ON A CONJECTURE OF HAMMERSLEY

P . ERDÖS*.

[Extracted from the Journal of the London Mathematical Society, Vol . 28, 1953 .]

Denote by E n,, the sum of the products of the first n natural numbers
taken s at a time, i.e. the s-th elementary symmetric function of 1, 2, . . ., n .
Hammersleyt conjectured that the value of 8 which maximises E., ., for
a given n is unique . In the present note I shall prove this conjecture and
discuss some related problems .
We shall denote byfln) the largest value of s for which E n, $ assumes its
maximum value . As Hammersleyt remarks, it follows immediately from
a theorem of Newton that

I n,1 < Y-.,2 < . . . < F+n, f(n)-1 < En, An) > F'n, f(n)+1 > . . . > E n n = n! . (1)
Thus it follows from (1) that the uniqueness of the maximising s will
follow if we can prove that
Z n,f(n)-1 < F'n,f(n)- (2)
Hammersley proves (2) for 1 < n < 188 . He also proves that

f(n)=n-[log (n+1) +y-1-E log (3)

(n+ 1 )+y - z + (log (n+1) { y-z)2J'

* Received 27 February, 1952 ; read, 20 March, 1952 .

f J . M . Hammersley, Proc. London Math . Soc . (3), 1 (1951), 435-452 .
233 P. ERnos

where [x] denotes the integral part of x, y denotes Euler's constant, ~(k) is
the Riemann c-function and -1 . 1 <h< 1 .5. Thus for n > 188 > e5 we
obtain by a simple computation
[log n- a] < n-f (n) < [log n]. (4)
First we prove
THEOREM 1 . For sufficiently large n all the integers En, a, 1 < 8 < n,
are different .
We evidently have*
! n k ! k
En, n-k<ki (Ei L) <ki
(1+1ogn)k<n! {
k
(l+logn)} <n!=Ef,n (5)

for k > e(logn+l) . Thus from (1) and (5) it follows that to prove
Theorem 1 we have only to consider the values
0 < k < e(logn+1) . (6)
The Prime Number Theorem in its slightly sharper form states that
for every l
x
"(x)= (7)
2log
~0 y+0 ((log x)i)'
From (7) we have that for sufficiently large x there is a prime between x
and x+x/(log x)2. Thus we obtain that for n > no and k < e(log n+l)
there always is a prime pk satisfying
n n
k+1 <Pk< k .
We have
En, n-k # 0 (mod pk) . (8)

For En, n-k is the sum of ( k) products each having n-k factors . Clearly
only one of these products is not a multiple of Pk (viz ., the one in which
none of the k multiples not exceeding n of pk occur) ; thus (8) is proved .
For r < k all the (n ) summands of En,n_r are multiples of pk. Thus
r
En,n_r =-0 (mod Pk) • (9)
(8) and (9) complete the proof of Theorem 1 .
We now give an elementary proof of Theorem 1 which will be needed
in the proof of Hammersley's conjecture . Let
r < k < e(logn+1) . (10)

* The proof is similar to the one in a joint paper with Niven, Bull . Amer . Math . Soc .,
52(1946),248-251 . We prove there that for n > no, 0 .,, (mod n!) .
a
 ON A CONJECTURE OF HAMMERSLEY . 234

We shall prove that for n > 10 8

F'n, n-r : En, n-k . (11)

Let q be a prime satisfying n/2k < q < n/k. Assume that

<q< i, k<l<2k-1 .
l-{ 1
Clearly En, nJ, - 0 (mod ql-r) . (12)
Now we compute the residue of En, n-k (mod ql-k+l) . Clearly
Y'n, n-k = 0 (mod ql -k) . The only summands of En, n-k which are not
multiples of ql-k+l are those which contain H' t where the product is
extended over the integers 1 < t < n, t # 0 (mod q) . H' t contains n-l
factors, and the remaining l-k factors of the summands in question of
En, n-k must be among the integers q, 2q, . . ., 1q. Thus clearly

E n, n-k - El , t -k . Wt . ql-k (mod qt-k+1), (13)

Therefore if (11) does not hold we must have

l-k+1)l .
El, l_k = 0 (mod q) (i .e . En, n-k - En, n-r - 0 (mod q

Thus if (11) is false

2k-1
11 q F'l, l-k
n/2k<g<n/k l=k

Now evidently (we can of course assume that k > 2 for if k = 1 then (11)
clearly holds)
2k-1 2k-1 l 2k-1
fl O ll-k < H (2k) l < (2k)~'ka < k3ka < (3 log n)27(logn)2, (15)
< l=k
F'l,l-k
l=k k l=k
since for n > 108 > e 10, k < e (1 +log n) < 3 log n . Define

$(x) = E loge .
p<x

By the well-known results of Tchebycheff* we have

$(2x)-$(x) > 0-7 . x-3 . 4 . xi-4.5(logx) 2-24 logx-32 .
Thus for n > 10 4 we have by a simple computation

$(2x)-$(x) > Ix. (16)

For n > 10 8, we have n/2k > n/(6 log n) > 104. Thus from (16) we have
q > en/4k > e n/(12 logn) • (17)
nl2k<q<n/k

* E . Landau, Verteilung der Primzahlen, I, 91 .

235 P.ERDÖS

From (14), (15) and (17) we have

(3 log ")27009 n? n/(121ogn) .
>1 e

Thus on taking logarithms and using log (3 log n) < log n for n > 108,
27 (logn)3 > n/(121ogn) or 324 (logn)4 > n,
which is false for n > 108. Thus the proof of Theorem 1 is complete .

THEOREM 2 (Hammersley's conjecture) . The value of s which maxi-

mises In, g is unique ; in other words
I n, f(n)-1 0 In, f(n)- (18)
It follows from the second proof of Theorem 1 that Theorem 2 certainly
holds if for n > 108. Thus since Hammersley proved Theorem 2 for n G 188
it suffices to consider the interval 188 < n < 10 8.
Put n-f(n) = t . We have, from (4),
logn-2 <t <logn. (19)
As was shown in the first proof of Theorem 1, (18) certainly holds if
there is a prime satisfying
n/(t+2) < p <n/(t+1) . (20)
It follows from (19) that if 1500 < n < 10 8
150 < n/(t+2) < 107.
The tables of primes* show that for 150 < x < 10 7 there always is a
prime q satisfying x < q < x+x}. For n > 1500 we have
l 8
t+2 + \t+2) < t+1 ,
( i
since
(t+l)(t+2) > `t+2) ,
or, by using (19), n > (1+ log n) 2 (2+logn),
which holds for n > 1500 . Thus for 1500 < n < 10 7 there always is a
prime in the interval (20) and thus Theorem 2 is proved for n > 1500 .
To complete our proof we only have to dispose of the n satisfying
188 < n < 1500 . Hammersley f showed that for n < 1500 the only
doubtful values of n are : 189 < n < 216, 539 < n < 580 . He also showed
that if 189 < n < 216 and (18) does not hold, then t = 5 . But then p = 31
is in the interval (20), which shows that (18) holds in this case . If
539 < n < 590 and (18) does not hold, he shows that t = 6 . But then

* A . E . Western, Journal London Math . Soc., 9 (1934), 276-278 .

t See footnote t, p . 232 .
 ON A CONJECTURE OF HAMMERSLEY . 236

either p = 73 or p = 79 lies in the interval (20) . Thus (18) holds here too,
and the proof of Theorem 2 is complete .
By slightly longer computations we could prove that for n > 5000
Theorem 1 holds . Theorem 1 is certainly not true for all values of n since
1 3,1 = E3,3' Hammersley proved that for n < 12 this is the only case
for which Theorem 1 fails, and it is possible that Theorem 1 holds for all
n > 3. The condition n > 5000 could be considerably relaxed, but to prove
Theorem 1 for n > 3 would require much longer computations .
Let ul < U2 < . . . be an infinite sequence of integers . Denote again by
In,, the sum of the products of the first n of them taken s at a time . It
seems possible that for n > n„ (n„ depends on the sequence) the maximising
s is unique and even that for n > nl all the n numbers In, ,, 1 < s < n are
distinct . If the u's are the integers -=a (mod d) it is not hard to prove
this theorem.
Stone and I proved by elementary methods the following
THEOREM . Let U1 < U2 < . . . be an infinite sequence of positive real
numbers such that
E 1
1= co and E U, 2
-<00
2 .
ui

Denote by En s the sum of the product of the first n of them taken s at a time and
denote by f (n) the largest value of s for which In, 8 assumes its maximum value .
Then
f(n) =n-C E 1 - E (I+
b =1 ui i=1 ui ui

Department of Mathematics,
University College, London .

Printed by C . F . Hodgson &Son, Ltd ., Pakenham Street, London, W .C.I.