A Sieve Method for Factoring Numbers

of the Form n2 + l
By Daniel Shanks
1. Introduction. Factorizations of numbers of the form n + 1, [1], [2], are of
mathematical interest for at least four reasons: 1) a possible insight into the un-
settled question as to whether there are infinitely many primes of this form; 2) a
possible insight into the unsettled question as to whether the reducible numbers
(explained below) have a definite density; 3) the relation of these factorizations
to the p-adic square roots of —1; and 4) the relation of these factorizations to the
Gaussian primes. The purpose of this paper is to describe a sieve method for factor-
ing these numbers and to present and discuss some empirical results bearing on
1), 2), 3), and 4) which were obtained by its use. The method can be considered
to be based, in part, on the p-adic square roots of —1, but it is also possible to
avoid the use of this language.
A program based on this sieve method was written for an IBM 704 with a
32,768-word high-speed memory, and with this program all n + 1 from n = 1
to 180,000 were completely factored in about 10 minutes. Since these factoriza-
tions of n + 1 exceed those in existing published tables, (82 percent of these
numbers are greater than a billion), a short summarizing statistical table should
be of interest. In the table below, P(N) is the number of primes of the form n + 1
for 1 ¿ n ^ N, and for comparison, ir_(N) is the number of primes of the form
4m — 1 for 1 < 4m — 1 g N. Further, R(N) is the number of reducible numbers
£N, r(N) = R(N) - R(N - 10,000), and SR(N) = R(N)/N, the mean density
of the reducibles.

Statistical Table of Primes and Reducibles

RW) tRW) PW) 7T-W) W)A-(iV)
10,000 2898 2898 .28980 841 619 1.3586
20,000 2935 5833 .29165 1559 1136 1.3724
30,000 2930 8763 .29210 2268 1633 1.3889
40,000 2918 11681 .29203 2952 2117 1.3944
50,000 2959 14640 .29280 3613 2583 1.3988
60,000 2912 17552 .29253 4252 3038 1.3996
70,000 2965 20517 .29310 4888 3485 1.4026
80,000 2881 23398 .29248 5513 3933 1.4017
90,000 2947 26345 .29272 6084 4364 1.3941
100,000 2875 29220 .29220 6656 4808 1.3844
110,000 3009 32229 .29299 7239 5247 1.3796
120,000 2934 35163 .29303 7795 5675 1.3736
130,000 2938 38101 .29308 8369 6103 1.3713
140,000 2888 40989 .29278 8944 6531 1.3695
150,000 2983 43972 .29315 9505 6941 1.3694
160,000 2952 46924 .29328 10072 7361 1.3683
170,000 2932 49856 .29327 10658 7770 1.3717
180,000 2981 52837 .29354 11223 8178 1.3723
Received March 10, 1959.
78

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 SIEVE METHOD FOR FACTORING NUMBERS 79

2. The Sieve. The sieve method (substantially more complicated than that of
Eratosthenes) is based on the facts listed in the following theorem. Since many
of these are well known and the rest can be easily verified, no proof need be given.
Theorem. For every prime p of the form 4m + 1 and every positive integer k
there are two and only two positive solutions of:
| n <pk

n2 + 1 = 0 (mod pk).
If we call these Ak and Bk, then
(2) Ak + Bk = p",
and if
(3) hm\(p+l)Ai (modp),
and
(4) Ck m h(Ak2 + l)/p* (modp),
the Ak for k = 2, 3, • • • may be computed recursively from Ai by
(5) At+i = Ak + Ckp\
Further, for all positive n,
n + 1= 0 (mod p ),
if and only if n is given by one of the linear forms :
¡Ak + mpk
(6) „ = (m = 0,1,2, •••),
[Bk + mpk

and aside from these factors of n2 + 1 (obtained as p runs through all primes of
the form 4m +1) the only other prime-power factors are the obvious
(7) n + 1= 0 (mod 2) for n = 1 (mod 2).
We will adopt the convention that Ai is the smaller of the two roots of (1)
for k = 1, and note that this implies:
(8) A1 < \p < Pi,
but it does not imply that Ak < Bk if k > 1. For example, let p = 5. Then Ax = 2.
Thus Ä = 1, and for k = 1, 2, 3, • • ■ we compute:
Ak = 2, 7, 57, 182, 2057, 14557, • • •
Bk = 3, 18, 68, 443, 1068, 1068, • ■•
and note, At > B6.
We also have in this case the interesting degeneracy B& = P6. While similarly,
for p = 13, we have the degeneracy A3 = At, it can be seen from (1), and it is
important for the validity of the sieve method, that for every p:

(9)

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 80 DANIEL SHANKS

Now suppose we wish to factor every n + 1 for 1 ^ n g L. The sieve method

proceeds as follows:
A.) In L contiguous cells we write the corresponding values of n + 1.
B.) We now divide out the factor of 2 for each n = 2m + 1, (m = 1, 2, • • • )
and store the quotients back into the same cells.
C.) At n = 2, we find the number 5. This implies (see a general proof below)
that 5 is a p and, corresponding to it, A\ = 2. Using formulas (2) through (6),
we factor a 5 from each n'th cell, where
Í2 + mb (m = 1,2, ■■■)
n = {
[3 + m5 (m = 0,1,2, ■■■)
and store back the quotients. Then a second 5 is factored for each
Í 7 + m25
n = (m = 0, 1, 2, • • ■)
[18 + to25
and so on until both Ak and Bk are greater than L. By now all factors of 5 have
been removed.
D.) At n = 3, we now find the quotient 1. This means 3 is not an Ai number.
It is, in fact, reducible (see below). We so record it and move on.
E.) At n = 4, we find 17 and divide out all factors of 17 as in C.
F.) At n = 5, we find 13 and divide out all factors of 13.
G. ) Proceeding in this way we examine the contents of the n'th cell when we
get to it, and find one of two cases.
1. The quotient is 1 and no new prime factor is contained inn2 + 1.
2. The quotient is >1. In this case the quotient must be a new p = 4m + 1
and n must be its corresponding At number. Proof: Since the quotient cannot
have a factor of 2 or a prime of the form 4m — 1, every prime factor must be a
p = 4m + 1. But for such, a p, n must be A\, for, if p had occurred in an earlier
n + 1, its Ai would have been encountered earlier and p would have been fac-
tored out. By (9), n + 1 is not divisible by p2. Nor can n be simultaneously an
Ai for two distinct primes, since from (8) the quotient:

UÏ + l)/p á lAi,
and therefore cannot contain a second prime which (again by (8)) must be >2Ai.
This completes the proof, and shows we are obtaining complete factorizations
without the use of any trial divisions and without the need to know in advance
the primes of the form 4m + 1. They are, in fact, being generated (though not
in numerical order) by the process itself.

3. The Program. In the 704 program mentioned above the following changes
were desirable or expedient.
a.) The sieve was started with — (n + 1) in the n'th cell and the absolute
value of each quotient was stored back in. If no division occurred in a particular
cell the contents of this cell remained negative and this indicated that n + 1 was
a prime.

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 sieve method for factoring numbers 81

b.) Since in a binary machine division by 2 is so simple, step B was combined

with step A.
c.) The factoring out of any prime p began when n reached its Bi instead of
its Ai. This saves time and memory and is equally valid.
d.) The linear forms, (6), are easily obtained by use of an index register. The
high efficiency of this method makes possible over 3000 divisions/sec.
e.) The sieve is done in several segments. In a memory of 32,768 words, with
4096 words reserved for an operating routine, a sieve segment of 20,000 words is
possible, since nearly 7000 words (see g) must be reserved to save the primes (and
corresponding Ai numbers) for later phases. The program itself, although quite
intricate, takes less than 300 words.
f.) An upper limit, L, of 180,000 (9 segments) was chosen since a tenth segment
would cause n + 1 to overflow the length of a 704 word. This is 2 (nearly 34.36
billion). Of course n could be made larger by using double precision.
g.) Although 127,162 of the 180,000 numbers are Ai numbers, in most cases
the corresponding p is enormous and is not used in sieving—the criterion being
p + Ai ^ L. The number of p's actually used was 6693.
h.) The program was checked by being repeated with a segment of 20,500
This changed all the phase relations in the linear forms (6). It also raised L to
184,500 and we find P(184,500) = 11,486; Ä( 184,500) = 54,162.
i.) If it were not for the "degeneracies" mentioned just before (9), a much
simpler sieve method would be possible. One would not compute the Ak and Bk
sequences and take out all factors of p at once but could instead re-factor the
A2 + nip , the B2 + nip , the A3 + mp , etc., when n came to A2, B2, A3, etc.,
respectively. This was in fact attempted, and some very rare errors ensued. For
example, one found R(20,000) = 5832 instead of the true 5833, the single error
stemming from the degeneracy B$ = Be (for p = 5) mentioned above.
j.) The main output of the program as described was the table shown above
but in a much greater detail (the interval in N being 100 instead of 10,000). The
program actually computed ir+(N) by counting the p's instead of the ir_(iV) shown.
A small modification of the program also produced a 360-page table showing the
largest prime factor in every n + 1 for n = 1 (1) 180,000. Small auxiliary tables
concerning the p-adic square roots (the Ak and Bk sequences above) and the dis-
tribution of the reducible numbers were also produced.

4. The Primes. Now consider P(N) in the statistical table. We see a steady
growth which is nearly proportional to w-(N). Since

2h mx
by the prime number theorem, the numbers P(N) are in excellent agreement with
the conjectured formula [3] of Hardy and Little wood:
fN dx
(10) P(N) -0.68641 J2
/ ~
In x

which therefore implies P(N)/w-(N) ~ 1.3728.

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 82 DANIEL SHANKS

The constant in (10) is equal to the infinite product, taken over all odd primes,

Iö['-(tK»-]'
where ( —1/p) is the Legendre symbol. It was computed by A. E. Western [4]
who also verified the substantial correctness of (10) to N = 15,000. Since the
infinite product converges too slowly, Western used a transformation of the product
due to Littlewood. The following related formula is simpler:

(10a) 0.68641-? 4 #£ II (W -^~t)

Gf(3) ,J£h \
6 ~ p(p, *- ,J)
p3 - 1/ \ l)2/
•
Here G is Catalan's constant, and the product is taken over the ir+ primes. More
rapid convergence may be obtained by multiplying through by the identity:

(10b) l-|Mk.&.(,+irh)I
where L(4) = X? ( —1)" (2n + l)-4. With this improvement, the first two p's,
i.e., 5 and 13, already suffice to yield the five decimal places shown.
Among the Gaussian integers, a + bi, the Gaussian primes on the positive real
axis are those of the form 4m — 1, and thus are those counted by w-(N). On the
other hand, n + i is a Gaussian prime if and only if n + 1 is a rational prime.
Therefore, the column headed "P(N)/w-(N)" shows that the +1 horizontal
line in the Gauss plane (and therefore also the —1) has about a 37 percent greater
density of primes than the axis has. (For an attractive picture of these primes see
van der Pol's Tea Cloth, [5].)
It is of interest to mention briefly the "Gaussian twin" primes, i.e., those where
(n — l)2 + 1 and (n + l)2 + 1 are both prime. They are not at all rare. In the
last block of 1000 numbers with L = 184,500, we find no less than 9 pairs, namely:
n = 183585;183635;183685;184055;184075;184145;184185;184325;and 184495.
The last pair, 34038036037and 34038774017,were the largest primes obtained
by this program. The conjecture suggests itself that there are infinitely many
such "twins."
It may also be mentioned that (although they were not particularly sought)
the program yielded a very large collection of "large" primes, i.e., those over 10
digits. Not only are 4830 of then2 + 1 "large" primes, but many others are equal
to twice a "large" prime. For example, 1844992 + 1 = 2-17019940501.

5. The Reducible and Irreducible Numbers. A reducible number, r, is a positive

integer whose arctangent is a linear combination, with integer coefficients, of the
arctangents of smaller positive integers:
r-l
(11) tan-1 r = X an tan-1 n.
n-I

If no such linear combination is possible, the number is irreducible [6]. For example,
since
tan- 3 = 3 tan- 1 — tan- 2,

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 SIEVE METHOD FOR FACTORING NUMBERS 83

3 is reducible. So are 7 and 8, while 1, 2, 4, 5, 6, 9, and 10 are irreducible. (If we

replace the arctangent function in (11) with the logarithm, we would be defining
composites and primes instead, so that the irreducibles can be thought of as the
quadratic analogue of the primes. ) John Todd showed [6] that r is reducible if and
only if the greatest prime factor of r2 + 1 is less than 2r. For each irreducible, n,
(other than 1), there is thus a unique prime p > 2n which is the greatest prime
factor of n + 1. [6, p. 526]; and this is the 1 — 1 correspondence indicated above
between A\ and p. Ax numbers are therefore irreducible, and all others except 1
are reducible. This includes all A2, A3, ■• •, all Bi, B2, • • •, and some numbers,
such as 21, which are neither Ak nor Bk.

6. The Density of Reducibles (Heuristic). If Ä(A^) is the number of reducibles

^N, and SR(N) = R(N)/N, and if Lim^-« BR(N) exists, we call this limit the
density of the reducible numbers, and write it 5R .
(12) bR = Lim*..«, R(N)/N, if it exists.
However, its existence is still unsettled, as is the older, and somewhat related
question :
(13) DoesP(A)^oo?
Early counts of the reducible numbers up to 5000, by J. C. P. Miller and Todd,
[6], [7], [8], were based on Todd's criterion [6] and Wrench's table [2], and indicated
that ôR(N) persisted in the vicinity of 0.29. The present paper extends N to 184,500
and shows a continuance of this state of affairs with a slow growth (in the mean)
to about 0.293.
Chowla and Todd [7] have proved that if C(N) is the number of numbers
n ^ N for which the greatest prime factor of n is greater than 2-\/n, then the
Limjv-,00C{N)/N exists and equals In 2 = 0.693- • ■. But note that here n ranges
through all numbers, not just those of the form m2 + 1. 2/ this latter class of
numbers constituted a good sample of all the numbers as regards factorization
properties, (with a rigorous definition!) it would follow from 2\/m2 + 1 ~ 2m
and Todd's criterion of reducibility that 5R does exist and is equal to 1 — In 2 =
0.30685-••.
But, if such a deduction were possible, it would probably also follow from the
same line of reasoning (good sample concept) that the mean local density of primes
of the form m + 1, like that of the ordinary prime sequence, would be
1_1_
In (to2 + 1) ~ 2 In to
and thus

2 J2 in m

Now in fact we have seen that P(N) remains persistently and significantly higher.
And this greater prevalence of primes is consistent with the smaller fraction of
reducibles (.293 instead of .307) which is observed. It is, of course, not precluded
that 5R(N) will rise to .307 for very much larger N. Its slow growth, mentioned

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 84 DANIEL SHANKS

above, seems to be of order of a/In N and to be associated with the falling off in
the mean local density of the excess primes, P(N) — t-(N). This (average)
increase of &r(N) hardly shows in the Table since In N changes so slowly and the
fluctuations are almost of the same size. Nonetheless it is there and may carry
SR(N) nearly up to 0.307.
The gross facts on which any attempted assessment of the "good sample"
concept must be based are these. The n + 1 numbers have only about 5 of all
primes, [9], as possible factors, but in "compensation" these factors occur twice
as often (see (6) above). While this may suggest a factorability of the same order
of magnitude, certainly no more exact equivalence is implied.
Assuming the future establishment of R(N) ~ Sa-N, the (deeper?) question
of 0(R(N) — Sr-N) will arise. Concerning this question—that of the uniformity
of the distribution of the reducibles—the following table and figure are informative.
Each of the 1800 intervals of 100 numbers:
100m < n ^ 100(m +1) (m = 0, 1, • • -, 1799)
has at least 17 reducible numbers and at most 42. The 52,837 reducibles ^ 180,000
are distributed as follows:
r, no.of reducibles 17] 18j 191 20121122123124125) 261 27| 281 29
v,no.of intervals 5| 6| 5| 13122129169[821931127115411471167
301 311 321 331341351361371381
39| 401 411 42
179114011471130193¡721411311161
16| 111 l| 4
In the Figure a bar graph of this distribution is compared with a binomial dis-
tribution, v (r):

(14) v - 1800 „„ff01 ,,(0.2935)r(0.7065)100-r

. r!(100 — r)\
where the "probability," 0.2935, was taken to be the final value of the mean
density, afi(180,000).

7. P-adic Numbers and Degeneracy. The sequence of the partial sums of the
infinite series:

(15) ¿i + XC*p*,
where p is a prime of the form 4m + 1 and Ai and the Ck are determined by (1)
through (4), is a convergent sequence in the p-adic sense. (See [10] for an elementary
account of p-adic valuation.) Since it converges, it represents a p-adic number,
namely, one of the two values of y/—l. This sequence, it is seen from (5), is the
sequence Ak. The other sequence, Bk, converges in the p-adic sense to the other
\/ —1 and their sum, Ak + Bk , converges in the p-adic sense to 0.
As an example, take p = 5, and Ak = 2, 7, 57, 182, etc., as above. If we write
them in the quinary system we see that they represent the sequence obtained
from the 5-adic number

(16) •••3140223032431212. = \f=\

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 SIEVE METHOD FOR FACTORING NUMBERS 85
200
The Continuous Curve is the
Binomial Distribution;

u=l800r°J(0.2935)
(0.7065)

Fig. 1.—Distribution of the reducible numbers between 1 and 180,000 into the 1800 in-
tervals of 100.

by starting at the quinary point and taking more and more places to the left. If
we take k places and carry the computation to at least k places we find that
Ak2 + 1 = 000 - • 00.
k zeroes
so that Ak 4- 1 is divisible by p and Ak is an approximation to y/ —1 correct
(p-adic sense) to k places. Similarly the Bk sequence gives the complement:

(17) •••1304221412013233. = - y/-í.

The point of this review is to indicate, first, that the degeneracy B$ = Bt —
1068 mentioned above (9) is associated with the last zero digit in (17). That is,
Bi = 13233 (quinary) and so is Be. Similarly from the zero digits of (16) we find
that Ar, = Ai and Au = Ai2. Since an irrational p-adic cannot have periodic
digits, [10, p. 196], it is reasonable to conjecture that the digits are "normal."
This would imply that degenerates appear infinitely often in every Ak and Bk se-
quence for every p, and further that we should expect multiple degenerates, Ak =
Ak+1 = Ak+2, etc'. For all that, aside from the Bt = B<¡= 1068 for p = 5 and
the Az = Ai = 239 for p = 13 already mentioned, there are no other degener-
ates ^180,000.

8. The Difficulty of the Unsettled Questions. It has often been remarked that
questions like (13) are "very difficult." The intent of this section is to assess this
difficulty. We do this by comparing the very simple Eratosthenes sieve for the

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

 86 DANIEL SHANKS

ordinary prime sequence with the present one for n 4- 1 and note that the latter
is more complicated in the following three ways:
1.) Instead of one linear form, mp, for each prime, we have a double infinity:
Ak + mpk, Bk 4- mpk.
2. ) Instead of a zero origin we have Ak and Bk origins, which are not related to
p in any simple fashion. While A2, A3, ■• • and Bt , B2, • • • can be computed by
the more or less complicated relations (2) through (6), A\ can arise at any n
satisfying \/p — 1 ;S n ^ (p — l)/2. Further, we have the complications of
occasional degeneracy.
3.) Finally, whereas in the Eratosthenes sieve it is not necessary to divide, but
it suffices to scratch the cells in the linear form, here we must divide the p out.
Otherwise, we would not obtain the new prime hidden in each vli 4- 1 which is
not itself prime.

9. Generalization. In conclusion, it should be stated that while we have con-

fined ourselves here to n 4- 1 the same type of sieve method is applicable to
n2 4- a, for a = ±2, ±3, etc. The main change is that these programs would be
based on the p-adic square roots of —a The sieve itself would generate those
primes for which —a is a quadratic residue, that is, all possible divisors. It is thought
that comparable statistics will be found for the primes of these forms and for the
"generalized irreducibles," that is, those n which yield a new prime factor. This is
now being investigated, [11].
Applied Mathematics Laboratory,
David Taylor Model Basin,
Washington, District of Columbia
1. L. E. Dickson, History of the Theory of Numbers, Stechert, New York, 1934, v. 1, Ch.
XVI. For example, Euler (1752) gave P (1500) = 161. See also D.H. Lehmer, Guide to Tables in
the Theory of Numbers, National Research Council, Washington, D. C, 1941, p. 31-32 and p. 45.
2. The most extensive table of all the prime factors of n2 + 1 (up to n = 31,622) is the un-
published table of J. W. Wrench, Jr. See UMT 1, MTAC, v. I, 1943,p. 26. Recently a 704 pro-
gram by the author in collaboration with Dr. Wrench raised this limit to 50,000 for a table of
the greatest prime factor. However, we now consider that type of program (with trial divisions)
to be superseded by the present sieve method.
3. G. H. Hardy & J. E. Littlewood, "Partitio numerorum III: On the expression of a
number as a sum of primes," Acta Math., v. XLIV, 1923, p. 48.
4. A. E. Western, "Note on the number of primes of the form n2 + 1," Cambridge Phil.
Soc., Proc, v. XXI, 1922, p. 108-109. Western assumes P(15000) = 1199 following Cunningham,
who omits 2 = l2 + 1. The correct value of P(15000) is 1200.
5. Fortune, June, 1958, p. 140.
6. John Todd, "A problem on arc tangent relations," American Math Monthly, v. LVI,
1949, p. 517-528.
7. S. D. Chowla & J. Todd, "The density of reducible integers," Canadian Jour, of Math,
v. I, 1949,p. 297-299. The table of R(N) to N = 5000 has many errors. It indicates Ä(5000) =
1453. A mimeographed errata sheet later circulated stated ñ(5000) = 1458, but the correct
value is 1467.
8. John Todd, Table of Arctangents of Rational Numbers, NBS, Applied Math. Series 11,
Washington, D. C, 1951.
9. While 7T+(N) and ir_(N) are asymptotically equal, x_(N) is larger for about 99.6 per cent
of the N less than 10s. See a forthcoming paper of the author, "Quadratic residues and the dis-
tribution of primes," for the statistics of this and related phenomena. That this weakness of
ir+(N) will tend to raise P(N) and lower R(N) seems clear, but no serious attempt has been
made to evaluate its effectiveness.
10. C. C. MacDuffee, An Introduction to Abstract Algebra, Wiley, New York, 1940, p. 193-
202.
11. Note added in proof, May 7, 1959. The number of primes of the forms n2 + 2, n2 — 2,
ra2+ 3, and n2 - 3 up to n = 180,000 are 5847, 15134, 9240, and 11354, respectively. These
numbers are all in good agreement with Conjecture F, [3], of Hardy and Littlewood. Fuller
details will be published later.

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use