(k -2) 0 1 (k -2)
g= g(k-4) ,M= 0 0 1 2 (k4 )l2

0 0 0 1..
and

We initially solve the p

and columns. In a simil
It is worth pointing ou
given I then this M71'
case 1 = 4, then we obta

1 2( 3 1 22 (4)1i2 4 3 126

(3.10) 0 k2)3l (k4 52)7

0 0 1 (k- 4)

0 0 0 1

This content downloaded from

194.117.40.104 on Sun, 14 Jun 2020 18:44:03 UTC
All use subject to https://about.jstor.org/terms
 1984] SUMS OF POWERS OF INTEGERS 399

so that

(3.11) S4 = MV9g4.
From this, we can obtain a 4-term approximation to S(k) given by

(3.12) S4(k) = g(k) -() 12g(k - 2) + () 7 4g(k - 4) -() 31 6g(k - 6).

Equation (3.12) agrees with the previous results (equations (2.8) and (2.10)) obtained when k was
a positive integer.
Again the numerical accuracy of the result may be demonstrated by taking the first few terms
as in the next example.

Now, E35 jr25 = 76138.722369 and the table below gives the successive approximations for
various values of 1.

1 Approximation
1 76160.721097
2 76138.725044
3 76138.722204
4 76138.722459

Note that the approximation 35.535/3.5 which is analogous to (2.12) is in error by

The matrix technique used in this section can be applied in the case where k is a positive
integer.
M will now be of finite order m and the previous analysis will still apply provided that the
restricted order 1 is less than m.

4. Modifications when k is Negative. In practice, when k is negative, the accuracy of (3.12) is

improved by starting the summation at r = p. This is equivalent to redefining S(k) by
n

(4.1) S(k) = E rk,

r=p

with the corresponding modifications

(4.2) g(k (( +)+ - p k k+

and

(4.3) g(-1) = log(n + - log( P

For example, in the case k = -1, we have the explicit form

(4.4) S: - = log(n + - log(p - + n + V (p-

- 960 + 2) -(p - 2 + 632 n + 2) (p - )

Thus for p = 3 and n = 50, we have used the inverse matrix technique described previously to
compute a 1-term approximation of 3.0056826 and a 5-term approximation of 2.9992059. This can
be compared with the exact result of 2.9992053 to 7 decimal places.
It is also interesting to compare the well-known approximation
n r

(4.5) E logn + -y,

r=l

This content downloaded from

194.117.40.104 on Sun, 14 Jun 2020 18:44:03 UTC
All use subject to https://about.jstor.org/terms
400 B. L. BURROWS AND R. F. TALBOT [August-September

where y = 0.57721566... is Euler's constant, with our first approximation

(4.6) r1 I + log { }
(4.6) ~~~~~~n p- 1 n + 2

r r=r

and our second approximation

(4.7) 1 -+ {log + 2 (( 2 ( 2
r=1 r r=1 r1 2

For n = 50, equation (4.5) leads to the approximation 4.48924 compared with the exact value of
4.499205, an error of .22%. The following table compares this with the approximations (4.6) and
(4.7) for various values of p.

Approximation p = 2 p = 3 p = 4
(4.6) 4.51651 4.40468 4.50254
(4.7) 4.49801 4.49903 4.49916

So far we have only considered examples of finite series of the f

the sum to infinity does not exist. However, equations (3.12) and (4
ZO=Irk when k < -1. In this case we may let n -s o and obtain
00 Pl1k1 -

(4.8) E rk= E rk p + k )p _
r=1 r=1 k 1 22 1 k 1 2

4 240 k - 3 -2 ) +(6) 1344 k - 5 (-2) -

Equation (4.8) can be written

00 P-1 1 /l\k+I k, ik-1 i 1k-3

(4.9) rk r r~- I - + - -( - -I'-i'
)rEik rE1r k + 1 (P 2) 24(P 2 3960 2

5 8064 ( 2 - 2) -
We can easily adapt the proof of convergence of (3.6) to the case where
course k < - 1). From a modified form of (3.7) we have

lRqI < (k (r -
2q 1) q 2r=p /

Since k < - 1, the infinite sum is convergent so that

lim Rq= 0.
q -o0

Computations using the terms given in (4.9) have been carried out for the infinite series where
k = -2 and k = - 3.5 with various values of p. The results are shown below.

k p = 2 p = 3 p = 4 Exact
- 2 1.644466 1.644928 1.644934 1.644934
- 3.5 1.124871 1.126723 1.126733 1.126730

This content downloaded from

194.117.40.104 on Sun, 14 Jun 2020 18:44:03 UTC
All use subject to https://about.jstor.org/terms
 1984] SUMS OF POWERS OF INTEGERS 401

5. Convergence. In the previous sec

terms of the series
00

(5.1) E m1g(k - 21),

1=0

where ml is the Ith entry in the first row of Ml71 (see (3.9)-(3.12)) and g(k) is given
In fact (5.1) converges only when k is a positive integer and in this case the series breaks off
after finitely many terms.
In all other cases the series (5.1) diverges and we now show that its terms eventually have
increasing modulus.
We first note that from (4.2)

g(k - 21) = - (n + 1 )k + ( - )2) )}

so that for sufficiently large,

(5.2) Ig(k - 21)1 >1k - 21+ 11 21k1 forp ?1.

The matrix Ml can be written in the form

(5.3)
2 1-2
1 a(k) X1a(k)a(k - 2) X2 17 a(k - 2i) ... X1-2 T a(k - 2i)

1- 2

0 1 a(k-2) X1a(k-2)a(k-4) ... X-3 H a(k-2i)

Mla(k-4)
0 0 1 ~~~~~~~~~~~~~~~~~~1- 2
... Xl14 Y a(k-2i)

a(k - 21 + 4)
0 0 0 0 1

where

a(k) =(213 222 and Xr (2r+ 3)!

Consider the equation

(5.4) Mix= e,
where e is the Ith column of t
Cramer's Rule and noting that
(5.5)
2 1-2
0 a(k) Xla(k)a(k - 2) 2 171a(k - 2i) ... X2 U a(k - 2i)
1- 2

0 1 a(k-2) X1a(k-2)a(k-4) .. X3Yrl1a(k-22i)

0 0 a(k- 21+ 4)
1 0 1

This content downloaded from

194.117.40.104 on Sun, 14 Jun 2020 18:44:03 UTC
All use subject to https://about.jstor.org/terms
402 B. L. BURROWS AND R. F. TALBOT [August-September

2 1-2

a(k) Xla(k)a(k - 2) X2 H a(k - 2i) ... X-2 El a(k - 2i)

i=O i=0

1 a(k-2) Xla(k-2)a(k - 4) ...

0 1 a(k-4) ...

0 0 0 a(k-21+4)

Use induction on I to prove that

(5.6) m= (-1)'+la(k)a(k - 2) * a(k - 21 + 4)A,

where

Ix 2 ... 1I-2
1 1 xi1 ... X1-3
0 1 1
(5.7) AI= 0
: ; 1 X1

0 0 ... 1

The value of A, can be found by successi

subtracting a multiple of the previous row. This produces a triangular determinant given by

1 X1 X -2
0 f,
(5.8) A,= 0 f2

0 0 ... f1-2

whose value is f112 1.. f1-2- Increasing / does not change th

example

(5 -9) Al+ I = flf2 .. fl t-2 fl-l1-

A numerical calculation of the sequence { fi } shows that we h

sequence tending rapidly to a limit. The table below gives various v

i fi
1 0.7
2 0.63265306
13 0.60792713
14 0.60792711
15 0.60792710
16 0.60792710

Defining this limit to be 1/3, we see that A, 0 as I oo. However, from t

can conclude that

(5.10) A (A)

This content downloaded from

194.117.40.104 on Sun, 14 Jun 2020 18:44:03 UTC
All use subject to https://about.jstor.org/terms
1984] SUMS OF POWERS OF INTEGERS 403

Now H-I-2a(k - 2i) can be written as

(5.11) k(k-1) ... (k-21+

241-1
3)
Thus combining the results of this section, we have shown that the general term of (5.1) satisfies

(5.12) m,g(k - 21)j >k(k - 1) ... (k - 2/ + 3)1 p k- 21+ 1

(5.12) lmlg(k - 21)1 > (24/3)1-1 k - 21 + 1
Since p and k are fixed, (5.11) shows that eventually the terms of the series (5.1) increase in
magnitude. This shows that these approximations will always get worse for sufficiently large 1.
The reason why the estimates given in this paper are good can be seen by regarding equations
(3.8) truncated in the form

(5.13) MIS/ = g, + ?1,

where SI is the first I values of the exact solution
the first I equations (see (3.7)). The exact expression for S(k) is then a finite sum of terms in the
form

(5.14) m,(g(k - 2t) + t),g

where m, are the elements of e/. Taking the partic

(5.15) = 7k(k - 1)(k - 2)(k - 3)

M3 ~4! .15 -24
Since % is small and for positive k the sums are of the order (n + k)k+1/(k + 1), then for large
n the relative error

(5.16) Im3A3I/S( k)
is small. For negative k (particularly k < -1) the approximations are formed by using p > 1.
Since X93 decreases with p, then it is always possible to reduce the size of ImA331. These argume
hold for any t with sufficiently large n or p so that it is always possible to obtain good estimates
from the initial terms of (5.1). This gives a heuristic argument why we have been able to obtain
surprisingly good approximations to S(k) from a series of terms (5.1) that actually diverges!

References

1. E. T. Bell, Men of Mathematics, Penguin, New York, 1937.

2. D. E. Smith, A Source Book in Mathematics, McGraw-Hill, New York, 1929.
3. A. W. F. Edwards, Sums of powers of integers: a little of the history, Math. Gaz. (1982) 22-28.
4. N. G. de Brujn, Asymptotic Methods in Analysis, North-Holland, Amsterdam, 1958.

130. MISCELLANEA

These forms were like the cunning tables used by mathematicians, which may be entered from
top, bottom, right, and left, which entrances consist of scores of lines and dozens of columns, and
from which may be drawn, without reasoning or thinking, thousands of different conclusions, all
unchallengeably precise and true.

-Martin Eden, by Jack London,

Macmillan, New York, 1973, p. 226.

This content downloaded from

194.117.40.104 on Sun, 14 Jun 2020 18:44:03 UTC
All use subject to https://about.jstor.org/terms