Approximate Formulas For Some Functions of Prime Numbers
Approximate Formulas For Some Functions of Prime Numbers
Approximate Formulas For Some Functions of Prime Numbers
PRIME NUMBERS
Dedicated to Hans Rademacher
on the occasion of his seventieth birthday
BY
J. BARKLEY ROSSER AND LOWELL SCHOENFELD
1. Acknowledgments
The maior portion of Rosser’s work on this paper was done at Cornell
University with support from the Air Force Office of Scientific Research. His
recent work was done as an employee of the Communications Research Divi-
sion (Proiect FOCUS) of the Institute for Defense Analyses. Schoenfeld’s
work was begun at the Westinghouse Research Laboratories and was com-
pleted at the Pennsylvania State University.
Three maior computations were performed to obtain various of the data
listed herein. The first was performed some nine years ago by workers under
the direction of Mr. Marvin Howard at the Institute for Numerical Analysis,
then part of the National Bureau of Standards, with support from the Office
of Naval Research. The second was performed some four years ago with the
cooperation of Professor R. J. Walker on the IBM 650 at the Cornell Com-
puting Center, with support from the National Science Foundation. The
third was performed recently with the cooperation of Dr. Kenneth I. Appel
at the Computing Center of and with the support of Proiect FOCUS.
2. Introduction
Counting 2 as the first prime, we denote by (x), (x), and (x), respec-
tively, the number of primes =< x, the logarithm of the product of all primes
=< x, and the logarithm of the least common multiple of all positive integers
-< x; if x < 2, we take v(x) O(x) (x) 0. We also let pn denote the
nth prime, and (n) denote the number of positive integers __< n and rela-
tively prime to n. Throughout, n shall denote a positive integer, p a prime,
and x a real number. We shall present approximate formulas for (x),
(x), (x), pn, (n), and other functions related to prime numbers.
In 1808, on the basis of attempting to fit known values of r(x) by an em-
pirical formula, Legendre coniectured an approximation very similar to that
given below in (2.19). In 1849, again on the basis of counts of the number
of primes in various intervals, Gauss communicated to Encke a coniecture
that in the neighborhood of the number x the average density of the primes
is 1/log x. On this basis, if one should wish an estimate for the sum of f(p)
over all primes p -< x, the natural approximation would be
Received February 13, 1961.
64
FORMULAS FOR SOME FUNCTIONS OF PRIME NUMBERS 65
(2.1)
<=x
.
’ fxf(P) log y
f(y) dy
1-----f
-
dy
(2.2) r(x)
_<_ log y
(2.4)
_<_
1
,fx dy
y log y
loglogx-l-B,
(2.5) log P
p
/ dyy
.2
log x + E,
II
(2.6)
exp cl (a) a
v<_-x c()
(log x) ’
where is a real constant, usually taken to be unity.
a
In (2.4) we have indicated a "constant of integration," B, whose value is
taken to be
lim __< 1/p log log z}.
Because this limit exists, the absolute error in (2.4) tends to zero as x tends
to infinity. In (2.5) and (2.6), we have indicated constants E, cl(a), nd
c(a) for analogous reasons. In (2.2) nd (2.3), no constants are indicated
because the limits to which they would correspond do not exist.
The validity of the approximations (2.2) through (2.6) was rigorously
established iust before the turn of the century by Hadmrd and de la Valle
Poussin. A very excellent account of these matters is given in Ingham [6],
together with extensive references to the literature.
From Ingham [6], we can get alternate expressions for B, E, and c(1) as
follows:
(2.7) B C -t-- {log [1 (l/p)] + (l/p)},
(2.s) E C n= (log p)/pn,
(2.9) c(1) e-c,
where C is Euler’s constant. We find (2.7) and (2.9) on pp. 22-23 of Ing-
ham [6], and cn derive (2.8) from a formula near the top of p. 81. Approxi-
mate numerical values re"
B 0.26149 72128 47643,
(2.11) E 1.33258 22757 33221,
66 $. BARKLEY ROSSER AND LOWELL SCHOENFELD
(2.18) -(x) =
log log
Using a closer approximagion for li(z) gives ghe sharper resulg"
Undoubtedly this is not the best possible result, but the precise behavior of
r(x) li(x) depends on the location of the zeros of the Riemann zeta func-
tion, and cannot be determined until we have more precise information about
them than we have now. A discussion of this point appears in Chap. IV of
Ingham [6].
Even in our present ignorance about the zeros of the zeta function, it can
be shown that li(x) alone cannot be a wholly satisfactory approximation to
r(x). Specifically, it has been shown that there is a sequence of values of x,
tending to infinity, at which alternately
(x) li(x) > xl/2/1Og X
and
(x) li(x) < xl/2/log x.
Indeed, Theorem 35 on p. 103 of Ingham [6] states a significantly stronger
result. An analogous result for O(x) x follows from Theorem 34 on p.
100 of Ingham [6] by means of the relations
(2.24) k(x)
(2.25) O(x) Zn=l p(n)b(xl/n).
Each of these summations is in fact only finite, since the summands become
zero as soon as n > (log x)/(log 2). The first of these equations is derived
on p. 12 of Ingham [6], and inversion of the first gives the second, in which
is the MSbius function defined on p. 567 of Landau [7], vol. 2.
Once one has a good estimate for r(x) li(x), one can get an approxima-
tion for sums of functions of primes as follows. Using the Stieltjes integral,
one has
.,
p_
f(p) f(y) dr(y).
f’(y) li(y) dy
f f’(y) r(y) ]i(y)} dy.
(2.26)
_,
_<_
f(p)
f f(y) dy
log y -t-
f(2) li(2)
which is the precise version of (2.1). If the integral in (2.28) below con-
8 J. BARKLEY ROSSER AND LOWELL SCHOENFELD
<__ log y
(2.27)
-t-f(x)/r(x)- li(x)}-t- f’()/’() li()} d,
where K is the eonsgant given by
(2.28) g] f(2) li(2) f’(y){.(y) li(y)} dy.
Using these, we can get sharper forms of (2.3) through (2.6). Thus,
from (2.21) und (2.26), we get
(2.29) O(x) x + O(x exp {-a(log z) /} ).
From (2.21) and (2.27), we get
(2.30) _,,<= lip log log x 4- B 4- 0(exp {-a(log x) /} ),
(2.31) _<_ (log p)/p log x 4- E 4- 0(exp {-a(log x) 1/}).
From (2.30) we proceed as in (2.6) to get
0(exp {-a(log x) }).
,<_<_ (log x) 4-
From (2.29) and (2.24), one can get a formula for (x) analogous to (2.29).
By starting from (2.23) rather than (2.21), one can get even sharper results
than (2.29) through (2.32).
Though results like those above are interesting, and are difficult to prove,
they are of little use for getting dependable numerical approximations unless
values of a, b, and X in (2.22) are furnished; this is seldom done. In Rosser
[12], explicit bounds were presented for the errors in our approximations.
More recently, much better bounds have been obtained by using modern
computing machinery and taking advantage of new information about the
zeros of the zeta function. These results will be stated in the early part of
the present paper, with the proofs being mainly withheld until the later
sections.
3. Widely applicable approximations
For a very sharp approximation, one must either use complicated formulas
or be satisfied with validity over a limited range. In this section, we shall
list approximations which combine the advantages of being reasonably simple,
reasonably precise, and valid for nearly all values. Note that Theorem 1
below will replace (2.18) by closely related and specific inequalities, while
Theorems 2-7 will do the same for (2.19), (2.20), (2.3), (2.4), (2.5), and
(2.6) respectively. Theorem 8 is a variant of Theorem 7 which is sometimes
more convenient.
FORMULAS FOR SOME FUNCTIONS OF PRIME NUMBERS 69
THEOREM 1. We have
(3.1) x
log x (1- 2
1
)
log------ <r(x) for59 < x,
(3.2) (x) x(
<l-ogx 1 +21ogx 3) for l<x.
THEOREM 2. We have
(3.3) x/(log x 1/2) < (x) for 67 _-< x,
(3.4) x
r(x) < (log x ) for e a/2 < x
(and hence for 4.48169 <_- x).
COROLLARY 1. We have
(3.5) x/log x < v(x) for 17 <= x,
(3.6) (x) < 1.25506 x/log x for 1 < x.
COnOLLAnY 2. For 1 < x < 113 and for 113.6 =< x
(3.7) (x) < 5x/(4 log x).
COROLLARY 3. We have
(3.8) 3x/(5 log x) < r(2x) (x) for 201/2 =< x,
(3.9) 0 < r(2x) r(x) < 7x/(5 log x) for 1 < x.
For the ranges of x for which these corollaries do not follow directly from
the theorem, they can be verified by reference to Lehmer’s table of primes
[10]. A similar remark applies to all corollaries of this section unless a proof
is indicated.
The inequality (3.8) improves a result of Finsler [3]. The left side of
(3.9) is just the classic result, conjectured by Bertrand (and known as
Bertrand’s Postulate) and proved in Tchebichef [14], that there is at least
one prime between x and 2x. The right side of (3.9) gives a result of Finsler
[3], with Finsler’s integral n replaced by our real x. Finsler’s elementary
proofs are reproduced in Trost [15] on p. 58. The relation (3.12) below
states a result of Rosser [11].
THEOREM 3. We have
(3.10) n(log n+ log log n -) < Pn for 2-<n,
(3.11) p, < n(log n -+- log log n 1/2) for 20 <__ n.
COROLLARY. We have
(.12)
(.1) p, <
n log n
n(log n
-
<: Pn
log log n)
forl <_ n,
for6 <__n.
70 J. BARKLEY ROSSER AND LOWELL SCHOENFELD
THEOREM 4. We have
(3.14) x(1 1/(2 log x)) < O(x) for 563 _<_ x,
(3.15) O(x) < x( + /(2 og x)) for 1
_
CoaoaY. We have
(3.16) x(1 1/log x) < O(x) for41 <=x.
THEOREM 5. We have
(3.17)
(3.18)
(3.19)
(3.20)
_,
log logx +B- 1/(2logsx) < _<xl/p
<= 1/p < log log x B 1/(2 log x)
COROLLARY. We have
THEOREM 6. We have
(3.21) log x + E
_,
log log x <
for l < x,
for l
COrOllArY. We have
e-C
(3.27) (1 for1 <x.
THEOREM 8. We hve
(3.28) e (logx) 1--2logsx p--1 for 1 < x,
P
(3.29) II
_<_p 1
<e e(logx) 1-t-21ogsx for 286 -< x.
COROLLARY 1. We have
(3.30) < e (log ) ( + o1)x for1 <x.
FORMULAS FOR SOME FUNCTIONS OF PRIME NUMBERS 71
COROLLARY 2. We have
1
(3.31) II p P
,_ 1
< eC E
1< n forl <__ x.
THEOREM 9. We have
(3.32) O(x) < 1.01624 x for 0 < x.
For better bound for 0(x) when x _<_ 10s, note Theorem 18 below.
THEOREM 10. For d <-_ x, we have cx < O(x) for each of the following pairs
of values of c and d"
c
d
.00.0,01.0__01
7481
.01
53811 34571 2657
.0o,.0l
148114331
.
o1.01.0_ -1"
-
1427[ 853
809l
1" 1"
3491
;"
227[
;
599[ 557[
1491 101
-
1/(n- 1) __< n/(n);
72 J. BARKLEY ROSSER AND LOWELL SCHOENFELD
THEOREM 22. For 0 < x < 113 and for 113.8 <__ x <= 10
(4.9) __< (log p)/p < log x + E 2Ix 1/2.
THEOREM 23. For 0 x <- l0
c
(4.10) e c log x < ]I_ P/(P 1) < e log x 2eC/x1/2.
THEOREM 24. We have
(4.11) x 1/ < (x) O(x) for 121 __< x __< 1016
1/2
(4.12) (x) O(X) < X -Jl- 3X 1/ for 0 < x <= 1016.
One immediately wonders if the results of Theorems 20-23 could be valid
for all x. It is known that each of (4.1), (4.2), and (4.5) fails infinitely
often for large x, and indeed each side of (4.5) fails infinitely often (see
Theorems 35 and 34 on pp. 103 nd 100 of Inghm [6]). Perhnps one cn
extend these results to show that ech of (4.7) through (4.10) fails for large
x; we hve not investigated the matter.
Theorem 18 gives sharp bounds for O(x) for 0 < x <__ 10 For larger s.
values of x, sharp bounds for O(x) can be obtained by use of Theorem 14 and
its corollary provided sharp bounds for k(x) are known. For 10 =< x =<
5000
e sharp bounds for k(x) can be obtained from our Table I, in which we
tabulate against b values of such that for e _<_ x
(-)x<(x) <(+)x.
The values of m listed pertain to the computations by which Table I was
established, all of which will be explained later.
Finally, Theorem 11 can be used to get close approximations to both (x)
and 0(x) for large x beyond the range of existing tables. Although Theorem
11 is valid for small values of x as well as large, for x below about e3 it gives
poorer estimates for 0(x) than can be obtained from Theorems 4, 9, 10, 18,
and 19. From e ls’4 to e4s, Theorem 11 gives poorer estimates than can be
obtained from Table I with the help of Theorem 14.
We can use our sharp estimates for 0(x) to get sharp estimates for other
functions depending on primes. Using the Stieltjes integral, we can write
E f(p)
_<_ f f(Y)
log y
dO(y).
dy.
,_<_ logx \iog
Alternatively, one can derive (4.13) by use of Theorem A on p. 18 of Ingham
[6]. From (4.13), as in the derivation of (2.26), we get
f(y) dy 2f(2)
(4.14)
f(p)
__< f log y log 2
+ f(x) {()
log x
x} f {O(y) y}- \log
For suitable f, we can write (4.14) as
c f(y) dy
F_,
_<_
f(P) [
J2 log y + Lf
-
(4.15)
+ f(x) {O(x)
log x
x} lO(Y) (f(y)y) dy,
d
Y}- \log
where Lf is the constant given by
2f(2) c
(4.16) L/ / {O(y) y} \log/dy.
log 2 Jo.
(4.18)
i (x)
+ f* O(y)(1 -+- log y) dy,
_< p x log x y log y
5-’ log p O(x) f* (y) dy
x +
(4.19) p y.
p<__
(4.21)
__<
log p
p
log x -t- E + O(x)x x
/i O(y)y y
dy.
differs by unity from ours. Just recently the Rand Corporation has prepared
a list of primes, published on microcards in Baker and Gruenberger [1], giv-
ing all primes up to slightly beyond 108 Within the large range of these
tables, one can read off exact values of r(x) and p. Their count agrees with
Lehmer’s rather than ours.
By use of Lehmer’s Table, Rosser verified Theorem 16 and the right half
of (4.5) for x __< 106. An account of his methods is given in Rosser [12].
By the same methods, these results were extended to 107 in the first major
computation cited in Section 1.
We turn next to the calculations required to establish (2.10) through (2.14).
The values for c(1) and 1/c(1) given there were computed from (2.9) and
the known value of C. The values for B and c(2) given in (2.10) and (2.14)
have been taken from pp. 43 and 44 of Rosser [11]. At the end of the foot-
note on p. 43 of this reference, there is given a twenty-four-decimal value of
C B; this value corrects the slightly erroneous value of B C given in
Table I of Gram [5]. Incidentally, Gram reproduces M:errifield’s incorrect
value of ] p-3 in his Table I; a correct value is given on p. 249 of Davis [2].
To determine the remaining quantity, E, we use the relation
’(s)/(s) (log p)/(p8 1) rl (log
which is given near the bottom of p. 17 of Ingham [6]. From this we obtain
forn > 1
(5.1) (log p)/p’ Em=l (m) ’(mn)/(mn).
Substituting this into (2.8) gives
(5.2) E C
=
t(m) ’(m)/(m).
We first computed i"(n) by using an electronic computer to sum the first
500 terms of the series
"(n) ]= (log m)/m’;
the remainder of the series was computed by the Euler-Maclaurin sum
formula. As a cheek, this was repeated with the first 1000 terms. Using
the values of i’(n) given on p. 244 of Davis [2], values of
’(n)/(n) and ] (log p)/pn
were computed for 2 =< n -< 56 by using (5.1). These, together with the
values of -i"(n) are listed for 2 =< n =< 29 in Table IV. In this table the
values given for -’(n) are in error by less than 10-17. Division by (n)
could cause slightly greater errors in -’(n)/(n). As we used seventeen-
decimal values in (5.1), the errors in p-n log p could be a bit greater. For
n > 29, the three functions tabulated do not differ in the first seventeen
decimals, and each is given to better than seventeen decimals by
(log 2)/2 + (log 3)/3 n.
76 . BARKLEY ROSSER AND LOWELL SCHOENFELD
-
THEOREM 25. For 1 <- x
(5.3) (x) -< (x) < 1.2 ax + (3 log x + 5)(log x + 1),
where a denotes the constant
v log 2 + log 3 + log 5 0.92129 ...,
according to the definition on p. 88 of Landau [7], vol. 1.
Using (5.3) with Theorem 18 gives the following weakened form of Theorem
9"
(5.4) (x) < 1.11 x for 0 < x.
Using values of O(x 1/2) from the Rosser-Walker computation, we verified
(3.38) for x < 50,653 by using Table III. For 50,653 -<_ x < 1024, one can
verify (3.38) by (2.24) and the right side of (4.5) to do this, we proceed by
cases such as 2 __< x < 2 N where M and N are conveniently chosen integers.
-
Finally, for 1024 <__ x, (3.38) holds by (2.24) and (5.4). Thus (3.38) has
been completely established. From it, by (5.4) one can infer the following
weakened form of (3.39)"
(5.5) (x) (x) < 1.11 x /2 3x / for 0 < x.
As (3.38) and the right side of (4.5) imply (4.12), we cn complete the
proof of Theorem 24 by establishing (4.11). This is readily done for 121 _<_
x < 50,653 by means of Table III. By using Theorems 18 and 19 with (2.24),
we can finish the proof of (4.11) nd hence of Theorem 24.
By comparison with Gram’s Table [5] of (x), Theorem 12 was verified
for x =<2000, nd it ws ascertained that in this rnge the maximum value
of (x)/x lies between 1.03882 and 1.03883. Thus one can complete the
proof of Theorem 12 by proving that (x) _<_ 1.03882 x for 2000 x. This =<
follows for x < 50,653 by Theorem 18 and Table III. For 50,653 -< x _-< 10
it follows by Theorem 18 nd (4.12). It will follow for 11 greter x by Table
78 ,1. BARKLEY ROSSER AND LOWELL SCttOENFELD
I as soon as we have justified the values in this table, which we will do in the
next section.
We can verify Theorem 13 in the range 0 < x < 50,653 by reference to
Table III. It then suffices to verify b(x) (x) -<_ 1.42619 x 1/ for 50,653 __<
x. For 106 _<_ x, we infer this by (5.5). For x < 106 we use (2.25); the
facts that h(Y) 0 for y 2 and that is monotone let us conclude
(x) (x) ()
s.
+ () + (x) (x)
+ (x") (x") + (x) + (x")
1/2 1/3 /5
since (3.35) holds for x _<_ 10 This suffices to complete the proof.
-
/(X TM) /(X 1/15) + ](X 1/17) /(X 1/19)
__< (x ") + x1/) + t(x") + (x ")
< 1.04 (x + x -t- x -t- x 1/1)
At this point, therefore, we have established (2.10) through (2.14), Theo-
rems 16 through 24, Theorem 12 except for 10 < x, Theorem 13, and (3.38)
of Theorem 14. We have also verified Theorems 1 through 10 and their
corollaries for x __< 16,000 or for pn <_- 16,000.
6. Sharpening of some results of Rosser, with application to
several proofs
In the preceding section, we carried our proofs as far as is practicable without
appealing to very deep results. From here on, we shall be mainly concerned
with invoking certain deep results to validate Table I and complete the proofs
of the results stated in Section 3. Space does not permit us to give proofs in
full, so that we shall assume that the reader is quite familiar with Ingham
[6] and Rosser [12], from which we shall use notation and results with a
minimum of reference.
The most significant sharpening of results from Rosser [12] arises from the
fact that it is now known that the first 25000 zeros of the zeta function have
real part equal to 1/2, as shown in Lehmer [8] and Lehmer [9]. This enables
us to replace the A on p. 223 of Rosser [12] by A e9"99. We do not now have
-
N(A) F(A), which will make a slight change in a key formula, as we note
below.
Observing that
322 W 546 cos -t- 329 cos 2 W 130 cos 3 25 cos 4
2(1 + cos ) (3 + 10 cos )2 >__ 0,
we can modify the proof of Theorem 20 of Rosser [12] to get a proof of
THEOREM 26. For A <- ,,
we have < 1 1/ R log /).
The R here is that defined in Theorem 11. In other places, as here, it re-
places the number 17.72 appearing in Rosser [12]. Thus, we are conforming
FORMULAS FOR SOME FUNCTIONS OF PRIME NUMBERS 79
with the notation of Rosser [12] when we temporarily abrogate the usual
denotation of (n) and define
(6.1) (-y) (m, x, /) 5’
--m--I e-(logx) (Rlogs,)
-
However, for the purposes of the next theorem we consider (6.1) as de-
fining (,) for arbitrary positive R.
THEOREM 27. If (’y) is defined as in (6.1) with m and R positive numbers,
if 2 <= K, and if 0 <= log x __< (m 1 R log K, then
(6.2) (,) < 2R(K)(K) +Q (y) log Y dy,
where
1 0.137 log K 0.443
(6.3) Q -!- K log K log (K/2z)
Proceed as in the proof of Lemma 18 of Rosser [12].
COROLLARY. If in (6.1) we take R to be the R of Theorem 11, and if A <= K,
0 < m, and
(6.4) 0 _-< log x < 1748(m -4- 1),
then
(6.5)
’
Proof. As we
(,) < 2R(K)O(K) + 0.1592 O(y) log Y dy.
are here using the R of Theorem 11, (6.4) verifies the final
hypothesis of Theorem 27. In (6.5), the coefficient in front of the integral
is got by taking K e9’99 in (6.3), which is permissible since e
9"99
A =< K.
In Rosser [12], in the situation corresponding to taking K A in the corol-
lary, the coefficient 2 did not appear in the first term on the right of (6.5).
This is because N(A) F(A) in that paper. Except for that, we now pro-
ceed as in the proofs of Lemma 19 and Theorem 21 of Rosser [12] to derive
THEOREM 28. If m is a positive integer,
1 < k loga < m 1748
0.123’
mT1
0.0003647 1.298 m 0.1592 /+)
--.3 mAa, ll I
m}2
+
and1 + ma < a, then for a x
x(1-)-.s4<(x) <x(l+)-log(1-x-).
80 J. BARKLEY ROSSER AND LOWELL SCHOENFELD
lation. For 16,000 -<- x _<_ 108, these are established by Theorems 18 and 19.
For 108 < x < 1016 we use Theorem 24 and Table I. For 1016 < x < e5
we use the corollary of Theorem 14 and Table I. Finally, above e5 we use
Theorem 11.
From these, Theorem 4 is an easy consequence.
7. Proof of Theorems 1 through 3
We start with five lemmas. As their proofs are similar, we state all five
lemmas first before giving the proofs. We first make the definition
(7.1)
J(x,a) r(1451) 0(1451)
og 5 og
x
( + og
f(1-[-loYa)dyy"
-i- log
LEMMA 1. For e -<- x,
(7.2) li(x) x(1-k 21ogx
<logx
LEMMA 2. ForlOs-<- xanda 0.31,
(7.a)
L
L 4. Fore
. z,
J(z, a) <
lo
For e N z and a 0.47, he ineqalig (7.) i vlid.
lo
(7.4)
L
(7.5)
. z/(log z ) < li(z)
ForlO N zaed -0.47,
x/(log x ) < J(x, a).
li(zn).
For each of these lemmas, the proof is in two parts. First, one verifies
that in the stated range of x, the derivative of the left side is less than that
of the right side. Second, one verifies that at the lower limit of x the left
side is less than the right side. To perform the needed calculations, one can
use the reduction formula
(7.6) log
a dy
+1 y
b
log b
x
log x t- f dy
log Y
to express the various integrals in terms of li(x) and elementary functions.
For x <- e 1, one can get numerical values of li(x) from the tables of El(y)
given in [18]. Outside this range, one can appeal to the following result.
THEOREM 32. If m is a positive integer and m <= y <= m + 1, then
-
82 BARKLEY ROSSER AND LOWELL SCHOENFELD
and
1/2
(7.8) 2(m+ )2
1
1.06
m+ 1
e
(j- 1)
y
are lower and upper bounds respectively for Ei (y).
This is a consequence of equations (58), (68), and (69), and of Lemma 3
and Theorem 7 of Rosser [13].
We note the value (1451) 230 and the approximation 0(1451) 1396.4
taken from the Rosser-Walker tabulation.
We now turn to (3.2), which has already been established for x N 16,000
by the Rosser-Walker tabulation. By Lemma 1 and Theorem 16, we verify
s.
(3.2) up to x 10 We have by (4.17) that
(x) (1451)- (1451)
log 1451 + 0()
log x +
By Theorem 29, we can conclude v(x) < J(x, 0.31) if 1451 x e
by Lemma 2 we can infer (3.2) for l0 x e
bine Lemma 3 and Theorem 31 to complete the proof of (3.2).
. So
In a similar way we com-
dy
y log y
.
In a similar way, we combine Lemma 4 and Theorem 16 to verify (3.3) for
x s,
10 and complete the verification by combining Lemmu 5 nd Theorem
31.
We get (3.1) from (3.3), and (3.4) from (3.2), by applying the inequality
x
logx (1+ ) <logx-a
logx
a x
forea<x
in the two cases a and a .
As preliminaries for the proof of Theorem 3 we undertake the proof of (3.12)
and
(7.9) pn < n(log n -t- 2 log log n) for 4 -< n.
These were proved in Rosser [11], but can be derived so readily from the strong
results now available that it seems worthwhile to indicate new proofs by this
method. For instance, suppose if possible that pn <- n log n. Then
n -< r(n log n).
So by (3.2), we have
n < log n nlogn
-t- log log n (1
log n -t- log log n 1.5)
a result which certainly fails if e __< n. So (3.12) holds for e =< n, and a
trivial computation verifies it for smaller n. The proof of (7.9) is analogous.
From these, we can now infer Lemmas 9 and 10 of Rosser [12], using the
proofs given there. Using Lemma 9 of Rosser [12] with Theorem 18 gives
nlogn-nloglogn- n- li(n)
FORMULAS FOR SOME FUNCTIONS OF PRIME NUMBERS 83
for 5 <= n -< r(10s). Then (3.10) follows for e -< n -< r(10 s) by Lemma 7
of Rosser [12]. Now use Lemma 9 of Rosser [12] with Theorem 30, and infer
o. o
nlogn+nloglogn--n--li(n) < p 1 ( +logp,]
for 140 _<_ n __< e s. As li(n) < 0.1 n for r(10s) =< n, by Lemma 7 of Rosser
[12], and
(7.10) pn/(log p,) < n
by (3.5), we infer (3.10) for n =< e s.
We use Theorem 31 in a similar manner
to complete the proof of (3.10).
We next prove (3.11). We note first of all that it has been verified for
n -<_ 1862 by the Rosser-Walker tabulution. Now let 1862 <- n, and suppose
that (3.11) has been verified for all integers less than n. Then the hypothesis
of Lemma 10 of Rosser [12] is verified, and we conclude that
n log log n
(7.11) log n + n log log n n -}-
log n
By Theorem 19, if n =< 7r(10s), then
n log log n
p,- 2(p,) :/ < n log n -t-n log log n- n log n
By (7.9),
2(p,) 1/2 < 2(n log n + 2n log log n) < 0.2 n,1/2
so that (3.tl) is verified. Now let r(10s) <= n -< e369. Then (7.11) and
Theorem 29 give
(
p= I o.31
1- =] < n log n -f- n log log n n -4-
n log log n
log n
Using (7.10), we again infer (3.11 ). In a similar way, we can use Theorem
31 to verify (3.11) for e69 _<_ n.
K(1, x) <
(8.1)
0.0463
X1]2
Taking m
8. Proof of Theorems 5 through 8
We require several lemmas.
2R(K)
-t- K -{- 0.1592
1
1
<= K,
i
x in (6.5) gives
2R (K)
K2 -t-0.1592
+ log (K/2r)
K
1 -b log (K/2’)
_ e-9.61 X--l/(logK).
K< 2< g
84 J. BARKLEY ROSSER AND LOWELL SCHOENFELD
< X 1/2
+ A<7EK + Z:
2
x-1/ 11:)
We next state two lemmas whose proofs are so similar that we give only the
-t- log x
log x
(K(1
\
_
1.84 0.31
__x
1
x) + 1.8_4 + 0.31
x x ]
Proof. We have
yp-2 (1 q- log y) dy XP--1
log y (p 1) log x (p-- 1) 2,log2x
2
log ayd
Now
x- f y- dy x- y-2
-
< -1-2
logx
2
J logay logx log ady
,/--1
So
log x "-I- 2x f dy
y log log x
[ yp-2 (1 + log y) dy
-
1
Jx log y
2-+-logx x
log z 7
Hence
(s.a) log y)
y
dy <2+ log x
log x
K(1, x).
So by Theorem 29 on p. 77 of Ingham [6], we have
(y (y)) (1 + log y) dy 2 + log x K(1, x) -f- I,
y2 log. y log x
FORMULAS FOR SOME FUNCTIONS OF PRIME NUMBERS 85
where
(2 log2r--log(1--y-2)) (1.logy)
1
f 2y log y
dy
q- log 2 log 2r
1
log 2 \ x + (x--2r--11)j
x
x rl r
-
n log y
ya--n--1 1 dy <
n x
log2y n--a logx"
Proof. We have
1 q- n 10g y d y-’
Y
a--n--1
log y
dy ya lo
dy
x
log x -l-
a
foo ya--n--1
log y
dy
< x a
ya-n-1 dy
logx log
-
n x
n a log x
Let us define
L(x) 2q-logx
og x K(1 x) q-1"84q- 0.31
(8.4)
1.02
+ 2.04
x- + 4.5 x-2 + x--l’
(8.15) M(z, a) (log x) log 1 q-
2 log z log x"
LEMMt 10. If1 < A <= B and a < 1/2 nd
and L(A < M(B, a), then L(x) < M(x, a) for A <- x <- B.
Proof. We readily verify that for A __< x both L(x) and M(x, a) are de-
creasing functions of x. So for A -< x _<_ B
86 $. BARKLEY ROSSER AND LOWELL SCttOENFELD
- - )
1 1 1.02
(8.9) _-_ log log x B log 1W21ogx
P (x- 1) logx"
Proof. By (4.20), Lemma 8, and (3.39), we have
1
log log x- B
<_
I’(x) -x
x log x + (y- (y)) (1 -t-log y)
yS logs y
dy
We apply Lemma 9 with n 1, once with a 1/2 and once with a 1/2. Then,
with the help of (8.4), the above inequality reduces to
(8.10) e
log x 1+ 2 log x > --))exp((x-- 1) .02
logx
/
where
9. Proof of Theorem 15
In this section, (n) again denotes the Euler totient function.
LEMMA 14. If m and n are positive integers, and n < exp 0(p+), then
{9.1) hie(n) p/(p 1).
Proof. Let q, q, q be the distinct primes, in increasing order, which
divide n. Then
exp (p+) > n q q p... p exp (Pr).
So r m. Consequently
n/(n) := q/(q- 1) := p/(p- 1) p/(p.- 1),
which is the sme as (9.1)
By means of Lemm 14 we can verify (3.41) numerically for succession
of intervMs if log n is not gret. For instance, we readily verify numericMly
that
p/(p 4.375 < e c log log 210 +
5/(2 log log 210).
So we cn tke m 4 in Lemm 14, nd conclude that if 210 n < 2310
exp 0(p+), then
c
n/(n) < e c log log 210 + 5/(2 log log 210) e log log n 5/(2 log log n)
With the help of vMues of 0(x) nd of
TABLE I
(1- e) x <(x) < (1 + e) xforeb_ x
b m lO’e n b m lO’e n b m lO’e n
474.55444 41547
956.24526 51201
1462.14165 18014
TABLE II
p_ P
_1
2.09670 95528
2.19808 01272
2.25562 82528
_
FORMULAS FOR SOME FUNCTIONS OF PRIME NUMBERS
log p
P
4.94448 99600
5.60951 04754
6.01869 61634
91
11.15950 15857
12.35097 56739
13.08291 09945
2000 1939.83920 03026 2.29244 84920 6.29327 07024 13.57375 00182
2500 2433.60275 29800 2.32105 31990 6.51384 37141 13.96771 93817
3000 2932.35921 18787 2.34404 93716 6.69584 35999 14.29270 53203
3500 3409.45718 45205 2.36222 13278 6.84275 32932 14.55484 67736
4000 3911.14539 95812 2.37858 30199 6.97729 51026 14.79498 00928
4500 4412.18831 05019 2.39276 53465 7.09571 03205 15.00632 74987
50O0 4911.69535 17069 2.40518 86577 7.20087 63227 15.19393 85100
5500 5391.37223 83531 2.41586 31315 7.29230 15305 15.35701 01674
6000 5893.29745 72481 2.42598 40781 7.37988 11065 15.51324 05518
6500 6408.90736 71752 2.43543 75880 7.46249 21364 15.66060 23831
7000 6920.42102 99437 2.44401 57706 7.53814 09953 15.79552 97610
7500 7364.85741 60237 2.45091 39779 7.59945 43711 15.90487 48550
80OO 7875.15038 47974 2.45829 17384 7.66550 12982 16.02265 87938
85OO 8343.99966 34035 2.46460 59355 7.72243 05072 16.12415 52820
9O00 8870.37499 26578 2.47124 44465 7.78267 62189 16.23155 79133
9500 9418.36877 33985 2.47772 62760 7.84187 45368 16.33711 55421
10000 9895.99137 91570 2.48305 99472 7.89086 36043 16.42448 96322
1O5O0 10403.90704 75207 2.48842 73950 7.94043 00603 16.51288 85620
11000 10877.34163 04695 2.49317 02420 7.98445 72244 16.59139 63457
11500 11362.43971 33403 2.49778 96786 8.02755 03003 16.66821 99665
12000 11840.48575 38722 2.50212 24249 8.06816 24606 16.74059 88963
12500 12348.83694 44657 2.50652 81281 8.10963 67280 16.81451 87368
13000 12868.72809 74239 2.51084 52555 8.15044 40734 16.88726 89363
13500 13371.76845 32826 2.51484 72175 8.18842 74850 16.95498 91186
14000 13867.29252 76925 2.51862 72664 8.22444 83834 17.01920 34304
14500 14307.28400 32521 2.52185 38317 .8.25531 02383 17.07420 76210
15000 14844.79169 21653 2.52565 30642 8.29177 62756 17.13920 20969
155OO 15384.23856 36932 2.52932 29775 8.32712 81756 17.20221 91093
16000 15886.79246 84213 2.53262 50069 8.35904 03921 17.25911 70367
92 ;. BARKLEY ROSSER AND LOWELL SCHOENFELD
TABLE III
(.) () (n)
TABLE IV
V() V()/’() log p
BIBLIOGRAPHY
1. C. L. BAKER AND :F. J. GRUENBERGER, The first six million prime numbers, Microcard
Foundation, West Salem, Wisconsin, 1959.
2. H. T. DAWS, Tables of the higher mathematical functions, Vol. II, Bloomington,
Indiana, The Principia Press, Inc., 1935.
3. P.
yon
x
the denominators being the
series of prime numbers, Messenger of Mathematics, vol. 28 (1898), pp. 1-17.
5. J. P. GRAM, Undersgelser angaaende Maengden af Primtal under en given Graense,
K. Danske Vidensk. Selskabs Skrifter, Naturv. og Math. Afd., ser. 6, vol. 2
(1881-86), pp. 183-308.
6. A. E. INGHAM, The distribution of prime numbers, Cambridge Tract No. 30, Cam-
bridge University Press, 1932.
7. E. LANDAU, Handbuch der Lehre yon der Verteilung der Primzahlen, 2 vols., Leipzig,
Teubner, 1909. Reprinted in 1953 by Chelsea Publishing Co., New York.
8. D.H. LEHMER, On the roots of the Riemann zeta-function, Acta Math., vol. 95 (1956),
pp. 291-298.
9. , Extended computation of the Riemann zeta-function, Mathematika, vol. 3
(1956), pp. 102-108.
10. D. N. LEHMER, List of prime numbers from 1 to 10,006,721, Carnegie Institution of
Washington, Publication No. 165, 1914.
11. BARKLEY ROSSER, The n-th prime is greater than n log n, Proc Lond. Math. Soc. (2),
vol. 45 (1939), pp. 21-44.
12. --, Explicit bounds for some functions of prime numbers, Amer. J. Math., vol. 63
(1941), pp. 211-232.
13. J. BARKLEY ROSSER, Explicit remainder terms for some asymptotic series, Journal of
Rational Mechanics and Analysis (J. Math. Mech.), vol. 4 (1955), pp. 595-626.
14. P. TCHEBCHEF, M$moire sur les nombres premiers, J. Math. Pures Appl. (1), vol. 17
(1852), pp. 366-390.
15. E. TROST, Primzahlen, Basel, Birkhiuser, 1953.
16. A. WALTHER, Anschauliches zur Riemannschen Zetafunktion, Acta Math., vol. 48
(1926), pp. 393-400.
17. I. M. VINOGRADOV, Novaya ocenka funkcii (1 it), Izv. Akad. Nauk SSSR, Ser.
Mat., vol. 22 (1958), pp. 161-164.
18. Tables of sine, cosine and exponential integrals, 2 vols., W.P.A. Tables prepared in
1940 under the sponsorship of the National Bureau of Standards.
19. KENNETH I. APPEL AND J. BARKLEY ROSSER, Table for estimating functions of primes,
IDA-CRD Technical Report Number 4, September, 1961 (available on request
from Communications Research Division, Institute for Defense Analyses,
yon Neumann Hall, Princeton, New Jersey).