Probability of digits by dividing random numbers: a ψ

and ζ functions approach

Alessandro Gambinia , Giovanni Mingari Scarpellob , Daniele Ritellia
a Dipartimento di matematica per le scienze economiche e sociali, viale Filopanti, 5 40126
Bologna, Italy
b Via Negroli, 6 20133 Milan, Italy

Abstract
This paper begins with the statistics of the decimal digits of n/d with (n, d) ∈ N2
randomly chosen. Starting with a statement by E. Cesàro on probabilistic number
theory, see [3] and [4], we evaluate, through the Euler ψ function, an integral
appearing there. Furthermore the probabilistic statement itself is proved, using
a different approach: in any case the probability of a given digit r to be the first
decimal digit after dividing a couple of random integers is
   
1 1 r 11 r
pr = + ψ + −ψ +1 .
20 2 10 10 10

The theorem is then generalized to real numbers (Theorem 1, holding a proof of

both dn results) and to the αth power of the ratio of integers (Theorem 2), via
an elementary approach involving the ψ function and the Hurwitz ζ function.
The article provides historic remarks, numerical examples, and original theoretical
contributions: also it complements the recent renewed interest in Benford’s law
among number theorists.
Keywords: Elementary probability, Euler ψ function, Hurwitz ζ function
2000 MSC: 33B15,
2000 MSC: 11M35,
2000 MSC: M11K99

n
1. Introduction: the d theorems

At the end of the 18th century, Gauss was one of the first to use probabilis-
tic arguments for investigating the number of products consisting of exactly k

Preprint submitted to Expositiones Mathematicae November 19, 2011

distinct prime factors below a given bound. The case k = 1 led to the prime num-
ber theorem (proved in 1896), a milestone of analytic number theory. Instead
of probabilistic number theory one should better speak about studying arithmetic
functions with probabilistic methods. Anyway, the first approach in this direc-
tion dates back to Gauss, and to Ernesto Cesàro1 who observed, 1881, that the
probability that two randomly chosen integers are coprime is 6/π 2 . Substantial
developments in probabilistic number theory started with the prime divisor count-
ing function Ω(n) of a positive integer n, but only in 1917 Hardy and Ramanujan
discovered the first deep results on such a function. The first of Cesàro’s papers on
the theory of numbers was followed, 1885, by a 117 pages book issued in Paris,
[3], entitled: Excursions arithmetiques a l’infini, which reproduces nine articles
from Annali di matematica pura ed applicata about some relevant problems in
arithmetic. They looked at problems concerning the number of common divisors
of two integers, determination of the values of the sum of their squares, the proba-
bility of the incommensurability of three arbitrary numbers, and so on; to these he
attempted to apply results obtained in the theory of Fourier series. In this paper we
focus on that collection’s third article (pp. 35–56), entitled Eventualités de la di-
vision arithmétique originally appeared in [4]. In addition to the topics mentioned
above, Cesàro contributed to founding probabilistic number theory studying the
distribution of primes and trying to improve results obtained there by Chebyshev.
1.1. Cesàro’s statements and the aim of the paper
In [3] p. 49, section 19, or [4], after a sequence of frequentistic arguments, he
notes that
On trouve facilement que, si L et Λ sont les frèquences d’une certaine
condition dans 
les 
séries  
n n n n
f , f , f ,... f
1  2  3 
 n
(1)
1 2 3 n
f , f , f ,... f
n n n n

1 E. Cesàro (1859-1906) entered the École des Mines where he studied mathematics with Cata-
lan. In Paris he was particularly struck by lectures of Darboux on geometry which led him to
his own studies of intrinsic geometry. Supported by Cremona, Battaglini, and Dini, he undertook
research at the University of Rome from 1884 where the doctorate was awarded to him in 1887.
He also developed the study of divergent series; furthermore his interest in mathematical physics
is evident from two successful calculus texts he wrote. Cesàro was offered the chair of mathemat-
ics at Palermo where he remained until 1891, moving then to Naples where he held the chair of
mathematical analysis until his death in 1906.

2
 la probabilité que, en prenant u et v au hasard, f uv satisfasse à

la même condition, est ègale à la moyenne arithmétique des deux

fréquences. Ainsi:
1
P= (L + Λ) (2)
2
At this point, taking f (x) = x in the first of two series (1) he states that
D’autre part, dans la seconde des séries (1), chaque chiffre est également
fréquent. Donc, d’après (2), on peut affirmer que: Dans une division
quelconque, la probabilité que le premier chiffre décimal soit r, est:
1
Z 1 1−ϕ
+5 10
ϕ 9+r dϕ. (3)
20 0 1 − ϕ
Next he states that, performing an arbitrary division of two random integers,
the ith decimal digit has a probability to be r given by
10i
Z 1
1 1−ϕ i
+ 10
ϕ 10 −1+r dϕ. (4)
20 2 0 1−ϕ
Our contribution starts by evaluating the integrals (3) and (4). Such integrals,
concerning a ratio of rational functions,

1 − ϕ 9+r ϕ 9+r
ϕ = ,
1 − ϕ 10 (ϕ + 1)P(ϕ)
where
1√ 1√
    
2 2
P(ϕ) = ϕ − 5−1 ϕ +1 ϕ + 5−1 ϕ +1 ×
2 2
√  √ 
  
2 1 2 1
× ϕ − 1+ 5 ϕ +1 ϕ + 1+ 5 ϕ +1 ,
2 2
are not easy. It is then better to compute them by standing on a higher viewpoint,
i.e., introducing the Euler ψ function. Such a computation was not performed by
Cesàro, who probably judged it too involved. But through an alternative repre-
sentations of ψ, we realized that those integrals could be expressed by means of
infinite series. This inspired an elementary proof of the same statement of Cesàro,
followed by some generalizations to be found in our Sections 3 and 4.
A theorem of Gauss on ψ of rational arguments simplifies the process, avoid-
ing decomposition into simple fractions.

3
1.2. Formal probability definition
(i)
In his arithmetic papers Cesàro freely mentions the probability pr that the ith
decimal digit of the ratio of two integers to be a given digit, say r ∈ {0, . . . , 9}, but
quite informally. For better clarity and chiefly to highlight the difference between
the definition in our paper (two real positive numbers randomly chosen) and that
implicitly followed by Cesàro (two integer numbers randomly chosen), we define
this discrete version: if di (v) is the ith decimal digit of the number v, we put:

(i) #{(n1 , n2 ) ∈ N2 | n j ≤ T for j = 1, 2, di (n1 /n2 ) = r}

pr := lim
T →∞ #{(n1 , n2 ) ∈ N2 | n j ≤ T for j = 1, 2}

whenever such a limit exists. Here #A stands for the cardinality of A. When i = 1
(1)
we write pr instead of pr .

2. Cesàro’s integrals evaluation through the ψ function

The main properties of Euler’s ψ function will be recalled, other details being
available in [1] and [6]. We have

Γ0 (x) 1 ∞
x
ψ(x) := = −γ − + ∑ , (5)
Γ(x) x n=1 n(x + n)

where γ is the Euler-Mascheroni constant,

  Z ∞ 
1 1 1 1
γ = lim 1 + + · · · + − ln n = − dx.
n→∞ 2 n 1 bxc x

where bxc is the floor of x. From (5) follows the recursive relation

1
ψ(x + 1) = ψ(x) + . (6)
x
Furthermore these integral representations for x > 0 will be useful:
Z ∞ −t
e − e−xt
ψ(x) = −γ + dt (7)
0 1 − e−t
1 − ux−1
Z 1
ψ(x) = −γ + du (8)
0 1−u

4
 Of course (8) is obtained from (7) by putting e−t = u. From the second integral
formula, taking a, b > 0, one obtains:
Z 1 a
y − yb
dy = ψ(b + 1) − ψ(a + 1)
0 1−y
of which the Cesàro integral in (3) is a special case:
Z 1 r r+1
1−ϕ 1 1 y 10 − y 10
Z
9+r
ϕ dϕ = dy
0 1 − ϕ 10 10 0 1−y
     (9)
1 r + 11 r + 10
= ψ −ψ
10 10 10
In such a way the probability in (3) is given by
   
1 1 r 11 r
pr = + ψ + −ψ +1 . (10)
20 2 10 10 10
Gauss, see his Gesammelte Werke, Bd. 3, pp. 155–156, (Göttingen, 1866), found
for ψ of rational arguments an expression involving elementary functions only. In
fact, for any p, q ∈ N with 1 ≤ p < q,
    q−1     
p 1 pπ 2kpπ kπ
ψ = −γ − π cot − ln q + ∑ cos ln 2 sin . (11)
q 2 q k=1 q q
A relevant proof, more elementary than the original one, can be found in [6].
Joining (6), (10), and (11) we have a closed form expression of all the required
probabilities:
√ √  √ i
q h
101
p0 = 20 − 4 5 + 2 5 − 18 ln(800 000) + 5 ln 9 + 4 5
π
" #
√ √ √ 
r   
1
p1 = 20 −49 + π 10 5 + 5 + 5 ln 16 + 5 ln 3+2 5

√ √ √
 q 
1 3+ 5
p2 = 60 −47 + 6π 25 − 10 5 + 30 5 ln 2 − ln 4
√  √ √ √
  q 
1 3+ 5
p3 = 60 −22 + 3π 5 − 5 5 − 2 5 − 15 5 ln 2 − ln 16
√ √ 
 q 
1
p4 = 40 −8 + 2π 25 − 10 5 − 40 + 5 ln 2+
√  √ i
+ 25 ln 5 + 5 ln 123 − 55 5

5
 √ √ √ 
 q  
1 3 125 1+ 5
p5 = 120 −14 + 6π 25 − 10 5 − 15 ln 256 − 2 5 ln 2
q √ √ √ h √  √ √ i
29 π 25−11 5 1+ 5 1 5 3+ 5
p6 = − 420 + 2 10 + ln 4 + 2 ln 5 − 1 + 2 ln 2
q √ √
11
p7 = − 280 + π2 1 − √25 + ln 2 + 25 ln 3−2 5
!
√ √ √
r  
1
p8 = 360 −7 + 18π 10 5 + 5 − 360 ln 2 − 90 5 ln 3+2 5

√ √  √ i
 q h 
1
p9 = 360 −2 − 90π 5 + 2 5 + 45 ln(800 000) + 5 ln 9 + 4 5 .

With the help of Mathematica R it is immediate to verify that

9
∑ pk = 1.
k=0

The correspondent numerical values are

p0 = 0.126730362245; p1 = 0.117357521909; p2 = 0.109924503863;

p3 = 0.103903172141; p4 = 0.098937259282; p5 = 0.094778739397;
p6 = 0.091250211050; p7 = 0.088221779210; p8 = 0.085596363935;
p9 = 0.083300086967.

The computed first digit probability is then decreasing with increasing digits, so
that the highest is ‘zero’, and the lowest is ‘nine’.
Let us quote the recent paper of Qiu and Vuorinen, [11], where, formula (2.3)
p. 727, it is proved that for any x > 1,
1 1 1 2γ − 1
`(x) := ln x − − 2
< ψ(x) < ln x − − := u(x).
2x 12x 2x 2x2
So, we see that
7
sup (u(x) − `(x)) = − γ ' 0.00611767
x>1 12

and then the use of either u(x) or `(x) instead of ψ(x) produces an error less than
1/100.

6
2.1. An experimental check of the Cesàro’s law
Before going ahead with the generalizations, the reader should know a possi-
ble experimental check on the above law’s reliability. In order to do it, following
the frequentist approach, after a very high number of divisions of two positive
integers, chosen randomly, we will extract the first digit, recording the occurrence
of each of ten digits. In such a sense we will follow the approach of the empir-
ical checks, performed via Mathematica R which is the same as our tool, on the
Weisstein web site [12] about Benford’s law.
It will be recalled that the leading digit in tables and physical data is not evenly
distributed among the digits 1 and 9. The first known written reference is a two-
pages article by Newcomb, [10] 1881, who stated2 that
 
1
Prob(first significant digit = r) = log10 1 + , r = 1, . . . , 9.
r

In such a way he conjectured that the digit ‘9’ occurs about 4.6% of the time. In
1938, Benford published a paper [2] describing how well his 20,229 observations
were fit by the logarithmic law, nowadays called the Benford Law stating that
for a dimensionless data sequence, the digit ‘1’ tends to occur with probability
' 30%, almost three times greater than the naive value3 , ' 11.1%. There is also
a general significant digit Benford law including not only the first digit, but also
the second, which may be zero, and all higher significant digits. For instance in
[5], the probability that the first three significant digits are 3, 1, 4 in that order, is
given by  
1
log10 1 + ' 0.14%
314
and similarly for other significant digit patterns.
Cesàro’s law has a different object than Benford’s: in fact Cesàro’s argument
is theoretical, founded upon number theory; on the contrary in the beginning the
Benford law was nothing but a empirical statement about real-world datasets, ap-
plicable to some datasets but not all, and only approximate. The things are quite
different nowadays, the law having been embodied as a theoretical mathematical

2 Assuming that a universal probability distribution exists over such numbers, it will be in-
variant under a change of scale. Normalizing and differentiating such a invariance one will find
a hyperbolic probability distribution, so that by integration the logarithmic phenomenon appears
known as Benford’s law.
3 Namely the percentage one would get if the digits were equidistributed.

7
result (see for instance [9] and [7] and the references therein). Accordingly, a
sequence of positive numbers xn is defined Benford (base b) if the probability of
observing the first digit of xn in base b is j ∈ { j = 1, 2, . . . b − 1} is given by:
 
#{n < N : first digit of xn in base b is j} 1
lim = logb 1 +
N→∞ N j

From this definition, one can, for instance, see that the first digit of 2n is Benford
in base 10, but not Benford in base 2 because the first digit is always 1 in this
second case. In [9] (Theorem 9.3.1 p. 220) some conditions are given in order
that a recurrence relation of assigned length meets Benford; it is also shown why
the first digit of a geometric Brownian motion is Benford.
But any sequence of digit obtained with the Cesàro procedure takes into ac-
count the ‘zero’ digit’s occurrences, so this phenomenon is not relevant for Ben-
ford’s, dealing with sequences of positive numbers.
Anyway: we did an experimental check by means of Mathematica R , ran-
domly taking couples of positive integers between 1 and 107 , taking the quotient,
repeating 107 times, and recording the relative frequency of each digit. Compar-
ing such values to the computed ones, we found for each digit a difference on the
order of magnitude of ±10−6 , which has been successively confirmed by a similar
test on the second place also.

2.2. Further digits

The integral in (4) can be again expressed through ψ:

10i r + 10i + 1 r + 10i

    
(i) 1
pr = + ψ −ψ . (12)
20 20 10 10

Like in the previous analysis of the first digit, one can compute via elementary
functions the probability that, dividing two randomly chosen integers, the ith digit
will be r = 0, 1, . . . , 9. But, due to the recursive relationship on ψ,
1 1 1
ψ(x + n) = ψ(x) + + +···+ , n ∈ N,
x x+1 x+n−1

8
such formulae quickly become intractable, having to be evaluated with n = 10i−1 .
For instance the probability the second decimal digit to be zero is
√
√
q
(2) 2 991 620 812 234 909 π 5
p0 = − 5+2 5+ ln 5
303 502 130 011 080 2 4
√ √ !
5 ln 20 5−1 1 5
− − ln +
4 2 2 2
1 √
h  √   √  √  √ i
+ 5 ln 7 − 3 5 + ln 5 − 1 − 5 ln 3 5 − 5 .
2
As i → ∞, the events become equiprobable. Roughly speaking, if, instead of
taking the first, we take for instance the millionth digit, then the probability distri-
bution of the ten occurrences (0, . . . , 9) is uniform.
In fact, see [6] p. 165, the following series expansion holds:
∞
(r)n
ψ(s) − ψ(s − r) = ∑ n(s)n ,
n=1

where (s)n = s(s+1) · · · (s+n−1) is the Pochhammer symbol, so that for k, α, β >
0,
 
α − β (α − β )(α − β + 1)
k [ψ(α + k) − ψ(β + k)] = k + +··· ,
α + k 2(α + k)(α + k + 1)

and then
10i r + 10i + 1 r + 10i
    
(i) 1 1 1 1
lim pr = + lim ψ −ψ = + = .
i→∞ 20 i→∞ 20 10 10 20 20 10

n
3. An elementary proof of both d theorems
In this section Cesàro’s statements will be proved by methods of elementary
probability theory, and then they will be extended to the case of the quotient of
positive real numbers with respect to an arbitrary base b ≥ 2. Choose at random
two real positive numbers n, d. We provide an alternative definition of the prob-
ability that the ith digit of the ratio n/d will be r ∈ {0, . . . , b − 1}. There being a
choice made in the whole continuum of possible outcomes, we cannot make use,
as for the discrete case, of the ratio of two cardinal numbers. In such a way the
definition can be applied to an arbitrary base b and not only to a simple ratio but

9
also to a power α > 0 of such a ratio. Given two numbers (n, d) ∈ R+ 2 , we need
(i)
α
to define the probability bα pr that the ratio dn has, to base b ≥ 2, the ith digit
r. To do this first we point out formally how to extract the ith digit in base b of a
given real number. Given an integer b ≥ 2 called the base and using the b digits
ci ∈ {0, 1, . . . , b − 1}, we have a representation of a positive real number ν in base
b as ∞
ν= ∑ ci bi ,
i=−∞
Observe that there exists N ∈ N such that i > N =⇒ ci = 0 thus
N −1
i
∑ cib and ∑ ci bi
i=0 i=−∞

are respectively the integer part and the fractional part of ν. The coefficient c−i in
front of b−i , i = 1, 2, . . . is called the ith digit in the base b representation and will
be denoted dib (ν). Thus

dib (ν) = r ∈ {0, 1, . . . , b − 1}

if and only if there exists k = 0, 1, . . . such that bi−1 ν ∈ k + br , k + r+1

 
b .

Definition 1. Let dib (ν) be the ith decimal digit for the real number ν; then, given
(i)
two real positive numbers n, d and α > 0 we define the probability bα pr that the
ith decimal digit to base b of the ratio (n/d)α to be r ∈ {0, . . . , b − 1} as
 α
µ{(n, d) ∈ R2+ | n, d ≤ T, dib dnα = r}
b (i)
α pr = lim
T →∞ µ{(n, d) ∈ R2+ |n, d ≤ T }

if the limit exists. Here µ is the Lebesgue measure on R2+ .

(i)
We will keep the notation pr whenever α = 1 and b = 10 and the notation pr
if i = 1, α = 1, b = 10.
As a first step we will, by means of the series expansion (5) of ψ, rearrange
(10). In fact from (5) we deduce the following relation, which holds for b, x > 0 :
  ∞
x+1 x 1
ψ −ψ =b∑ (13)
b b n=0 (x + nb)(x + 1 + nb)

10
In such a way, using (13) we see that (10) can be written using a numerical series:
∞
1 1
pr = +5 ∑ . (14)
20 k=0 (10 + 10k + r)(11 + 10k + r)

This fact inspired us in looking for an alternative proof to the Cesàro statement,
we are going to present in the following theorem. We take hereby the opportunity
to thank the anonymous referee for his/her fruitful suggestions which simplify and
clarify our original proof.
Theorem 1. For a couple of real positive numbers n, d taken at random and rep-
(i)
resented in base b, the probability b pr that the ith decimal digit of n/d will be
r ∈ {0, 1, . . . , b − 1} can be computed by the formula

(i) 1 1 ∞ bi
b pr = + ∑ i
2b 2 k=0 (b + bk + r)(bi + bk + r + 1)
(15)
bi−1
  i   i 
1 b +1+r b +r
= + ψ −ψ .
2b 2 b b
Proof. We have
∞  
2 kb + r n kb + r + 1
n o [
2 b
(n, d) ∈]0, T ] | di (n/d) = r = (n, d) ∈]0, T ] | ≤ < .
k=0
bi d bi

This union is split as U ∪ L, where

bi−1
[−1
 
2 kb + r n kb + r + 1
U= (n, d) ∈]0, T ] | ≤ <
k=0
bi d bi

and ∞  
[
2 kb + r n kb + r + 1
L= (n, d) ∈]0, T ] | ≤ <
bi d bi
k=bi−1

Let us explain the split. Reasoning in the plane (n, d) for 0 ≤ k ≤ bi−1 − 1, each of
the subsets forming U succeeds in being a triangle with the base on the horizontal
straight line of equation d = T bounded by the straight lines with slopes ≥ 1. If,
on the contrary, k ≥ bi−1 , each of the subsets forming V is a triangle based on
the vertical straight line n = T and sided by the straight lines with slopes < 1, as
shown in the figure below.

11
 d

T
n

Figure 1: Reference (n, d): the shaded areas sketch the standard subsets making up the set U (on
the left) and the set L (on the right).

For k = 0, . . . , bi−1 − 1 the kth set in U is the triangle with vertices in the (n, d)-
plane    
kb + r kb + r + 1
(0, 0), T, T , T, T ,
bi bi
hence its Lebesgue measure is T 2 /(2bi ).
For k = bi−1 , . . . the kth set in L is the triangle with vertices

bi bi
   
(0, 0), T, T , T, T ,
kb + r kb + r + 1

with Lebesgue measure

bi T 2
.
2(kb + r)(kb + r + 1)

12
Therefore,
bi−1 −1
(i) 1 ∞
bi
b pr = ∑ + ∑
k=0 2bi k=bi−1 2(kb + r)(kb + r + 1)
1 bi ∞ 1
= + ∑ i
2b 2 n=0 (b + bn + r)(bi + bn + r + 1)

This shows the first of (15), while the second relation follows from (13).
Remark 1.1. The theorem given above doesn’t prove the statement of Cesàro,
which concerns integer variables (discrete case) whilst our outcome concerns all
the real ones (continuous case). Nevertheless it would be not difficult to retrace the
original statement in combination with Gauss lattice point counting technique, see
[8]. In fact, considering the lattice’s ”integer” points inside the triangles described
by the sets U and L, following the approach given in [8] section 1.1, theorems 1.1
and 1.4 , one can prove the number of integer points inside the triangle equates
its area plus a term which is growing as the perimeter of the area itself: and then
negligible with respect to the area when T → ∞. In Theorem 1 the first part of the
sum is found to be:
bi−1 −1  2 
T
∑ 2bi + O(T )
k=0

dividing to T 2 , it goes to 1/2b for T → ∞. On the contrary, the second part of the
sum requires further attention, being an addition of infinitely
√ many terms. There
is no restriction in limiting the series from bi−1 to b T c and then to put T to
infinite. In such a way, see e.g. the Theorem 1, the number of the points inside the
lattice inside the triangles of the set L will be:
√
b Tc
bi T 2
 
lim ∑ + O(T )
T →∞
k=bi−1
(kb + r)(kb + r + 1)

When dividing to T 2 the expression

√
b Tc
3
∑i−1 O(T ) = O(T 2 )
k=b

passing to the limit, tends to 0. We take the opportunity to thank again the anony-
mous referee, for having pointed us this remark.

13
4. The ( dn )α theorems
In this section we can see how the same approach can easily lead to the any
decimal digit of nα /d α for α real and positive.
Let us state in advance something about the Hurwitz zeta function. It will be
recalled that the Riemann zeta function ζ (s) is defined by
∞
1
ζ (s) = ∑ ns , Re(s) > 1
n=1

where s is a complex variable.

The Hurwitz zeta function ζ (s, x), generalizing Riemann’s, is defined by
∞
1
ζ (s, x) := ∑ (n + x)s (16)
n=0

where Re s > 1 and Re x > 0. We have ζ (s) = ζ (s, 1). Recall the Hurwitz ζ
integral representation
Z ∞ −xt
1 e
ζ (s, x) = t s−1 dt, (17)
Γ(s) 0 1 − e−t
which holds for Re s > 1 and Re x > 0. It is known that the Hurwitz zeta function
can be meromorphically continued to be defined for all complex numbers s with
s , 1. At s = 1 it has a simple pole with residue 1. Its meromorphic continuation
becomes explicit for s , 1 and x > 0 by means of Hermite’s integral representation
theorem:
1 1 1−s sin(s arctan xt ))
Z ∞
ζ (s, x) = x−s + x +2 dt (18)
2 s−1 0 (x2 + t 2 )s/2 (e2πt − 1)
More information about ζ (s, x) can be found in [13] chapter 13 or in [1].
The Hurwitz zeta function is linked to the polygamma function if its first vari-
able is a positive integer, i.e., n ∈ N:
dn
ψn (z) := ψ(z) = (−1)n+1 n! ζ (n + 1, z). (19)
dzn
In what follows, we will need the integration formula
Lemma 1. Let s , 1 and a ∈ R+ , b ∈ R be such that a + b ∈ R+ . Then
Z 1
1
ζ (s + 1, a + bx) dx = {ζ (s, a) − ζ (s, a + b)} . (20)
0 sb

14
Proof. Formula (20) follows by:
1
Z
ζ (s, x)dx = ζ (s − 1, x)
1−s
which, in turn, follows from the relationship:

∂ζ
(s − 1, x) = (1 − s)ζ (s, x),
∂x
which is a consequence of (16)
Finally for b > 0, Re(s) > 1, Re(x) > 0 we will employ this relation which is
immediate consequence of the definition of ζ (s, x)
∞
1 1 x
∑ (bn + x)s = bs ζ (s, b ) (21)
n=0

We then can present our generalization of Cesàro’s results.

Theorem 2. For a couple of positive real numbers n, d taken at random and rep-
(i)
resented in base b, the probability bα pr , with α > 0, that the ith decimal digit of
nα /d α will be r ∈ {0, 1, . . . , b − 1} is given by
i−1
α (i) 1 b −1 h 1/α 1/α
i
b pr = √ α ∑ (kb + r + 1) − (kb + r) )
2 bi k=0
(22)
1√ ∞  
α
i
1 1
+ b ∑ √ −√
2 k=0
α i
b + bk + r α i
b + bk + r + 1

Moreover when α < 1 in equation (22) the infinite series can be written in terms
of Hurwitz zeta function:
i−1
α (i) 1 b −1 h 1/α 1/α
i
b pr = √ α ∑ (kb + r + 1) − (kb + r) )
2 bi k=0
√ (22b)
1 bi + r 1 bi + r + 1
α i−1     
b
+ ζ , −ζ ,
2 α b α b

15
Proof. The proof is similar to Theorem 1. In this case we have
n o
(n, d) ∈]0, T ]2 | dib (nα /d α ) = r
∞  
[
2 kb + r nα kb + r + 1
= (n, d) ∈]0, T ] | ≤ α <
k=0
bi d bi
( r r )
∞
kb + r n kb + r + 1
(n, d) ∈]0, T ]2 |
[ α α
= ≤ < .
k=0
bi d bi

As for the previous case in Theorem 1, this union can be split as U ∪ L, where

bi−1
( )
[−1
r r
kb + r n kb + r + 1
(n, d) ∈]0, T ]2 |
α α
U= ≤ <
k=0
bi d bi
and ( r r )
∞
kb + r n kb + r + 1
(n, d) ∈]0, T ]2 |
[ α α
L= ≤ <
bi d bi
k=bi−1

For k = 0, . . . , bi−1 − 1 the kth set in U is the triangle with vertices in the (n, d)-
plane r ! r !
α kb + r α kb + r + 1
(0, 0), T, T , T, T ,
bi bi
hence its Lebesgue measure is
√ √
T 2 ( α kb + r + 1 − α kb + r)
√
α
.
2 bi

For k = bi−1 , . . . the kth set in L is the triangle with vertices

 s   s 
b i b i
α α
(0, 0), T, T  , T, T,
kb + r kb + r + 1

with Lebesgue measure

s s 
1 α bi α bi T2
−
2 kb + r kb + r + 1

16
Therefore summing the measures of the disjointed events in order to detect the
(i)
required probability we find the first formula for bα pr given in (22). To end the
proof and get the formula (22b) for α < 1, first notice that the series (22) converges
also for this set of α’s, then use the following identity a, k, β > 0
Z 1
1 1 β
− = dx
(a + k)β (1 + a + k)β 0 (1 + a + k − x)β +1
then summing on k and interchanging summation and integration, we see that
∞  
1 1
∑ √α bi + bk + r − √α bi + bk + r + 1
k=0
1 1 ∞ 1
Z
= ∑ dx
α 0 k=0 (bi + r + 1 + kb − x) α1 +1
1 bi + r − x + 1
Z 1  
1
= 1 ζ 1+ , dx
αb α +1 0 α b
Formula (22b) now follows using (21) and (20).
Remark 2.1. As for Theorem 1, we proved in such a way the theorem in the case
of a continuous variable. For passing to the discrete one, we will refer once again
to the theory of lattice points in plane domains, [8]. We get:
bi−1 −1 
√ √ 
2 kb + r + 1 − kb + r
α α

∑ T √
α i
+ O(T )
k=0 2 b
and again dividing to T 2 , the last term will vanish for T → ∞.
The same argument concerning the infinite series spent about Theorem 1 can
be applied to the set L of Theorem 2.
Le us close with some points to note.
If in (22) we take the limit α → 1, (15) shall by necessity be obtained, and
this is done easily using the expressions containing infinite series. If one instead
wants to use the special functions representations, the pole at s = 1 of ζ (s, x) can
be bypassed: in fact if x, y are two positive real numbers, then
Axy (s) := ζ (s, x) − ζ (s, y)
is continuous in s = 1 attaining there the value
Axy (1) = ψ(y) − ψ(x). (23)

17
In fact, expressing Axy by means of the integral representation (17) of ζ (s, ·), we
have Z ∞ −xt
1 e − e−yt s−1
Axy (s) = ζ (s, x) − ζ (s, y) = t dt
Γ(s) 0 1 − e−t
the limit under the integral for s → 1 is allowed due to the integrand’s summability
(including the origin). So that, by the first part of (7):
Z ∞ −xt
e − e−yt
Z ∞ −xt
e − e−t + e−t − e−yt
Axy (1) = dt = dt = ψ(y) − ψ(x).
0 1 − e−t 0 1 − e−t
Then, by means of (23), we can put α = 1 in (22), finding (15) again.
Whilst in Cesàro’s formulae the digamma appears, in ours we use the Hurwitz
ζ function: this depends on the non-integrality of the first argument of the gen-
eralized zeta function. Therefore if in (22b) one puts α = 1/n with n ∈ N, then
thanks to (19) it would be possible to express (22b) by a digamma.

Conclusions
In 1885 E. Cesàro considered in [4] the probability that a whichever digit r,
zero included, could occur as the first when dividing two integers taken at random,
say n and d. Such a probability was stated without proof as depending on the
integral
Z 1
1 − ϕ 9+r
10
ϕ dϕ, r = 0, 1, . . . , 9
0 1−ϕ
which is a non-trivial function of r. In our article, an elementary proof is given of
such a formula in the case of two arbitrary choosen positive real numbers; next,
the integral is computed by reducing it to the Euler ψ function, namely the log-
arithmic derivative of the Γ function. Such a law has been tested by computer,
taking random couples of positive integers between 1 and 107 , performing the
division, repeating 107 times, recording the relative frequency of each digit. By
comparison of such values to the computed ones, we found a difference for each
digit of the order of ±10−6 , which has been successively confirmed when arrang-
ing a similar test of the second digit. In fact further theorems have been obtained
generalizing such a result to the ith decimal digit of n/d and to the case of num-
bers in a whichever base b. Further generalizations, concerning the probabilistic
occurrence to the second power and finally to the αth power of n/d, are proved
by using the Hurwitz ζ function by means of theorems about that function whose
role was long ago established through number theory.

18
References
[1] G. E. Andrews, R. Askey, R. Roy, Special Functions, Encyclopedia of Math-
ematics and its Applications, 71. Cambridge University Press, Cambridge,
1999.

[2] F. Benford, The law of anomalous numbers, Proc. Amer. Phil. Society 78
(1938) 551–572.

[3] E. Cesàro, Excursions arithmetiques a l’infini, Hermann, Paris, 1885.

[4] E. Cesàro, Eventualités de la division arithmétique, Annali di Matematica

Pura ed Applicata, 13 (1885) 269–290.

[5] T. P. Hill, A statistical derivation of the significant-digit law, Statistical Sci-

ence 10 (1995) 354–363.

[6] J. L. W. V. Jensen, An elementary exposition of the theory of the gamma

function, Ann. Math. 17 (1915) 124–166.

[7] A. V. Kontorovich, S. J. Miller, Benfords law, values of L-functions and the

3x + 1 problem, Acta Arith. 120 (2005) 269–297.

[8] E. Krätzel, Lattice points. Kluwer Academic Publishers Group, Dordrecht,

1988.

[9] S. J. Miller, R. Takloo-Bighash, An invitation to modern number theory.

Princeton University Press, Princeton, 2006.

[10] S. Newcomb, Note on the frequency of use of the different digits in natural
numbers, Proc. Amer. Journal of Math. 4 (1881) 39–40.

[11] S. L. Qiu, M. Vuorinen, Some properties of the Gamma and Psi Functions,
with Applications. Math. Comp. 74 (2004) 724–742.

[12] E. W. Weisstein, Benford’s Law. In: MathWorld A Wolfram Web Resource

mathworld.wolfram.com/BenfordsLaw.html.

[13] E. T. Whittaker, G. N. Watson, A Course of Modern Analysis. Cambridge

University Press, Cambridge, 1927.

19