EGYPTIAN FRACTIONS WITH EACH DENOMINATOR HAVING

THREE DISTINCT PRIME DIVISORS

Steve Butler
Department of Mathematics, Iowa State University, Ames, IA 50011, USA
[email protected]

Paul Erdős1
Rényi Institute, Budapest, Hungary

Ron Graham
Department of Computer Science and Engineering,
UC San Diego, La Jolla, CA 92037, USA
[email protected]

Abstract
P
Any natural number can be expressed as an Egyptian fraction, i.e., 1/ai with
a1 < a2 < · · · < a` , where each denominator is the product of three distinct primes.

1. Introduction

Egyptian fractions date back over 3500 years to the Rhind papyrus [5] (making them
among the oldest mathematics still extant). They are a way to express rational
numbers in a very specific form, namely
m 1 1 1
= + + ··· + , where a1 < a2 < · · · < a` .
n a1 a2 a`
No one is quite certain why ancient Egyptians chose to express fractions this way
(though André Weil offered the following explanation: “It is easy to explain. They
took a wrong turn! ” [1]).
The existence of Egyptian fractions for any rational number has been known
since at least Fibonacci (for example, the greedy algorithm will always produce a
solution, though other methods are known). However, one can place additional
constraints on the allowable ai and then interesting questions arise as to what is
possible. In this note we will establish the following result.

Theorem 1. Any natural number can be written as an Egyptian fraction where

each denominator is the product of three distinct primes.
1 Deceased.
 A stronger version of this result for rational numbers with square-free denom-
inators was previously mentioned by Guy [2, D11] and attributed to two of the
authors (Erdős & Graham); however a proof of this result was never published! To
begin correcting this situation, we give a proof of the natural number case here. We
note that similar arguments could be used to show that any natural number can be
written as an Egyptian fraction where each ai is the product of ω distinct primes,
for ω ≥ 4. We also conjecture that a similar result holds for ω = 2. As an example
Johnson [3] showed how 1 can be expressed as the sum of 48 unit fractions with the
following denominators each of which is the product of two primes (taken from [2]):

6 21 34 46 58 77 87 115 155 215 287 391

10 22 35 51 62 82 91 119 187 221 299 689
14 26 38 55 65 85 93 123 203 247 319 731
15 33 39 57 69 86 95 133 209 265 323 901

2. Large block in the sums of products of primes

The main work in establishing Theorem 1 will be to show that there is a large
contiguous block of integers composed of sums of products of primes. We will need
the following definitions.

Definition 1. Let pn denote the n-th prime. Let


Sn (k) = pi1 pi2 · · · pik : 1 ≤ i1 < i2 < · · · < ik ≤ n},

i.e., the set of nk products of k distinct primes from among the first n primes.

Given a set X = {x1 , x2 , . . . , xm }, we let

m
X 
P (X) = εi xi : εi ∈ {0, 1} ,
i=1

i.e., the set of all
possible subset sums involving elements of X. Finally, we let
P
Ln (k) = P Sn (k) and σn (k) = s∈Sn (k) s; note that σn (k) is the maximal element
of Ln (k).

From the definition we have Ln (k) is symmetric, and for some small values of
n and k it can be shown by computation that there are large contiguous blocks of
elements in the middle of Ln (k). For example, we have the following:

• L8 (4) has σ8 (4) = 414849 and contains i for 3482 ≤ i ≤ 411367.

• L8 (5) has σ8 (5) = 2429223 and contains i for 54728 ≤ i ≤ 2374495.

• L8 (6) has σ8 (6) = 8130689 and contains i for 1750114 ≤ i ≤ 6380575.

2
Lemma 1. For n ≥ 5, Ln+3 (n) contains i for (1/6)σn+3 (n) ≤ i ≤ (5/6)σn+3 (n).

Before we begin the proof we state some results which will be needed.

Fact 1 (Chebyshev). For n ≥ 1, pn+1 /pn ≤ 2.

Theorem 2 (Olson [4]). Let X be a set of distinct nonzero numbers modulo p,

where p is a prime. If p < (|X|2 + 3)/4, then P (X) contains all residues modulo p.


Fact 2. For n ≥ 3, the set P Sn+2 (n + 1) contains all residues modulo pn+4 .

Proof. For 3 ≤ n ≤ 13 this was verified computationally. For n ≥ 14 we apply

Theorem 2. We note that elements in Sn+2 (n + 1) are of the form p1 · · · pbi · · · pn+2 ,
i.e., the product of all but one of the first n + 2 primes. Further, these entries are
all distinct (p1 · · · pbi · · · pn+2 ≡ p1 · · · pbj · · · pn+2 (mod pn+4 ) if and only if pj ≡ pi )
and nonzero (since pn+4 does not divide any of these terms). Finally we note that
|Sn+2 (n + 1)| = n + 2. Thus we obtain all residues if pn+4 < (n2 + 4n + 7)/4, which
holds for all n ≥ 14 (since pn is bounded above by n ln n + n ln ln n (c.f. [6]) which
grows much more slowly than (n2 + 4n + 7)/4).

Fact 3. For n ≥ 3, p1 p2 · · · pn pn+3 ≤ σn+2 (n + 1).

Proof. Using the definition of σn+2 (n + 1) and dividing both sides by p1 p2 · · · pn ,

this is equivalent to showing that
 
1 1
pn+1 pn+2 + ··· + + pn+1 + pn+2 ≥ pn+3 .
p1 pn

Since (1/2) + (1/3) + (1/5) > 1 we have

 
1 1
pn+1 pn+2 + ··· + + pn+1 + pn+2 ≥ pn+1 pn+2 + pn+2 ≥ 2pn+2 ≥ pn+3 ,
p1 pn
where at the last step we used Fact 1.

Fact 4. For n ≥ 1, 3σn+2 (n + 1) ≤ pn+4 σn+3 (n).

Proof. A computation verifies this for n ≤ 6, so we can assume n ≥ 7. Consider

the following (subscripts in the denominator being taken modulo n + 2):
X pn+3 pn+4 X pn+4 X  pn+4 pn+4 pn+4

+ > + +
pi pj pk pi pj pi pi+1 pi pi+2 pi pi+3
1≤i<j<k≤n+2 1≤i<j≤n+2 1≤i≤n+2
 
X 1 1 1 X 3
> + + = .
pi pi pi pi
1≤i≤n+2 1≤i≤n+2

For the first inequality, we drop our first term and most of the summands in the
second, making sure that what remains is a group of distinct summands. For the

3
second inequality, we note that each summand involves a ratio of pn+4 and a smaller
prime, which is greater than 1. Multiplying the left and right by p1 p2 · · · pn+2 this
becomes pn+4 σn+3 (n) > 3σn+2 (n + 1), as desired.

We are now ready to proceed with the proof of the main lemma.

Proof of Lemma 1. By the previously noted computation, this is true for n = 5 (in
fact it is also true for n = 3 and 4).
We now proceed by induction and note that

Ln+4 (n + 1) = Ln+3 (n + 1) + pn+4 Ln+3 (n), (1)

i.e., the first term on the right are all the terms not involving the largest available
prime while the second term are all the terms which do involve the largest available
prime (here the “+” indicates we take all pairwise sums of the two sets). By
our induction hypothesis we have that pn+4 Ln+3 (n) contains all multiples of pn+4
between 61 pn+4 σn+3 (n) and 65 pn+4 σn+3 (n).
For the first term we decompose it in the following way:


Ln+3 (n + 1) = P Sn+2 (n + 1) + P (Tn ), (2)


where Tn = Sn+3 (n + 1) \ Sn+2 (n + 1). By Fact 2 we have that P Sn+2 (n + 1)
contains all of the residues modulo pn+4 .
Now let us focus on Tn , with elements t1 < t2 < · · · < tm . We have

Tn = {p1 p2 · · · pn pn+3 = t1 , t2 , . . . , tm = p3 p4 · · · pn+3 }.

Note that we can go from t1 to tm by a sequence of elements of T where at each

stage we replace some pi by pi+1 . From Fact 1 we can conclude that the ratio of
two elements in this chain is at most 2 and therefore we have ti+1 /ti ≤ 2 for all i.
Because the ratios between consecutive ti are bounded by 2 we claim

tj+1 − (t1 + t2 + · · · + tj ) ≤ t1 for j ≥ 1.

To see this we use the ratio bound to observe that

2t1 + t2 + · · · + tj ≥ t1 + (1/2) t2 + · · · + tj+1 .

Multiplying through by 2 and rearranging then gives us the desired inequality.

We now claim that
the largest gap in P (Tn ) is t1 . We show this
by an induction
on i for P {t1 , . . . , ti } . First note that for i = 1 we have P {t1 } = {0, t1 }, which

has a gap of size t1 . Now suppose that the result holds for P {t1 , . . . , ti } and
consider






P {t1 , . . . , ti , ti+1 } = P {t1 , . . . , ti } ∪ ti+1 + P {t1 , . . . , ti } .

4



In other words P {t1 , . . . , ti
, ti+1 } consists of the union of P {t1 , . . . , ti } and a
translation of P {t1 , . . . , ti
} by ti+1 . By our induction hypothesis we know that
each copy of P {t1 , . . . , ti } has gaps bounded by at most t1 . So if there is a larger
gap it must come between the two copies. In particular, if the two copies overlap,
then we are done. Otherwise the gap between the two copies is ti+1 −(t1 +t2 +· · ·+ti )
(minimal element in the second minus the maximal element in the first), and by
the preceding claim this is at most t1 and we are again done.
We can now rewrite (2) in the following way:
[


Ln+3 (n + 1) = vi + P Sn+2 (n + 1) .
vi ∈P (Tn )

In particular we have that Ln+3 consists of shifted copies of P Sn+2 (n + 1) .
We know by
Fact 3 that t1 < σn+2 (n + 1). By Fact 2 we know that each copy of
P Sn+2 (n + 1) contains all of the residues modulo pn+4 .
Let u1 , u2 , . . . be the elements of Ln+3 (n + 1) congruent to r0 (mod pn+4 ) for
some fixed r0 . Then the preceding observations allow us to conclude the following:


• u1 ≤ σn+2 (n + 1), since there is
a copy of P Sn+2 (n + 1) shifted by 0 in
Ln+3 (n + 1) and P Sn+2 (n + 1) has all residues modulo pn+4 .


• ui+1 − ui ≤ 2σn+2 (n + 1), since the length of P Sn+2 (n + 1) is more than the
distance between any two consecutive values used for translation (i.e., this
distance is at most t1 ), every copy other than the last must intersect with
the succeeding copy. Therefore in any interval of length twice the width,

i.e.,
2σn+2 (n + 1), there must be at least one full copy of P Sn+2 (n + 1) , and so
also an element uj .

Now we finally return to (1). We proceed by placing many copies of pn+4 Ln+3 (n)
(where the middle consists of a long run of terms which differ by pn+4 ), one for each
element of Ln+3 (n + 1) which act as translates.
We claim that we have every element of Ln+4 (n + 1) which is congruent to r0
(mod pn+4 ) in the interval (1/6)σn+4 (n + 1) and (5/6)σn+4 (n + 1). Since this will
hold for any arbitrary r0 then we can finally conclude that Ln+4 (n + 1) contains i
for (1/6)σn+4 (n + 1) ≤ i ≤ (5/6)σn+4 (n + 1), which will finish the proof.
In particular we will focus on the translations by ui (forcing everything to be
congruent to r0 (mod pn+4 )), and show that the run of consecutive multiples of
pn+4 will connect together and start soon enough to satisfy what we need.
To see that they connect (i.e., one run will start before the previous run finishes)
we need to have for i ≥ 1,
5 1
ui + pn+4 σn+3 (n) ≥ ui+1 + pn+4 σn+3 (n),
6 6

5
or equivalently,
2
ui+1 − ui ≤ pn+4 σn+3 (n).
3
We previously noted that ui+1 − ui ≤ 2σn+2 (n + 1) and so this last relationship will
hold if
2
2σn+3 (n + 1) ≤ pn+4 σn+3 (n).
3
But this is precisely Fact 4. Therefore we have that these runs connect.
To see that they begin soon enough we need to have
1 1
pn+4 σn+3 (n) ≤ σn+4 (n + 1).
6 6
This is true since the left hand side can be found by summing a subset of the terms
which sum to give the right hand side. Therefore we have that we will begin soon
enough. (By symmetry if we start soon enough, we will also end late enough.)
We now have the needed elements modulo r0 in the middle of Ln+4 (n+1). Since,
as already noted, r0 was arbitrary this concludes the proof of the lemma.

3. Estimating the number of terms

By showing that Ln+3 (n) has a contiguous long middle section we can now finish
the proof of the main result.

Proof of Theorem 1. Let m be a natural number. If for some n

1 5
Ln+3 (n) ≤ p1 p2 · · · pn+3 m ≤ Ln+3 (n),
6 6
then by Lemma 1 we can conclude that
X
p1 p2 · · · pn+3 m = pi1 pi2 · · · pin
S

where the sum on the right hand side is over a collection of distinct sets S having
the form S = {i1 , . . . , in } ⊆ {1, . . . , n+3}. Now dividing both sides by p1 p2 · · · pn+3
we get our desired expression of m as an Egyptian fraction where each denominator
is the product of three distinct primes.
So it remains to understand the behavior of
Ln+3 (n) X 1
= .
p1 p2 · · · pn+3 pi pj pk
1≤i<j<k≤n+3

But we now note that X 1

∼ log log n
pi
1≤i≤n

6
(see [6]), and so
 X 3
1
3
∼ log log n .
pi
1≤i≤n

But we also have that

 X 3
1 X 1 X 1 X 1
= 3 +2 +6 .
pi pi p2i pj pi pj pk
1≤i≤n 1≤i≤n i6=j 1≤i<j<k≤n+3

The first term on the right is bounded by some constant c1 , the second term will
have growth bounded by c2 log log n for some constant c2 , and so we can conclude
that the third term drives the growth. In particular we have that the sum will go
to infinity, but do so slowly and so that for any m we will find some suitable value
of n so
1 X 1 5 X 1
≤m≤ ,
6 pi pj pk 6 pi pj pk
1≤i<j<k≤n+3 1≤i<j<k≤n+3

establishing the result.

We note the proof also helps give a bound for the number of terms that will be
needed. In particular, for a large m, we choose n so that
5
3 p

log log n ≈ m or n ≈ exp exp( 3 36m/5) ,
36
i.e., so that m falls into the appropriate
interval to be able to express its location.
Once n is chosen then we will have n+3 n ∼ n3 /6 possible terms to work with and
in the worst case scenario we will need all of them so that we will only need at most
1 p

exp 3 exp( 3 36m/5)
6
terms to express m as a sum of unit fractions whose denominators are products of
three primes. Even for m = 1 we are unlikely to ever find the actual decomposition
as it will possibly involve over one hundred million unit fractions!
We also note that the proof shows that we can write many rational numbers as
Egyptian fractions with denominators the product of three primes. In particular,
we have that if q is a rational number with q 0 = p1 p2 · · · pn+3 q a natural number
with 16 σn+3 (n) ≤ q 0 ≤ 65 σn+3 (n) (or more generally as long as q 0 ∈ Ln+3 (n)), then
we can also write q 0 as an Egyptian fraction where denominators are the products
of three distinct primes.
For example we have that
1
·2·3·5·7·11·13 = 4290 ∈ L(6, 3),
7

7
so we can write 1/7 as an Egyptian fraction whose denominators are the prod-
ucts of three primes. This can be done with 12 unit fractions with the following
denominators:

30 42 66 70 78 105 110 154 165 195 273 286

The approach we have given here will not be enough to establish the stronger
result that every rational number of the form m/n, where n is square-free, can be
expressed as an Egyptian fraction whose denominators are the product of three
primes. The problem is that the structure of Ln+3 (n) is not well understood at
the ends of the interval. So the main challenge becomes in recovering very small
rational numbers.2
Understanding the structure of Ln (k) is related to the following problem: For
each k, determine the largest number, a(k), which cannot be written as the sum
of distinct numbers each having k distinct prime divisors. For example, we have
a(1) = 6, a(2) = 23, a(3) = 299, a(4) = 3439, a(5) = 51637, and a(6) = 894211.
The large contiguous block in the center of Ln (k) will always start after a(k) and so
understanding this behavior gives some indication on what is happening on the end.
(Indeed, one can think of a(k) + 1 as marking the start of the infinite contiguous
block in L∞ (k).)
Another variation that can be considered is to not use the small primes. In
particular, suppose we want to only use the primes pr , pr+1 , . . . for the prime divisors
of the denominators. The same arguments carry through, we only need to verify
the initial conditions. As an example, let

Sn0 (k) = pi1 pi2 · · · pik : 2 ≤ i1 < i2 < · · · < ik ≤ n},



and L0n (k) = P Sn (k) and σn0 (k) = s∈Sn0 (k) s, i.e., these are the analogous def-

P

initions of Sn (k) and Ln (k) except we do not use the prime 2. We then have the
following:

• L07 (4) has σ70 (4) = 340419 and contains i for 20877 ≤ i ≤ 319542.

This is more than sufficient to begin the induction and establish the following result.

Theorem 3. Any natural number can be written as an Egyptian fraction where

each denominator is the product of three distinct odd primes.

We look forward to seeing further progress in the area of Egyptian fractions.

2 One of the authors believes that all rational numbers can be expressed in this form, another
author has doubts that every rational number can be expressed in this form, and the third author,
already having looked in The BOOK at the answer, remains silent on this issue.

8
References

[1] R. Graham, Paul Erdős and Egyptian fractions, in Erdős Centennial, L. Lovász,
I. Ruzsa, V. T. Sós and D. Palvolgyi eds., Bolyai Society Mathematical Studies
25, Springer-Verlag, Heidelberg (2013), 289–309.

[2] R. K. Guy, Unsolved Problems in Number Theory, Third Edition, Springer-

Verlag, New York (2004), 437pp.

[3] A. W. Johnson, Letter to the Editor, Crux Mathematicorum 4 (1978), 190.

[4] J. E. Olson, An addition theorem modulo p, J. of Combin. Theory 5 (1968),

45–52.

[5] G. Robins and C. Shute, The Rhind Mathematical Papyrus: an ancient Egyp-
tian text, Dover, New York (1987), 88pp.

[6] J.B. Rosser and L. Schoenfeld, Approximate formulas for some functions of
prime numbers, Illinois J. Math., 6 (1962), 64–94.