O N S OME I NTEGRAL R EPRESENTATION O F ζ(n) I NVOLVING

N IELSEN ’ S G ENERALIZED P OLYLOGARITHMS A ND T HE

R ELATED PARTITION P ROBLEM

A P REPRINT
arXiv:2101.03421v1 [math.NT] 9 Jan 2021

Xiaowei Wang(Potsdam)∗

January 12, 2021

A BSTRACT
In this paper, we study a family of single variable integral representations for some products of
ζ(2n + 1), where ζ(z) is Riemann zeta function and n is positive integer. Such representation
1
R1 a b
involves the integral Lz(a, b) := (a−1)!b! 0 log (t) log (1−t)dt/t with positive integers a, b, which
is related to Nielsen’s generalized polylogarithms. By analyzing the related partition problem, we
discuss the structure of such integral representation, especially the condition of expressing products
of ζ(2n + 1) by finite Q(π)-linear combination of Lz(a, b).

Keywords Riemann zeta function at integers · integral representation · Nielsen’s generalized polylogarithms

1 Introduction
It’s well known that many number theoretic properties of ζ(2n + 1) are nowadays still unsolved mysteries, such as
the rationality (only known ζ(3) is irrational), transcendence and existence of closed-form functional equation that
satisfied by ζ(2n + 1). Thanks to the basic functional relation between gamma function Γ(z) and sine function sin(z),
πz
i.e. Γ(z)Γ(1 − z) = sin(πz) = πz csc(πz), one can explicitly express ζ(2n) by r2n π 2n , where r2n is some rational
number related to Bernoulli number B2n . Unfortunately, to find such a simple analogous formula for ζ(2n + 1) is
considered to be impossible. Studying the integral and series representations for ζ(2n + 1) is somehow an important
way to analyze the number theoretic properties of ζ(2n + 1), for instance, F. Beukers’ work [1] is an excellent
example. In this paper we discuss a class of integral representation with
Z 1
1 loga−1 (t) logb (1 − t)
Lz(a, b) := dt
(a − 1)!b! 0 t
Theoretically, many polynomials of ζ(n) can be represented by this integral. In fact, only ζ(2n + 1) or ζ(2n + 1)d
are interesting. Among all family of single variable integral representations that can represent polynomials of ζ(n),
Lz(a, b) is likely the most simple one. Via establishing some linear combination of Lz(a, b) on Q(π), we are even
able to express some ζ(2n + 1)d . However, this method is in somehow restricted, which we shall discuss in the last
section.
In fact, (−1)a+b−1 Lz(a, b) is exactly a special value Sa,b (1) of Nielsen’s generalized polylogarithm Sa,b (z), which
was introduced by N. Nielsen[2].
Z 1
a+b−1 1 loga (t) logb (1 − zt)
Sa,b (z) = (−1) dt
(a − 1)!b! 0 t
In most cases, this function is known for mathematical physicists in the context of quantum electrodynamics. Only
few literatures [3][4][5] concerned about the special case Sa,b (1). However, number theoretic properties of Sa,b (z)
∗
This paper is written in November 2019
 A PREPRINT - JANUARY 12, 2021

have seldom been studied.

R1 R 1−
Throughout the paper, integrals 0 f (t)dt may be regarded as improper, namely 0+ f (t)dt. ζ(s) denotes Riemann
P∞
zeta function ζ(s) := k=1 k1s . In general, N, m, n, k, i, j, l, a, b denote nonnegative integers.

2 Preliminaries
Our main result is based on the following well-known formula([6], p45).
Theorem 1. For z ∈ C and |z| < 1, we have
∞
X ζ(k) k
Γ(1 + z) = exp(−γz + (−1)k z )
k
k=2
Proof. By the definition of Gamma function
∞
1 −γz Y z
Γ(z) = e (1 + )−1 ez/n
z n=1
n
we can rewrite it as

∞
X z z
log Γ(1 + z) = −γz − log(1 + )−
n=1
n n
When |z| < 1, then |z/n| < 1 for all n = 1, 2, ..., thus we have
X∞ ∞
z zk z X zk
log(1 + )= (−1)k+1 k = + (−1)k+1 k
n kn n kn
k=1 k=2
Therefore

∞ X
X ∞
zk z z
log Γ(1 + z) = −γz − (−1)k+1 + −
n=1 k=2
knk n n
X∞ X ∞
zk
= −γz + (−1)k
n=1 k=2
knk
∞
X ∞
zk X 1
= −γz + (−1)k
k n=1 nk
k=2
X∞
ζ(k) k
= −γz + (−1)k z
k
k=2
Taking exp on both sides we get what we need to prove. The validity of changing the order of double sum is based on
the normal convergence of Γ(1 + z).

About the partition problem, the related notations we adopt are following.

For any positive integer N , a partition of N is a way written N into the sum of positive integers. Two sums that
differ only in the order of their summands are considered the same partition. For given N , each partition of N can be
regarded as a finite multiset M = ([n], µM ) in which the underlying set is [n] = {n ∈ Z : 1 ≤ n ≤ N }. Therefore M
is determined by the multiplicity function µM : [n] → Z≥0 that satisfies
X
nµM (n) = N
n=1

2
 A PREPRINT - JANUARY 12, 2021

Now let X = (x1 , ..., xN ), where xn = µM (n), then X totally determines M = ([n], µM ). Therefore the alternative
way to define the partition of N is by

Definition 2.
N
X
P(N ) := {X = (x1 , x2 , ..., xN ) ∈ ZN : nxn = N, 0 ≤ xn ≤ N }
n=1
P(N ) is called the partition set of N . Its element is called a partition element of N , which denoted by X.

For given X ∈ Pts (N ), the support and the norm of X = (x1 , x2 , ..., xN ) are defined by

Definition 3.
Supp(X) := {(n, xn ) : xn > 0}
N
X
kXk := xn
n=1

The set of restricted partition of N that has exactly t parts and the size of each part is not less than s(s > 1), is
denoted by Pts (N ), namely

Definition 4.
Pts (N ) := {X = (0, ..., 0, xs , ..., xN ) ∈ P(N ) : kXk = t}
On could similarly define Ps (N ), Ps≥t (N ), Ps≤t (N ) and so on. In this paper, P2 (N ) is particularly more often used
than others. That is
N
X
P2 (N ) := {X = (0, x2 , ..., xN ) : nxn = N, ∀xn ∈ Z, 0 ≤ xn ≤ N }
n=2

Further, the odd partition set and even partition set are defined by following
Definition 5.
POts (N ) := {X = (x1 , ..., xN ) ∈ Pts (N ) : x2m = 0 for all m ∈ Z}
PEts (N ) := {X = (x1 , ..., xN ) ∈ Pts (N ) : ∀x2m−1 = 0 for all m ∈ Z}
Note that for odd N , PEts (N ) = ∅ for all t, POts (N ) = ∅ for all even t. Similarly, For even N , POts (N ) = ∅ for all
odd t.

Before discussing the relation between Lz(a, b) and ζ(n), we shall introduce Lz(a, b) and lz(a, b).
Definition 6. For nonnegative integers a, b, define
Z 1
lz(a, b) := log a (t)log b (1 − t)dt
0
For positive integers a, b, define
1
Lz(a, b) := (lz(a, b) + alz(a − 1, b) + blz(a, b − 1))
a!b!
It’s obvious to see that both lz and Lz are symmetric, namely lz(a, b) = lz(b, a), Lz(a, b) = Lz(b, a). In fact, we have

Proposition 1.
Z
1 1
loga−1 (t) logb (1 − t)
Lz(a, b) = dt
(a − 1)!b! 0 t
or by the symmetry,
Z 1
1 logb−1 (t) loga (1 − t)
Lz(a, b) = dt
a!(b − 1)! 0 t

3
 A PREPRINT - JANUARY 12, 2021

Proof. Only need to prove that

Z 1
logb−1 (t) loga (1 − t)
lz(a, b) + alz(a − 1, b) + blz(a, b − 1) = b dt
0 t
With integration by parts, one can see

Z 1
1 1
lz(a, b) = − t(a loga−1 (t) logb (1 − t) − b loga (t) logb−1 (1 − t))dt
0 t 1−t
Z 1
t
= −alz(a − 1, b) + b loga (t) logb−1 (1 − t)dt
0 1−t

Still by substituting x = 1 − t, we obtain immediately

Z 1
logb−1 (x) log a (1 − x)
lz(a, b) = −alz(a − 1, b) − blz(b − 1, a) + b dx
0 x
That is what we need.

3 The relation between lz(a, b) and ζ(n)

Recall the simple relation between gamma function and beta function with x, y ∈ C, |x|, |y| < 1.
Γ(1 + x)Γ(1 + y)
f (x, y) := (1 + x + y)B(1 + x, 1 + y) = (1)
Γ(1 + x + y)
Since Re(1 + x) > 0, Re(1 + y) > 0, the left hand side of the equation has the integral representation
Z 1
B(1 + x, 1 + y) = tx (1 − t)y dt
0
Our strategy is follow: Applying Taylor’s theorem for multivariate functions f (x, y). On the one hand, at
R1
(x, y) = (0, 0) any all partial derivatives of (1 + x + y) 0 tx (1 − t)y dt can be evaluated explicitly. On the other
hand, applying Theorem 1 for Γ(1 + x), Γ(1 + y) and Γ(1 + x + y), then evaluate the expansion coefficients of
Γ(1+x)Γ(1+y)
Γ(1+x+y) , namely
∞
X
Γ(1 + x)Γ(1 + y) ζ(n) n
= exp( (−1)n (x + y n − (x + y)n ))
Γ(1 + x + y) n=2
n
If we denote (x + y)n − xn − y n by Pn (x, y) or Pn . Let

ζ(n)k
Dn,k := (−1)k(n+1)
k!nk
Thus we can rewrite the above formula as

Y∞
Γ(1 + x)Γ(1 + y) ζ(n)
= exp((−1)n+1 Pn )
Γ(1 + x + y) n=2
n
∞
Y ∞
X ζ(n)k k
= (1 + (−1)k(n+1) P )
n=2
nk k! n
k=1
Y∞ X∞
= (1 + Dn,k Pnk )
n=2 k=1

4
 A PREPRINT - JANUARY 12, 2021

Due to the normal convergence we can expand the last infinite product and rearrange terms with the order up to
deg(Pnk ), where deg(.) is the total degree of polynomial. Since Pn (x, y) is homogeneous polynomial of x, y,
therefore obviously Pnk (x, y) is also homogeneous polynomial of x, y with deg(Pnk ) = k deg(Pn ) = nk.
Q∞ P∞
That is, we can expand and rearrange n=2 (1 + k=1 Dn,k Pnk ) as follow

∞
Y ∞
X
(1 + Dn,k Pnk ) =1 + (D2,1 P2 ) + (D3,1 P3 )
n=2 k=1
+ (D4,1 P4 + D2,2 P22 )
+ (D5,1 P5 + D3,1 D2,1 P3 P2 )
+ (D6,1 P6 + D4,1 D2,1 P4 P2 + D3,2 P32 + D2,3 P23 ) + ...

or
∞
Y ∞
X ∞
X X Y
f (x, y) = (1 + Dn,k Pnk ) =1+ Dα,β Pαβ (2)
n=2 k=1 N =2 X∈P2 (N ) (α,β)∈Supp(X)

Notice that
Y X
deg( Dα,β Pαβ ) = αβ = N
(α,β)∈Supp(X) (α,β)∈Supp(X)

For fixed positive

P Q integers a, b, the term xa y b has degree a + b. Therefore it only appears in
β
X∈P2 (a+b) (α,β)∈Supp(X) Dα,β Pα . Now we can assume that

X Y N
X −1
Dα,β Pαβ = ρN −j,j xN −j y j (3)
X∈P2 (N ) (α,β)∈Supp(X) j=1

Now we can evaluate ρa,b in two ways. The first one is integral representation.

Lemma 1. For positive integers a, b, we have

1 ∂ a+b
ρa,b = |(0,0) f (x, y) = Lz(a, b)
a!b! ∂xa ∂y b
∂g ∂g
Proof. Let g(x, y) = 1 + x + y, h(x, y) = B(1 + x, 1 + y), notice that ∂x = ∂y = 1 for all (x, y), therefore any one
2 2 2
∂ g ∂ g ∂ g
of second-order partial derivatives ∂x2 , ∂y 2 , ∂x∂y vanishes. Using Leibniz rule for x-component

∂a ∂a ∂ a−1
a
gh = g a h + a a−1 h
∂x ∂x ∂x
and then for y-component, we have

∂ a+b ∂b ∂a ∂ a−1
a b
gh = b
(g a h + a a−1 h)
∂x ∂y ∂y ∂x ∂x
a+b
∂ ∂ a+b−1 ∂ a+b−1
= g a b h + b a b−1 h + a a−1 b h
∂x ∂y ∂x ∂y ∂x ∂y

Its value at the point (x, y) = (0, 0) is

∂ a+b ∂ a+b ∂ a+b−1 ∂ a+b−1

| (0,0) gh = ( h + b h + a h)|(0,0)
∂xa ∂y b ∂xa ∂y b ∂xa ∂y b−1 ∂xa−1 ∂y b

5
 A PREPRINT - JANUARY 12, 2021

On the other hand notice that for positive integer a, b

Z 1
∂ a+b ∂ a+b
|(0,0) B(1 + x, 1 + y) = |(0,0) tx (1 − t)y dt
∂xa ∂y b 0 ∂xa ∂y b
Z 1
= loga (t) logb (1 − t)dt
0
= lz(a, b)
Therefore,

∂ a+b
|(0,0) gh = lz(a, b) + alz(a − 1, b) + blz(a, b − 1)
∂xa ∂y b
Namely,

1 ∂ a+b
|(0,0) f (x, y) = Lz(a, b)
a!b! ∂xa ∂y b

Now we discuss the second approach.

Lemma 2. Assume that nj ≥ 2, kj ≥ 1, let

K
Y X
Pnkjj = Cλ,µ xλ y µ
j=1
PK PK
If λ + µ 6= j=1 nj kj or λµ = 0, then Cλ,µ = 0. Otherwise if µ = j=1 nj kj − λ then Cλ,µ is given by
kj 
K Y 
X Y nj
Cλ,µ =
ℓji
ℓ∈S(µ) j=1 i=1
with S(µ) as follow
kj
K X
X
S(µ) = {ℓ : ℓji = µ, ∀ℓji ∈ Z, 1 ≤ ℓji ≤ nj }
j=1 i=1
k Q kj
Proof. Firstly, it is obviously that Pnjj is homogeneous polynomial of degree nj kj for each j ,therefore K j=1 Pnj is
Q kj PK
also a homogeneous polynomial with deg( K j=1 Pnj ) = j=1 nj kj . On the other hand, notice that for all nj ≥ 2,
PK
the coefficient of the terms xnj and y nj in Pnj are both 0. Hence only xλ y µ with the conditions λ + µ = j=1 nj kj
and λµ 6= 0 has nonzero coefficient.
Secondly, Assume that for each j we have

X  nj  nj −1
Pnkjj
=( xnj −ℓji y ℓji )kj
ℓji
ℓji =1
λ µ
QK k nj


In this way, the coefficient of x y in the expansion of j=1 Pnjj should be the sum of all product of ℓji that by
b
choose kj coefficients from Pnjj respectively and satisfying that
kj
K X
X
ℓji = µ, ∀ℓji ∈ Z, 1 ≤ ℓji ≤ nj
j=1 i=1
We denote such constraint by S(µ). In fact it is coincide with
kj
K X
X
nj − ℓji = λ
j=1 i=1

6
 A PREPRINT - JANUARY 12, 2021

PK
since µ = j=1 nj kj − λ. Now for fixed nj , kj , (1 ≤ j ≤ K), C only determined by λ or µ, we can form now on
simplify this notation by Cµ

Above Lemma is for general integers aj , bj , now we reformulate it and only aim to Supp(X). Assume that
X ∈ P2 (N ), let
Y X
Pnk = Cb (X)xa y b
(n,k)∈Supp(X)
P
Then if b > N − kXk or b < kXk, then Cb (X) = 0. Otherwise, it is given by

X Y k  
Y n
ℓ
ℓ∈S(b) (n.k)∈Supp(X)
Q
Now we are able to represent ζ(N ) and some P ζ(nj ) by Lz(a, b) with a + b = N . Following we provide
nj =N
the representation structure of Lz(a, b).

Theorem 7. (The Partition represented Relation) Assume that a ≤ b, then

X Y
Lz(a, b) = (cb (X) ζ(n)k ) (4)
X∈P2 (N ),kXk≤b (n,k)∈supp(X)

where cX is rational number related to X, and it can be evaluated by Lemma 2

Proof. By the expansion (3) and Lemma 1, we have

N
X −1 N
X −1
Lz(N − j, j)xN −j y j = ρN −j,j xN −j y j
j=1 j=1
X Y
= Dn,k Pnk
X∈P2 (N ) (n,k)∈supp(X)
X Y Y
= Dn,k Pnk
X∈P2 (N ) (n,k)∈supp(X) (n,k)∈supp(X)

It remains to show that the coefficient of xN −b y b on the right handQside has the form (4)
On the other hand, by Lemma 2, we notice that for any term Q of (n,k)∈supp(X) Pnk
X
degy (Q) ≥ k = kXk
(n,k)∈supp(X)

and
X
degy (Q) ≤ (n − 1)k = N − kXk
(n,k)∈supp(X)

Therefore
N −kXk
Y X
Pnk = Cj (X)xN −j y j
(n,k)∈supp(X) j=kXk

That is to say,
N −1 N −kXk
X X X
Lz(N − j, j)xN −j y j = qX Cj (X)xN −j y j
j=1 X∈P2 (N ) j=kXk

7
 A PREPRINT - JANUARY 12, 2021

Q
where qX = (n,k)∈supp(X) Dn,k . Now compare the coefficients on both sides, we have
X
Lz(N − b, b) = Cb (X)qX
X∈P2 (N ),kXk≤b
k
Recalling that Dn,k := (−1)k(n+1) ζ(n)
k!nk
e
, so there is rational number C(X) such that
Y Y
qX = e
Dn,k = C(X) ζ(n)k
(n,k)∈supp(X) (n,k)∈supp(X)

Therefore
X Y
Lz(N − b, b) = e
Cb (X)C(X) ζ(n)k
X∈P2 (N ),kXk≤b (n,k)∈supp(X)

e
Let Cb (X)C(X) denoted by cb (X). It’s obviously rational. We finally have
X Y
Lz(a, b) = (cb (X) ζ(n)k )
X∈P2 (N ),kXk≤b (n,k)∈supp(X)

The expression of cb (X) is given by following

Theorem 8. Assume that Supp(X) = {(nj , kj ) : 1 ≤ j ≤ J}, then

J kj  
Y 1 X Y nj
cb (X) = (−1)N +kXk k
j=1 kj !nj j i=1
ℓji
ℓ∈S(b)

or rewrite as
Y k
1 X Y (n − 1)!
cb (X) = (−1)N +kXk
k! ℓ!(n − ℓ)!
(n,k)∈supp(X) ℓ∈S(b)

where S(b) is given by

kj
K X
X
S(b) = {ℓ ∈ Z : ℓji , 1 ≤ ℓ ≤ b − 1}
j=1 i=1

e
Proof. The proof is straightforward. Recall that cb (X) = Cb (X)C(X), now on the one hand we have,

Y 1
e
C(X) = (−1)k(n+1)
k!nk
(n,k)∈supp(X)

PJ J
Y
kj (nj +1) 1
= (−1) j=1
k
j=1 kj !nj j
J
Y 1
= (−1)N +kXk k
j=1 kj !nj j
On the other hand, by Lemma 2
kj 
J Y 
X Y nj
Cb (X) =
P ℓji
ℓ=b j=1 i=1

e
Multiply Cb (X) and C(X) together, then we have what we need.

8
 A PREPRINT - JANUARY 12, 2021

4 Properties of Lz(a, b) and lz(a, b)

Proposition 2. Z ∞
(−1)a za
Lz(a, 1) = dz = (−1)a ζ(a + 1).
a! 0 ez − 1

The proof is also easy, consider the substitution of t = 1 − e−z . This formula connects Lz(a, 1) to the well-known
formula about ζ(n).

Proposition 3.
n
X
Lz(2n − 1, 2) = nζ(2n + 1) − ζ(j)ζ(2n − j + 1)
j=2

Xn
1
Lz(2n − 2, 2) = (n − )ζ(2n) − ζ(j)ζ(2n − j)
2 j=2

or mixing them, as
a
⌊ 2 ⌋+1
a+1 X
Lz(a, 2) = ζ(a + 2) − ζ(j)ζ(a + 2 − j)
2 j=2

Proof. By Theorem 7, only need to consider the restricted partition that has merely one or two parts. For the case I.
N = 2n + 1
For exactly one-part partition, there is only one element X1 ∈ P2 (N ), and Supp(X1 ) = {(N, 1)}. Similarly,
for two-part partition, there are n − 1 elements: X2,j ∈ P2 (N ) with Supp(X2,j ) = {(j, 1), (N − j, 1)}, where
j = 2, ..., n. Therefore
n
X
Lz(2n − 1, 2) = c2 (X1 )ζ(2n + 1) + c2 (X2,j )ζ(j)ζ(2n − j + 1)
j=2

By Theorem 8, it easy to get the coefficients

 
N 1
c2 (X1 ) = =n
2 N
  
j N −j 1 1
c2 (X2,j ) = − = −1
1 1 j N −j
Therefore
n
X
Lz(2n − 1, 2) = nζ(2n + 1) − ζ(j)ζ(2n − j + 1)
j=2

case II. N = 2n.

For exactly one-part partition, there is only one element X1 ∈ P2 (N ), and Supp(X1 ) = {(N, 1)}. Similarly,
for two-part partition, there are n − 1 elements: X2,j ∈ P2 (N ) with Supp(X2,j ) = {(j, 1), (N − j, 1)}, where
j = 2, ..., n. Therefore
n
X
Lz(2n − 2, 2) = c2 (X1 )ζ(2n) + c2 (X2,j )ζ(j)ζ(2n − j)
j=2

By Theorem 8, it easy to get the coefficients

 
N 1 2n − 1
c2 (X1 ) = =
2 N 2
  
j N −j 1 1
c2 (X2,j ) = − = −1
1 1 j N −j

9
 A PREPRINT - JANUARY 12, 2021

Therefore

Xn
1
Lz(2n − 2, 2) = (n − )ζ(2n) − ζ(j)ζ(2n − j)
2 j=2

Let a = 2n − 1 or a = 2n − 2, rewrite those two formulas we have

a
⌊ 2 ⌋+1
a+1 X
Lz(a, 2) = ζ(a + 2) − ζ(j)ζ(a + 2 − j)
2 j=2

there is a relation between Lz(a, b) and multiple zeta function. Following example shows that the symmetric
property of Lz(a, b) implies a nontrivial relation between Lz(a, b) and multiple zeta function for argument of integers
ζ(n1 , n2 ). For larger a, b, we need more techniques.

Example 9. Consider Lz(2, 1) = Lz(1, 2) . One the one hand,

Z 1
log(t) log(1 − t)
Lz(2, 1) = dt
0 t
Z 1 X∞
tn
=− log(t)(1 + )dt
0 n=1
n+1
Z 1 X∞ Z 1
1
= −( log(t)dt + tn log(t)dt)
0 n=1
n+1 0
∞
X 1
= = ζ(3)
n=0
(n + 1)3

However, on the other hand, with the similar trick, for |t| < 1, we have the expansion
∞
log2 (1 − t) X (2) n
= Sn+1 t
t n=1
(2) P 1
where Sn denotes i+j=n,i,j≥1 ij . then

Z ∞ (2)
1 1
log2 (1 − t) 1 X Sn
Lz(1, 2) = dt =
2! 0 t 2 n=2 n
(2)
Notice that for Sn
n−1 n−1 n−1
1X n 1X 1 1 2X1
Sn(2) = = ( + )=
n k(n − k) n k n−k n k
k=1 k=1 k=1

Therefore, above Lz(1, 2) can be reformulated as

∞ (2)
1 X Sn X 1
Lz(1, 2) = = = ζ(2, 1)
2 n=2 n n2 m
n>m≥1

Hence we prove Euler identity ζ(3) = ζ(2, 1) by using Lz(a, b) = Lz(b, a).
Remark 1. There is another proof using a series involving Striling numbers, see[7].

In fact, generally we have

10
 A PREPRINT - JANUARY 12, 2021

Theorem 10. (Symmetry of Series Representation) For any positive integer a, b, we have
∞ (a) ∞ (b)
1 X Sn 1 X Sn
=
a! n=a nb b! na
n=b
where

X k
Y
Sn(k) = m−1 +
j , mj ∈ Z , j = 1, ..., k,
P
mj =n j=1
(1)
In particular,Sn = n1
Proof. Given positive integer a, b,

Z
1 loga−1 (t) logb (1 − t)
1
Lz(a, b) = dt
(a − 1)!b!0 t
Z 1 X∞ n
1 t b1
= loga−1 (t)(− ) dt
(a − 1)!b! 0 n=1
n t
Z 1 X∞
(−1)b
= loga−1 (t) Sn(b) tn−1 dt
(a − 1)!b! 0
n=b
∞ Z
(−1)b X (b) 1
= Sn loga−1 (t)tn−1 dt
(a − 1)!b! 0
n=b

By the substitution t = e−z , it turns out to be

∞ Z
(−1)b X (b) +∞
Lz(a, b) = Sn (−z)a−1 e−nz dz
(a − 1)!b! 0
n=b
∞ (b)
(−1)a+b−1 X Sn
=
b! na
n=b

On the other hand, by the similar method, we have

∞ (a)
(−1)a+b−1 X Sn
Lz(b, a) =
a! n=a
nb
Since Lz(a, b) = Lz(b, a) holds for all positive integers a, b, therefore
∞ (b) ∞ (a)
(−1)a+b−1 X Sn (−1)a+b−1 X Sn
=
b! na a! n=a
nb
n=b
namely
∞ (a) ∞ (b)
1 X Sn 1 X Sn
=
a! n=a nb b! na
n=b

A straightforward corollary is that, if b = 1, we have

∞ (a)
1 X Sn
ζ(a + 1) =
a! n=a n

On the other hand, by the substitution of t = sin2 (θ) we have

11
 A PREPRINT - JANUARY 12, 2021

Proposition 4.
Z π
2
lz(a, b) = 2 a+b+1
sin(θ) cos(θ) loga sin(θ) logb cos(θ)dθ
0
Z π
2a+b 2
Lz(a, b) = cot(θ) logb−1 sin(θ) loga cos(θ)dθ
a!(b − 1)! 0

5 Examples of Establishing the Integral Representations

• N =3

As the first example, for N = 3, the integral representation is trivial. But as it already demonstrated in the last section,
those two equivalent representations imply some nontrivial relation ζ(3) = ζ(2, 1).
Z
1 1 log2 (1 − t)
ζ(3) = Lz(1, 2) = dt
2 0 t
Z 1
log(t) log(1 − t)
ζ(3) = Lz(2, 1) = dt
0 t
• N =4
Z
1 1
log2 (t) log(1 − t)
ζ(4) = −Lz(3, 1) = − dt
2 0 t
Z 1
log(t) log2 (1 − t)
ζ(4) = −4Lz(2, 2) = −2 dt
0 t
Z
1 1
log3 (1 − t)
ζ(4) = −Lz(1, 3) = − dt
6 0 t
(3) P 1
They correspond to following series representations respectively. Let Sn denotes n1 +n2 +n3 n1 n2 n3 , then

X∞
1
ζ(4) = 4
n=1
n
X∞ (2)
Sn
ζ(4) = 2 = 4ζ(3, 1)
n=2
n2
∞ (3)
1 X Sn
ζ(4) =
6 n=3 n
Following we only concern about the integral representations.

• N =5

Lz(4, 1) = ζ(5)
Lz(3, 2) = 2ζ(5) − ζ(2)ζ(3)
By mixing them, we obtain
Z 1
1 1 log2 (t) log(1 − t) t4
ζ(3) = (2Lz(4, 1) − Lz(3, 2)) = log dt
ζ(2) 2π 2 0 t (1 − t)3
This is a new nontrivial integral representation of Apéry’s constant. On the other hand, we have another integral
representation for ζ(5).
Z
1 1 log2 (1 − t) π2
ζ(5) = (log2 (t) + )dt
8 0 t 3

12
 A PREPRINT - JANUARY 12, 2021

Once notice that

Z
1 1
log2 (1 − t) 2π πζ(4)
(log(t) √ )dt = − √
8 0 t 3 8 3
By plus to above formula, we obtain
Z √
πζ(4) 1 1
log2 (1 − t) log2 (eπ/ 3
t)
ζ(5) − √ = dt
8 3 8 0 t

• N =6

P2 (6) has 4 elements:

X1 = (0, 0, 0, 0, 0, 1), X2 = (0, 1, 0, 1, 0, 0), X3 = (0, 0, 2, 0, 0, 0), X4 = (0, 3, 0, 0, 0, 0)

Supp(X1 ) = {(6, 1)}

Supp(X2 ) = {(2, 1), (4, 1)}
Supp(X3 ) = {(3, 2)}
Supp(X4 ) = {(2, 3)}
By (3)

Lz(5, 1)x5 y + Lz(4, 2)x4y 2 + Lz(3, 3)x3 y 3 = D6,1 P6 + D2,1 D4,1 P2 P4 + D3,2 P32 + D2,3 P23
Comparing the coefficients, we have

Lz(5, 1) = −ζ(6)
1 5 1 π6
Lz(4, 2) = ζ(3)2 + ζ(2)ζ(4) − ζ(6) = ζ(3)2 −
2 2 2 1260
3 10 1 23π 6
Lz(3, 3) = ζ(3)2 + ζ(2)ζ(4) − ζ(6) − ζ(2)3 = ζ(3)2 −
2 3 6 15120

By mixing above two equations, we can rewrite a more interesting but sophisticated formula.
Z 1
log2 (t) log2 (1 − t) (1 − t)12
log dt = 6ζ(3)2
0 t t23

• N =7

Lz(6, 1) = ζ(7)
Lz(5, 2) = 3ζ(7) − ζ(2)ζ(5) − ζ(4)ζ(3)
5
Lz(4, 3) = 5ζ(7) − 2ζ(2)ζ(5) − ζ(4)ζ(3)
4
reformulate them, we have interesting similar representation of ζ(3) and ζ(5).

3 4
ζ(2)ζ(5) = Lz(6, 1) + Lz(5, 2) − Lz(4, 3)
5 5
3
ζ(4)ζ(3) = Lz(6, 1) − 2Lz(5, 2) + Lz(4, 3)
4
• N =8

13
 A PREPRINT - JANUARY 12, 2021

Lz(7, 1) = −ζ(8)
π8
Lz(6, 2) = ζ(3)ζ(5) −
7560
π 2 ζ(3)2 61π 8
Lz(5, 3) = 3ζ(3)ζ(5) − −
12 226800
π 2 ζ(3)2 499π 8
Lz(4, 4) = 4ζ(3)ζ(5) − −
6 1814400
• N =9

Lz(8, 1) = ζ(9)
Lz(7, 2) = 4ζ(9) − ζ(2)ζ(7) − ζ(4)ζ(5) − ζ(6)ζ(3)

28 7 7 1 1
Lz(6, 3) = ζ(9) − 3ζ(2)ζ(7) − ζ(4)ζ(5) − ζ(6)ζ(3) + ζ(2)ζ(3)ζ(4) + ζ(2)2 ζ(5) + ζ(3)3
3 2 2 2 6
ζ(3)3 28ζ(9) π 6 ζ(3) π 4 ζ(5) π 2 ζ(7)
= + − − −
6 3 540 40 2

35 5 1 1
Lz(5, 4) = 14ζ(9) − 5ζ(2)ζ(7) − 6ζ(4)ζ(5) − ζ(6)ζ(3) + ζ(2)ζ(3)ζ(4) + ζ(2)2 ζ(5) + ζ(3)3 − ζ(2)3 ζ(3)
6 2 2 6
ζ(3)3 π 6 ζ(3) 7π 4 ζ(5) 5π 2 ζ(7)
= + 14ζ(9) − − −
2 432 180 6

When N become larger, a, b become closer, the expression of Lz(a, b) would be more complicated. In fact, one can
prove following statement
Q
Theorem 11. If integer N > 20, then there always exist partition X ∈ P2 (N ), such that (n,k)∈Supp(X) ζ(n)k
cannot be represented by finite Q(π)-linear combination of Lz(a, b) for all a, b ∈ Z+ with a + b ≤ N by using the
partition represented relation (Theorem 7).

Remark 2. It’s still unknown whether there is any functional equation in closed form that satisfied by ζ(n) for any
different n, therefore such Q(π)-linear combination of Lz(a, b) may be constructed in other ways that differ from the
partition represented relation. Hence all the representations in the following proof are referred to the representations
by only using the partition represented relation.
Proof. Firstly, for fixed N , let Y
Π : P(N ) → R; X 7→ ζ(n)k
(n,k)∈supp(X)

Assume that X = X1 + X2 ∈ P2 (N ) with X1 ∈ PE2 (N ), X2 ∈ PO3 (N ). Since for all positive even number 2n,
ζ(2n) can be represented as qn π 2n with qn ∈ Q, then in other words Π(X1 ) ∈ Q(π). Conversely, if Π(X2 ) cannot be
represented as finite Q(π)-linear combination of Lz(a, b) for a + b ≤ N , then Π(X) neither.
Therefore, it remains to prove that there always exist X ∈ PO3 (N ) for sufficiently large N , such that Π(X) cannot
be represented by finite Q(π)-linear combination of Lz(a, b). For even or odd N , such X is constructed by different
way.

• Case I. Suppose that N = 2M + 1.

Let
∞
G
T (N ) = PO2t+1
3 (N )
t=1

14
 A PREPRINT - JANUARY 12, 2021

Notice that for Y ∈ P2 (N )\T (N ), we can always assume that Π(X) e can be represented by finite Q(π)-linear combi-
nation of Lz(a, b) for a + b < N . Because otherwise, we come to a smaller N1 < N , and therefore we could repeat
the process by starting from N1 instead of N . That is. it reasonable to assume that if Y ∈ P2 (N )\T (N ), then
X X
Y = pj Lz(aj , bj )

with aj + bj < N . Now Suppose that

(2t+1) (2t+1)
PO2t+1
3 (N ) = {X1 , X2 , ..., Xr(2t+1)
2t+1
}

(2t+1)
with |PO2t+1
3 (N )| = r2t+1 . According to theorem 7, Π(Xi ) only appears in the representation formulas
Lz(N − 2t − 1, 2t + 1), Lz(N − 2t − 2, 2t + 2), ..., Lz(2t + 1, N − 2t − 1). In fact, due to the symmetric property
of Lz(a, b), only Lz(N − 2t − 1, 2t + 1), Lz(N − 2t − 2, 2t + 2), ..., Lz(M + 1, M ) provide the representations
that differ from each other. Therefore by Theorem 7 following linear equations system are derived, if we regard
(2t+1)
Π(Xi ) as unknowns, Lz(a, b) as coefficients

X r3
X (3) (3)
Lz(N − 3, 3) − c3 Lz(N − 1, 1) = p3j π 2j Lz(aj , bj ) + q3i Π(Xi )
i=1
X r3
X (3) (3)
Lz(N − 4, 4) − c4 Lz(N − 1, 1) = p4j π 2j Lz(aj , bj ) + q4i Π(Xi )
i=1
X r3
X r5
X
(3) (3) (5) (5)
Lz(N − 5, 5) − c5 Lz(N − 1, 1) = p5j π 2j Lz(aj , bj ) + q5i Π(Xi ) + q5i Π(Xi )
i=1 i=1
...
X r3
X rτ
X
(3) (3) (τ ) (τ )
Lz(M + 1, M ) − cM Lz(N − 1, 1) = pMj π 2j Lz(aj , bj ) + q5i Π(Xi ) + ... + qτ i Π(Xi )
i=1 i=1

where c, p, q ∈ Q, aj + bj < N , τ is the largest 2t + 1 such that PO2t+1

3 (N ) 6= ∅. There are M − 3 + 1 equations. It
obvious, the number of unknowns |T (N )| = r3 + ... Q + rτ ≥ r 3 . Therefore if M − 3 + 1 < r3 , due to the unknowns are
more than the number of equations, then Π(X) = (n,k)∈supp(X) ζ(n)k cannot be solved by above equations system.
It’s well-known that r3 = |P3O,3 (2M + 1)| increases faster than M . Hence there exist M1 , such that if M > M1 , then
M1 − 3 + 1 < |P3O,3 (2M1 + 1)|. In fact, it’s not hard to find out, if M > 9, then M − 3 + 1 < |T (2M + 1)|.

• Case II. Suppose that N = 2M

Let

∞
G
T (N ) = PO2t
3 (N )
t=1

Similar to the case of odd N , it reasonable to assume that if Y ∈ P2 (N )\T (N ), then

X X
Y = pj Lz(aj , bj )

with aj + bj < N . Now Suppose that

(2t) (2t)
PO2t
3 (N ) = {X1 , X2 , ..., Xr(2t)
2t
}

15
 A PREPRINT - JANUARY 12, 2021

with |PO2t
3 (N )| = r2t . According to theorem 7, analogous linear equations system are derived

X r2
X (2) (2)
Lz(N − 2, 2) − c2 Lz(N − 1, 1) = p2j π 2j Lz(aj , bj ) + q2i Π(Xi )
i=1
X r2
X (2) (2)
Lz(N − 3, 3) − c3 Lz(N − 1, 1) = p3j π 2j Lz(aj , bj ) + q3i Π(Xi )
i=1
X r2
X r4
X
(2) (2) (4) (4)
Lz(N − 4, 4) − c4 Lz(N − 1, 1) = p4j π 2j Lz(aj , bj ) + q4i Π(Xi ) + q4i Π(Xi )
i=1 i=1
X r2
X r4
X
(2) (2) (4) (4)
Lz(N − 5, 5) − c5 Lz(N − 1, 1) = p5j π 2j Lz(aj , bj ) + q5i Π(Xi ) + q5i Π(Xi )
i=1 i=1
...
X r2
X rτ
X
(2) (2) (τ ) (τ )
Lz(M, M ) − cM Lz(N − 1, 1) = pMj π 2j Lz(aj , bj ) + q5i Π(Xi ) + ... + qτ i Π(Xi )
i=1 i=1

where c, p, q ∈ Q, aj + bj < N , τ is the largest 2t such that PO2t

3 (N ) 6= ∅. There are M − 2 + 1 equations. It obvious,
the number of unknowns |T (N )| = r2 +...+rτ ≥ r2 +r4 if N is largeQenough. Therefore if M −2+1 < r2 +r4 , due to
the unknowns are more than the number of equations, then Π(X) = (n,k)∈supp(X) ζ(n)k cannot be solved by above
equations system. It’s well-known that r4 = |P4O,3 (2M )| increases faster than M . Hence there exist M2 , such that
if M > M2 , then M2 − 2 + 1 < |P2O,3 (2M2 )|. In fact, it’s not hard to find out, if M > 10, then M − 2 + 1 < |T (2M )|.

Finally, by above discussion we obtain N0 = 20. If N > N0 , whenever N is odd or even, there always exist
X ∈ P2 (N ) such that X cannot be represented by finite Q(π)-linear combination of Lz(a, b) with a + b ≤ N .

References
[1] F. Beukers, A note on the irrationality of ζ(2) and ζ(3), Bull. London Math. Soc. 11(1979), 268–272.

[2] N. Nielsen, Der Eulersche Dilogarithmus und seine Verallgemeinerungen, Nova Acta Leopoldina, 90(1909), pp.
123-211.

[3] K. S. Kölbig, Nielsen’s Generalized polylogarithms, SIAM J. MATH. ANAL. Vol. 17, No. 5, September 1986
R1
[4] K. S. Kölbig, Closed Expressions for 0
t−1 logn−1 t logp (1 − t)dt, Mathematics of Computation, Vol. 39, No.
160 (Oct., 1982), pp. 647-654

[5] Leonard C. Maximon, The dilogarithm function for complex argument, Proc. R. Soc. Lond. A (2003) 459,
2807-2819

[6] H. Bateman, A. Erdélyi, Higher Transcendental Functions (Volume 1), Krieger Publishing Company, 1985

[7] Victor Adamchik, On Stirling numbers and Euler sums, Journal of Computational and Applied Mathematics 79
(1997) 119-130

Xiaowei Wang(王骁威)
Institut für Mathematik, Universität Potsdam, Potsdam OT Golm, Germany
Email: xiawang@gmx.de

16