THE 392 PROBLEM

Aaron Meyerowitz
1
Department of Mathematics, Florida Atlantic University
[email protected]
John Selfridge
Abstract
We examine a conjecture from [E] which asserts that for all non-square i > 8 ,
except i = 392, there are integers x, y with ixy a square and i < x, y < j where
j > i is minimal such that ij is a square. That paper reduces the conjecture to
the case i = 2s
2
N
where S
N
is a sequence with S
N
6
N
. We give theoretical
and computational results which establish the result for N < 20, 000 and for several
innite classes.
1. Introduction
The following products are square
15 60 = 30
2
15 20 27 = 90
2
15 18 20 24 = 360
2
.
It turns out that u
15
(2) = 45, u
15
(3) = 12, and u
15
(f) = 9 for all f 4 where
u
i
(f) is minimal such that there is a square product from [i, i +u
i
(f)] with at most
f factors including i. Let i = rs
2
with r > 1 and square-free. It is immediate that
i + u
i
(2) = r(s + 1)
2
. Erdos et al [E] have the
Conjecture 1 (The 392 Problem) When i > 8 is not a square, u
i
(3) < u
i
(2),
except for i = 392.
They reduce the problem dramatically. They show that u
i
(3) < u
i
(2) for all
nonsquare i > 8 except i = 392 and perhaps some (other) numbers of the form
i = 2s
2
N
where s
N
is the Nth term of the sequence 0, 2, 14, 84, 492, 2870, 16730 . . .
with recurrence
s
N+2
= 6s
N+1
s
N
+ 2
1
Supported by a grant from the Number Theory Foundation
2
Thus the conjecture reduces to answering the question Is it the case that u
i
(3) <
u
i
(2) for i = 2s
2
N
when 3 N?
In this paper we investigate these remaining open cases. The conjecture is still
open but we will show among other things
u
i
(3) < u
i
(2) if i = 2s
2
N
for 3 N 20, 000.
u
i
(3) < u
i
(2) if i = 2s
2
N
and N mod 21 0, 7, 8, 10, 11, 15, 16, 20 or N mod 15
0, 6, 7, 8, 14 .
There is no diculty in principle in extending the rst result well past N =
20, 000. The second result can be expanded to include many more moduli. However,
the next smallest moduli are 68, 205 and 253
The main focus of this paper is the theoretical results behind some extensive
calculations (many of which are of independent interest) and the most interesting
ndings.
Here are the key steps of [E]: Recall that i = rs
2
with r > 1 and square-free.
If r is not prime then u
i
(3) < u
i
(2) . Thus we assume i = ps
2
with p prime
If p ,= a
2
+ 1 then u
i
(3) < u
i
(2) .
If p = a
2
+ 1 and p > 2 then u
i
(3) < u
i
(2) .
This reduces the conjecture to showing that for every s > 2 except s = 14 there
are integers x, y such that 2s
2
< x, y < 2 (s + 1)
2
and 2s
2
xy is a square. Evidently
we must have x = Mu
2
and y = 2Mv
2
for some M > 1. This is most likely to
work when

u
2
2v
2

is small. Indeed we can almost always nd a solution with

u
2
2v
2

= 1, i.e. one such that

u
v
is a convergent to

2. For example
7
5
and
17
12
are two successive convergents and
_
u
2
, 2v
2
_
= (49, 50) can be used for 15 s 83
while
_
u
2
, 2v
2
_
= (289, 288) works for 85 s 491. However, neither works for s =
84 The only solutions for
_
2 84
2
, 2 85
2

are (M, u, v) = (3, 69, 49) and (7, 45, 32) .

For
_
2 492
2
, 2 493
2

there is just one solution: (M, u, v) = (7, 263, 186) .

2. Notation and some estimates
Let = 1 +

2 2.41421 and = 1/ = 1

2 . 41421. Also,
n
=
x
n
+ y
n

2.
the Nth critical interval is CI
N
=
_
2s
2
N
, 2 (s
N
+ 1)
2
_
where
s
N

2, (s
N
+ 1)

2 =
y
2N+1
1
2

2 =

2N+1

2N1
2

2
4
3
Thus the integers in CI
N
have about 1.531 N decimal digits.
We will sometimes write [a, b] to mean [min (a, b) , max (a, b)] This should cause
no confusion.
We reduce the problem to its remaining cases:
Conjecture 2 (The 392 Problem) For N > 2 there are integers M, u, v such
that the interval
_
Mu
2
, 2Mv
2

is properly contained in the Nth critical interval

_
2s
2
N
, 2 (s
N
+ 1)
2
_
.
A core is an interval
_
u
2
, 2v
2

(order not important) with relatively prime ends.

The gap of the interval is g =

u
2
2v
2

. We associate an element of Z

2 with
this core:
_
u
2
, 2v
2

u + v

2
We will call u + v

2 a starter if gcd (u, 2v) = 1 and either u > 2v > 0 or v > u.

i.e. if
_
u + v

2
_ _
2 1
_
= (2v u) +(u v)

2 has a negative coecient. Then

we call [2v u[ +[u v[

2 = u

+ v

2 the paired starter.

Claim 3 Let = u + v

2 be a starter with gap g =

u
2
2v
2

and dene

=
[2v u[ +[u v[

2 = u

+ v

2. Then

u v

=
g

=
g

is a starter
Both and

have the same gap g =

u
2
2v
2

u
2
2v
2

.
These denitions given do not quite t the case of gap g = 1. The strict inequali-
ties do not allow either 1+0

2 or 1+

2 to be a starter. This will not be a problem

since we are considering precisely the few cases which have no solution with gap
g = 1.
Consider an interval [2s
2
, 2(s+1)
2
] such as the Nth critical interval
_
2s
2
N
, 2 (s
N
+ 1)
2
_
.
A solution of this interval is a smaller interval
_
Mu
2
, 2Mv
2

properly contained
in it. We assume that gcd (u, 2v) = 1. Then the interval
_
u
2
, 2v
2

is called the
core of the solution and the numbers , g =

u
2
2v
2

and M are called the gap

and the broad multiplier of the solution. We reserve the term multiplier (or
narrow multiplier) for m, the largest odd square-free divisor of M. Note that
_
Mu
2
, 2Mv
2

=
_
mU
2
, 2mV
2

where
(U, V ) =
_
_
_
__
M
m
u,
_
M
m
v
_
M
m
is square
__
2M
m
v,
_
M
2m
u
_
M
2m
is square
.
4
We have veried the conjecture up to critical interval 20, 000. The integers in this
interval have magnitude about 10
30620
. We have two approaches to nd solutions.
The multiplier method seeks solutions with relatively small (narrow) multiplier
and the starter method seeks solutions with a relatively small gap (more precisely
with u + v

2 relatively small). The starter method can be carried out with large
integer arithmetic or with calculations mod 2g
2
combined with some oating point
computations.
It would be nice to be able to x m and nd many or all U, V, N such that
2s
2
N
< mU
2
, 2mV
2
< 2 (s
N
+ 1)
2
.
It is highly plausible that there solutions for about

2
m
of the intervals but we can
neither prove this nor nd cases short of doing the obvious large integer calculations.
For a given starter u + v

2 we dene a sequence of cores

_
u
2
n
, 2v
2
n

by
(u
0
, v
0
) = (u, v)
(u
n
, v
n
) = (u
n1
+ 2v
n1
, u
n1
+ v
n1
)
So
u
n
+ v
n

2 =
_
u
0
+ v
0

2
__
1 +

2
_
n
=
n
u
n
= u
0
x
n
+ 2v
0
y
n
v
n
= u
0
y
n
+ v
0
x
n
Also,

u
n
v
n

u
0
v
0

_
1

2
_
n

= g
1

n
_
[u
n
[ ,

v
n

_
=
_

_
u
n
+ v
n

2
_

_
u
n
v
n

2
_

_
=

n
g
1

n
2

u
2
n
+ 2v
2
n
=

2

2n
+ g
2

2n
2
We call this sequence of cores a family. All the cores in a family will have the
same gap

u
2
n
2v
2
n

. Each core belongs to a unique family whose starter can be

found by iterating (u
n1
, v
n1
) = (2v
n
u
n
, v
n
u
n
) until a starter results. Here
are (the) two families with gap 89:
n u
n
v
n
u

n
v

n
2 19 15 17 10
1 11 4 3 7
0 3 7 11 4
1 17 10 19 15
2 37 27 49 34
3 91 64 117 83
5
One starter is 3 + 7

2 [9, 98] and the paired starter is 11 + 4

2 [32, 121] .
Note that
_

n1

n1

_
= ([u
n
[ , [v
n
[) .
This can be explained by recalling that
_
[u
n
[ ,

v
n

_
=

n
g
1

n
2

_
[u

n
[ ,

_
=

n
g
1

n
2

Where

=
g

An integer g > 1 occurs as a gap exactly when all of its prime divisors are of
the form 8k 1. The number of families corresponding to this gap is 2
d
where d
is the number of distinct prime divisors. If desired, it is not too hard to use the
factorization routine over Z
_
2

of Maple to nd the starters corresponding to a

given gap.
Claim 4 Let = u+v

2 be a starter with gap g =

u
2
2v
2

> 1. Then the ratio

=
g

2
satises 1 > > 3 2

2 =
2
. 171 57
Proof. Let

=
g

2
where

= u

+ v

2 =
g

is the paired starter. Surely

=
g/

=
[uv

2[
u+v

2
< 1 and similarly

< 1 . We nish by noting that

=
g

2
g

2
=
1

2
since

= g.
We mention, but do not prove or use, these sharper estimates
1
2

2

_
3 2

2
_
_
1 +
2

_
.
Any potential solution of the critical interval CI
N
has the form
2s
2
N
< MU
2
, 2MV
2
< 2 (s
N
+ 1)
2
.
Here
U + V

2 =
_
u
k
+ v
k

2
_
=
_
u + v

2
__
1 +

2
_
k
for some starter
_
u + v

2
_
and k > 0. Thus,
2s
2
N
< M (u
k
)
2
, 2M (v
k
)
2
< 2 (s
N
+ 1)
2
6
We call this a type j solution where j = k N. We will describe it with the code
N, u + v

2, j, m
where m is the narrow multiplier. It turns out that j = 0 is by far the most common
case.
Given a solution N, u+v

2, j, m we set n = N j so k = N +j = n+2j. Hence

for a type j solution we are considering possible integer values of M such that:
2s
2
n+j
< M (u
n+2j
)
2
, 2M (v
n+2j
)
2
< 2 (s
n+j
+ 1)
2
Since we have exact expressions for all these quantities we can given exact ex-
pressions for the (perhaps empty) interval of rational values M

such that
2s
2
n+j
< M

(u
n+2j
)
2
, 2M

(v
n+2j
)
2
< 2 (s
n+j
+ 1)
2
.
Putting these expressions into useful form and getting useful information is some-
what more work.
Consider a xed starter = u + v

2. For brevity we also dene

= = u v

2
=
[[

g =

u
2
2v
2

= [[ =
2

f
1
= u
2
+ 2v
2
=

2
+
2
2
=
2
1 +
2
2
f
2
= 4uv =

2
_

2
=
2
1
2

2
Claim 5 For the xed starter = u + v

2, the exact interval of rational values

M

such that
2s
2
n+j
< M

(u
n+2j
)
2
, 2M

(v
n+2j
)
2
< 2 (s
n+j
+ 1)
2
.
has center at C
n
+ O
_

2n
_
where
C
n
=
x
2n+2
f
1
+ y
2n+2
f
2
2g
2
=

2n+2
+
2n2

2
4
2
and radius r
j
+ O
_

4n
_
where
r
j
=
1

2j1
_

2
g
2
2j1

2
_
=
1

2j1
_

2

2
2j1
_
7
Also, throughout this interval

=

n+1
2
+ O
_

2n
_
= x
n+1
+ O
_

n
_
.
In contrast to the exact interval we call (C
n
r
j
, C
n
+ r
j
) the estimated in-
terval. We take as a
Guiding Principle The error terms in the claim are so small that we can discard
them and still detect all valid solutions while rejecting all non-solutions.
This is well supported by experience but not rigorously established. There are
strong results bounding the errors. We present only some of them here.
We wish to determine if there is an integer in the exact interval. Using the
Guiding Principle we instead examine the estimated interval. This gives a nice
advantage. Note that C
n
is a rational number with odd numerator. To determine
if there is some integer in this estimated interval it is enough to nd the remainder
of the numerator mod 2g
2
.
In discarding the error terms there are two potential types of error. We might
incorrectly include an integer which is not in the interval or exclude one which is.
However any presumed solution can be checked with exact large integer arithmetic
and near misses can be noted and checked similarly. The Guiding Principle is sup-
ported by the fact that no errors of either type showed up in extensive calculations
including an independent exact check for the rst 25 critical intervals. In practice
the errors are exceedingly small even for small n and decrease rapidly. A typical
example is:
Example 6 The starter 25 + 12

2 has gap g = 25
2
2 12
2
= 337 and provides
solutions for intervals 6, 7 and 8 all with broad multiplier M = 189 (so the narrow
multiplier is m = 21).
critical type estimated center radius center error radius error
interval for broad multiplier
6 1
_
189
41
2g
2
_

123.808
2g
2
5. 4910
3
2g
2
1. 6310
7
2g
2
7 0
_
189
41
2g
2
_

368.353
2g
2
1. 025 10
4
2g
2
9. 0710
11
2g
2
8 1
_
189
41
2g
2
_

73.417
2g
2
6. 2310
4
2g
2
5. 3610
12
2g
2
Note that x
8
= 577 and

189
_
25 + 12

2
_
= 576. 999 842. In all these solutions the
error in the estimated center and radius for the interval of rational multipliers is
insignicant for the conclusion that the interval contains an integer. In each case
we estimate the center to be at 189
41
227 138
= 189
41
2g
2
. The range of broad rational
8
multipliers for a solution 9, 25 + 12

2, 2, m is about
_
189
41
2g
2
_

12.89
2g
2
and any
small error terms do not alter the conclusion that there are no integer multipliers
in this range.
Consider using the same starter but increasing N. The estimated radii r
j
are
the same and the estimated centers C
N
are always rational numbers with odd
numerator and denominator 2g
2
. But each time we increase from N to N + 1 the
already minuscule error terms decrease by an order of magnitude. This is further
illustrated by the tables of center and radius errors in section 5. In some cases
we will give a short justication of some result using the Guiding Principle and
elsewhere give a more delicate and unconditional proof. An example is the remark
following this:
Lemma 7 For a xed starter = u + v

2
r
j
< 0 for j < 1.
r
1
< 0 unless < 20 14

2 . 201 010 126 8

r
1
< r
0
r
0
> r
1
> > 0

1
2g
2
>
1
g
2
> r
j
for j > 1 +
log

g
2
.
Proof. Recall that
r
j
=
1

2j1
_

2
g
2
2j1

2
_
=
1
2
2

2j1
_
2

2j1
_
and that the ratio =
g

2
satises
2
< < 1. If j < 1 then
2

2j1
2

2
5
< 2

2
5
_

2
_
=
_
7 + 3

2
_
.
For j = 1
2

2j1
= 2

2
3
=
3
_
20 14

2
_
.
For any j,
r
j+1
r
j
=
1
2
2

2j+1
_
2

2j+1
_

2j1
_
2

2j1
_
=
1
2
2

2j+1
_
_
1
2
_
2

2j+1
_
1
4
_
_
=
_
1
2
_
2
2

2j+1
_
2

2

_
1 +
2
_

2j+1
_
.
9
Thus when j = 1
r
0
r
1
=

_
1
2
_
2
2
_
2

2
_
1 +
2
_
_
and hence r
0
> r
1
since
2

2
_
1 +
2
_
< 2

2
1
_
1 +
2
_
= 0.
For j 0 we have r
j+1
< r
j
since
2

2

_
1 +
2
_

2j+1
> 2

2
_
1 +
2
_

2j+1
2

2
_
1 +
2
_

= 0.
Finally, if j > 1 +
log

g
2
then
2j2
> g so
r
j
=
1
2
2

2j1
_
2

2j1
_
<

2j1
=

2
g
2j2
<
1

2
g
2j2
<
1
g
2
.
Remark 8 This exact result about the estimated radii r
j
can be combined with
the Guiding Principle to conclude that, as in the example above, every solution
N, u + v

2, j, m of type j > 0 is connected in a string of related solutions, all with

the same multiplier, to a solution of type 0.
N i, u + v

2, j i, m for 0 i j.
Similarly, a solution N, u + v

2, 1, m is always accompanied by a solution N +

1, u + v

2, 0, m.
3. Algorithms
All computations were done with Maple.
Multiplier searches
The algorithm used is the obvious one. It uses the relatively fast integer quotient
and integer square root operations. Naive but plausible arguments predict that a
narrow multiplier m will provide solutions for about

2
m
of the intervals (
1

m
for
y and

m
for x). This agrees well with our computations although the intervals
solved are not well distributed. The phenomenon of a string of consecutive solutions
all with the same multiplier, as in the example above with m = 21 and M = 189,
can cause clustering followed by relatively long empty gaps.
10
starter searches
A straightforward algorithm using large integer arithmetic is easy to give. Here
is a version of the modular starter algorithm. Given a starter u
0
+v
0

2 it will nd
all solutions involving cores u
k
+v
k

2 from the associated family. The version given

here runs from 3 to some top n. It is not hard to alter it to simply check a specied
interval. The running time and expected frequency of solutions are somewhat more
complicated to describe than for the multiplier method.
Compute the gap g =

u
2
0
2v
2
0

, the magnitude = u
0
+ v
0

2 and the
constants f
1
=
_
u
2
0
+ 2v
2
0
_
mod 2g
2
and f
2
= 4u
0
v
0

2 mod 2g
2
.
Dene the positive integers
H
j
=
_
2g
2
r
j
_
odd
=
_
2g
2
_

2
(
j
)
2


2
g
2 (
j
)
4
__
odd
for j = 1, 0, 1, 2, . . . , j
top
Here |
odd
means truncate to an odd integer. We only consider H
1
when it
might be positive (Namely if > 20 14

2). H
0
H
1
> H
2
> > H
jtop
.
Here j
top
is the largest j with H
j
> 0.
For n 3 consider the estimated center C
n
=
x2n+2f1+y2n+2f2
2g
2
and compute
c
n
= x
2n+2
f
1
+y
2n+2
f
2
mod 2g
2
(note that 2g
2
and f
2
are even while x
n
, f
1
and c
n
are odd)
1. If min
_
c
n
, 2g
2
c
n
_
H
1
then u
n2
+ v
n2

2 gives a solution for

CI
n1
2. If min
_
c
n
, 2g
2
c
n
_
H
0
then u
n
+ v
n

2 gives a solution for CI

n
3. If min
_
c
n
, 2g
2
c
n
_
H
1
then u
n+2
+v
n+2

2 gives a solution for CI

n+1
4. etc.
If min
_
c
n
, 2g
2
c
n
_
< H
j
then u
n+2j
+ v
n+2j

2 gives a solution for I

n+j
.
we called this a type j solution. After step 1) we stop the rst time we get
failure, although we may save near misses for future examination. All the
solutions which occur before failure will have the same multiplier.
Check all solutions and near misses for possible errors of inclusion or omission
.
A naive exact integer routine is faster for checking , say, the rst 25 intervals.
Also, there are no round o concerns. The given algorithm uses integer arithmetic
mod 2g
2
and is eective for, say, the 10
6
th critical interval (modulo some attention
to the error terms). One might make two passes, rst looking only for type 0
11
solutions and then returning to check each one found for related solutions of type
j ,= 0. This avoids computing H
1
etc. for starters which give no solutions in the
range under consideration. The numerators satisfy the recurrence
c
n+1
= 6c
n
c
n1
mod 2g
2
.
Thus each can be computed from the previous two without a computation of x
2n+2
and y
2n+2
.
The proportion of intervals having a solution arising from a xed starter =
u
0
+v
0

2, is, as we have said, a fraction c

with denominator P
g
. It is plausible (in
a sense which can be made precise) that (on the average, when is not too small)
c

is about
1

and that there about

2

2
s starters with

u
0
+ v
0

2 s

<
1
2
.
4. The rst 25 and the rst 20000 critical intervals
To nd all the solutions for CI
3
through CI
25
we rst nd all the type 0 solutions
for each interval then go back and check which ones give solutions of type j = 1
or j > 0 for adjacent intervals. Every type 0 solution for CI
n
will have a starter
= u + v

2 and broad multiplier M with

M almost exactly x
n+1
Hence it
will suce to do a multiplier search of odd squarefree numbers m m
top
and a
starter search for starters = u + v

2 (with gcd (u, 2v) = 1) such that [[

top
where the upper bounds are chosen so that m
top

top
> x
n+1
+ 1. In this range it
turns out to be faster (in Maple) to do the starter search using exact large integer
arithmetic. (This avoids any concerns about roundo error as well)
This approach does not account for the type 1 solutions of the top interval. We
deal with this by considering the starters which are larger than
max
(25) and have
small enough to give such a solution and are small enough that the multiplier
search could miss them (It took 20 hours to show that there were none).
We used the following procedure to nd some solutions for each critical interval
from 3 up to 20000
Use the simple (but carefully written) exact integer Maple algorithm to nd
all solutions with narrow multiplier m < m
top
(n). This gets harder as n
increases so m
top
(n) was smaller for large n than for small.
Attempt to carry the multiplier search further if needed in order to nd the
smallest multipliers.
Find all solutions with starter less than 200.
Attempt to nd the smallest starter among the solutions (if it is larger).
continue both searches far enough to nd at least one solution.
12
The results of the modular computations were veried with exact integer arith-
metic. No near misses were found to be solutions and no incorrect solutions were
oered.
For critical intervals 3 to 20000 all have either a minimal multiplier below 1000
or a solution with starter u
0
+ v
0

2 such that max(u

0
, v
0

2) < 70 EXCEPT:
interval starters mmin
639 [511, 69] 1963
9624 [45, 127], [279, 37], [695, 69], [525, 557] ?
16677 [71, 35], [161, 73], [211, 94], [455, 227] ?
16835 [25, 74], [229, 38], [1, 331], [83, 483] ?
(1)
Considering starters with 906 v > u or 1283 u > 2v, solutions were found for
all intervals except:
Interval 2386 2537 7120 7430 10047 14297 18063 19993
multiplier 7 5 3 5 91 413 3 215
(2)
We also know all multipliers up to 30000 for the rst 137 critical intervals.
The smallest known multipliers for critical interval 590 are 1294449, 4977589 and
16073971.
5. Periodic solutions
This section is concerned with results like the following which provides solutions for
10 out of every 21 critical intervals.
Claim 9 Let be one of the two starters 1 + 2

2 and 3 +

2 for gap 7. Then

gives a type j solution of CI
N
exactly if N 1 and
= 1 + 2

2, j = 0 and N mod 21 0, 7, 8, 10, 15, 20

= 1 + 2

2, j = 1 and N mod 21 11, 16

= 3 +

2, j = 0 and N mod 21 0, 5, 10, 12, 20

= 3 +

2, j = 1 and N mod 21 = 11
More generally
Theorem 10 For a xed starter with gap g there is an integer N
0
= N
0
() and
a period P = P
g
such that for all N N
0
the truth of the statement gives a type
j solution of CI
N
depends only on j and the congruence class of N mod P. The
period P, initial value N
0
and appropriate congruence classes can all be determined.
13
It follows from the Guiding Principle that N
0
() = 3 for all . We rst discuss
the general situation. Then we give a careful treatment of the two starters with gap
7 followed by results for other starters.
Given a xed starter = u
0
+v
0

2 the modular starter algorithm rst computes

several radii
Hj
2g
2
with H
j
an odd integer and the constants f
1
= u
2
0
+ 2v
2
0
and
f
2
= 4u
0
v
0
. It then decides if u
n+2j
+v
n+2j

2 gives a type j solution for critical

interval CI
n+j
by considering the odd integer x
2n+2
f
1
+y
2n+2
f
2
and checking if the
distance from
x2n+2f1+y2n+2f2
2g
2
to the nearest integer is
Hj
2g
2
. This distance depends
only on the congruence class of x
2n+2
f
1
+y
2n+2
f
2
mod 2g
2
. This in turn depends
only on the congruence classes of x
2n+2
and y
2n+2
. There are only
_
2g
2
_
2
possible
values for (x
k
, y
k
) mod 2g
2
since each pair determines the next:
(x
n
, y
n
) =
_
(1, 0) n = 0
(x
n1
+ 2y
n1
, x
n1
+ y
n1
) n > 0
there is some period P = P
g
4g
4
such that for all n
(x
2n+2P+2
, y
2n+2P+2
) (x
n
, y
n
) mod 2g
2
.
We may take P
g
to be the smallest positive integer with
(x
2P
, y
2P
) (1, 0) mod 2g
2
.
Then, with one restriction, the truth of u
n+2j
+ v
n+2j

2 gives a type j solution

for critical interval CI
n+j
depends only on j and the congruence class of n mod
P. The restriction is that the O
_

2n
_
errors in approximating radii and centers
do not cause an error of either type. This is why the restriction N = n + 2j > N
0
is included. The Guiding Principle is the claim that the errors are insignicant for
N 3.
Actually, P
g
is much less than 4g
4
. In fact P
g
is either equal to or a divisor of
g

p1
2
where the product is over the prime divisors of g. Equality is fairly common.
We will summarize the results of claim 7 for the two starters with g = 7 as
follows:
7 21
1 + 2

2 0, 7, 8, 10, 11
1
, 15, 16
1
, 20
3 +

2 0, 5, 10, 11
1
, 12, 20
With this notation we have
17 136
5 + 2

2 0, 9, 13, 15, 16
1
, 42, 60,
68, 77, 81, 83, 84
1
, 110, 128
1 + 3

2 0, 7, 25, 26
1
, 52, 53
1
, 54
2
, 54, 57, 58, 67,
68, 75, 93, 94
1
, 120, 121
1
, 122
2
, 122, 125, 126, 135
14
23 253
3 + 4

2 0, 8, 55, 69, 94, 95

1
, 96
2
, 107, 111, 126,
127, 147, 148
1
, 174, 179, 200, 208, 209
1
, 227, 240, 243, 249
5 +

2 0, 3, 4
1
, 9, 12, 25, 44, 45
1
, 52, , 53
1
, 73, 78, 79
1
, 105, 106
1
, 116, 126,
127
1
, 140, 141, 145, 158, 159
1
, 160
2
, 183, 191, 197, 203, 244, 252
31 15
7 + 3

2 0, 7
1 + 4

2 6, 7, 8
1
, 14
41 205
7 + 2

2 0, 3, 53, 61, 81, 102, 103

1
, 104
2
, 114, 125, 204
1 + 4

2 0, 79, 90, 102, 103

1
, 104
2
, 123, 143, 151, 201, 204
217 105
17 + 6

2 0
11 + 13

2 0, 52
15 + 2

2 34, 51, 52, 53

1
, 104
5 + 11

2 0, 104
A few remarks
1. In the details above for g = 7 we see that
_
1 +

2
_
6
= 1 mod 2 7 while
_
1 +

2
_
42
= 1 mod 2 7
2
. This is typical of many primes g where P
g
=
g(g1)
2
.
2. Note that P
41
is only 205. This is related to the fact that x
5
= 41
3. The period P
31
= 15 is surprisingly small. This is because of a very special
property of 31. As above
_
1 +

2
_
30
= 1 mod 2 31 and this is the smallest
power which works. However, it is also true that
_
1 +

2
_
30
= 1 mod 231
2
.
(x
30
, y
30
) = (152 139 002 499, 107 578 520 350)
=
_
3
2
11 19 59 601 2281, 2 5
2
7 29 31
2
41 269
_
At least up to 100000, no other prime has this property that the rst k > 0
such that (x
k
, y
k
) = (1, 0) mod p is also the rst such that (x
k
, y
k
) = (1, 0)
mod p
2
.
4. The starter 1 + 3

2 for g = 17 gives two solutions for CI

54
. One of type 0
and one of type 2.
5. From the table above it seems as though P
17
should be 68. There is actually
a central symmetry which occurs when P
g
is even (as for g = 73, 89, 97).
15
6. This chart shows how many out of each 105 consecutive intervals have a
solution with the indicated gap. The third column shows how many of these
do not have a solution from a previous row. Thus the four starters for g = 217
each have period 105 and between them solve 6 of every 105 intervals. Of these
6 all but one has a solution with gap 7 or 31. This accounts for 69 of every
105 critical intervals.
gap solved new
7 50 50
31 35 18
217 6 1
7. Here is the same information for 7, 31, 17 and their products (in that order).
The numbers are now out of every 14280 consecutive intervals
gap solved new
7 68000 6800
31 4760 2448
17 3360 1184
119 1045 278
217 816 95
527 294 72
3689 55 12
So gaps 7, 31 and 17 account for just over 73% of the intervals and this
increases to just over 76.25% if we include the other four gaps with period
dividing 14280.
8. It does not seem promising to cover all the integers by such arithmetic progres-
sions. There are critical intervals which have no solutions with small starters.
For example critical interval 639 has no starters smaller than 511 + 69

2
with gap g = 511
2
2 69
2
= 251 599 = 311 809 and period P
251599
=
3938782345 = 15 655 251 599 = 5 31 101 311 809
Reference
[E] P. Erdos, J.L. Malouf, J.L. Selfridge and E. Szekeres, Subsets of an interval
whose product is a power, Disc. Math. 200 (1999) 137-147