1

2 Journal of Integer Sequences, Vol. 24 (2021),

3
47 6
Article 21.9.5
23 11

Sum of Reciprocals of Germain Primes

Samuel S. Wagstaff, Jr.

Center for Education and Research in Information Assurance and Security
and
Department of Computer Sciences
Purdue University
West Lafayette, IN 47907-1398
USA
[email protected]

Abstract
A prime p is a Germain prime if 2p + 1 is also prime. We compute the sum of
the reciprocals of the Germain primes and related sequences. Since we do not know
whether there are infinitely many Germain primes, all we can do is bound the sum in
an interval.

1 Introduction
In 1919, Brun [4] proved that the sum
1 1 1 1 1 1 1 1 1 1
B= + + + + + + + + + + ···
3 5 5 7 11 13 17 19 29 31
of the reciprocals of all twin primes either converges or is a finite sum. Various authors
[17] have estimated Brun’s constant B. Sebah [16] computed the sum of the reciprocals of
all twin primes < 1016 and found that B > 1.830. Klyve [8] showed that B < 2.347 as
reported in Crandall and Pomerance [5]. Platt and Trudgian [14] used the number of twin
primes < 4 · 1018 computed by Oliveira e Silva [13] and found that 1.840503 < B < 2.288513.
Several authors [3, 2, 10, 11, 13, 14, 16, 17] agree that the most probable value of B is about
1.90216, but this estimate is not rigorous. These results show how little we know about the
distribution of twin primes, even whether there are infinitely many of them. The twin primes
are Sequence A001097 in the OEIS [12].

1
 A prime p is a Germain prime if 2p + 1 is also prime. They are named after Sophie
Germain, who proved nearly 200 years ago that the first case of Fermat’s Last Theorem is
true when the exponent is a Germain prime. The Germain primes are Sequence A005384
in the OEIS. Computing the sum of the reciprocals of the Germain primes is similar to
computing Brun’s constant B. First, it is easy to prove using a sieve that this sum either
converges or is a finite sum. Second, we do not know whether there are infinitely many
Germain primes, so the best lower bound for the sum is the partial sum up to the limit to
which we can compute it directly. One can compute a most probable value for the sum just
as for B.
Little extra work is required to study the primes p for which 2k p + 1 is also prime for any
fixed k ≥ 1. The same methods apply to the sum of the reciprocals of the primes p for which
p + 2k is also prime. See Lee and Park [9] for the case of p, p + 8 both prime, for example.

2 Notation and easy results

Throughout this work p denotes a prime. Let
Y 1

c2 = 1− ≈ 0.6601618158468695739278121100
p>2
(p − 1) 2

be the twin prime constant as computed by Harley (See Lee and Park [9]).
Let a and b be positive integers with gcd(a, b) = 1 and 2 | ab. Let Sa,b = {p : ap +
b is prime}. For real x > 0, let Sa,b (x) denote the sum of p1 over all primes p ≤ x with
′
p ∈ Sa,b . Also let Sa,b 1
(x) be the sum of p1 + ap+b over all primes p ≤ x with p ∈ Sa,b .
Clearly one sum is finite if and only if the other is finite. Let Sa,b = limx→∞ Sa,b (x) and
′ ′
Sa,b = limx→∞ Sa,b (x).
′
Thus S1,2 = B, Brun’s constant, and

1 1 1 1 1 1 1 1
S2,1 = + + + + + + + + ···
2 3 5 11 23 29 41 53
is the sum of the reciprocals of the Germain primes. Also,
1 1 1 1 1 1 1 1
S4,1 = + + + + + + + + ···
3 7 13 37 43 67 73 79
and
1 1 1 1 1 1 1 1
S8,1 = + + + + + + + + ··· .
2 5 11 17 29 71 101 107
See Sequences A023212 and A023228 in the OEIS. The case of S1,8 was studied by Lee and
Park [9].
′
We show first that the limits Sa,b and Sa,b always exist. Let πa,b (x) be the number of
primes p ≤ x for which ap + b is also prime.

2
 ′ ′
Theorem 1. Suppose gcd(a, b) = 1 and 2 | ab. Define Sa,b (x), Sa,b (x), Sa,b and Sa,b as
′
above. Then the limits in the definitions of Sa,b and Sa,b exist.

Proof. Since Sa,b (x) is the sum of positive terms we need only show that it is bounded.
According to Theorem 3.12 of Halberstam and Richert [7], we have
  
cx log log x
πa,b (x) ≤ 1+O .
(log x)2 log x

for some constant c. Therefore,

x
X 1 X πa,b (t) − πa,b (t − 1)
Sa,b (x) = =
p≤x,p∈Sa,b
p t=1
t
x  
X 1 1
≤ − πa,b (t) (since πa,b (1) = 0)
t=2
t t+1
x   
X 1 ct log log t
≤ 1+O = O(1)
t=2
t(t + 1) (log t) 2 log t

′
as x → ∞. The proof for Sa,b is identical.
Fix x0 . Since 1/p > 0 always, we have Sa,b ≥ Sa,b (x0 ) for every (a, b). These are the best
lower bounds for Sa,b that we know how to compute. Table 1 shows these lower bounds for a
few values of x0 . We computed lower bounds for Brun’s constant for comparison, although
a better lower bound is found in Platt and Trudgian [14]. The sums were computed in
double precision, but only nine decimal places are shown in the table. Table 2 shows the
corresponding values of πa,b . The data for twin primes is exactly the same as Nicely [10]
found.
The most probable value for Sa,b is based on the Hardy-Littlewood [6] (or see Bateman
and Horn [1]) approximation to πa,b (x). This heuristic estimate is
Z x
dt Y p−1
πa,b (x) ≈ 2c2
2 (log t) p−2
2
p>2, p|ab

for fixed coprime integers a, b. When ab is a power of 2, which we always assume in this work,
the product over primes p > 2 is 1. Assuming this heuristic and using partial summation,
one finds Z ∞
dt 2c2
Sa,b − Sa,b (x0 ) ≈ 2c2 = .
x0 t(log t) log x0
2

Thus, after computing Sa,b (x0 ), the most probable value for Sa,b is Sa,b (x0 ) + 2c2 / log x0 . The
′
constant 2c2 is replaced by 4c2 for S1,2 because we add the reciprocals of two (consecutive)
primes in that case. These values are shown in Table 3.

3
 ′
x0 S1,2 (x0 ) S2,1 (x0 ) S4,1 (x0 ) S8,1 (x0 )
102 1.330990366 1.268745760 0.654987903 0.898299886
103 1.518032463 1.350207371 0.718127473 0.985924806
104 1.616893558 1.395040698 0.758673876 1.021448995
105 1.672799585 1.422243022 0.783466978 1.045080342
106 1.710776931 1.440222094 0.800867822 1.061448591
107 1.738357044 1.453223143 0.813394490 1.073447436
108 1.758815621 1.463079361 0.822866962 1.082581870
109 1.774735958 1.470771610 0.830302440 1.089777846
1010 1.787478503 1.476946485 0.836290746 1.095594911
1011 1.797904311 1.482013752 0.841220054 1.100393501
1012 1.806592419 1.486246659 0.845347507 1.104420686
2 · 1012 1.808931050 1.487387756 0.846461664 1.105509109
5 · 1012 1.811852563 1.488814135 0.847855248 1.106871419
1013 1.813943761 1.489835793 0.848854028 1.107848344
′
Table 1: Some values of S1,2 (x) and Sa,b (x).

x0 π1,2 (x0 ) π2,1 (x0 ) π4,1 (x0 ) π8,1 (x0 )

102 8 10 9 6
103 35 37 31 34
104 205 190 176 161
105 1224 1171 1057 1019
106 8169 7746 7422 7090
107 58980 56032 53709 51464
108 440312 423140 407198 392990
109 3424506 3308859 3198946 3095744
1010 27412679 26569515 25773602 25030013
1011 224376048 218116524 212205881 206597032
1012 1870585220 1822848478 1777532673 1734464714
2 · 1012 3552770943 3464491157 3380477421 3300430590
5 · 1012 8312493003 8112446112 7921805792 7740062386
1013 15834664872 15462601989 15107599980 14768799353

Table 2: Some values of πa,b (x0 ).

4
 ′
x0 S1,2 S2,1 S4,1 S8,1
102 1.904399633 1.555450394 0.941692537 1.185004520
103 1.900305309 1.541343794 0.909263896 1.177061229
104 1.903598191 1.538393015 0.902026193 1.164801312
105 1.902163292 1.536924875 0.898148831 1.159762196
106 1.901913353 1.535790305 0.896436033 1.157016803
107 1.902188263 1.535138753 0.895310100 1.155363045
108 1.902167938 1.534755520 0.894543121 1.154258029
109 1.902160239 1.534483751 0.894014581 1.153489986
1010 1.902160356 1.534287412 0.893631672 1.152935837
1011 1.902160541 1.534141868 0.893348170 1.152521616
1012 1.902160630 1.534030764 0.893131612 1.152204792
2 · 1012 1.902160522 1.534002492 0.893076401 1.152123846
5 · 1012 1.902160560 1.533968133 0.893009246 1.152025417
1013 1.902160571 1.533944198 0.892962433 1.151956749
′
Table 3: Most probable values of S1,2 and Sa,b .

The consistency of the estimates in the columns instills confidence, although there is no
proof that they are anywhere near the true values. One might consider “higher order” terms
in this approximation, but Shanks and Wrench [17] tell why this is unlikely to provide a
closer estimate; see Sections 3 and 5 of their paper.

3 Upper bound on the limit

The upper bound is based on Lemma 5 of Riesel and Vaughan [15], which we quote here
using Inequality (3.20) from their proof, as was done by Klyve [8].

Theorem 2. Let a and b be integers with a > 0, b 6= 0 and gcd(a, b) = 1. Let

X
R(x, a, b) = sup 1,
I
p∈I,ap+b prime

where the supremum is taken over all intervals of length x. Suppose that L and C = C(L)
are related by Table 4. Then, whenever x ≥ eL we have
√
  Y
16c2 x p−1
R(x, a, b) < +2 x .
(log x)(C + log x) p−2
p>2, p|ab

5
 L C L C
24 0.97 48 8.2054
25 2.31 60 8.302
26 3.40 82 8.3503
27 4.28 100 8.3708
28 5.00 127 8.3905
29 5.58 147 8.404
31 6.45 174 8.4102
34 7.24 214 8.4201
36 7.56 278 8.4301
42 8.04 396 8.44004
44 8.11 690 8.45001
Table 4: L and C of Theorem 2.

The product in Theorem 2 is 1 when ab √is a power of 2, as we always assume. Klyve [8]
found it more convenient not to have the 2 x term in the theorem. He gave an alternate
version, which we quote here.

Theorem 3. Let a and b be coprime positive integers with ab a power of 2. Suppose that L
and D = D(L) are related by Table 5. Then, whenever x ≥ eL we have
16c2 x
πa,b (x) < .
(log x)(D + log x)

The same bound applies when πa,b (x) is replaced by the count over any other interval of
length x ≥ eL .

In Table 5 we correct a tiny error in Klyve [8]. He had D = 2.30 when L = 25; it should
be D = 2.296.
L D L D
24 0.95 48 8.20
25 2.296 60 8.30
26 3.39 82 8.35
27 4.27 100 8.37
28 4.99 127 8.39
29 5.57 147 8.40
31 6.44 174 8.41
34 7.23 214 8.42
36 7.55 278 8.43
42 8.03 396 8.44
44 8.10 690 8.45

Table 5: L and D of Theorem 3.

6
 Note that the upper bound in Theorem 3 is independent of a and b, so long as ab is a
power of 2.
Theorem 4. We have 1.4898 < S2,1 < 1.8027, 0.8488 < S4,1 < 1.1617 and 1.1078 < S8,1 <
1.4208.
Proof. The lower bounds come from Table 1 with x0 = 1013 .
Let a and b be positive integers with gcd(a, b) = 1 and 2 | ab. Let 0 < M < N . Stieltjes
integration by parts yields
N Z N
X πa,b (t) − πa,b (t − 1) πa,b (N ) πa,b (M ) πa,b (t)
= − + dt . (1)
t=M
t N M M t 2

Now ∞
X πa,b (t) − πa,b (t − 1)
Sa,b = Sa,b (x0 ) + .
t=x
t
0

We will divide the interval [x0 , ∞) into segments with boundaries at eL for L in Table 5
to take advantage of the constants D(L) in that table. We will use Stieltjes integration to
bound Sa,b (N ) − Sa,b (M ) on each interval [M, N ).
Use Theorem 3 to bound the integral above. If L, L′ are consecutive entries in Table 5,
′
then the integral over [M, N ) = [eL , eL ) becomes
′ ′
eL eL
πa,b (t) 16c2 t
Z Z
dt ≤ dt
eL t2 eL t (log t)(D(L)
2 + log t)
L′
ds
Z
= 16c2 (s = log t)
L s(s + D(L))
L′
16c2
= (log s − log(s + D(L))
D(L)
 ′  L
16c2 L (L + D(L))
= log .
D(L) L(L′ + D(L))
Note that log x0 ≈ 29.993606. If L ≤ log x0 < L′ in Table 5, then L = 29, L′ = 31 and the
first integral is bounded by
 ′ 
16c2 L (log(x0 ) + D(L))
log ≈ 0.010262 .
D(L) log(x0 )(L′ + D(L))
The last integral is bounded by
 
16c2 8.45 + 690
log ≈ 0.015216 .
8.45 690
The values of these upper bounds are shown in Table 6.

7
 M N Upper Bound M N Upper Bound
x0 e31 0.0102619142161059 e100 e127 0.0208935182022260
e31 e34 0.0250838160282704 e127 e147 0.0106594237333466
e34 e36 0.0143023911370928 e147 e174 0.0105916018816358
e36 e42 0.0350847550483840 e174 e214 0.0108704126795364
e42 e44 0.0096317734983264 e214 e278 0.0109807270912696
e44 e48 0.0170051336041258 e278 e396 0.0110369750969163
e48 e60 0.0381486495943858 e396 e690 0.0111777020604581
e60 e82 0.0421828971523529 e690 ∞ 0.0152151240070501
e82 e100 0.0212203397984129
RN
Table 6: M , N and the upper bound for M
πa,b (t)t−2 dt.

The sum of all upper bounds gives

Z ∞
πa,b (t)
dt < 0.3143472 ,
x0 t2

which is the same for every (a, b) with ab a power of 2. The sum of the terms πa,b (N )/N −
πa,b (M )/M in Equation (1) telescopes to give −πa,b (x0 )/x0 . Adding these values to the lower
bounds from Table 1 gives

S2,1 < 1.4898358 − 0.0015463 + 0.3143472 = 1.8026367,

S4,1 < 0.8488541 − 0.0015108 + 0.3143472 = 1.1616905, and
S8,1 < 1.1078484 − 0.0014769 + 0.3143472 = 1.4207187.

Better upper and lower bounds would result from using a larger x0 . One probably can
derive slightly smaller upper bounds (with the same x0 ) by assuming the ERH as Klyve [8]
′
did for S1,2 .

4 Acknowledgments
I thank Harold Diamond and Carl Pomerance for valuable correspondence concerning this
research. I am grateful to Nick Egbert for noticing the typo in Klyve’s thesis [8].

References
[1] P. T. Bateman and R. A. Horn, A heuristic asymptotic formula concerning the distri-
bution of prime numbers, Math. Comp. 16 (1962), 363–367.

8
 [2] J. Bowman, Some computational results regarding the prime numbers below
2,000,000,000, Nordisk Tidskr. Informationsbehanding (BIT) 13 (1973), 242–244.

[3] R. Brent, Irregularities in the distribution of primes and twin primes, Math. Comp. 29
(1975), 43–56. Corrigendum, ibid. 30 (1976), 1183.

[4] V. Brun, La série 1/5 + 1/7 + 1/11 + 1/13 + 1/17 + 1/19 + 1/29 + 1/31 + 1/41 +
1/43 + 1/59 + 1/61 + · · · , ou les dénominateurs sont ‘nombres premieres jumeaux,’ est
convergente ou finie, Bull. Sci. Math. 43 (1919), 100–104, 124–128.

[5] R. Crandall and C. Pomerance, Prime Numbers: A Computational Perspective,

Springer-Verlag, 2001.

[6] G. H. Hardy and J. E. Littlewood, Some problems of ‘partitio numerorum,’ III: on the
expression of a number as a sum of primes, Acta Math. 44 (1923), 1–70. Reprinted in
Collected Papers of G. H. Hardy, Clarendon Press, Vol. 3, 1966, pp. 561–630.

[7] H. Halberstam and H.-E. Richert, Sieve Methods, London Mathematical Society Mono-
graphs, No. 4. Academic Press, 1974.

[8] D. Klyve, Explicit bounds on twin primes and Brun’s constant, Ph.D. thesis at Dart-
mouth College, Hanover, NH, 2007. Available at https://collections.dartmouth.
edu/archive/object/dcdis/dcdis-klyve2007.

[9] H. S. Lee and Y. Y. Park, On the primes with pn+1 − pn = 8 and the sum of their
reciprocals, J. Appl. Math. Comput. 22 (2006), 441–452.

[10] T. R. Nicely, Enumeration to 1014 of the twin primes and Brun’s constant, Virginia J.
Sci. 46 (1996), 195–204.

[11] T. R. Nicely, A new error analysis for Brun’s constant, Virginia J. Sci. 52 (2001), 45–55.

[12] OEIS Foundation Inc., The On-Line Encyclopedia of Integer Sequences, 2021. Available
at https://oeis.org.

[13] T. Oliveira e Silva, Tables of values of pi(x) and of pi2(x), 2015. Available at http://
sweet.ua.pt/tos/primes.html.

[14] D. Platt and T. Trudgian, Improved bounds on Brun’s constant. Springer Pro-
ceedings in Mathematics and Statistics 313 (2020), 395–406. https://doi.org/10.
1007/978-3-030-36568-4_25. Also available at http://arxiv.org/abs/1803.01925,
March 5 2018.

[15] H. Riesel and R. C. Vaughan, On sums of primes, Ark. Mat. 21 (1983), 46–74.

9
[16] P. Sebah and X. Gourdon, Introduction to twin primes and Brun’s constant com-
putation, 2002. Computation reported at http://numbers.computation.free.fr/
Constants/constants.html.

[17] D. Shanks and J. W. Wrench, Jr., Brun’s constant, Math. Comp. 28 (1974), 293–299.
Corrigendum, ibid., 28 (1974), 1183.

2010 Mathematics Subject Classification: Primary 11N05. Secondary 11L20, 11N37, 11N55.
Keywords: Brun’s constant, Germain prime, twin prime, sum of reciprocals.

(Concerned with sequences A001097, A005384, A023212, and A023228.)

Received August 16 2021; revised version received October 2 2021. Published in Journal of
Integer Sequences, October 2 2021.

Return to Journal of Integer Sequences home page.

10