On the counting function of semiprimes

Dragos, Cris, an and Radek Erban

Abstract
A semiprime is a natural number which can be written as the product of two primes. The
asymptotic behaviour of the function π2 (x), the number of semiprimes less than or equal to x,
is studied. Using a combinatorial argument, asymptotic series of π2 (x) is determined, with all
the terms explicitly given. An algorithm for the calculation of the constants involved in the
asymptotic series is presented and the constants are computed to 20 significant digits. The
errors of the partial sums of the asymptotic series are investigated. A generalization of this
approach to products of k primes, for k ≥ 3, is also proposed.

1. Introduction
For a positive integer k and a positive integer (or real number) x, let πk (x) be the number
of integers less than or equal x which can be written as the product of k prime factors.
The behaviour of πk (x) has been extensively studied during last two centuries, with the
main focus on the case k = 1, where π1 (x) is the prime counting function, denoted π(x)
in the rest of this paper. The prime Rx number theorem states that π(x) ∼ li(x), where the
logarithmic P integral function li(x) = 0 log−1 t dt can be written as an asymptotic expansion
∞
li(x) ∼ logx x n=0 (logn!x)n . Bounds on the error term have been established in the literature,
including the recent work of Trudgian [22], who proved that, for sufficiently large x,
r !
 
π(x) − li(x) ≤ 0.2795 x log x
exp − .
(log x)3/4 6.455
This implies the existence of constants d1 and d2 such that
 
π(x) − li(x) ≤ d1 x  p 
exp −d 2 log x , for all x ≥ 2. (1.1)
(log x)3/4
Assuming the Riemann hypothesis, Rosser and Schoenfeld [19, 20] established even sharper
bounds on the error term, including
√
x log x
|π(x) − li(x)| <
8π
for large enough x. Other explicit estimates of π(x), in terms of x and log x are achievable, as
proved by Axler [1].
In this paper, we focus on the case k = 2, where the numbers written as products of two
(not necessarily distinct) primes are called semiprimes. In this case, Ishmukhametov and
Sharifullina [14] recently used probabilistic arguments to approximate the behaviour of π2 (x)
as
x log(log x) x x
π2 (x) ≈ + 0.265 − 1.540 . (1.2)
log x log x (log x)2

2000 Mathematics Subject Classification 11N37 (primary), 11K65, 11Y60 (secondary). .

Page 2 of 13 DRAGOS, CRIS, AN AND RADEK ERBAN

The first term of (1.2) has already been known to Landau [15, §56], with his result stated, for
general k ∈ N, as
1 x (log(log x))k−1
πk (x) ∼ . (1.3)
(k − 1)! log x
Delange [6, Theorem 1] obtained the asymptotic expansion of πk (x) in the form
∞
x X Pn,k (log(log x))
πk (x) ∼ , (1.4)
log x n=0 (log x)n
where Pn,k are polynomials of degree k − 1, with the leading coefficient equal to n!/(k − 1)! .
Tenenbaum [21] proved a similar result, giving an Q expression
for the

coefficients in the
1 z 1 z
polynomial P0,k in terms of the derivatives of Γ(z+1) p 1 + p−1 1 − p evaluated at z = 0.
Considering k = 2 in (1.4), we can write an asymptotic expansion for π2 (x) as
∞ ∞
X x log(log x) X x
π2 (x) ∼ (n − 1)! n
+ Cn−1 . (1.5)
n=1
(log x) n=1
(log x)n
In Theorem 2.3, we prove that C0 = M , where M = 0.261497... is the Meissel–Mertens constant
defined by
!
X 1
M = lim − log(log x) , (1.6)
x→∞ p
p≤x

where we sum over all primes such that p ≤ x. In Section 3, we calculate the rest of constants
Cn appearing in equation (1.5). They are given in Table 1 and obtained by the formula
n n n
! !
X Bi X 1 X Bi
Cn = n! − = n! − Hi , (1.7)
i=0
i! i=1
i i=0
i!
where Hi is the i-th harmonic number, B0 = M and constants Bi are defined using the
asymptotic behaviour of sums [15, §56]
X (log p)i (log x)i  14√ 
= + Bi + O e− log x , for i ∈ N. (1.8)
p i
p≤x

Constants Bi are given as limits (3.1) in Section 3, where we present an algorithm to efficiently
calculate them to a desired accuracy. They are computed in Table 1 to 20 significant digits.
Rosser and Schoenfeld [18] prove that the error term in (1.8) can be given explicitly in terms
of an integral, which contains the error terms in the prime number theorem. For the case i = 1
in equation (1.8), explicit estimates of this sum and, in particular, of the constant B1 involved,
were recently obtained by Dusart [8].
A related arithmetic function, Ω(m), is defined to be the number of prime divisors of m ∈ N,
where prime divisors are counted with their multiplicity. Considering fixed x in equation (1.3),
we can view this approximation of πk (x)/x as the probability mass function of the Poisson
distribution with mean log(log(x)). Erdős and Kac [9] showed that the distribution of Ω(x)
is Gaussian with mean log(log(x)) (see also Rényi and Turán [17] and Harper [13] for
generalizations and better bounds). Diaconis [7] obtained the asymptotic expansions for the
average number of prime divisors as (see also Finch [11, Section 1.4.3])
n−1
∞
!
1 X X X γi (n − 1)!
Ω(m) ∼ log(log x) + 1.0346538818 · · · + −1 + , (1.9)
x n=1 i=0
i! logn x
m≤x

where the constants γi are the Stieltjes constants, numerically computed in [3] to 20 significant
digits. An asymptotic series for the variance of Ω have also been obtained [11, Section 1.4.3].
 ON THE COUNTING FUNCTION OF SEMIPRIMES Page 3 of 13

The Stieltjes constants γi are used in Section 3 during our calculation of the values of constants
Bn and Cn , for n ∈ N.
This paper is organized as follows. Section 2 begins with a counting lemma for expressing
the semiprime counting function π2 in terms of the prime counting function π. Using this
lemma, the main results on the asymptotic behaviour of π2 are stated and proved in Section 2
as Theorem 2.3 and Theorem 2.5. While Theorem 2.3 only gives the first two terms, its proof
is more coincise than the proof of Theorem 2.5, which gives the full asymptotic series of π2 .
The constants Cn which appear in this asymptotic series are computed in Section 3, where we
present an efficient approach to calculate both constants Bn and Cn , based on the differentiation
of the prime zeta function. In Section 4, we investigate the behaviour of the error terms given
by the partial sums of the asymptotic series of π2 . We conclude with a generalization of the
counting argument in Section 5, discussing the extensions of the presented results to the general
case of counting functions πk for k ≥ 3.

2. Asymptotic behaviour of the counting function of semiprimes

As in equation (1.6), we denote primes by p and the sums over p shall be understood as sums
over all primes satisfying the given condition. In the case of summing over primes twice, we
denote the corresponding prime summation indices by p1 and p2 . We begin with a simple
counting formula [14], that gives a way of computing π2 (x).

Lemma 2.1. For a positive integer x, the following holds

√ √
π( x) − π( x)2 X x
π2 (x) = + π . (2.1)
2 √ p
p≤ x

Proof. By the definition of counting functions π2 and π, we have

X X X X  x 
π2 (x) = 1 = 1 = π − π(p1 ) + 1
√ x √ p1
p1 ≤p2 p1 ≤ x p1 ≤p2 ≤ p1 p1 ≤ x
p1 p2 ≤x

and formula (2.1) follows by renaming p1 to p in the first term and observing
√ that the rest of
the right hand side is the sum of all natural numbers from 1 up to π( x) − 1.

Formula (2.1) gives an expression of π2 (x) in terms of the prime counting function π(x), which
can be approximated using the prime number theorem [21] as
 
x
π(x) = αn (x) + O , (2.2)
(log x)n+1
where n ∈ N and
n−1
!
x X i!
αn (x) = . (2.3)
log x i=0
(log x)i
Using Landau [15, §56], we can rewrite equation (1.6) for any integer n ∈ N as
X 1  
1
= log(log x) − log 2 + M + o , (2.4)
√ p (log x)n
p≤ x

where we use the little o asymptotic notation [5], as opposed to the big O asymptotic notation
used in the prime number theorem (2.2). First we use this result to approximate the sum on
the right hand side of equation (2.1).
Page 4 of 13 DRAGOS, CRIS, AN AND RADEK ERBAN

Lemma 2.2. Let n ∈ N and αn (x) be defined by (2.3). Then we have

X x X  
x

x log(log x)

π = αn +O . (2.5)
√ p √ p (log x)n+1
p≤ x p≤ x

Proof. Using equation (2.2), we have

 
     
X x X x  X  x  
x  X 1
 √ π p − √ αn p  ≤ √ π p − αn p  ≤ c x √ p (log x − log p)n+1 ,
 
p≤ x p≤ x  p≤ x p≤ x

where c > 0 is a constant. Equation (2.5) then follows by estimating the right hand side by
c x 2n+1 X 1
 
x log(log x)
= O ,
(log x)n+1 √ p (log x)n+1
p≤ x

where the last equality follows from equation (2.4).

2.1. The first two terms of the asymptotic series for π2 (x)
Using (2.5) for n = 1, we obtain
X x X x

x log(log x)

π = +O . (2.6)
√ p √ p (log x − log p) (log x)2
p≤ x p≤ x

Using Landau [15, §56], we can rewrite equation (1.8) for any integers i ∈ N and n ∈ N as
X (log p)i (log x)i

1

= + B i + o . (2.7)
√ p i 2i (log x)n
p≤ x

We will use this to prove the first theorem of this section.

Theorem 2.3. Let M be the Meissel–Mertens constant defined by (1.6). Then

 
x log(log x) x x
π2 (x) = +M +o . (2.8)
log x log x log x

Proof. Using equations (2.1), (2.2) for n = 1, and (2.6), we obtain

X x 
x

x X log x

x

π2 (x) = π +o = +o . (2.9)
√ p log x log x √ p (log x − log p) log x
p≤ x p≤ x

We have the following identity

 
log x 1 log p log x
= + .
p (log x − log p) p log x p (log x − log p)
Substituting the left hand side into the right hand side, we obtain, for any natural number
n ∈ N, that
n  i  n+1  
log x 1 X log p log p log x
= + .
p (log x − log p) p i=0 log x log x p (log x − log p)
√
Summing over all primes p ≤ x , we get
 
n i
X log x X
 1
X (log p)  1 X (log p)n+1
= i
+ n
.
√ p (log x − log p) (log x) √ p (log x) √ p (log x − log p)
p≤ x i=0 p≤ x p≤ x
 ON THE COUNTING FUNCTION OF SEMIPRIMES Page 5 of 13

Each term on the right hand side can be evaluated using equations (2.4) and (2.7), with o(1)
accuracy, as
n
X log x X 1
= log(log x) − log 2 + M + + f (n) + o(1), (2.10)
√ p (log x − log p) i=1
i 2i
p≤ x

where f (n) is a decreasing function of n satisfying f (n) → 0 as n → ∞. This can be deduced

by applying equation (2.7) to the error term
1 X (log p)n+1 2 X (log p)n+1 1
n
≤ n+1
= + o(1).
(log x) √ p (log x − log p) (log x) √ p (n + 1) 2n
p≤ x p≤ x

Since we have
n
X 1
− log 2 + + f (n) → 0, as n → ∞,
i=1
i 2i

equation (2.10) implies that

X log x
= log(log x) + M + o(1).
√ p (log x − log p)
p≤ x

Substituting into equation (2.9), we obtain formula (2.8).

2.2. Asymptotic series for the counting function of semiprimes

To derive formulas for all terms in the asymptotic series of π2 , we first define an auxiliary
sequence of numbers qn for n ∈ N by
n−1
X 2i − 1
qn = , for n ≥ 2, and q1 = 0 . (2.11)
i=1
i
Then qn is an increasing sequence of rational numbers with the first few terms given as q1 = 0,
q2 = 1, q3 = 5/2, q4 = 29/6, q5 = 103/12 and q6 = 887/60, which satisfies the following identity.

Lemma 2.4. Let n ∈ N and let qn be given by equation (2.11). Then we have
∞  
X n+i−1 1
= qn + log 2 . (2.12)
i=1
n−1 i 2i

Proof. Considering the binomial series

∞  
X n+i
(1 − t)−n = 1 + t ti , for t ∈ (−1, 1),
i=0
n−1
we can rewrite it as
∞  n
(1 − t)−n − 1 X

X n+i i
t = = (1 − t)−i .
i=0
n−1 t i=1

Integrating, we get
∞  n
n + i ti+1 (1 − t)−i+1 − 1
X  X
= − log(1 − t) + ,
i=0
n−1 i+1 i=2
i−1

which holds for t ∈ (−1, 1). Substituting t = 1/2, we obtain (2.12).

Page 6 of 13 DRAGOS, CRIS, AN AND RADEK ERBAN

We will use Lemma 2.4 in the proof of the following theorem, giving the asymptotic series for
the semiprime counting function π2 (x).

Theorem 2.5. The constants Cn appearing in the asymptotic expansion (1.5) are given
by equation (1.7) for n ∈ N and as C0 = B0 = M for n = 0.

Proof. The case n = 0 is studied in Theorem 2.3, which states that C0 = M. To derive
equation (1.7), we again use formula (2.1) from Lemma 2.1 and approximate each term using
the prime number theorem (2.2). We need to analyze sums of the form
−n
X (log x)n X 1 log p
Sn (x) = n
= 1− . (2.13)
√ p (log x − log p) √ p log x
p≤ x p≤ x

Using the binomial series on the right hand side, we get

X 1X ∞   i ∞   X (log p)i
n+i−1 log p X n+i−1 1
Sn (x) = = i
. (2.14)
√ p n−1 log x n−1 (log x) √ p
p≤ x i=0 i=0 p≤ x

Substituting x2 for x, we obtain

∞   X (log p)i
X n+i−1 1
Sn (x2 ) = . (2.15)
i=0
n−1 2i (log x)i p
p≤x

To estimate the sums over primes on the right hand side, we apply the result of Rosser and
Schoenfeld [18, equation (2.26)], which can be formulated as
X1 X (log p)i (log x)i
= log(log x) + L0 (x), = + Li (x),
p p i
p≤x p≤x

where the error terms Li (x) are defined by

Zx
li(2) π(x) − li(x) π(y) − li(y)
L0 (x) = − log(log 2) + + + dy,
2 x 2 y2
(log 2)i (log 2)i li(2) (log x)i

Li (x) = − + + π(x) − li(x)
Zxi 2 x
(log y − i) (log y)i−1

+ 2
π(y) − li(y) dy, for i ∈ N.
2 y
Using this notation and identity (2.12) in Lemma 2.4, we rewrite equation (2.15) as
∞  
2
X n + i − 1 1 Li (x)
Sn (x ) = log(log x) + qn + log(2) + . (2.16)
i=0
n−1 2i (log x)i

Using the inequality (1.1), there exist constants d1 and d2 such that
Zx
|i − log y| (log y)i−1 
 
 Li (x)  2 |π(x) − li(x)| 1 
 (log x)i  (log x)i +
 < + π(y) − li(y) dy
x i
(log x) 2 y 2
√

2 exp −d2 log x i I(i − 1, x) + I(i, x)
≤ + d1 + , (2.17)
(log x)i (log x)3/4 (log x)i
where we define (note that we allow the second argument to be ∞ in this definition):
Zx √

i−3/4 exp −d2 log y
I(i, x) = d1 (log y) dy .
2 y
 ON THE COUNTING FUNCTION OF SEMIPRIMES Page 7 of 13

Choose ℓ ∈ N. Our goal is to use (2.17) to estimate the rate of convergence of the sum on the
right hand of equation (2.16). To do this we first observe that, for i ≥ ℓ, we have
Zx √

I(i, x) d1 ℓ−3/4 exp −d2 log y I(ℓ, ∞)
i
≤ ℓ
(log y) dy ≤ . (2.18)
(log x) (log x) 2 y (log x)ℓ
Using inequalities (2.17) and (2.18) and assuming log(x) ≥ 1, we can estimate the remainder
of the series on the right hand of equation (2.16) as
 ∞   √
! ∞ 
 X n + i − 1 1 L (x) 
i 2 exp −d2 log x X n + i − 1 1
 ≤ + d1
 
n−1 2i (log x)i  (log x)ℓ+1 (log x)3/4 n−1 2i


i=ℓ+1 i=ℓ+1
∞   ∞  
I(ℓ, ∞) X n + i − 1 i I(ℓ + 1, ∞) X n + i − 1 1
+ + .
(log x)ℓ+1 n−1 2i (log x)ℓ+1 n−1 2i
i=ℓ+1 i=ℓ+1

Since all three sums on the right hand side
converge independently of x, we deduce that the
remainder is of the order O (log x)−(ℓ+1) ) . Therefore, equation (2.16) becomes
ℓ    
2
X n + i − 1 1 Li (x) 1
Sn (x ) = log(log x) + qn + log(2) + +O .
i=0
n−1 2i (log x)i (log x)ℓ+1

This means that an asymptotic expansion of Sn (x2 ) in terms of negative powers of log x is
given by the sum of the asymptotic series of terms in equation (2.15). The same is true for
Sn (x) in equation (2.14). Thus, using equations (2.4), (2.7), (2.12) and (2.14), we obtain
ℓ   X (log p)i  
X n+i−1 1 1
Sn (x) = +O
i=0
n−1 (log x)i √ p (log x)ℓ+1
p≤ x
ℓ    
X n+i−1 Bi 1
= log(log x) + M + qn + +O . (2.19)
i=1
n−1 (log x)i (log x)ℓ+1

Using equations (2.3) and (2.5), we have

X x   X ℓ ℓ
x X x (n − 1)! X x Sn (x)
π +o ℓ
= n
= (n − 1)! ,
√ p (log x) n=1
√ p (log x − log p) n=1
(log x)n
p≤ x p≤ x

where we used the definition (2.13) of Sn (x) to get the second equality. Using equation (2.19)
and notation B0 = M , we obtain
ℓ ℓ n−1
!
X x X x log(log x) X X Bi x

x

π = (n − 1)! + (n − 1)! q n + + o .
√ p n=1
(log x)n n=1 i=0
i! (log x)n (log x)ℓ
p≤ x

Thus, using formula (2.1) and the prime number theorem (2.2), we obtain the asymptotic
expansion (1.5), where we have
n n
!
X Bi X
Cn = n! qn+1 + − 2n (i − 1)! (n − i)! .
i=0
i! i=1

This can be further simplified by using definition (2.11) of qn . We get

n n n−1 n n n
! !
X 2i − 1 X Bi 2n X 1 X 2i − 1 X Bi X 2i
Cn = n! + −
= n! + − .
i=1
i i=0
i! n i=0 n−1 i i=1
i i=0
i! i=1
i

Subtracting the first and the third sum, we obtain (1.7).

Page 8 of 13 DRAGOS, CRIS, AN AND RADEK ERBAN

3. Computing the constants

In this section, we use a fast converging series to determine the values of the constants Bn and,
as a result, of the constants Cn given by equation (1.7). The first constant, B0 = C0 = M , is
the well-studied Meissel–Mertens constant, so we will focus on constants Bn in the case n ≥ 1.
They have been defined by equations (1.8) or (2.7), which can be rewritten as
   
X (log p)n n
(log x)  X (log p) n n
(log x) 
Bn = lim  − = lim  − . (3.1)
x→∞ √ p n 2n x→∞ p n
p≤ x p≤x

To derive a formula for evaluating Bn on a computer, we use the prime zeta function [12]
defined by
X 1
P (s) = , for s ∈ (1, ∞). (3.2)
p
ps

Differentiating equation (3.2), we get the formula for the n-th derivative of the prime zeta
function as
X (log p)n X (log p)n X (log p)n
P (n) (s) = (−1)n = (−1) n
+ (−1) n
. (3.3)
p
ps ps p>x
ps
p≤x

The prime zeta function P (s) can also be related to the Riemann zeta function ζ(s) through
the formula [12]
∞

X log ζ(i s)
P (s) = µ(i) , for s ∈ (1, ∞),
i=1
i
where µ(n) is the Möbius function. Taking the derivative of order n of this expression, we
obtain
∞  ′ (n−1)
(n)
X
n−1 ζ
P (s) = µ(i) i (i s) .
i=1
ζ
Substituting into (3.3), we obtain
X (log p)n ∞  ′ (n−1) X (log p)n
n
X
n−1 ζ
s
= (−1) µ(i) i (i s) − . (3.4)
p i=1
ζ p>x
ps
p≤x

Using integration by parts, we obtain

Z∞ (n−1)
(log u)n−1 (−1)n

(n − 1)!
du = =− .
1 us (s − 1)n s−1
Thus, the second sum on the right hand side of equation (3.4) can be approximated by
X (log p)n Z∞ (n−1) Z x
(log u)n−1 (−1)n (log u)n−1

= du + o(1) = − − du + o(1) .
p>x
ps x us s−1 1 us

Substituting into equation (3.4), we get

X (log p)n Z x (log u)n−1 X∞  ′ (n−1)
ζ

(−1)n
(n−1)
n n−1
− du = (−1) µ(i) i (i s) + + o(1) .
ps 1 us i=1
ζ s−1
p≤x

Taking the limit as s → 1 and substituting into equation (3.1), we obtain

∞  ′ (n−1)  ′ (n−1)
n
X
n−1 ζ n ζ (s) 1
Bn = (−1) µ(i) i (i) + (−1) lim + , (3.5)
i=2
ζ s→1 ζ(s) s−1
 ON THE COUNTING FUNCTION OF SEMIPRIMES Page 9 of 13

n Bn Cn

0 0.26149721284764278375 0.26149721284764278375
1 −1.3325822757332208817 −2.0710850628855780875
2 −2.5551076154464547041 −7.6972777412176108802
3 −10.253827096911327612 −35.345660320564161516
4 −59.332397971808450296 −206.71503925406509339
5 −453.62459086132753356 −1.5111997871316530251 × 103
6 −4.3591249600559955673 × 103 −1.3546323682845914021 × 104
7 −5.0684840978914262902 × 104 −1.4622910675883565523 × 105
8 −6.9270677393697978276 × 105 −1.8675796280076650637 × 106
9 −1.0884508606344556845 × 107 −2.7733045258413542557 × 107
10 −1.9329009099289751454 × 108 −4.7098342357703294361 × 108

Table 1. Table of constants Bn and Cn , for n = 0, 1, 2, . . . , 10, defined by equations (2.7) and (1.7),
which appear in the asymptotic expansion of π2 (x). The values of constants Bn are computed by
formula (3.5) using the Laurent series (3.6). The values of constants Cn are computed by
equation (1.7).

where the first term on the right hand side is a quickly converging series and the limit in the
second term can be evaluated using the Laurent expansion of ζ(s) around s = 1. This is given
by
∞
1 X (−1)n
ζ(s) = + γn (s − 1)n ,
s − 1 n=0 n!
where the Stieltjes constants γn are computed to 20 significant digits in [3]. Then, the Laurent
expansion of the logarithmic derivative [4] of the Riemann zeta function is
ζ ′ (s)
 
1 3
=− + γ0 + (−2γ1 − γ02 )(s − 1) + γ2 + 3γ0 γ1 + γ03 (s − 1)2
ζ(s) s−1 2
 
2
+ − γ3 − 2γ0 γ2 − 2γ12 − 4γ02 γ1 − γ04 (s − 1)3 + . . . . (3.6)
3
Constants Bn computed by formula (3.5) using the Laurent series (3.6) are presented in Table 1.
Once we know constants Bn to the desired accuracy, we can use equation (1.7) to calculate
constants Cn . They are also presented in Table 1 to 20 significant digits.

4. Computational results: behaviour of error terms

In this section, we illustrate the accuracy of the asymptotic series (1.5) by calculating its error
terms at each order. Since (1.5) is a sum of two formal asymptotic series, we have two ways to
define its errors. First, we can truncate both sums after the same number, ℓ, of terms to define
the relative error
 
ℓ ℓ
!
1  X x log(log x) X x 
ε2ℓ (x) = (n − 1)! + Cn−1 − π2 (x) , for ℓ ∈ N, (4.1)

π2 (x)  n=1

(log x) n (log x) n 
n=1

which is plotted in Figure 1(a) for x = 106 , x = 108 and x = 1010 . Combining both sums in
the asymptotic expansion (1.5) into one, we can write it as
x log(log x) x x log(log x) x 2 x log(log x)
π2 (x) ∼ + C0 + 2
+ C1 2
+ + ...,
log x log x (log x) (log x) (log x)3
Page 10 of 13 DRAGOS, CRIS, AN AND RADEK ERBAN

(a) (b)

10 -2 10 -1

10 -2
10 -3
10 -3
10 -4
10 -4
10 -5 10 -5

1 2 3 4 5 6 7 8 9 10 1 4 7 10 15 20

Figure 1. (a) Relative errors calculated by (4.1) for x = 106 , x = 108 and x = 1010 .
(b) Relative errors calculated by (4.5) for x = 108 and x = 1010 . The relative error of Landau’s
approximation (4.2) is plotted as ε1 (x). The relative error of the approximation given in Theorem 2.3
is plotted as ε2 (x).

so we can define the n-th approximation, an (x), by

x log(log x)
a1 (x) = , (4.2)
log x
ℓ ℓ
X x log(log x) X x
a2ℓ (x) = (i − 1)! i
+ Ci−1 , for ℓ = 1, 2, 3, . . . , (4.3)
i=1
(log x) i=1
(log x)i
ℓ+1 ℓ
X x log(log x) X x
a2ℓ+1 (x) = (i − 1)! + Ci−1 , for ℓ = 1, 2, 3, . . . , (4.4)
i=1
(log x)i i=1
(log x)i
where approximation a1 (x) is used in equation (1.3) and α2 (x) is the approximation given in
equation (2.8). With definitions (4.2)–(4.4), we can define the relative error by
|an (x) − π2 (x)|
εn (x) = . (4.5)
π2 (x)
This definition is consistent with (4.1) for even values of n. The results computed by (4.5) are
plotted in Figure 1(b) for x = 108 and x = 1010 .
In Figure 1, we study the behaviour of the relative error (4.5) for a fixed value of x. We
observe that the relative error εn (x) is initially a decreasing sequence (for smaller values of n).
It reaches its minimum and then it starts increasing again. Denoting the value of n where the
relative error reaches its minimum as nmin (x), we observe that nmin (x) is an increasing function
of x, at least for the three values of x considered in Figure 1(a).
Since x is fixed in Figure 1, the relative error (4.5) is a constant multiple of the absolute
error |an (x) − π2 (x)|. In particular, the same conclusion about nmin (x) could be reached when
considering the absolute errors instead of relative errors. Since we plot results for two or three
different (fixed) values of x in Figure 1, we can also observe that the relative error is (for many
of the values of n) a decreasing function of x, although this does not look to be true for n = 2
in Figure 1.
To confirm this observation, we plot the behaviour of the relative error εn (x) as a function
of x in Figure 2. We use the first eight values of n and confirm that the relative error εn (x)
resembles a decreasing function of x for n = 1, 3, 4, . . . , 8, while ε2 (x) reaches its minimum
between 2 × 105 and 3 × 105 . For smaller values of x, the approximation a2 (x) overestimates
 ON THE COUNTING FUNCTION OF SEMIPRIMES Page 11 of 13

(a) (b)

10 -1
10 -2

10 -2

10 -3
10 -3

10 -4
5 6 7 8 9 10
10 10 10 10 10 10 10 5 10 6 10 7 10 8 10 9 10 10

Figure 2. (a) Relative errors εn (x), for n = 1, 2, 3 and 4, calculated by (4.5) as a function of x. The
relative error of Landau’s approximation (4.2) is plotted as ε1 (x). The relative error of the
approximation given in Theorem 2.3 is plotted as ε2 (x).
(b) Relative errors εn (x), for n = 5, 6, 7 and 8 as a function of x.

π2 (x), while it is an underestimate for larger values of x. The graphs of a2 (x) and π2 (x) cross
for the values of x between 2 × 105 and 3 × 105 , so in this interval we can get very close to
the correct answer, which results in the minimum of ε2 (x) in Figure 2(a), because our error
definition (4.5) includes the absolute value. In Figure 2, we also observe that the other errors
are not strictly decreasing, but they fluctuate with a decreasing trend, see, for example, the
plot of ε8 (x) in Figure 2(b).

5. Discussion
In this paper, we have studied the behaviour of the semiprime counting function π2 (x), which
is a special case (k = 2) of the k-almost prime counting function πk (x). To generalize the
presented results to the case k ≥ 3, we need to first generalize the counting Lemma 2.1. Using
the inclusion-exclusion principle, it is possible to deduce the following counting formula
k  
X X x
πk (x) = (−1)i−1 πk−i , (5.1)
i=1
√k
p 1 p2 . . . pi
p1 <p2 <...<pi ≤ x

where we define function

√ π0 (x) to be identically equal to 1, i.e. π0 (x) = 1, and the sum over
p1 < p2 < . . . < pi ≤ k x means that we are summing i-times over all primes satisfying the
given condition. Substituting k = 2 into equation (5.1), we obtain
X x X 
x

π2 (x) = π1 − π0 .
√ p1 √ p1 p2
p1 ≤ x p1 <p2 ≤ x

Using π0 (x) = 1, we deduce equation (2.1). Thus, equation (5.1) provides a generalization of
equation (2.1), which expresses the k-almost prime counting function πk (x) in terms of the
counting functions π1 (x), π2 (x), . . . , πk−1 (x). It can be inductively used to derive forms of
coefficients of polynomials Pn,k in the asymptotic series (1.4). In addition to constants Bn and
Cn , certain new
P constants will appear in such calculations, including the (converging) sums
of the form p (log p)i p−ℓ with ℓ ≥ 2 and i ∈ N. For a detailed discussion of the asymptotic
behaviour of these sums for ℓ = 1, see Axler [2]. Substituting n = π(x) in [2, Theorem 5] gives
Page 12 of 13 DRAGOS, CRIS, AN AND RADEK ERBAN

a different expansion for the sums in (1.8), which may be further examined using the prime
number theorem.
There are, also, other possible approximations for πk (x). For example, Erdős and Sárkőzy [10]
prove that
k−1

c(δ) x (log(log x))

, for 1 ≤ k ≤ (2 − δ) log(log x);
πk (x) < log x (k − 1)!
 4 −k

c k 2 x log x , for k ≥ 1;
for some constants c(δ) and c. Other approximations, relating the function πk to some other
products over primes are possible to obtain, as explained in [21].
Functions πk (x) and Ω(x), used in expansion (1.9), count the prime divisors with their
multiplicity. Another possible generalization is to investigate the related functions Nk (x) and
ω(x), counting prime divisors without multiplicity. That is, functions Nk (x) and ω(x) are
defined to be the number of natural numbers n ≤ x which have exactly k distinct prime divisors
and the number of distinct prime divisors of x, respectively. Finch [11, page 26] shows that
n−1
∞
!
1 X X X γi (n − 1)!
ω(m) ∼ log(log x) + 0.2614972128 · · · + −1 + ,
x n=1 i=0
i! logn x
m≤x

which has the higher order terms in the same form as in √the expansion
√ (1.9). Using the prime
number theorem, we also observe that N1 (x) = π (x) + π ( x) + π ( 3 x) + . . . ∼ li(x) admits an
identical asymptotic expansion as π(x). Delange [6, Theorem 1] and Tenenbaum [21] obtained
the asymptotic expansion of Nk (x) in the form
∞
x X Qn,k (log(log x))
Nk (x) ∼ ,
log x n=0 (log x)n

where Qn,k are polynomials of degree k − 1. Here, the expansion is similar to the expansion (1.4)
for πk (x), but the polynomials Pn,k and Qn,k are different. Results about the leading terms of
polynomials Qn,k and about Q0,k have also been obtained, as in the case of πk . Several different
approximations for Nk (x) are also possible to derive, as shown in Tenenbaum [21], who points
out that the function Nk is easier to analyse than πk , for larger values of k, relative to log(log x).
For example, the following holds uniformly for x ≥ 3 and (2 + δ) log(log x) ≤ k ≤ A log(log x):
x log x  
−δ 2 /5

Nk (x) = C 1 + O (log x) ,
2k
where A > 0, 0 < δ < 1 and C ≈ 0.378694. Similar results, but for larger values of k, can be
obtained for functions πk as well [16].

Acknowledgements. Authors would like to thank Julia Stadlmann for helpful comments
during the preparation of this manuscript.

References
1. C. Axler. New bounds for the prime counting function. Integers, 16:A22, 2006.
2. C. Axler. On a family of functions defined over sums of primes. Journal of Integer Sequences, 22(5):Article
19.5.7, 2019.
3. J. Bohman and C. Fröberg. The Stieltjes function – definition and properties. Mathematics of Computation,
51(163):281–289, 1988.
4. K. Broughan. Extension of the Riemann ζ-functions logarithmic derivative positivity region to near the
critical strip. Canadian Mathematical Bulletin, 52(2):186–194, 2009.
5. N. de Bruijn. Asymptotic Methods in Analysis. North-Holland Publishing Co., Amsterdam, 1958.
6. H. Delange. Sur des formules de Atle Selberg. Acta Arithmetica, 19:105–146, 1971.
 ON THE COUNTING FUNCTION OF SEMIPRIMES Page 13 of 13

7. P. Diaconis. Asymptotic expansion for the mean and variance of the number of prime factors of a number
n. Technical Report 96, Department of Statistics, Stanford University, Stanford, California, December
1976.
8. P. Dusart. Explicit estimates of some functions over primes. Ramanujan Journal, 45:227–251, 2018.
9. P. Erdős and M. Kac. The Gaussian law of errors in the theory of additive number theoretic functions.
American Journal of Mathematics, 62(1):738–742, 1940.
10. P. Erdős and A. Sárkőzy. On the number of prime factors of integers. Acta Scientiarum Mathematicarum,
42:237–246, 1980.
11. S. Finch. Mathematical Constants II. Cambridge University Press, Cambridge, 2018.
12. C. Fröberg. On the prime zeta function. BIT, 8:187–202, 1968.
13. A. Harper. Two new proofs of the Erdős-Kac theorem, with bound on the rate of convergence, by Stein’s
method for distributional approximations. Mathematical Proceedings of the Cambridge Philosophical
Society, 147(1):95114, 2009.
14. S. Ishmukhametov and F. Sharifullina. On distribution of semiprime numbers. Izvestiya Vysshikh
Uchebnykh Zavedenii. Matematika, 8:53–59, 2014.
15. E. Landau. Handbuch der Lehre von der Verteilung der Primzahlen. Druck und Verlag von B.G.Teubner,
Lepzig und Berlin, 1909.
16. J. Nicolas. Sur la distribution des nombres entiers ayant une quantité fixée de facteurs premiers. Acta
Arithmetica, 44(3):191–200, 1984.
17. A. Rényi and P. Turán. On a theorem of Erdős-Kac. Acta Arithmetica, 4(1):71–84, 1958.
18. J. Rosser and L. Schoenfeld. Approximate formulas for some functions of prime numbers. Illinois Journal
of Mathematics, 6(1):64–94, 1962.
19. J. Rosser and L. Schoenfeld. Sharper bounds for the Chebyshev functions θ(x) and ψ(x). Mathematics of
Computation, 29(129):243–269, 1975.
20. L. Schoenfeld. Sharper bounds for the Chebyshev functions θ(x) and ψ(x). II. Mathematics of
Computation, 30(134):337–360, 1976.
21. G. Tenenbaum. Introduction to Analytic and Probabilistic Number Theory, volume 163 of Graduate
Studies in Mathematics. American Mathematical Society, 2015.
22. T. Trudgian. Updating the error term in the prime number theorem. Ramanujan Journal, 39:225–234,
2016.

Dragos, Cris, an Radek Erban

Merton College Mathematical Institute
Merton Street University of Oxford
Oxford OX1 4JD Woodstock Road
United Kingdom Oxford OX2 6GG
United Kingdom
[email protected]
[email protected]