Bernoulli Polynomials

Talks given at LSBU, October and November 2015

Tony Forbes
Bernoulli Polynomials
The Bernoulli polynomials Bn (x) are defined by

B0 (x) = 1,
Z 1
Bn0 (x) = nBn−1 (x) and Bn (x) dx = 0, n ≥ 1. (1)
0
Thus
1 1
B1 (x) = x − , B2 (x) = x2 − x + ,
2 6
2
3x x 1
B3 (x) = x3 − + , B4 (x) = x4 − 2x3 + x2 − ,
2 2 30
5x4 5x3 x 5x4 x2 1
B5 (x) = x5 − + − , B6 (x) = x6 − 3x5 + − + ,
2 3 6 2 2 42
6 5 3 6 4
7x 7x 7x x 14x 7x 2x2 1
B7 (x) = x7 − + − + , B8 (x) = x8 − 4x7 + − + − ,
2 2 6 6 3 3 3 30
8 5 8 2
9x 21x 3x 15x 3x 5
B9 (x) = x9 − +6x7 − +2x3 − , B10 (x) = x10 −5x9 + −7x6 +5x4 − + ,
2 5 10 2 2 66
11x10 55x9 11x3 5x
B11 (x) = x11 − + − 11x7 + 11x5 − + ,
2 6 2 6
8 4
33x 33x 691
B12 (x) = x12 − 6x11 + 11x10 − + 22x6 − + 5x2 − , ....
2 2 2730
The constant term gives the nth Bernoulli number, Bn = Bn (0),
1 1 1 1 1
B0 = 1, B1 = − , B2 = , B3 = 0, B4 = − , B5 = 0, B6 = , B7 = 0, B8 = − ,
2 6 30 42 30
5 691 7 3617
B9 = 0, B10 = , B11 = 0, B12 = − , B13 = 0, B14 = , B15 = 0, B16 = − ,
66 2730 6 510
43867 174611
B17 = 0, B18 = , B19 = 0, B20 = − , ....
798 330
There is some merit in defining what one might call the alternative Bernoulli polynomials:

B
b0 (x) = 1,
Z 1
bn0 (x) = nB
B bn−1 (x) and B
bn (x) dx = 1, n ≥ 1.
0

bn (x) = Bn (x) + nxn−1 = (−1)n Bn (−x). The corresponding alternative Bernoulli

Then B
numbers are B
bn = Bbn (0) = Bn (1). The only difference occurs when n = 1:

b1 = 1 = − B1 ,
B B
bn = Bn , n = 0, 2, 3, 4, . . . .
2

We won’t investigate the alternative polynomials, but watch out for B

bn in what follows.

1
The graphs show that Bn (1 − x) = (−1)n Bn (x). From the definition, the
coefficient of x
0 m n
is Bn (0) = nBn−1 . More generally, the coefficient of x in Bn (x) is m Bn−m ; hence
n  
X n
Bn (x) = Bn−m xm .
m=0
m

Moreover, for x ∈ [0, 1] we have these approximations for large n:

Bn (x) ≈ Bn cos 2πx, n even, Bn (x) ≈ Bn (1/4) sin 2πx, n odd,
and Z 1
Bn (x)Bm (x) dx = 0 when n + m is odd.
0
R1
Exercise for reader: compute 0
Bn (x)Bm (x) dx when n + m is even.

0.15

0.10

0.05

0.2 0.4 0.6 0.8 1.0

-0.05

Bernoulli polynomials Bn (x) for n = 2, 3, . . . , 8

The generating functions for the Bernoulli polynomials and the Bernoulli numbers are
respectively
∞ ∞
text X tn t X tn
= Bn (x) and = Bn . (2)
et − 1 n=0
n! et − 1 n=0
n!
These could be used as definitions; then the left-hand equality in (1) is recovered by
differentiating the left-hand equality in (2) with respect to x:
∞ ∞ 0 ∞
t2 ext X
0 tn text X
0 tn−1 X Bn+1 (x) tn
= B (x) ⇒ = B (x) = · .
ex − 1 n=1 n n! ex − 1 n=1 n n! n=0
n+1 n!
From the right-hand equality in (2) we obtain two facts: (i) B2n+1 = 0 for n ≥ 1 since the
function t/(et − 1) + t/2 is even, and (ii) multiplying the series by the Taylor expansion
of (et − 1)/t gives
n  
X n+1
Bk = 0n .
k=0
k
This last equality can be used to compute the Bn recursively; see Lovelace [3], for example.

2
Stirling numbers of the second kind
The Stirling number of the second kind S(n, k) is the number of ways to partition a set
of n objects into k non-empty subsets:
k k
(−1)k X
   
1 X k−j k n j k
S(n, k) = (−1) j = (−1) j n,
k! j=0 j k! j=0 j

or they can be defined recursively

S(0, 0) = 1, S(n, 0) = 0 for n > 0, S(0, k) = 0 for k > 0,

S(n + 1, k) = kS(n, k) + S(n, k − 1) for k > 0.

k
n 0 1 2 3 4 5 6 7 8 9 10
0 1 0 0 0 0 0 0 0 0 0 0
1 0 1 0 0 0 0 0 0 0 0 0
2 0 1 1 0 0 0 0 0 0 0 0
3 0 1 3 1 0 0 0 0 0 0 0
4 0 1 7 6 1 0 0 0 0 0 0
5 0 1 15 25 10 1 0 0 0 0 0
6 0 1 31 90 65 15 1 0 0 0 0
7 0 1 63 301 350 140 21 1 0 0 0
8 0 1 127 966 1701 1050 266 28 1 0 0
9 0 1 255 3025 7770 6951 2646 462 36 1 0
10 0 1 511 9330 34105 42525 22827 5880 750 45 1

It is clear from the recursive definition that the S(n, k) are integers, P
that S(n, n) =

1 and
that S(n, k) = 0 if k > n. Interesting problem: prove directly that kj=0 (−1)j kj j n = 0
whenever k > n > 0.

1000

800

600

400

200

k
50 100 150 200 250 300
Plot of log S(n, k) + 1 for k = 5, 10, . . . and n = 1, 2, . . . , 300

3
Euler–Maclaurin summation
Let f (x) be a sufficiently well-behaved function. Suppose we want to sum f (x) over
integer values of x and that we know how to integrate f (x). Let m < n be integers. Then
n Z n r bk Z n br (bxc − x)
X X
(k−1) (k−1)

B B
f (r) = f (x) dx + f (n) − f (m) − f (r) (x) dx,
k=m+1 m k=1
k! m r!

or, making use of the fact that Bn = 0 for odd n ≥ 3,

n Z n s
X f (n) − f (m) X (2k−1)
B2k
f (k) = f (x) dx + + f (n) − f (2k−1) (m)
k=m+1 m 2 k=1
(2k)!
Z n
B2r (x − bxc)
− f (2r) (x) dx, (r ≥ 1).
m (2r)!
If f (x) is a polynomial, the last term will vanish for sufficiently large r. For example, to
get the formula for the sum of the first n squares put m = 0 and f (x) = x2 :
n Z n
X
2 n2 B2 n3 n2 n n(2n + 1)(n + 1)
k = x2 dx + + 2n = + + = .
k=1 0 2 2! 3 2 6 6

Poly-Bernoulli numbers
(k)
If we define the more general poly-Bernoulli numbers Bn by
∞ ∞
X tm Lik (1 − e−t ) X
(k) t
n
Lik (t) = , = Bn ,
m=1
m k 1 − e−t n=0
n!
(1)
then k = 1 gives Li1 (x) = − log(1−x) and the generating function t/(1−e−t ) of Bn = B
bn .

Properties of the Bernoulli numbers

Theorem 1 (Explicit formula for the Bernoulli numbers) We have
n k n
(−1)k
 
X 1 X j k n
X
Bn = (−1) j = k! S(n, k). (3)
k=0
k + 1 j=0 j k=0
k + 1

Proof We refer to the generating function:

∞ ∞
x log(1 + (ex − 1)) 1 X x
k+1 (e − 1)
k X x
k (e − 1)
k
= = (−1) = (−1) .
ex − 1 ex − 1 ex − 1 k=1 k k=0
k + 1

Now expand the binomial (ex − 1)k and then expand ex :

∞ k ∞ k ∞
(−1)k X (−1)k X j n xn
   X
x X
k−j k jx
X
k−j k
= (−1) e = (−1)
ex − 1 k=0
k + 1 j=0 j k=0
k + 1 j=0 j n=0 n!
∞ ∞ k
xn X (−1)k X
 
k−j k
X
= (−1) jn
n=0
n! k=0
k + 1 j=0
j
∞ ∞
X xn X (−1)k
= k!S(n, k).
n=0
n! k=0
k+1

4
But S(n, k) = 0 when k > n. Hence
∞ n
x X xn X (−1)k
= k!S(n, k),
ex − 1 n=0
n! k=0 k + 1

as required to get the right-hand expression in (3). The other expression follows by
substituting for S(n, k). 

Theorem 2 (von Staudt – Clausen) If n is a positive even integer, then

X 1
Bn + ≡ 0 (mod 1). (4)
p
p−1|n, p prime

Proof Suppose n ≥ 2 is even. We consider each of the terms in the sum over k in either
of the two expressions of (3).
If k + 1 is not prime and k > 3, then k!/(k + 1) is an integer, as also is S(n, k); so the
term corresponding to this k is an integer. Also one can verify directly that the terms
corresponding to k = 0 and k = 3 are integers.
Now suppose k + 1 is prime and k does not divide n. Write n = qk + r with 0 < r < k.
qk
Pk for 0j <k
j r≤ k, j ≡ 1 (mod k + 1). Hence modulo k + 1 the sum over j becomes
Then
j=0 (−1) j j = ±k!S(r, k) = 0 since k > r.
Finally, suppose k + 1 is
Pprime and k
divides n. Then for 0 < j ≤ k, j n ≡ 1 (mod k + 1),
k
and 0 = 0. Therefore j=0 (−1)j kj j n ≡ −1 (mod k + 1) and hence modulo 1 there is a
n

contribution of −1/(k + 1) to the sum. 

When n is odd the sum is zero modulo 1; the k = 1 term contributes 1/2, as before, but
now the k = 3 term also contributes 1/2. All other k terms are integers.
The Von Staudt–Clausen theorem allows you to compute the fractional part of Bn for
some quite large n. For example, if n = 28000000000 , the divisors of n are powers of two and
the denominators of the fractions in (4) are just 2 and the Fermat primes, 3, 5, 17, 257 and
65537. On the other hand, the calculation is currently impossible for, say, n = 29000000000 .
It is convenient to have an expression for the absolute value of Bn . So we rearrange (4)
to get, writing {x} for x − bxc,
 nP o


 p−1|n, p prime 1/p , n > 0, n ≡ 0 (mod 4),

 n P o
 1−
p−1|n, p prime 1/p , n ≡ 2 (mod 4),

|Bn | = b|Bn |c + (5)

 0, n = 0,
1/2, n = 1,




0, n > 1, n odd.


Thus for n = 2, 4, . . . , we get

1 1 1 1 5 691 1 47 775 41 17 691 1 59 12899
, , , , , , , , , , , , , , ,
6 30 42 30 66 2730 6 510 798 330 138 2730 6 870 14322
47 1 638653 1 2011 1 53 41 14477 5 83 775 59 53 22298681
, , , , , , , , , , , , , , ,
510 6 1919190 6 13530 1806 690 282 46410 66 1590 798 870 354 56786730
1 47 62483 1 289 48540859 1 1 37 47717
, , , , , , , , , ,
6 510 64722 30 4686 140100870 6 30 3318 230010
77 1058237 1 5407 230759 77 1 1450679 1 4471
, , , , , , , , ,
498 3404310 6 61410 272118 1410 6 4501770 6 33330
5
Theorem 3 The Bernoulli numbers are related to the Riemann zeta-function by

bn = − nζ(1 − n), n ≥ 0; 2 n!
B Bn = − ζ(n), even n ≥ 0. (6)
(2πi)n

Proof (Titchmarsh [6, section 2.4]) Let s be a complex veriable with Re s > 1. Using
Z ∞
Γ(s)
xs−1 e−kx dx =
0 ks
we get
∞ Z ∞ ∞ ∞ ∞
xs−1
X Z X Z
s−1 −ks s−1 −kx
Γ(s)ζ(s) = x e dx = x e dx = dx.
k=1 0 0 k=1 0 ex − 1

Let Cρ , ρ > 0, denote the contour that goes from +∞ to ρ on the real axis, circles the
origin once anticlockwise and then returns to +∞. Let ρ → 0. Since the integral along
the circle tends to zero as ρ tends to zero, we have
Z 0 s−1 Z ∞ 2πi s−1
z s−1
Z
z (e z)
z
dz = z
dz + z −1
dz = (e2πis − 1)Γ(s)ζ(s).
C0 e − 1 ∞ e − 1 0 e

Using the well-known formula Γ(1 − z)Γ(z) = π/(sin πz), we obtain

z s−1 e−πis Γ(1 − s) z s−1

Z Z
1
ζ(s) = dz = dz,
(e2πis − 1)Γ(s) C0 ez − 1 2πi z
C0 e − 1

which extends the domain of ζ(s) to the whole complex plane except for s = 1. Now let
n ≥ 1 be an integer and put s = 1 − n. Then
∞
(−1)n−1 (n − 1)! z −n (−1)n−1 (n − 1)!
Z Z X
Bk k−n−1
ζ(1 − n) = z
dz = z dz.
2πi C0 e − 1 2πi C0 k=0 k!

But the value of integral on the right is 2πi times the residue of the pole at z = 0, and
this is just the coefficient of 1/z in the sum. Hence, recalling that B1 = −B
b1 and Bn = 0
for odd n ≥ 3,

(−1)n−1 (n − 1)! Bn (−1)n−1 Bn B

bn
ζ(1 − n) = · 2πi = = − .
2πi n! n n
This is the left-hand equality in (6) for n ≥ 1. For the remaining case, recall that
1
ζ(s) = − γ + O(s − 1) as s → 1.
s−1

The right-hand equality follows from the functional equation of ζ(s),

 
s
−s/2 1−s
Γ π ζ(s) = Γ π −(1−s)/2 ζ(1 − s).
2 2

(See Edwards [1], for example.) 

Of course we can discard the sign if we are interested only in the absolute value:
 
2 n! 2 n! 1 1
|Bn | = ζ(n) = 1 + n + n + . . . , even n ≥ 2. (7)
(2π)n (2π)n 2 3

6
We can use (5) and (7) to determine |Bn | exactly. Write b(n) = (2 n!)1/n /(2π). Then split
the sum into three parts,

|Bn | = In + En + Rn ,
 
db(n)e   n  db(n)e  n ∞  n
 X b(n)  X b(n) X b(n)
In =  , En = − In , Rn = .
k=1
k k=1
k k
k=db(n)+1e

We compute In exactly and En approximately, we assume that Rn is small enough to be

ignored, and we get the exact value of Fn , the fractional part of |Bn |, using the von Staudt
– Clausen formula, (5). Then for positive even n, we have |Bn | = In +  + Fn , where  = 0
if En < Fn ,  = 1 otherwise. Example:

db(1000)e = 59, I1000 = 5318704469 . . . 6955251716 ≈ 5.318704469415522 × 101769 ,

55743421
E1000 ≈ 0.1625177182 and F1000 = ≈ 0.1625177220 > E1000 .
342999030
There is an interesting application concerning prime generating functions, [4], and also
an interesting problem. Is there an n that requires  = 1? By simple-minded direct
calculation there are no instances for even n up to 3200. But could it possibly happen
that En is very nearly equal to but less than 1, Fn is very small, and Rn , which is very small
anyway, is just large enough so that En + Rn > 1? Then we would have Fn = En + Rn − 1
and hence b|Bn |c = In + 1. A similar problem is presented in M500 [2, 5].

References
[1] H. M. Edwards, Riemann’s Zeta-Function.

[2] TF, Problem 267.4 – Bernoulli numbers, M500 267 (December 2015).

[3] A. A. Lovelace, Note G in: L. F. Menabrea, Sketch of The Analytical Engine Invented
by Charles Babbage With notes upon the Memoir by the Translator, Ada Augusta,
Countess of Lovelace, Bibliothèque Universelle de Genève 82 (October 1842).

[4] R. Thompson, Bernoulli numbers and prime generating functions, M500 267 (De-
cember 2015).

[5] R. Thompson, Solution 267.4 – Bernoulli numbers, M500, to appear.

[6] E. C. Titchmarsh, The Theory of the Riemann Zeta-Function.