PACIFIC JOURNAL 0~ MATHEMATICS

Vol. 82, No. 2, 1979

ON THE ASYMPTOTIC BEHAVIOR OF LARGE

PRIME FACTORS OF INTEGERS
K. ALLADI AND P. ERD&

We prove results on the asymptotic behavior of large

prime factors of the integers. The basic idea of the paper
is that if k is any fixed integer, then the kth largest prime
factor of n, denoted by P,(lz) is generally much bigger than
CjJ’Jn). We give precise estimates of this phenomenon.
This paper is a sequel to an earlier paper by the authors.

1. Notations and definitions‘ Throughout this paper the letters

p and q, with or without subscript will denote primes.
Let 12 = IIrE1 p?, p, > p, > a.. > p, be the canonical decomposi-
tion of an integer n > 1 into primes. We set

and

Cl*% Q(n) = 2 ai , o(n) = 7’ .

*=I
Let A(1) = A*(l) = Q(l) = o(l) = 0.
We may define the lcth largest prime factor in two ways depending
on whether we want to count prime factors according to multiplicity
or not. To be more precise set

P:(n) = p, for k I o(n)

(1.3) = 0 for k > o(n) .

We may also define

MC = 24
(1.4) 1 < k d Q(n)
pk(a) = ‘4 PI(n) .P,(n;“. . 0 Pk-l(n) ) ’
PJn) = 0 for k > Q(n) .

Observe that P,(n) = P:(Iz).

The terms “average order” and “normal order” will mean the
following: Let f be an arithmetic function and set

(1.5)

Suppose g is a monotonic function such that

295
296 K. ALLADI AND P. ERDt%

lim --Wx) - 1
(l-6) 2-m F(x)
where

then f has average order g. Next, we say that two functions f and
g are “nearly the same almost always” if for each E > 0

where

If in (1.9) and (1.8), the function g is monotonic, we say that f has

normal order g.
Consider the sum
(1.10)

If a I 1 is a real number and 3 = @, it is well known (see [lo])

that
p(a) = lim ?Ip(‘P x1’a)
(1.11) 2-m X

exists. The limit in (1.11) is also defined if - 00 < a < 1 and

1 O,cCr<l
(1.12) da> =
i 0 --<a<O.

The function p(a) is a monotonic decreasing continuous function of

a for a 2 1.
Finally we define the sums

A*(n) - P:(n) - . . . - PL(%) , k> 1

Sz(x,k) = c
2insz R(n) -
(1.13)
t&(x, k) = c G@L , k&l
256~ PI(n)
 ASYMPTOTIC BEHAVIOR OF LARGE PRIME FACTORS 297

The aim of this paper is to obtain estimates for these sums S&C, k),
i = 1, 2, 3, 4.

2. General background and main theorem. The results in

this paper are in continuation of those in 9 2 of [2].
It is a well known theorem of Hardy and Ramanujan [6], [7]
that the functions Q(n) and o(n) both have average and normal order
log log 9%= g(n). This means that a number YL usually has log log n
prime factors and most of them occur square free. Thus it is natural
to expect the large prime factors to occur with multiplicity one,
most of the time. So one should be able to show that the functions
A and A* have the same average order. In an earlier paper [2] we
showed this to be true and much more.
Not only do A and A* have the same average order, but the
function P1(rz) dominates the sums in (1.1) to such an extent that
A, A* and P, have the same average order. More generally A(n) -
P,(n) - +*a - P,-,(W) and PJn) have the same average order. It was
observed in [l] that the functions P,*(n) and A*(n) - P:(n) - .* - l

P,*_,(N) also have the same average order as Pk(n), since the asymptotic
analysis in [2] remains unaffected if the weak inequalities are re-
placed by strict ones. Thus we restate (without proof) the main
theorem in [2] in a more complete form:

THEOREM A. If k is a .fixed positive integer then

where ak is a comtant depending o&v on k, alzd is a rational

multiple of C(1 i- lJk) where t: is the Riemann zeta function. In
addition for each k 2 1

(2.2) ls&z{A(n) - A*(n)) = x log logx + Q(X) = o(15~zp~(n))

- .
Theorem A says that the average order in (2.1) is g(n) =
a,*.n’lk/(log n)’ where a,Y = a,‘(1 + l/k). An average is essentially
influenced by two things-(i) the abnormally large values of a func-
tion, which certainly contribute to (2.1) and (ii) the values a function
takes most often.
The question now arises whether A, A*, and Pl are nearly the
same almost always. The main theorem stated below answers this
question in the affirmative.
298 K. ALLADE AND P.ERDt)S

THEOREM B. If k is a Jixed positive integer then

(2.3) Si(X, k) - &(x, k) - s&z, w - WC, k) - c&

(log:)“-’
where a: = 1 and a; for k > 1 is a constant depending only on k,
and is a ratiolzal multiple of er where Y is Euler’s constant. In
addition for each k 2 1

,&,4%)
C2w4)- d*(n)
E(n) =O( x e”Vlog x log log x )=oL&$$f) 1
where c is an absolute constant >O.

3, Consequences and motivation. Statements (2.3) and (2.4)

may be looked upon as analogues to (2.1) and (2.2). Theorem A said
that A, A* and P, have the same average order, x%/6 log n, (a, =
7?/12, see [2]). We can deduce from Theorem B the following,

COROLLARY. The functions A, A* and P, are all nearly the

same almost always. Also all three fumtions fail to possess a normal
order.

Proof. Consider two arithmetic functions f, g satisfying f(m) 2

g(n) > 0. Suppose that

(3.1)

We rewrite (3.1) as

,,FJgf - I} = 44 .

Since f/g 2 1 we infer from (3.2) that

+&4
I--+~ -0 as x-w
(3.3) X .
for each E > 0, where $,(x) is as in (1.8). So f and g are nearly
the same almost always. (We can deduce (3.3) also if f(n) s g(n) for
all 12).
Setting k = 1 in (2.3) we see that (3.1) is true with f = d(n)
and g(n) = PI(%). Therefore A and P, are nearly the same almost
always. Since A 2 A* 2 P,, the same is true for all three functions.
Now to show that these three functions do not have normal
orders it suffices to show that one of them does not. It follows
easily from a theorem of Elliott [5] on additive functions
 ASYMPTOTIC BEHAVIOR OF LARGE PRIME FACTORS 299

(3.4)

that A* does not have a normal order. That proves the corollary.

REMARK. Since A(s) 2 log n, it follows from (2.2) that

A(N) - A*(w)
log n
(3.5)
ZZ

From (3.2), (3.3) and (3.5) we can deduce that A and A* are nearly
the same almost always.
Let us look a little more closely at (2.3) which for f = A or A*
and g = P, is a more accurate form of (3.1). We may rewrite (2.3)
as

(3.6)

where

We show in 0 5 that

(3.8) a; = ys - k)sk-2ds
1
where p is defined in (1.11). We deduce from (3.8) in 56 that a; is
a rational multiple of er for k > 1. The integral representation is
investigated in 5 6 and this leads to pretty connections with some
related problems.
The next section is devoted to obtaining upper and lower bounds
for S&r, k), i = 1, 2, 3, 4. This enables us to deduce the first four
asymptotic relations in (2.3). It is only 8 5 that we prove (3.7) and
(3.8). But the upper bound method in 8 4 is used in 5 5 to take care
of the error terms arising out of (3.6) and (2.3). For the reader
who does not want to go through the detailed proof, see [l], where
some of the ideas of this paper and an earlier paper by the authors
[2] are summarized.
We now move on to the proofs of our results.

4. Upper and lower bounds. In what follows, cl, cz, cB, a*-
300 K. ALLADI AND P.ERDC)S

denote absolute positive constants whose precise values will not be

our concern. Also exp {z} = e”. We begin by proving

THEOREM 1. There exists for each positive integer k a constant

b, and a real number x0 = x,(k) such that if x 2 x0 the% SJx, k) >
b,* x/(log x)A-f for i = 1, 2, 3, 4.

To prove this we need

LEMMA 1, Let s be a positive real number. Then

1
c 1
P>Zp(log p)” = s(log x)#
+ O(expf -c,V&Gj) .

Proof. We use the Prime Number Theorem [4], [9] in the form

(4.1) 1n(x) - h(x) 1 = 0(x exp { - c,VGjG}) .

Now write

(4.2)

Lemma 1 follows from (4.1) and (4.2).

Proof 1. It suffices to prove Theorem 1 for the

of Theorem
smallest of the four sums &(x, k).
Assume first that k > 1. For x sufficiently large choose a prime
p, in the interval

(4.3) k! x1/k+15 p1 6 x’lk ,

Now choose primes pz, p,, . .., pk satisfying

Consider any multiple m 5 x of p,p, l e. pk

(4.5) m = n’p1p2 l 8 l pk .

Because of (4.3) and (4.4) we have

(4.6) P,$h ’ ’ ’ pk 5 x
 ASYMPTOTIC BEHAVIOR OF LARGE PRIME FACTORS 301

and

(4.7)

By (4.7) and (4.4)

(4.8) Pl(n’) 2 n’ < 2 uk+l r (k - l)! x”k+’ < p, *
Thus by (4.5) and (4.8) we see that P,*(m) = pk. So any multiple
5 x of p, ’ e. p& has p,$ as its kth largest prime factor (Pk*). So

&(x, k) = t: p,*(rt> 2 c PZ(n)

(4.9) 2sd2 Pi(n) zs,?sss -XT’
?3=7&‘p1...p&
t.9,; mtisfying (4.4))

We can estimate the second sum in (4.9) by using the well known
result [ll]

(4.10) *Gk$- = log log x + c3 + O(exp { - c,y/logx}) .

Observe that the second sum in (4.9) is

by virtue of Lemma 1 and (4.10). Theorem 1 follows from (4.9) and

(4.11), for k > 1. For k = 1, Theorem 1 is trivially true.
Now for an upper bound.

THEOREM 2. All four sums S&z, k), i = 1, 2, 3, 4 are O(x/(log x)“-I)

where k is an integer 2 1, and the O-constant depends only on k.

We need a few preliminary results before proving Theorem 2.

LEMMA 2. Let k be a nonnegative integer and

Sk*@) =&log log x - log log p)” .

Then
302 K. ALLADI AND P. ERDOS

Proof. If we write S:(x) as a Stieltjes integral, use the fact

that

ddy) = & + d{n$y) - WY)} 9

integrate the second integral by parts and then use (4.1) we get

(4.12) Sk*(x) = \I (log log xlo; 7 log Y)kdy + 0 (c10g”,,k+3) .

Next

= (log log x - log log #Zi(y) /1

(4.13)
+ /( ~i(Y)(log ‘“~lo; ;” log dk-‘dy

= @(log log X)k) + $ zi(?d(log ‘““,“1, ;” log VI>“-’ dy .

But

(4.14) G(y)= i&+“k&-~*

So the integral in (4.13) becomes
k z(log log 2 - log log Y)!+l * (log log $ - log log y)“-’ dy
(4.15) s4 w Y
+ O(5 4 h3s Y >

= I, + Is .

We split I, into
*/(log .9+3
(4.16) I, = k Sk*
4 sz/(log .)k+3 -

Clearly in (4.16)

(4.17)
z/(log.)k+”
s =o(
4
x(log log @-l = 0
(log X)k+3 > ( (log~y ) ’
Regarding the second integral in (4.16) we observe that

(4.18) n (log log % - log log VY-’ dY ,

X
s e/mg.)k+3 log Y
 ASYMPTOTIC BEHAVIOR OF LARGE PRIME FACTORS 303

Now the last integral in (4.18) is

x(log log)k-1 _
(4.19) T!+-l(x) + O( (log x)k+3 ) - Tdx) + O ((10g;)“+2) *

From the definition of Tk we have

(4.20) To(x)= x+0x.

log x ( lo& x >
Now make the induction hypothesis that for k 2_ 1

(4.21) = (k - l)!x + 0 ( z(log log x)

Trc--1(x)
m- 2)” (log x)k+’ > l

Then from equations (4.16) through (4.21) we deduce that

I1 = k!x + 0 xloglogx
(4.22) (log X)k+’ ( (log Z)k+” 1 m
By analysis very similar to the above one can show that

I, = 0
( (log:)k+s > -

So from (4.22), (4.23), (4.15) and (4.13) we see that (4.21) is true for
Tk(x) and so by induction for all k 2 1. Lemma 2 follows from (4.12)
and (4.21).

LEMMA 3. Let x, g 2 4 be real numbers awi k 2 0 an integer.

Then

c (log log x - log log p)” _- (log log 2 - log log 7d)kfl
YdP$Z P lc+l
+ O,((log log x - log log v))” exp { - c,7/log Y}) .

Proof. As in the beginning of the proof of Lemma 2 we convert

the above sum into a Stieltjes integral and replace &c(g) by &/log v.
Lemma 3 can be easily proved by making the substitution log log x -
log log u = t. We do not go through the details.

Proof of Theorem 2. It suffices to prove Theorem 2 for the

largest of the four sums E&(x, k). That is we will show

(4.24) f&(X, k) = c A(m) - “‘n&-n; ’ - - pk-1(n) = o( (log;)l-l)

zsasz 1
for lc 2 1 an integer. We claim that it s&lees to prove (4.24), for
k 3 1 because for k = 1 we have
304 K. ALLADI AND P. ERDUS

A(n)pyn-cn)
1
(4.25)
=x - 1 + Sl(X, 2)
= O(x)
assuming that (4.24) is true for k = 2. So from now on we assume
that k > 1.
We write
A(n) - PI(w) - Pz(a) - l . . - P.&n) - Pk(n)
P,(n) PI(n)
(4.27)
+ pk+l(n) + .,.
P,(n) *
Let us denote a general nonzero term of (4.2’7) by pJp,. We would
like to know how often this term occurs in &(x, k). The term p,Jp,
occurs as often as we can find integers n = pip2 - - pk.--ll pkrn =( x l

where the p, satisfy Pi r p,-, 5 . 5 p, and PI(m) d pkml. If we

l l

fix the primes pi to satisfy these conditions then the number of such
n is given by

(4.28) , Pk-1)

where ?~pis defined in (1.10).

Thus we may rewrite (4.27) as

We first consider a subsum of (4.29) with a restriction on p,. That

is we choose ,B with 0 < ,6 < 1, whose vaIue will be specified later,
and consider p, in (4.29) satisfying $6 r p1 5 x. We shall get an
upper bound for this sum.
Observe that the sum in (4.29) with this extra condition on p, is

(4.30)

(Note: If FG= 2 in (4.30) we have only

(4.31) XC
dLPl’.$ 1P2Z, 1
 ASYMPTOTIC BEHAVIOR OF LARGE PRIME FACTORS 306

and no other terms, For k > 2, there is no confusion in (4.30).)

Because of this difference assume for the moment that k > 2. Then
if we use Lemma 3 we infer

(4.32) EL = O(log log p, - log log 1)s) .

l+@PLp2
Again by Lemma 3 and (4.32)

c - c L = O((log log p, - log log pJ”) *

(4.33) p,1 PQBP~SB1
P4SPSBBl p,
Iterating this process we get in (4.30) for k > 2

c 1. c (loglogz - log log &)‘-‘)

&PlsZ p: Pks;Pl
by repeated use of Lemma 3. Now observe that because of (4.31)
we see that (4.34) is true even for Fc= 2. Thus for Ic 2 2, we may
replace (4.30) by (4.34). Thus from now on we drop the assumption
k > 2, but of course still assume k > 1.
To estimate (4.34) we use Lemma 2 which gives

(4.35)

Finally Lemma 1 and (4.35) impIy that the sum in (4.35) and hence
in (4.30) is

0
( pk-‘& xy )
where the constant on the O-term in (4.36) depends only on k and
not on 6.
So (4.36) gives a bound for the sum in (4.29) with the condition
xp S p, 5 x. For the sum corresponding to p1 (= zfl we write

(4.37)

To estimate (4.37) we use the following result of de Bruijn [3]; If

II =x: lla then
(4.38) 4-.(x, 21) = 0(x em I- wl> .
In (4.37) consider the case
(4.39) xBfi*+l <
- p1 5 xfilk” ,
Then in (4.29) with the restriction (4.39) on p1 we have from (4.38)
306 X. ALLADI AND P. ERDUS

the following:

We choose p = ,8(k), depending on k, so small that

a > 2” - kp > -.2”~’
(4.41)
B P
Then by (4.38), (4.39), and (4.41) we will have in (4.29) for the
subsum corresponding to (4.39)

(4.42)

If we substitute (4.42) in (4.29) and analyze this sum just the way
we derived (4.36) we get

(4.42) O( {log(xb:m‘.‘)}“-’exp { - ~~~“-1/4

=o( X(2m+y/3)k-1
(log x)“-l exp {c,2”-l/B} > *
But then

(4.43)

This means that (4.43), (4.42), and (4.36) imply that in (4.29)
&(x, k) = O(x/(log X>k-l>
for k > 1. That completes the proof of Theorem 2.
It is interesting to note that Theorems 1 and 2 actually imply
the first four asymptotic relations in Theorem B, as will be shown
below. Before establishing this we prove the last part of Theorem
B namely

THEOREM 3. For each positive integer k we have

c A(m) - A*(n) = 0 (x exp { - c,‘L/log 2 log log 2))

2s*sz PI(n)
=o c Pk*(n) .
( z<fisz P,(a) >

Proof. First let 1 s ?J5 x and YJ= cP. N. G. de Bruijn [3]

showed that if 3 < a < 4y”‘/log Y then
 ASYMPTOTIC BEHAVIOR OF LARGE PRIME FACTORS 307

(4.44) $(x, y) = 0(x log2 y exp {-a log a - a log log a + ~,a}) ,
Take g = exp {-/log x log log 2). Then from (4.44) we have
(4.45) +(x, y) = 0(x exp { - c,/log x log log x}) .
Next observe that

(4.46) A(N)p-@--*(n) 2 Q(n) = O(logPS) ,

1
We now split
c A(n) - A”(n) =
(4.47) 2snsz 22, + 2szL = F + F .
E(n) Pl(n)std P1(“)>zI
Clearly from (4.46) and (4.45)
(4.48) J$ = O(log x*$(x, 76))= 0(x exp { -cc,,‘l/log x log log x}) .

But then by Theorem A, (2.2), we have

(4.49) C2 5 exp {-‘L/log x log log x} C (A(n) - A*(a))
2srz

= 0(x exp { - c,,tilog x log log x}) .

The first equation in Theorem 3 folIows from (4.47), (4.48) and (4.49).
The second equation is a consequence of Theorem 1. That proves
Theorem 3.

THEOREM 4. For every integer k 2 1 we have

Sk& k) - S2(x, k) - ax, w - ax, k) *

Proof. The smallest of the four sums is S&r, k). By Theorem 1

(4.50) Sl(X, Ii) 2 S,(x, k) s b,x/(log xyl .
The largest of the four sums is &(x, Ic). Consider the difference

A(m) - PI(n) - . m- P,(m)

=25&z
l

P,(n)
(4.51)

By Theorem 2
308 K. ALLADI AND P. ERDBS

(4.52)
But then

So by (4.53) and Theorem 3 we have

= 0 (z exp { -c,z/log 2 log log x}) .

Clearly from (4.51), (4.52) and (4.54)
(4.55) s&E, k) - s&c, k) = O(z/(log 2y) .
Thus from (4.55) and (4.50) we deduce
(4.56) X(x, N - S,(x, w l

But since these are the smallest and largest sums, Theorem 4 follows
from (4.56).
While proving Theorem 2 we did not use Lemmas 1, 2, and 3 in
the forms in which they were stated, but used only the upper bounds
they implied. These lemmas will play a role in obtaining asymptotic
estimates, which we take up in the next section. We refer to the
method of proof of Theorem 2 (namely the choice of ,8 and the con-
vergence of the series (4.43)), as the “upper bound method” and use
this method to take care of the error terms arising out of the
asymptotic estimates in what follows.

5, Asymptotic estimates, Our goal in this section is to prove

THEOREM 5. Let k be a positive irzteger. Then, all the four

sums S&, k), i = 1, 2, 3, 4 are asymptotically equal to
a:x/(log Z)k-’

a: = r p(s - k)skm8ds.
5

We need some lemmas before we go to the proof.

LEMMA 4. If cx 2 1 and E > 0 then

 ASYMPTOTIC BEHAVIOR OF LARGE PRIME FACTORS 309

Proof. It is well known (see [lo]) that p satisfies

Furthermore (see [lo], [3])

Combining (5.1) and (5.2) we get

0 5 p(a) - p(a + E)= 5,;’ f@ ; I) dt 5 lea ; UE = o( r(ae+ 1))

because p is monotonic decreasing.

LEMMA 5. There exists constants c13,cl, and cl5 such that if

a 2 1 and y = xlla, x 2 1, then
3 Cdf=
I44x, x1’? - xp(a) I 2 2 max ( exp {c,,z/log] c,3xa

’ eaj4.log x > *

Proof. Lemma 5 is obtained by combining certain results of de

Bruijn [3]. For the function A(x, x’ia) defined by de Bruijn, it is
known
(5.3) 1q-(x, xi/a) - d(x, xl/a) I < c13xa2exp { - c,,m
and

(5.4) 1d(x, x1/q - xp(a) / < c,,xa/(e”‘4.10g x) .

Lemma 5 follows from (5.3) and (5.4).

Proof of Theorem 5. Because of Theorem 4 it suffices to prove

Theorem 5 for one of sums S,(x, k). We consider &(x, k). So we
start with (4.29). (We assume k > 1 since Theorem 5 is trivially true
for lc = 1. (See (5.1), (5.2) and Theorem 4.)
In (4.29) we first look at the contribution due to numbers for
which

We will get an upper bound for the contribution due to such numbers.
310 K. ALLADI AND P. ERDdS

Let 0 < p < 1 be a real number whose value will be specified later.
Then write

In the interval xfilzmfl 5 p, < xb/2mone has an upper bound for + given
in (4.42), while for xB 5 p, d x we use the trivial upper bound

Then for numbers satisfying (5.5) together with $8 5 p, B 2, we have

the following bound in (4.29)

Analysis similar to (4.32), (4.33) and (4.34) yields

c ii-p, (loglogp, - loglog2)k)‘-‘)

&log p,)k+’Pk$Pt
6%

using Lemma 1. To estimate the contribution due to integers satis-

fying (5.5) for the case p, (= z 8, we use the decomposition of the last
sum of (5.6). Then the upper bound method yields

(5.10)

provided ,B is suitably chosen. Thus from (5.9) and (5.10) we conclude

that the contribution due to terms satisfying (5.5) is given by (5.10),
and is smaller than the asymptotic term we are seeking.
Next we observe that the contribution due to terms for which
p1 = pi(n) is small is negligible, For that purpose set
(5.11) y = (exp {(log x>~/~}).
With g as in (5.11) we have by (4.38)
(5.12) +-(x, 21)= 0(x exp I-cdlw xY31) .
So, if p, = P,(n) 2 21, then
c A(m) - PI(n) - ..a - P,-,(n)
(5.13) zdnSz P,(n)
= O(log xg(z, 21))= 0(x exp { - c,,(log x)~/~}) .
 ASYMPTOTIC BEHAVIOR OF LARGE PRIME FACTORS 311

Because of (5.13) and (&lo), we assume from now on that

(5.14) I pk S p,; Pl(lz) = p, > exp {(log xY3} .

(log?)bJk+l -
Once we assume (5.14) we can rewrite Lemma 5 as

‘k(p,p,“. . pkt Pk-1

) = xP( ‘““‘;;;;;-: jpk’)
(5.15) x.pN4.a
4-O
Pl ’ ’ ’ Pk log Pk--l
>

where a = log (z/p1 00 pJlog l pk-1.

The idea is to substitute (5.15) in (4.29). It is then easy to take
care of the contribution due to the error term in (5.15) in (4.29) by
observing that (5.14)

x 2 x, .
(5.16) log Pk-, 2 log Pk - log Pl > $ log Pl ,

This means if we substitute the O-term of (5.15) in (4.29), and use

the upper bound method we get
(5.17) wm% Xlkl -
The convergence of a series like (4.43) is ensured this time by the
e-@ term in (5.15). Since (5.17) is smaller than the asymptotic term
we are seeking, we may forget the contribution of the O-term in
(5.15), in the sum (4.29).
As to the leading term of (5.15) we observe that

log (xh ’ ’ * pk) log % - xt, log p,

(5.18)
log Pk-, log Pk--l *

By (5.14) we have
(5.19) log Pi = log Pl + om log PJ , 1~iIk.

Substituting (5.19) in (5.18) we get

(5.20) p( log EFpi:,’ “) ) = p {s - k + 0 (log ‘iJTjrg ‘l)} .

Using Lemma 4 to estimate (5.20) we get

(5.21)

where a is as in (5.15).
312 K. ALLADI AND P. ERDOS

Thus the factor p in the leading term of (5.15) is equal to the

quantity in (5.21). Recall that our idea is to substitute (5.15) in
(4.29) and estimate the sum, The contribution of the O-terms in
(5.21) can be obtained by the upper bound method. There is a log II:
in the numerator, but a log” p, in the denominator. This time the
presence of r(a) in the denominator ensures convergence in a series
like (4.43). Thus the upper bound method yields

(5.22) 0(x log log z/(log X)k>

as the contribution due to the O-term of (5.21). Thus we deduce

that the main contribution from (4.29) comes by assuming (5.14) and
replacing +(x/P~ - -. P,, P~-J by

(5.23)

So we replace (4.29) by

(5.24)

To estimate (5.24) we use Lemma 3. First we get

(5.25) p,5gsp,k = (log log P, - log log p3) + O(exp I- W%x P3H .

The contribution due to the O-term in (5.25) in (5.24) is taken care

of by the upper bound method. This time the presence of p in (5.24)
ensures convergence of a series like (4.43), because of (5.2). Actually
every error term that arises in (5.24) by repeated use of Lemma 3
can be estimated by the upper bound method, yielding

(5.26) 0(x exp { - c,,tilog x}) .

So we need only look at the leading terms arising out of Lemma 3
in (5.24). After k-2 applications of the lemma we are left with

X c P((lW xmg PJ - k)
expl(log~Ptspt5z P;
(5.27)
(log log Pl - lofir log Pkjk+ .
X c
Pi\(1%PI)kflXPkSP1 (k - a!

In (5.27) we use Lemma 2 to get

 ASYMPTOTIC BEHAVIOR OF LARGE PRIME FACTORS 313

P((lw- x)/m Pl>> - w

X c
exp~~log 2)*/31s-Pl% P:

x
C (log;1)k-L + O(
Pl(lW

(log
log

Pd”
PJ
>I*
As before, the O-term in (5.28) contributes
(5.29) 0(x log log x/(log z)k>
by use of the upper bound method. Finally the leading term in
(5.28) is estimated by writing it as a Stieltjes integral, That is

X c P(UW xmg Pl) - k)

expl(logz)~~3l$P1s% P,(lW PP
zf
epx((log
z) ,3_ .PNlW xmg 24 - NW&
2: ?dlw 2/Y-’
(5.30) 2:
=x e9p((lo 21)‘
,3~P((lW x)lOw II) - k) (-jy
s c 2 zluw 7dY
Z+(log
+x sexp,z)2,3~-
P((lWMlw ?I)- k)
vex 2/Y-’
x 4eY)- Zi(rj/)}
= I3+ I, .
We can bound I, rather easily. First observe that 1p 1 5 1. Ignoring
o, we integrate by parts, and use (4.1) to deduce
(5.31) I4 = O((x exp {-&log x)lf3}) .
To estimate & write g = xl/‘. Then
(log Z) ‘/3
13 = p(s - k)&“ds
(lo&& s 1

(5.32) = (log$’ K - Lg.,131

=a;(logs)“-’
+ O( (lo[x)n >
because of (5.2). So Theorem 5 follows from (5.32) and the preceding
estimates.

REMARKS. Note that we have actually shown that

&(x, k) = a; + 0 xloglogx
(5.33) (log& ( (log x)” > -
Observe that S(z, k) is the largest of the four sums and S,(x, k) is
the smallest. Therefore, because of (4.55), we deduce a stronger
form of Theorem 5, namely
314 K. ALLADI AND P. ERDOS

E&(x,
k) = a:(log xX)k-l + 0 xloglogX
(5.34) ( (log 2y >
for i = 1, 2, 3, 4.
Thus we have proved all the statements of Theorem B, except
the relation between ub and e7. We do this in the next section.

6. The constants a:+ It is obvious from Theorem 4 or (5.1)

and (5.2) that a: = 1. So we suppose k 2 2. For k 2 2 write

(6.2) a: = y-p& - k)s”-‘ds = Ipp(s - k)sk-2ds = 1; p(t>(t + k)k-2dt

s s

where

In a recent paper, Knuth and Pardo [8], have studied the behavior of

In the course of their investigations they show

(6.5) fj = ergi
where r is Euler’s constant and the gj are. recursively defined by

(6.6) go = g1 = 1 ,
gj=-lc J‘
gj-i fr J’ZO.
j 1SiSj i
0

Combining (6.2), (6.5), and (6.6) we infer that ah is a rational

multiple of er for k 2 2. For instance
ai = f, = erg, = e’ .
That completes the proof of Theorem B.

Acknowledgments. K, Alladi is grateful to Professors R. Miech

and E. G. Straus for their advice and encouragement. They went
through an early draft of this paper and made helpful suggestions.
We also thank the referee for his critical comments and suggestions.

REFERENCES

1. K. Alladi, Analogues to the HardpRamanujan theorems, in Proceedings of the

 ASYMPTOTIC BEHAVIOR OF LARGE PRIME FACTORS 315

Conference of Number Theory and Computational Methods, Publications of “Matscience”,

Madras, India (1977).
2. K. Alladi and P. Erdiis, On a1z additive arithmetic function, Pacific J. Math., 71
(2), (1977), 55-294.
3. N. G. de Bruijn, On the number of positive integers 2 x and free of prime fac-
tors > y, Indag. Math., 13 (1951), 50-60.
4. K. Chandrasekharan, Arithmetical Functions, Springer-Verlag, New York, 1970.
5. P. D. T.A. Elliott, On a problem of Hardy and Ramanujan, Mathematika 23,
(1976), 10-17.
6. G. H. Hardy and S. Ramanujan, The normal number of prime factors of a number
n, Quarterly J. Math., 58 (1917), 76-92.
7. G. H. Hardy and E. M. Wright, An Introduction to the Theory of Numbers,
Oxford, Clarendon, 1960, 4-Ed.
8. D. E. Knuth and L. T. Pardo, Analysis of a simple factorisation algorithm, (to
appear).
9. W. J. LeVeque, Topics in Number Theory, Vol. 2, Addison-Wesley, Reading, Mass.,
1956.
10. K. Norton, Numbers with small prime factors and the least kth-power +zon-residue,
Memoirs of the A. M. S., 106, Providence, Rhode Island, 1971.
11. J. B. Rosser and L. Schoenfeld, Approximate formulas for some functions of
prime numbers, Illinois J. Math., 6 (1962), 64-94.

Received April 3, 1978 and in revised form October 16, 1978. This work was supported
in part under National Science Foundation Grants No. MCS 77-01780 and MCS 78-02685.

UNIVERSITY OF CALIFORNIA
DEPARTMENT OF MATHE~IATICS
Los ANGEI.ES, CALIFORNIA 90024
AND
MATHEMATICS INSTITUTE!
THE HUNGARIAN ACADEMY OF SCIENCES
BUDAPEST, HUNGARY

Current Address of jirst author: University of Michigan

Department of Mathematics
Ann Arbor, MI 48109