It Is Easy To Determine Whether A Given Integer Is Prime

BULLETIN (New Series) OF THE
AMERICAN MATHEMATICAL SOCIETY

Volume 42, Number 1, Pages 338
S 0273-0979(04)01037-7
Article electronically published on September 30, 2004
IT IS EASY TO DETERMINE WHETHER A GIVEN INTEGER

IS PRIME
ANDREW GRANVILLE
Dedicated to the memory of W. Red Alford, friend and colleague
Abstract. The problem of distinguishing prime numbers from composite

numbers, and of resolving the latter into their prime factors is known to be
one of the most important and useful in arithmetic. It has engaged the industry
and wisdom of ancient and modern geometers to such an extent that it would
be superuous to discuss the problem at length. Nevertheless we must confess
that all methods that have been proposed thus far are either restricted to very
special cases or are so laborious and dicult that even for numbers that do
not exceed the limits of tables constructed by estimable men, they try the
patience of even the practiced calculator. And these methods do not apply at
all to larger numbers ... It frequently happens that the trained calculator will
be suciently rewarded by reducing large numbers to their factors so that it
will compensate for the time spent. Further, the dignity of the science itself
seems to require that every possible means be explored for the solution of a
problem so elegant and so celebrated ... It is in the nature of the problem
that any method will become more complicated as the numbers get larger.
Nevertheless, in the following methods the diculties increase rather slowly
... The techniques that were previously known would require intolerable labor
even for the most indefatigable calculator.
from article 329 of Disquisitiones Arithmeticae (1801) by C. F. Gauss
There are few better known or more easily understood problems in pure mathematics than the question of rapidly determining whether a given integer is prime.
As we read above, the young Gauss in his rst book Disquisitiones Arithmeticae
regarded this as a problem that needs to be explored for the dignity of our subject. However it was not until the modern era, when questions about primality
testing and factoring became a central part of applied mathematics,1 that there
was a large group of researchers endeavoring to solve these questions. As we shall
see, most of the key ideas in recent work can be traced back to Gauss, Fermat and
other mathematicians from times long gone by, and yet there is also a modern spin:
With the growth of computer science and a need to understand the true diculty
of a computation, Gausss vague assessment intolerable labor was only recently
Received by the editors January 27, 2004, and, in revised form, August 19, 2004.
2000 Mathematics Subject Classication. Primary 11A51, 11Y11; Secondary 11A07, 11A41,
11B50, 11N25, 11T06.
Lauteur est partiellement soutenu par une bourse du Conseil de recherches en sciences naturelles et en genie du Canada.
1
Because of their use in the data encryption employed by public key cryptographic schemes;
see section 3a.
c
2004
American Mathematical Society
Reverts to public domain 28 years from publication
ANDREW GRANVILLE
claried by running time estimates, and of course our desktop computers are
todays indefatigable calculators.
Fast factoring remains a dicult problem. Although we can now factor an arbitrary large integer far faster than in Gausss day, still a 400 digit integer which is
the product of two 200 digit primes is typically beyond practical reach.2 This is just
as well since the safety of electronic business transactions, such as when you use
your ATM card or purchase something with your credit card over the Web depends
on the intractability of such problems!
On the other hand we have been able to rapidly determine whether quite large
numbers are prime for some time now. For instance recent algorithms can test an
arbitrary integer with several thousand digits for primality within a month on a
Pentium IV,3 which allows us to easily create the cryptographic schemes referred to
above. However the modern interpretation of Gausss dream was not realized until
August 2002, when three Indian computer scientistsManindra Agrawal, Neeraj
Kayal and Nitin Saxenaconstructed a polynomial time deterministic primality
test, a much sought-after but elusive goal of researchers in the algorithmic number
theory world. Most shocking was the simplicity and originality of their test ...
whereas the experts had made complicated modications on existing tests to
gain improvements (often involving great ingenuity), these authors rethought the
direction in which to push the usual ideas with stunning success.4 Their algorithm
is based on the following elegant characterization of prime numbers.
Agrawal, Kayal and Saxena. For given integer n 2, let r be a positive integer
< n, for which n has order > (log n)2 modulo r. Then n is prime if and only if
n is not a perfect power,
n does not have any prime factor r,
(x + a)n xn + a mod (n, xr 1) for each integer a, 1 a r log n.

We will discuss the meaning of technical notions like order and a little
later. For the reader who has encountered the terminology before, one word of
warning: The given congruences here are between polynomials in x, and not between
the two sides evaluated at each integer x.
At rst sight this might seem to be a rather complicated characterization of the
prime numbers. However we shall see that this ts naturally into the historical
progression of ideas in this subject, is not so complicated (compared to some other
ideas in use), and has the great advantage that it is straightforward to develop into
a fast algorithm for proving the primality of large primes.
Perhaps now is the moment to clarify our goals:
My denition will stretch the usual use of the word practical by erring on the cautious
side: I mean a calculation that can be done using all computers that have or will be built for the
next century, assuming that improvements in technology do not happen much faster than we have
seen in the last couple of decades (during which time computer technology has, by any standards,
developed spectacularly rapidly).
3
This is not to be confused with algorithms that test the primality of integers of a special
shape. For example, the GIMPS project (the Great Internet Mersenne Prime Search) routinely
tests Mersenne numbers, numbers of the form 2p 1, for primality which have millions of digits.
However these algorithms have very limited applicability.
4
Though some experts had actually developed many of the same ideas; see section 6.5 for
example.
IT IS EASY TO DETERMINE WHETHER A GIVEN INTEGER IS PRIME
1.1. Our objective is to nd a quick foolproof algorithm to determine whether

a given integer is prime. Everyone knows trial
division, in which we try to divide
n by every integer m in the range 2 m n. The number of steps in this
algorithm
will be at least the number of integers m we consider,
which is something
like n in the worst case (when n is prime). Note that n is roughly 2d/2 where d
is the number of digits of n when written in binary (and d is roughly log n where,
here and throughout, we will take logarithms in base 2).
The objective in this area has been to come up with an algorithm which works
in no more than cdA steps in the worst case, where c and A are some xed positive constants, that is, an algorithm which works in Polynomial Time (which is
often abbreviated as P). With such an algorithm one expects that one can rapidly
determine whether any reasonably sized integer is prime. This is the modern
interpretation of Gausss dream.
Before the work of Agrawal, Kayal and Saxena the fastest that we could prove
any primality testing algorithm worked was something like dc log log d steps,5 for
some constant c > 0. At rst sight this might seem far away from a polynomial
time algorithm, since log log d goes to innity as d . However log log d goes to
innity with great dignity (as Dan Shanks put it), and in fact never gets larger
than 7 in practice!6 By January 2004 such algorithms were able7 to prove the
primality of 10,000 digit primes (in base 10), an extraordinary achievement.
The algorithm of Agrawal, Kayal and Saxena works in about d7.5 steps;8 and
a modication by Lenstra and Pomerance in about d6 steps. This realizes Gausss
5Though it is believed, but unproven, that some of these tests always work in polynomial time.
6By this I mean that if all computers on earth were to focus simply on writing down as many
7
digits as possible for the next century, they would write down far less than 22 digits.
7
Using a version of the Elliptic Curve Primality Proving routine of Francois Morain and his
collaborators.
8And thus the legend Seven and a half logs should do it! on the Larry Gonick cartoon above.
Printed with permission.
ANDREW GRANVILLE
dream, or, in modern language, implies that Primes are in P. No one has yet
written a computer program implementing these algorithms which comes close to
proving primality of primes as large as those 10,000 digit primes discussed above.
Building on a clever idea of Berrizbeitia, Bernstein (and, independently, Mihailescu and Avanzi) gave a modication that will almost certainly run in around
d4 bit operations. There is hope that this can be used to write a computer program
which will prove the primality of 10,000 digit primes rapidly. This is a challenge
for the near future.
1.2. This article is an elaboration of a lecture given at the Current Events special session during the 2004 annual meeting of the American Mathematical Society.
The purpose is to explain the AKS9 primality test, with complete proofs, and to
put the result and ideas in appropriate historical context.
To start with I would like to discuss simpler ideas from the subject of primality testing, focusing on some that are closely related to the AKS algorithm. In
the process we will discuss the notions of complexity classes from theoretical computer science10 and in particular introduce the P=NP problem, one of the great
challenges of mathematics for the new millennium.
We will see that most of the key ideas used to prove the theorem above were
already in broad circulation, and so it is surprising that such an approach was not
successfully completed earlier. I believe that there were two reasons for this
rst, the way in which these classical ideas were combined was clever and original
in several aspects. Second the authors are not number theorists and came at it
from a little bit of a dierent angle; in particular not being so aware of what was
supposedly too dicult, they trod where number theorists fear to tread.
In the third section we will discuss running time of algorithms in some detail
and how they are determined, and so analyze the AKS algorithm.
In the fourth, and perhaps most interesting, section, we give the proof of the
main results. In fact Agrawal et al. have produced two manuscripts, the second
giving an even easier proof than the rst, and we shall discuss both these proofs
and relevant background information.
To prove the best running times for the algorithm it is necessary to employ tools
of analytic number theory. In section 5 we introduce the reader to some beautiful
theorems about the distribution of primes that should be better known and use
them to prove the claimed running times.
In section 6, we discuss the modied AKS algorithm of Berrizbeitia and Bernstein, as well as Lenstras nite eld primality test,11 and then, in section 7, the
AKS-inspired algorithm of Lenstra and Pomerance.
Since the rst announcement of this result in August 2002 there have been more
than a dozen preprints circulating containing interesting ideas concerning the AKS
algorithm, though none have yet appeared in print. I have thus succumbed to the
temptation to include several of these ideas in the nal section, in part because
they are quite accessible, and in part because they are too elegant to leave out.
9An abbreviation for Agrawal, Kayal and Saxena.
10The cost of conveying the essence, rather than the details, of these notions is that our
denitions will be a little awry, but not in a way that eects the key considerations in our context.
11Since this twenty-year-old test has much in common with the AKS test and has a running
time that is not far from polynomial.
Bernstein reckons that these and other ideas for improving the AKS algorithm
result in a speed up by a factor of about two million, although, he cautions, two
million times faster is not necessarily fast.
1.3. Undergraduate research experiences. Manindra Agrawal is a faculty
member in the Computer Science Department at the Indian Institute of Technology in Kanpur, India. The fundamental approach taken here to primality testing
was developed by Agrawal in conjunction with two bachelors theses which we will
discuss in section 8.5, the rst completed by Pashant Pandey and Rajat Bhattacharjee in 2001, the second completed by Neeraj Kayal and Nitin Saxena in 2002.
Later that summer they developed what they had done into a rst version of the
characterization of primes given above. There can have been few undergraduate
research experiences with such a successful outcome!
2. Primality testing and the Childs Binomial Theorem
2.1. Recognizing primes. To nd a primality test that works faster than trial
division we look for other simple characterizations of prime numbers which might
be used in a more ecient algorithm. If you have studied a little number theory,
then a simple characterization of the primes that comes to mind is:
Wilsons Theorem (1770). Integer n 2 is prime if and only if n divides
(n 1)! + 1.
In trying to convert this elegant characterization into a fast algorithm we run into
the problem that there is no obvious way to compute (n 1)! rapidly (or even
(n 1)! (mod n)).
Another idea is to use Matijasevics unbelievable polynomial (1970) which was
essential in resolving Hilberts Tenth Problem. He showed how to construct a
polynomial with integer coecients (in several variables) such that whenever one
substitutes in integers for the variables and gets a positive value, then that value is a
prime number; moreover, every prime number will be such a value of the polynomial.
(In fact one can construct such a polynomial of degree 10 in 26 variables.) However
it is far from evident how to quickly determine whether a given integer is a value
taken by this polynomial (at least no one to date has found a nice way to do so),
and so this seems to be a hopeless approach.
Primes come up in many dierent places in the mathematical literature, and
some of these suggest ways to distinguish primes from composites. Those of us who
are interested in primality testing always look at anything new with one eye open
to this application, and yet nding a fast primality testing algorithm has remained
remarkably elusive. The advent of the AKS algorithm makes me wonder whether
we have missed some such algorithm, something that one could perform in a few
minutes, by hand, on any enormous number.
Such speculation brings me to a passage from Oliver Sacks The man who mistook
his wife for a hat, in which he tells us of a pair of severely autistic twins with a
phenomenal memory for numbers and a surprising aesthetic. Sacks discovered the
twins holding a purely numerical conversation, in which one would mention a sixdigit number, the other would listen, think for a moment and then beam a smile of
contented pleasure before responding with another six-digit number for his brother.
After listening for a while, Sacks wrote the numbers down and, following a hunch,
determined that all of the numbers exchanged were primes. The next day, armed
ANDREW GRANVILLE
with a table of primes, Sacks butted into their conversation, venturing an eightdigit prime and eliciting, after a short pause, enthusiastic smiles from the twins.
Now the twins kept on going, increasing the number of digits at each turn, until
they were trading (as far as Sacks could tell) twenty-digit prime numbers. So how
did the twins do it? Perhaps we will never know, since the twins were eventually
separated, became socialized and forgot their amazing algorithm!
Since we do not know of any such shortcuts, I wish to move on to a property
of prime numbers that has fascinated mathematicians since antiquity and has been
developed into several key approaches to primality testing.
2.2. The Childs Binomial Theorem. The binomial theorem gives a formula
for expanding (x + y)n as a sum of multiples of terms xi y ni , namely
n

n i ni
(2.1)
(x + y)n =
xy
i
i=0

n!
where ni = i!(ni)!
and m! = m(m 1)(m 2) . . . 3.2.1. This is one of the rst
signicant formulas that students learn in high school, though, perhaps due to
its complicated structure, all too often we nd university undergraduates who are
unable to recall this formula and indeed who write down the (generally) incorrect
(2.2)
(x + y)n = xn + y n .
This is called, by some, the Childs Binomial Theorem. Despite (2.2) being wrong
in general, we shall be interested in those (unlikely) circumstances in which this
formula is actually correct!
One of the most amazing properties of prime numbers, discovered12 by Fermat
around 1637, is that if n is prime, then n divides an a for all integers a. This may
be rewritten13 as
(2.3)
an a
mod n
for all integers a and primes n; and thus

(2.4)
(x + y)n x + y xn + y n
(mod n)
for all integers x, y and primes n.

The integers form equivalence classes mod n, and this set of equivalence classes
forms the ring denoted Z/n (or Z/nZ). If n is prime, then Z/n is a eld, and so
the last equation may be rewritten as (2.2) in Z/n.
Actually (2.4) holds also for variables x and y when n is prime, as we shall see
later in section 2.8, a true Childs Binomial Theorem. It is easy to deduce Fermats
Little Theorem from the Childs Binomial Theorem by induction: Evidently (2.3)
holds for a = 1, and if (2.3) holds for all a < A, then taking x = A 1, y = 1
in (2.4) gives An (A 1)n + 1n (A 1) + 1 = A (mod n) by the induction
hypothesis.
2.3. Composite numbers may sometimes be recognized when they do not
have a particular property that primes have. For example, as we noted above
12There is evidence that this was known for a = 2 far earlier.
13For the uninitiated, we say that a b (mod m) if and only if m divides b a; the main
advantage of this notation is that we can do most regular arithmetic operations (mod m).
Fermats Little Theorem (1637). If n is prime, then n divides an a for all

integers a.
Therefore conversely, if integer n does not divide an a for some integer a, then n
is composite. For example, taking a = 2 we calculate that
21001 123 (mod 1001),
so we know that 1001 is composite.
We might ask whether this always works. In other words,
Is it true that if n is composite, then n does not divide 2n 2?
For, if so, we have a very nice way to distinguish primes from composites. Unfortunately the answer is no since, for example,
2341 2
(mod 341),
but 341 = 11 31. Note though that by taking a = 3 above we get

3341 168 (mod 341),
so we can still use these ideas to prove that 341 is composite.
But then we might ask whether this always works, whether there is always some
value of a that helps us prove a composite n is indeed composite.
In other words,
Is it true that if n is composite, then there
is some integer a for which n does not divide an a?
Again the answer is no since 561 divides a561 a for all integers a, yet 561 =
3 11 17. Composite integers n which divide an a for all integers a are
called Carmichael numbers, 561, 1105 and 1729 being the smallest three examples. Carmichael numbers are a nuisance, masquerading as primes like this, though
computationally they only appear rarely. Unfortunately it was recently proved that
there are innitely many of them and that when we go out far enough they are not
so rare as it rst appears.
2.4. Square roots of 1. In a eld, a non-zero polynomial of degree d has at most
d roots. For the particular example x2 1 this implies that 1 has just two square
roots mod p, a prime > 2, namely 1 and 1.
What about mod composite n? For the smallest odd composite n with more
than one prime factor, that is n = 15, we nd 12 42 112 142 (mod 15); that
is, there are four square roots of 1 (mod 15). And this is true in general: There are
at least four distinct square roots of 1 (mod n) for any odd n which is divisible by
two distinct primes. Thus we might try to prove n is composite by nding a square
root of 1 (mod n) which is neither 1 nor 1, though the question becomes, how do
we eciently search for a square root of 1?
Our trick is to again use Fermats Little Theorem, since if p is prime > 2, then
p1
p 1 is even, and so ap1 is a square. Hence (a 2 )2 = ap1 1 (mod p), so
p1
a 2 (mod p) is a square root of 1 mod p and must be 1 or 1. Therefore if
n1
a 2 (mod n) is neither 1 nor 1, then n is composite. Lets try an example: We
have 64948 1 (mod 949), and the square root 64474 1 (mod 949). Hmmmm,
we failed to prove 949 is composite like this, but, wait a moment, since 474 is
even, we can take the square root again, and a calculation reveals that 64237 220
(mod 949), so 949 is composite. More generally, using this trick of repeatedly taking
10
ANDREW GRANVILLE
square roots (as often as 2 divides n 1), we call integer a a witness to n being
composite if the nite sequence
an1
(mod n), a(n1)/2
(mod n), . . . , a(n1)/2
(mod n)
(where n 1 = 2u v with v odd) is not equal to either 1, 1, . . . , 1 or 1, 1, . . . , 1, 1,

, . . . , (which are the only two possibilities were n a prime). One can compute
high powers mod n very rapidly using fast exponentiation, a technique we will
discuss in section 3b.
It is easy to show that at least one-half of the integers a, 1 a n, are
witnesses for n, for each odd composite n. So can we nd a witness quickly if n
is composite?
The most obvious idea is to try a = 2, 3, 4, . . . consecutively until we nd a
witness. It is believed that there is a witness 2(log n)2 , but we cannot prove this
(though we can deduce this from a famous conjecture, the Generalized Riemann
Hypothesis).14
Pick integers a1 , a2 , . . . , ak , . . . from {1, 2, 3, . . . , n 1} at random until we nd
a witness. By what we wrote above, if n is composite, then the probability that
none of a1 , a2 , . . . , ak are witnesses for n is 1/2k . Thus with a hundred or so such
tests we get a probability that is so small that it is inconceivable that it could occur
in practice, so we believe that any integer n for which none of a hundred randomly
chosen as is a witness is prime. We call such n industrial strength primes.
In practice the witness test allows us to accomplish Gausss dream of quickly
distinguishing between primes and composites, for either we will quickly get a
witness to n being composite or, if not, we can be almost certain that our industrial
strength prime is indeed prime. Although this solves the problem in practice, we
cannot be absolutely certain that we have distinguished correctly when we claim
that n is prime since we have no proof, and mathematicians like proof. Indeed if you
claim that industrial strength primes are prime without proof, then a cynic might
not believe that your randomly chosen a are so random or that you are unlucky or
... No, what we need is a proof that a number is prime when we think that it is.
2.5. Proofs and the complexity class NP. At the 1903 meeting of the American Mathematical Society, F.N. Cole came to the blackboard and, without saying
a word, wrote down
267 1 = 147573952589676412927 = 193707721 761838257287,
long-multiplying the numbers out on the right side of the equation to prove that he
was indeed correct. Afterwards he said that guring this out had taken him three
years of Sundays. The moral of this tale is that although it took Cole a great deal
of work and perseverance to nd these factors, it did not take him long to justify
his result to a room full of mathematicians (and, indeed, to give a proof that he
was correct). Thus we see that one can provide a short proof, even if nding that
proof takes a long time.
In general one can exhibit factors of a given integer n to give a short proof that n
is composite (such proofs are called certicates). By short we mean that the proof
14We will not discuss the Riemann Hypothesis, or its generalizations, here. Suce to say
that this is one of the most famous and dicult open problems of mathematics, so much so
that the Clay Mathematics Insitute has now oered one million dollars for its resolution (see
http://www.claymath.org/millennium/).
11
can be veried in polynomial time, and we say that such problems are in class NP
(non-deterministic polynomial time).15 We are not suggesting that the proof can be
found in polynomial time, only that the proof can be checked in polynomial time;
indeed we have no idea whether it is possible to factor numbers in polynomial time,
and this is now the outstanding problem of this area.
What about primality testing? If someone gives you an integer and asserts that
it is prime, can you check that this is so in polynomial time? Can they give you
better evidence than their say-so that it is a prime number? Can they provide some
sort of certicate that gives you all the information you need to verify that the
number is indeed a prime? It is not, as far as I can see, obvious how to do so,
certainly not so obvious as with the factoring problem. It turns out that some old
remarks of Lucas from the 1870s can be modied for this purpose:
First note that n is prime if there are precisely n 1 integers a in the range
1 a n 1 which are coprime to n. Therefore if we can show the existence
of n 1 distinct values mod n which are coprime to n, then we have a proof that
n is prime. In fact if n is prime, then these values form a cyclic group under
multiplication and so have a generator g; that is, there exists an integer g for which
1, g, g 2, . . . , g n2 are all coprime to n and distinct mod n, so these are the n 1
distinct values mod n that we are looking for (note that g n1 1 (mod n) by
Fermats little theorem). Thus to show that n is prime we need simply exhibit g
and prove that g has order16 n 1 (mod n). It can be shown that any such order
must divide n1, and so one can show that if g is not a generator, then g (n1)/q 1
(mod n) for some prime q dividing n 1. Thus a certicate to show that n is
prime would consist of g and {q prime : q divides n 1}, and the checker would
need to verify that g n1 1 (mod n) whereas g (n1)/q 1 (mod n) for all primes
q dividing n 1, something that can be accomplished in polynomial time using fast
exponentiation.
There is a problem though: One needs certication that each such q is prime.
The solution is to iterate the above algorithm, and one can show that no more
than log n odd primes need to be certied prime in the process of proving that n
is prime. Thus we have a polynomial time certicate (short proof) that n is prime,
and so primality testing is in the class NP.
But isnt this the algorithm we seek? Doesnt this give a polynomial time algorithm for determining whether a given integer n is prime? The answer is no,
because along the way we would have had to factor n 1 quickly, something no one
knows how to do in general.
2.6. Is P=NP? The set of problems that are in the complexity class P are those
for which one can nd the solution, with proof, in polynomial time, while the set
of problems that are in the complexity class NP are those for which one can check
the proof of the solution in polynomial time. By denition PNP, and of course
we believe that there are problems, for example the factoring problem, which are
in NP but not in P; however this has not been proved, and it is now perhaps the
outstanding unresolved question of theoretical computer science. This is another
of the Clay Mathematics Institutes million dollar problems and perhaps the most
15Note that NP is not non-polynomial time, a common source of confusion. In fact it is
non-deterministic, because the method for discovering the proof is not necessarily determined.
16 The order of h (mod n) is the least positive integer k for which hk 1 (mod n).
12
ANDREW GRANVILLE
likely to be resolved by someone with less formal training, since the experts seem
to have few plausible ideas for attacking this question.
It had better be the case that P=NP, else there is little chance that one can have
safe public key cryptography (see section 3a) or that one could build a highly unpredictable (pseudo-)random number generator17 or that we could have any one of
several other necessary software tools for computers. Notice that one implication of
the P=NP question remaining unresolved is that no fast public key cryptographic
protocol is, as yet, provably safe!
2.7. Random polynomial time algorithms. In section 2.4 we saw that if n is
composite, then there is a probability of at least 1/2 that a random integer a is a
witness for the compositeness of n, and if so, then it provides a short certicate
verifying that n is composite. Such a test is called a random polynomial time test
for compositeness (denoted RP). As noted, if n is composite, then the randomized
witness test is almost certain to provide a short proof of that fact in 100 runs of the
test. If 100 runs of the test do not produce a witness, then we can be almost certain
that n is prime, but we cannot be absolutely certain since no proof is provided.
Short of nding a polynomial time test for primality, we might try to nd a
random polynomial time test for primality (in addition to the one we already have
for compositeness). This was achieved by Adleman and Huang in 1992 using a
method of counting points on elliptic and hyperelliptic curves over nite elds
(based on ideas of Goldwasser and Kilian). Although beautiful in structure, their
test is very complicated and almost certainly impractical, as well as being rather
dicult to justify theoretically in all its details. It does however provide a short
certicate verifying that a given prime is prime and proves that primality testing
is also in complexity class RP.
If this last test were practical, then you could program your computer to run
the witness test by day and the Adleman-Huang test by night and expect that you
would not only quickly distinguish whether given integer n is prime or composite,
but also rapidly obtain a proof of that fact. However you could not be certain that
this would workyou might after all be very unluckyso mathematicians would
still wish to nd a polynomial time test that would always work no matter how
unlucky you are!
2.8. The new work of Agrawal, Kayal and Saxena starts from an old beginning,
the Childs Binomial Theorem, in fact from the following result, which is a good
exercise for an elementary number theory course.
Theorem 1. Integer n is prime if and only if (x + 1)n xn + 1 (mod n) in Z[x].

Proof. Since (x + 1)n (xn + 1) = 1jn1 nj xj , we have that xn + 1 (x + 1)n

(mod n) if and only if n divides nj for all j in the range 1 j n 1.

If n = p is prime, then p appears in the numerator of pj but is larger than,

and so does not divide, any term in the denominator, and hence p divides pj for
1 j p 1.
17So-called random number generators written in computer software are not random since
they need to work on a computer where everything is designed to be determined! Thus what are
called random numbers are typically a sequence of numbers, determined in a totally predictable
manner but which appear to be random when subjected to randomness tests in which the tester
does not know how the sequence was generated.
13
If n is composite let p be a prime dividing n. In the expansion

n
n(n 1)(n 2) . . . (n (p 1))
=
p
p(p 1) . . . 1
we see that the only terms p divides are the n in the numerator and the p in the
pk1 is the
denominator; and so if pk is the largest power of p dividing n, then

n
n

largest power of p dividing p , and therefore n does not divide p .
This simple theorem is the basis of the new primality test. In fact, why dont
we simply compute (x + 1)n (xn + 1) (mod n) and determine whether or not
n divides each coecient? This is a valid primality test, but computing (x + 1)n
(mod n) is obviously slow since it will involve storing n coecients!
Since the diculty here is that the answer involves so many coecients (as the
degree is so high), one idea is to compute mod some small degree polynomial as
well as mod n so that neither the coecients nor the degree get large. The simplest
polynomial of degree r is perhaps xr 1. So why not verify whether18
(2.5)
(x + 1)n xn + 1
mod (n, xr 1)?
This can be computed rapidly (as we will discuss in section 3b.2), and it is true for
any prime n (as a consequence of the theorem above), but it is unclear whether this
fails to hold for all composite n and thus provides a true primality test. However,
the main theorem of Agrawal, Kayal and Saxena provides a modication of this
congruence, which can be shown to succeed for primes and fail for composites, thus
providing a polynomial time primality test. In section 4 we shall show that this is
so, but rst we discuss various computational issues.
3a. Computational issues: Factoring and primality testing as applied
to cryptography
In cryptography we seek to transmit a secret message m from Alice to Bob in
such a way that Oscar, who intercepts the transmission, cannot read the message.
The idea is to come up with an encryption key , an easily described mathematical
function, which transforms m into r := (m) for transmission. The number r
should be a seemingly meaningless jumble of symbols that Oscar cannot interpret
and yet Bob can decipher by computing (r), where = 1 . Up until recently,
knowledge of the encryption key would allow the astute Oscar to determine the
decryption key , and thus it was extremely important to keep the encryption key
secret, often a dicult task.
It seems obvious that if Oscar is given an encryption key, then it should be easy
for him to determine the decryption key by simply reversing what was done to
encrypt. However, in 1976 Die and Hellman postulated the seemingly impossible
idea of creating a public key , which Oscar can see yet which gives no hint in and
of itself as to how to determine = 1 . If feasible this would rid Alice of the
diculty of keeping her key secret.
In modern public key cryptosystems the diculty of determining from tends
to be based on an unsolved deep mathematical problem, preference being given
to problems that have withstood the onslaught of the nest minds from Gauss
18The ring of polynomials with integer coecients is denoted Z[x].
Then f (x) g(x)

mod (n, h(x)) for f (x), g(x), h(x) Z[x] if and only if there exist polynomials u(x), v(x) Z[x]
for which f (x) g(x) = nu(x) + h(x)v(x).
14
ANDREW GRANVILLE
onwards, like the factoring problem. We now discuss the most famous of these
public key cryptosystems.
3a.1. The RSA cryptosystem. In 1982, Ron Rivest, Adi Shamir and Len Adleman proposed a public key cryptosystem which is at the heart of much computer
security today and yet is simple enough to teach to high school students. The idea
is that Bob takes two large primes p < q, and their product n = pq, and determines two integers d and e for which de 1 (mod (p 1)(q 1)) (which is easy).
The public key, which Alice uses, will consist of the numbers n and the encryption
key e, whereas Bob keeps the decryption key d secret (as well as p and q). We
will suppose that the message m is an integer19 in [1, n 1]. To encrypt m, Alice
computes r := (m) me (mod n), with (m) [1, n 1], which can be done
rapidly (see section 3b.2). To decrypt Bob computes (r) := rd (mod n), with
(r) [1, n 1]. Using Fermats Little Theorem (for p and q) the reader can easily
verify that mde m (mod n), and thus = 1 .
Oscar knows n, and if he could factor n, then he could easily determine ; thus
the RSA cryptosystems security depends on the diculty of factoring n. As noted
above, this is far beyond what is feasible today if we take p and q to be primes that
contain more than 200 digits. Finding such large primes, however, is easy using the
methods discussed in this article!
Thus the ability to factor n gives Oscar the ability to break the RSA cryptosystem, though it is unclear whether the RSA cryptosystem might be broken much
more easily. It makes sense then to try to come up with a public key cryptosystem whose security is essentially as strong as the diculty of the factoring problem.
3a.2. The ability to take square roots (mod n) is not as benign as it
sounds. In section 2.4 we saw that if we could nd a square root b of 1 mod n
which is neither 1 nor 1, then this proves that n is composite. In fact b yields a
partial factorization of odd n, for
(3.1)
gcd(b 1, n) gcd(b + 1, n) = gcd(b2 1, n) = n,
(as b2 1 (mod n)) whereas gcd(b 1, n), gcd(b + 1, n) < n (since b 1 or 1

(mod n)), which together imply that 1 <gcd(b 1, n), gcd(b + 1, n) < n, and hence
(3.1) provides a non-trivial factorization of n.
More generally let us suppose that for a given odd, composite integer n with
at least two distinct prime factors, Oscar has a function fn such that if a is a
square mod n, then fn (a)2 a (mod n). Using fn , Oscar can easily factor n (in
random polynomial time), for if he picks integers b in [1, n 1] at random, then
gcd(fn (b2 ) b, n) is a non-trivial factor of n provided that b fn (b2 ) or fn (b2 )
(mod n); since there are at least four square roots of b2 (mod n), the probability
that this provides a partial factorization of n is 1/2.
Using this idea, Rabin constructed a public key cryptosystem which is essentially
as hard to break as it is dicult to factor n.
19Any alphabetic message m can be transformed into numbers by replacing A by 01, B
by 02, etc., and then into several such integers by cutting the digits (in binary representation)
into blocks of length < log n.
15
3a.3. On the diculty of nding non-squares (mod p). For a given odd
prime p it is easy to nd a square mod p: take 1 or 4 or 9, or indeed any
a2 (mod p). Exactly (p 1)/2 of the non-zero values mod p are squares mod
p, and so exactly (p 1)/2 are not squares mod p. One might guess that they
would also be easy to nd, but we do not know a surere way to quickly nd such
a value for each prime p (though we do know a quick way to identify a non-square
once we have one).
Much as in the search for witnesses discussed in section 2.4, the most obvious
idea is to try a = 2, 3, 4, . . . consecutively until we nd a non-square. It is believed
that there is a non-square 2(log p)2 , but we cannot prove this (though we can
also deduce this from the Generalized Riemann Hypothesis).
Another way to proceed is to pick integers a1 , a2 , . . . , ak , . . . from {1, 2, 3, . . . ,
n1} at random until we nd a non-square. The probability that none of a1 , a2 , . . . ,
ak are non-squares mod p is 1/2k , so with a hundred or so such choices it is
inconceivable that we could fail!
3b. Computational issues: Running times of calculations
3b.1. Arithmetic on a computer. Suppose a and b are two positive integers,
each with no more than digits when written in binary. We are interested in
the number of bit operations a computer takes to perform various calculations.
Both addition and subtraction can obviously be performed in O() bit operations.20
The most ecient method for multiplication (using Fast Fourier Transforms) takes
time21 O( log log log ). The precise log and log log powers in these estimates
are more or less irrelevant to our analysis, so to simplify the writing we dene O(y)
to be O(y(log y)O(1) ). Then division of a by b and reducing a (mod b) also take
time O().
Now suppose a and b are two polynomials, with integer coecients, of degree less
than r whose coecients have no more than binary digits. Adding or subtracting
will take O(r) operations. To multiply a(x) and b(x) we use the method of single
point evaluation which is so well exploited in MAPLE. The idea comes from the
observation that there is a natural bijection

r1
i
ai x Z[x] : A < ai A for all i
Z/(2A)r ,
a(x) =
i=0
where (a) = a(2A). To recover a0 , a1 , . . . ar1 successively from this value, note
that a0 a(2A) mod 2A and A < a0 A so a0 is uniquely determined. Then
a1 (a(2A) a0 )/(2A) mod 2A and A < a1 A, so a1 is uniquely determined,
and we continue like this. In an algorithm to determine c(x) := a(x)b(x), we rst
note that the absolute values of the coecients of a(x)b(x) are all < A := r22 .
Then we evaluate a(2A) and b(2A) and multiply these integers together to get
a(2A)b(2A), and then recover c(2A) (by a single point evaluation). Unsurprisingly
the most expensive task is the multiplication (of two integers which are each <
+ log r)).
(2A)r ) and so this algorithm takes time O(r(
In our application we also need to reduce polynomials a(x) mod (n, xr 1), where
the coecients of a have O( + log r) digits and a has degree < 2r. Replacing each
20The notation O() can be read as bounded by a xed multiple of .
21
In this context, read time as a synonym for bit operations.
16
ANDREW GRANVILLE
xr+j by xj and then reducing the coecients of the resulting polynomial mod n will
take time O(r(

+ log r)) by the running times for integer arithmetic given above.
3b.2. Fast exponentiation. An astute reader might ask how we can raise something to the nth power quickly, a problem which was beautifully solved by computer scientists long ago:22
We wish to compute (x+a)n mod (n, xr 1) quickly. Dene f0 (x) = (x+a) and
then fj+1 (x) fj (x)2 mod (n, xr 1) for j 0 (at each step we determine fj (x)2
and then reduce mod xr 1 so the degree of the resulting polynomial is < r, and
j
then reduce mod n to obtain fj+1 ). Note that fj (x) (x + a)2 mod (n, xr 1).
Writing n in binary, say as n = 2a1 + 2a2 + + 2a with a1 > a2 > >
a 0, let g1 (x) = fa1 (x) and then gj (x) gj1 (x)faj (x) mod (n, xr 1) for
j = 1, 2, . . . , . Therefore
a1
+2a2 ++2a
g (x) (x + a)2
= (x + a)n
mod (n, xr 1).
Thus we have computed (x + a)n mod (n, xr 1) in a1 + 2 log n such steps,

where a step involves multiplying two polynomials of degree < r with coecients in
{0, 1, . . . , n1} and reducing mod (n, xr 1), so has running time O(r(+log
r)).
3/2 (log n)3 ) bit operations, as we
3b.3. The AKS algorithm runs in O(r
will now show. To transform the theorem of Agrawal, Kayal and Saxena into an
algorithm we proceed as follows:
Determine whether n is a perfect power.
We leave this challenge to our inventive reader, while noting that this can be done
in no more than O((log

n)3 ) bit operations.
If n is not a perfect power, then we
Find an integer r for which the order of n (mod r) is > (log n)2 .
The obvious way to do this is to compute nj (mod q) for j = 1, . . . , [(log n)2 ] and
each integer q > [(log n)2 ] until we nd the rst value of q for which none of these
residues is equal to 1 (mod q). Then we can take r = q. This stage of the algorithm
takes O(r(log
n)2 ) bit operations.23
Determine whether gcd(a, n) > 1 for some a r, which will take O(r(log
n)2 )
bit operations using the Euclidean algorithm, provided r < n.
Finally we verify whether the Childs Binomial Theorem holds:
Determine whether (x+a)n xn +a mod (n, xr 1) for a = 1, 2, . . . , [ r log n];
each such congruence takes O(r(log

n)2 ) bit operations to verify by the previous
3/2 (log n)3 ).
section, and so a total running time of O(r
Adding these times up shows that the running time of the whole algorithm is as
claimed.
3b.4. The size of r. Since r is greater than the order of n (mod r) which is
> (log n)2 , therefore r > (log n)2 . This implies that the AKS algorithm cannot
run in fewer than O((log

n)6 ) bit operations. On the other hand we expect to be
able to nd many such r which are < 2(log n)2 (in which case r will be prime and
22Legendre computed high powers mod p by similar methods in 1785!
23A subtlety that arises here and elsewhere is that (log n)2 could be so close to an integer J
that it would take too many bit operations to determine whether [(log n)2 ] equals J 1 or J.
However, if we allow j to run up to J here, and in the AKS theorem if we allow a to run up
to the smallest integer that is obviously > r log n, then we avoid this diculty and do not
signicantly increase the running time.
17
the powers of n will generate all of the r 1 non-zero residues mod r), and this
is borne out in computation though we cannot prove that this will always be true.
Evidently, if there are such r then the AKS algorithm will run in O((log
n)6 ) bit
operations which, as we have explained, is as good as we can hope for.
We can show unconditionally that there are integers r for which the order of n
(mod r) is > (log n)2 and with r not too big. In section 4.3 we will use elementary
estimates about prime numbers to show that such r exist around (log n)5 , which
1
leads to a running time of O((log

n)10 2 ) (since 10 21 = 32 5 + 3).
In section 6 we will use basic tools of analytic number theory to show that such
r exist which are a little less than (log n)50/11 (using an old argument of Goldfeld),
9
which leads to a running time of O((log n)9 11 ).
It is important to note that the two upper bounds on the running time given
above can both be made absolutely explicit from the proofs; in other words all the
constants and functions implicit in the notation can be given precisely, and these
bounds on the running time work for all n.
Using much deeper tools, a result of Fouvry24 implies that such r exist around
1
(log n)3 , which leads to a running time of O((log n)7 2 ). This can be improved
using a recent result of Baker and Harman to O((log n)7.49 ). However the constants
implicit in the O(.) notation here cannot be given explicitly by the very nature
of the proof, as we will explain in section 5.4.
4. Proof of the theorem of Agrawal, Kayal and Saxena
We will assume that we are given an odd integer n > 1 which we know is not a
perfect power, which has no prime factor r, which has order d > (log n)2 modulo
r, and such that
(x + a)n xn + a mod (n, xr 1)
for each integer a, 1 a A where we take A = r log n. By Theorem 1 we know

that these hypotheses hold if n is prime, so we must show that they cannot hold if
n is composite.
Let p be a prime dividing n so that
(4.1)
(4.2)
(x + a)n xn + a
mod (p, xr 1)
for
each integer a, 1 a A. We can factor xr 1 into irreducibles in Z[x],
as d|r d (x), where d (x) is the dth cyclotomic polynomial, whose roots are the
primitive dth roots of unity. Each r (x) is irreducible in Z[x], but may not be
irreducible in (Z/pZ)[x], so let h(x) be an irreducible factor of r (x) (mod p).
Then (4.2) implies that
(4.3)
(x + a)n xn + a
mod (p, h(x))
for each integer a, 1 a A, since (p, h(x)) divides (p, xr 1).

The congruence classes mod (p, h(x)) can be viewed as the elements of the ring
F : Z[x]/(p, h(x)), which is isomorphic to the eld of pm elements (where m is the
degree of h). In particular the non-zero elements of F form a cyclic group of order
pm 1; moreover, F contains x, an element of order r, thus r divides pm 1. Since
F is isomorphic to a eld, the congruences (4.3) are much easier to work with than
(4.1), where the congruences do not correspond to a eld.
24Fouvrys 1984 result was at the time applied to prove a result about Fermats Last Theorem.
18
ANDREW GRANVILLE
Let H be the elements mod (p, xr 1) generated multiplicatively by x, x + 1,

x + 2, . . . , x + [A]. Let G be the (cyclic) subgroup of F generated multiplicatively by
x, x + 1, x + 2, . . . , x + [A]; in other words G is the reduction of H mod (p, h(x)). All
of the elements of G are non-zero, for if x + a = 0 in F, then xn + a = (x + a)n = 0
in F by (4.3), so that xn = a = x in F, which would imply that n 1 (mod r)
and so d = 1, contradicting
the hypothesis.
Note that if g(x) = 0aA (x + a)ea H, then
((x + a)n )ea

(xn + a)ea = g(xn ) mod (p, xr 1)
g(x)n =
a
by (4.2). Dene S to be the set of positive integers k for which g(xk ) g(x)k
mod (p, xr 1) for all g H. Then g(xk ) g(x)k in F for each k S, so that
the Childs Binomial Theorem holds for elements of G in this eld for the set of
exponents S! Note that p, n S.
Our plan is to give upper and lower bounds on the size of G to establish a
contradiction.
4.1. Upper bounds on |G|.
Lemma 4.1. If a, b S, then ab S.
Proof. If g(x) H, then g(xb ) g(x)b mod (p, xr 1); and so, replacing x by xa ,
we get g((xa )b ) g(xa )b mod (p, (xa )r 1), and therefore mod (p, xr 1) since
xr 1 divides xar 1. Therefore
g(x)ab = (g(x)a )b g(xa )b g((xa )b ) = g(xab ) mod (p, xr 1)
as desired.
Lemma 4.2. If a, b S and a b mod r, then a b mod |G|.
Proof. For any g(x) Z[x] we have that u v divides g(u) g(v). Therefore xr 1
divides xab 1, which divides xa xb , which divides g(xa ) g(xb ); and so we
deduce that if g(x) H, then g(x)a g(xa ) g(xb ) g(x)b mod (p, xr 1).
Thus if g(x) G, then g(x)ab 1 in F; but G is a cyclic group, so taking g to be
a generator of G we deduce that |G| divides a b.

Let R be the subgroup of (Z/rZ) generated by n and p. Since n is not a power
of p, the integers
ni pj with i, j 0 are distinct. There are > |R| such integers with

0 i, j |R| and so two must be congruent (mod r), say
ni p j nI p J
(mod r).
By Lemma 4.1 these integers are both in S. By Lemma 4.2 their dierence is
divisible by |G|, and therefore
|G| |ni pj nI pJ | (np) |R| 1 < n2 |R| 1.

(Note that ni pj nI pJ is non-zero since n is neither a prime nor a perfect power.)
We will improve this by showing that n/p S and then replacing n by n/p in the
argument above to get
(4.4)
|G| n |R| 1.
d
Since n has order d mod r, nd 1 (mod r) and so xn x (mod xr 1).

d
Suppose that a S and b a (mod nd 1). Then xr 1 divides xn x, which
19
divides xb xa , which divides g(xb ) g(xa ) for any g(x) Z[x]. If g(x) H, then
d
d
d
g(x)n g(xn ) mod (p, xr 1) by Lemma 4.1 since n S, and g(xn ) g(x)
d
d
mod (p, xr 1) (as xr 1 divides xn x) so that g(x)n g(x) mod (p, xr 1).
But then g(x)b g(x)a mod (p, xr 1) since nd 1 divides b a. Therefore
g(xb ) g(xa ) g(x)a g(x)b
mod (p, xr 1)
since a S, which implies that b S. Now let b = n/p and a = np(n 1)1 , so
that a S by Lemma 4.1 since p, n S. Also b a (mod nd 1) so b = n/p S
by the above.
d
4.2. Lower bounds on |G|. We wish to show that there are many distinct
elements of G. If f (x), g(x) Z[x] with f (x) g(x) mod (p, h(x)), then we can
write f (x) g(x) h(x)k(x) mod p for some polynomial k(x) Z[x]. Thus if f
and g both have smaller degree than h, then
k(x) 0 (mod p) and so f (x) g(x)
(mod p). Hence all polynomials of the form 1aA (x + a)ea of degree < m (the
degree of h(x)) are distinct elements of G. Therefore if m, the order of p (mod r),
is large, then we can get good lower bounds on |G|.
This was what Agrawal, Kayal and Saxena did in their rst preprint, and to
prove such r exist they needed to use non-trivial tools of analytic number theory.
In their second preprint, inspired by remarks of Hendrik Lenstra, they were able to
replace m by |R| in this result, which allows them to give an entirely elementary
proof of their theorem and to get a stronger result when they do invoke the deeper
estimates.
Lemma 4.3. Suppose that f (x), g(x) Z[x] with f (x) g(x) mod (p, h(x)) and
that the reductions of f and g in F both belong to G. If f and g both have degree
< |R|, then f (x) g(x) (mod p).
Proof. Consider (y) := f (y) g(y) Z[y] as reduced in F. If k S, then
(xk ) = f (xk ) g(xk ) f (x)k g(x)k 0 mod (p, h(x)).
It can be shown that x has order r in F so that {xk : k R} are all distinct roots
of (y) mod (p, h(x)). Now, (y) has degree < |R|, but |R| distinct roots mod
(p, h(x)), and so (y) 0 mod (p, h(x)), which implies that (y) 0 (mod p)
since its coecients are independent of x.

By denition R contains all the elements generated by n (mod r), and so R is
2
at least as large as d, the order of n (mod
r), which is > (log n) by assumption.
Therefore
A, |R| > B, where B := [ |R| log n]. Lemma 4.3 implies that the
products aT (x + a) give distinct elements of G for every proper subset T of
{0, 1, 2, . . . , B}, and so
|G| 2B+1 1 > n |R| 1,

which contradicts (4.3). This completes the proof of the theorem of Agrawal, Kayal
and Saxena.
4.3. Large orders mod r. The prime number theorem can be paraphrased25 as:
The product of the primes up to x is roughly ex . A weak explicit version states that
the product of the primes between N and 2N is 2N for all N 1.
25See section 5.1 for a more precise version, or the book [20].
20
ANDREW GRANVILLE
Lemma 4.4. If n 6, then there is a prime r [(log n)5 , 2(log n)5 ] for which the
order of n mod r is > (log n)2 .
Proof. If not, then the order of n mod r is I := (log
n)2 for every prime r
5
[N, 2N ] with N := (log n) , so that their product divides iI (ni 1). But then

5
2N
r
(ni 1) < n iI i < 2(log n) ,
N r2N
r prime
iI
for n 6, giving a contradiction.
The bound on r here holds for all n 6, and thus using this bound our running
time analysis of AKS is eective; that is, one can explicitly bound the running time
of the algorithm for all n 6. In some of the better bounds on r discussed in
section 3.4, the proofs are not eective in that they do not imply how large n must
be for the given upper bound for r to hold.
4.4. Large prime factors of the order of n (mod r). The other estimates
mentioned in section 3b.4 all follow from using deeper results of analytic number
theory which show that there are many primes r for which r 1 has a large prime
factor q = qr > (log n)2 . By showing that this large prime q divides the order of
n (mod r) for all but a small set of exceptional r, we deduce that the order of n
(mod r) is q > (log n)2 . (In the rst version of the AKS paper they needed m, the
order of p (mod r), to be > (log n)2 and obtained this through the same argument,
since the order of n (mod r) divides the product of the orders of p (mod r), where
the product is taken over the prime divisors p of n, and thus q must divide the
order of p (mod r) for some prime p dividing n.) We now describe our argument a
little more explicitly in terms of a well-believed
Conjecture. For any given in the range 0 < < 1/2 there exists c = c() > 0
such that there are at least 2cR/ log R primes r in [R, 2R] for which r 1 has a
prime factor q > r1/2+ , provided R is suciently large.
Lemma 4.5. Assume the conjecture for some , 0 < < 1/2. Suppose n is a
suciently large integer and that c()R2 log n. There are at least c()R/ log R
primes r in [R, 2R] for which the order of n (mod r) is > r1/2+ .
Proof. We will show that the number N of primes r given in the conjecture for which
q does not divide the order of n (mod r) is < c()R/ log R. Now, if r is such a prime,
then the order of n (mod r) divides (r1)/q, and
r/q r1/2 (2R)1/2 .
so is <
m
Therefore the product of such r divides
(n 1), which implies that
RN
m(2R)1/2
12
nm < nR
m(2R)1/2
and thus N < R12 (log n)/(log R) c()R/ log R as desired.
One easily deduces

Corollary 4.6. Assume the conjecture for some , 0 < < 1/2, and dene () :=
max{1/(2), 4/(1+2)}. There exists a constant c () such that if n is a suciently
large integer, then there exists a prime r < c ()(log n)() for which the order of n
mod r is > (log n)2 .
21
In section 5.3 we will prove the conjecture for some value of > .11 using basic
tools of analytic number theory, so that Corollary 4.6 shows us that we can take
r < (log n)50/11 . Fouvry proved a version of the conjecture for some > 1/6, and
so a modication of Corollary 4.6 shows us that we can take r < (log n)3 . These
estimates were used in section 3b.4.
5. Results about counting primes that everyone should know
5.1. Primes and sums over primes. The prime number theorem tells us that
(x), the number of primes up to x, is26 x/ ln x as x . This can be shown
to be equivalent to the statement

(5.1)
ln p x
px
p prime
mentioned in section 4.3. (Here and throughout section 5 we

will use p to denote
only prime numbers, and so if we rewrite the above sum as px ln p, then the
reader should assume it is a sum over primes p in this range.) There is no easy
proof of the prime number theorem, though there are several that avoid any deep
machinery.
In the proofs below we will deduce several estimates from (5.1).
There will also be estimates of less depth. For example we encounter several
sums, over primes of positive summands, which converge when extended to a sum
over all integers 2, so that the original sum converges.
We now give a well-known elementary
estimate. First observe
that since the

2m
,
which
is
2
,
thus
log p 2m;
primes in (m, 2m] always divide 2m
m<p2m
m

and summing over m = [x/2i ] + 1 for i = 1, 2, . . . we deduce that px log p
2(x + log x), a weak but explicit version of (5.1). Now

x
ln n =
ln p =
ln p
1=
ln p
pa
a
a
a
nx
nx
p |n
p x
nx
pa |n
p x

ln p
x
+ O(x),
+ O(1) ln p = x
=
a
p
p1
a
p x
since
px
(ln p)/pa (2/x)
px<pa
and
ln p = O(1)
px
ln p =
pa x

ln p = O(x),
a1 px1/a
using
our weak form of (5.1). One can use elementary calculus to show that

nx ln n = x ln x + O(x) and thus deduce from the above that
(5.2)
ln p
= ln x + O(1).
p1
px
26We write u(x) v(x) if u(x)/v(x) 1 as x .
22
ANDREW GRANVILLE
5.2. Primes in arithmetic progressions. We are interested in estimating

(x; q, a), the number of primes up to x that are a (mod q). We know that
(x; q, a) (x)/(q) as x if (a, q) = 1,
where (q) = #{a (mod q) : (a, q) = 1}. We wish to know whether something like
this27 holds for small x. Unconditionally we can only prove this when x > exp(q )
for any xed > 0 once q is suciently large, but x > exp(q ) is far too big to
use in this (and many other) applications. On the other hand we believe that this
is so for x > q 1+ for any xed > 0 once q is suciently large and can prove it
when x > q 2+ assuming the Generalized Riemann Hypothesis.
However there is an unconditional result that this holds for almost all q in this
range, where almost all is given in the following form:
The Bombieri-Vinogradov Theorem. For every A > 0 there exists B =
B(A) > 0 such that if x is suciently large and Q = x/(ln x)B , then

(x)
x
max (x; q, a)
.
(a,q)=1
(q) (ln x)A
qQ
This is believed28 to be true when Q = x1 for any > 0.

We also know good bounds on the number of primes in a segment of an arithmetic
progression.
Selbergs version29 of the Brun-Titchmarsh Theorem. For all positive X
and x > q, with a and q coprime integers, we have
2x
(X + x; q, a) (X; q, a)
.
(q) ln(x/q)
5.3. Proof of the conjecture (of section 4.4) for .11. (This proof is
essentially due to Goldfeld.) First note the identity

ln q {(2R; q a , 1) (R; q a , 1)} =
ln q
1
qa 2R
q prime, a1
qa 2R
q prime, a1
R<p2R
p1 mod qa
ln q
R<p2R qa |p1
(5.3)
ln(p 1) R,
R<p2R
in which the last estimate may easily be deduced

from (5.1). The contribution of
a
a
the prime powers q , with a 2 and q R, to the rst sum in (5.3) is

R
R
ln q
=
O
4
,
a1 (q 1) ln R
ln R
q
a
q R
a2
by the Brun-Titchmarsh Theorem for q a R, and then by extending the sum to

all integers q 2 and noting that this new sum is bounded. For prime powers q a ,
27That is, whether 1 < (x; q, a)/((x)/(q)) < 1 + for given small > 0.
28And if so we could prove the conjecture of section 4.4 for any < 1/2 and then deduce that
the AKS algorithm runs in O((log n)6+o(1) ) bit operations by Corollary 4.6.
29Though this rst appeared in a paper of Montgomery and Vaughan.
23
with a 2 and q a R, in the rst sum in (5.3) we bound the number of primes
by the number of terms in the arithmetic progression, namely 2R/q a . Now for
given q the number
of such a 2 is (ln R)/(ln q), and so their contribution to
the sum is 2 R(ln R), which totals to O(R/ ln R) when summing over all q
(2R)1/3
. For the remaining q we only have the a = 2 term which thus contributes
2R q>(2R)1/3 ln q/q 2 = O(R/ ln R). Thus the total contribution of all these
q a , a 2 terms in (5.3) is O(R/ ln R).
apply the Bombieri-Vinogradov Theorem with A = 2, Q =
Next we
R/(ln R)B(2) and x = R and 2R to obtain
ln q

R
R
+ O(1)
ln q((2R; q, 1) (R; q, 1))
ln R
q
1
2
qQ
qQ
q prime
q prime
by (5.2). Finally we apply the Brun-Titchmarsh Theorem to note that

2R
ln q
.
ln q {(2R; q, 1) (R; q, 1)}
(q
1)
ln(R/q)
1/2+
1/2+
Q<q(2R)
q prime
Q<q(2R)
To estimate this last sum we break the sum over q into intervals (QLi , QLi+1 ] for
i = 0, 1, 2, . . . , I 1. Here L is chosen so that QLI = (2R)1/2+ , and we shall require
that L as R does, but with L = Ro(1) . In such an interval we nd that each
ln(R/q) ln(R1/2 /Li ), and so the sum over the interval is 2R ln L/ ln(R1/2 /Li )
1 ln R
1

. Approximating
by (5.2). Thus the quantity above is 2R I1
i=0 2 ln L i
this sum by an integral and then taking u = t ln L/ ln R we get
1
I1

1 ln R
t
2R
dt 2R
(1/2 u)1 du = 2R ln(1/(1 2)).
2 ln L
t=0
0
Combining the above estimates for the various q and using the fact that log q
log 2R give

R
,
(5.4)
{(2R; q, 1) (R; q, 1)} {2c() + o(1)}
ln
R
1/2+
(2R)
<q
q prime
where c() := 1/4ln(1/(12)), which is > 0 provided <

Note that (5.4) implies the conjecture since r 2R.
1e1/4
2
.1105996084.
5.4. How can it be so hard to make some results explicit? In section 5.2 we
stated that one can unconditionally prove (x; q, a) (x)/(q) when (a, q) = 1
where (q) = #{a (mod q) : (a, q) = 1} when x > exp(q ) for any xed > 0 once
q is suciently large, but there is a catch. If 1/2, we do not know how large
we mean when we write q is suciently large (though we do for > 1/2). The
reader might suppose that this fault exposes a lack of calculation ability on the part
of analytic number theorists,30 but this is not the case. It is in the very nature of
the proof itself that, at least for now, an explicit version is impossible, for the proof
comes in two parts. The rst, under the supposition that the Generalized Riemann
30I know I did when I rst encountered this statement.
24
ANDREW GRANVILLE
Hypothesis is true, gives an explicit version of the result. The second, under the
supposition that the Generalized Riemann Hypothesis is false, gives an explicit
version of the result in terms of the rst counterexample. We strongly believe that
the Generalized Riemann Hypothesis is true, but it remains an unproved conjecture,
indeed one of the great open problems of mathematics. So, as long as it remains
unproved, we seem to be stuck with this situation, for how can we put numbers into
the second case in the proof when we believe that there are no counterexamples?
This result of Siegel underpins many of the key results of analytic number theory;
hence many of the results inherit this property of being inexplicit. Make Siegels
result explicit and you change the face of analytic number theory, but for now there
is no sign that this will happen and so we are lumbered with this considerable burden, particularly when trying to apply analytic number theory results to determine
complexity of algorithms.
6. A randomized algorithm with running time O((log n)4+o(1) )
Next we describe an algorithm that distinguishes primes from composites and
provides a proof within O((log n)4+o(1) ) steps. The only drawback is that this is
not guaranteed to work. Each time one runs the algorithm the probability that it
reports back is 1/2, but each run is independent, so after 100 runs the probability
that one has not yet distinguished whether the given integer is prime or composite
is < 1/2100 , which is negligible. This is arguably our best solution to at least part of
the problem that Gauss described as so elegant and celebrated. As we discussed
in section 1.1, this does not yet work in practice on quite such large numbers as
certain tests which are not yet proven but believed to run in polynomial time, but
I believe that it is only a matter of time before this situation is rectied.
The algorithm is a modication of AKS given by Bernstein, following up on
ideas of Berrizbeitia, as developed by Qi Cheng (and a similar modication was
also given by Mihailescu and Avanzi). This is an RP algorithm for primality testing which is faster, easier and more elegant than that of Adleman and Huang. In
practice this makes the original AKS algorithm irrelevant, for if we run the witness test, which is an RP algorithm for compositeness, half of the time and run
the AKS-Berrizbeitia-Cheng-Bernstein-Mihailescu-Avanzi RP algorithm for primality the other half, then a number n is, in practice, certain to yield its secrets faster
(in around O((log n)4+o(1) ) steps) than by the original AKS algorithm!
6.1. Yet another characterization of the primes. For a given monic polynomial f (x) with integer coecients of degree d 1 and positive integer n, we say
that Z[x]/(n, f (x)) is an almosteld with parameters (e, v(x)) if
(a) Positive integer e divides nd 1,
d
(b) v(x)n 1 1 mod (n, f (x)), and
d
(c) v(x)(n 1)/q 1 is a unit in Z[x]/(n, f (x)) for all primes q dividing e.
If n is prime and f (x) (mod n) is irreducible, then Z[x]/(n, f (x)) is a eld;
moreover any generator v(x) of the multiplicative group of elements of this eld
satises (b) and (c) for any e satisfying (a).
Bernstein. For given integer n 2, suppose that Z[x]/(n, f (x)) is an almosteld
with parameters (e, v(x)) where e > (2d log n)2 . Then n is prime if and only if

d
25
(t 1)n tn 1 mod (n, f (x), te v(x)) in Z[x, t] (that is, we work with
polynomials with integer coecients in independent variables t and x).
Proof. Write N = nd and v = v(x). If n is a perfect power, then n is composite. If n
is prime, then the second condition holds by the Childs Binomial Theorem. So we
may henceforth assume that n is not a perfect power and is not prime, and we wish
d
d
to show that (t 1)n tn 1 mod (n, f (x), te v(x)). Let p be a prime dividing
n and h(x) an irreducible factor of f (x) (mod p), so that F = Z[x]/(p, h(x)) is
isomorphic to a nite eld. Let P = |F| = pdeg h , and note that since p < n and
deg h deg f , hence P < N .
Let v (N 1)/e mod (p, h(x)) so that is an element of order e in F: To see
this, note that e v N 1 1 mod (p, h(x)) by (b); whereas if had order m, a
proper divisor of e, then let q be a prime divisor of e/m so that 1 e/q v (N 1)/q
mod (p, h(x)), contradicting (c).

i
ai
in F[t] are distinct, and so those of
The polynomials of the form e1
i=0 ( t 1)
e
degree e 1 are distinct in F[t]/(t v).
Now tN = tN 1 t v (N 1)/e t (mod te v), so that tN t mod (p, h(x), te v).
Thus our second criterion implies that (t 1)N t 1 mod (p, h(x), te v).
Moreover replacing t by i t gives ( i t 1)N i+1 t 1 mod (p, h(x), te v)
i
for any integer i 0 (since ( i t)e v = te v), and thus (t 1)N i t 1
e
mod (p, h(x), te v) by a suitable induction argument. Note that (t 1)N (t 1)
e
mod (p, h(x), t v).

i
iI N
Therefore
for
proper
subsets
I
of
{0,
1,
.
.
.
,
e
1}
the
powers
(t
1)

i
e
iI ( t 1) mod (p, h(x), t v) all have degree e 1 and so are distinct
polynomials, and thus there are at least 2e 1 distinct powers of (t 1) mod
(p, h(x), te v).
Now e is the order of an element of F , which is a cyclic group of order P 1,
and so P 1 is a multiple of e. Therefore v (P 1)/e is an eth root of 1 in F, so
must be a power of , say . Arguing as in two paragraphs above, but now with N
j
and replaced by P and , we see that (t 1)P j t 1 mod (p, h(x), te v).
i j
Combining these results we obtain that (t1)N P i+j t1 mod (p, h(x), te v)
for all integers i, j.
There are more than e pairs of integers (i, j) with 0 i, j [ e], and so there
exist two numbers of the form i + j (with i and j in this range) that are congruent
(mod e), say i + j I + J (mod e). Therefore if u := N i P j and U := N I P J ,
then (t 1)u i+j t 1 = I+J t 1 (t 1)U mod (p, h(x), te v). We will
v) so that we can deduce that there
show that t 1 is a unit mod (p,
h(x), te
e
are no more than |U u| < (N P ) 1 < N 2 e 1 distinct powers of (t 1) mod
(p, h(x), te v), and thus 2e < N 2 e , contradicting the hypothesis.
Now v(x) = 1 in F by (c), so that t 1 is not a factor of te v(x) in F[t]; in
other words t 1 is a unit in the ring F[t]/(te v(x), that is mod (p, h(x), te v).

6.2. Running this primality test in practice. We will show that if n is prime,
then an almosteld may be found rapidly in random polynomial time. The primality
test given by the statement of the theorem may evidently be implemented by similar
methods to those we discussed previously.
26
ANDREW GRANVILLE
Assume that n
is prime. By the inclusion-exclusion formula one can prove that
there are31 (1/d) |d (d/)n irreducible polynomials mod n of degree d. The
biggest term here is the one with = d; that is, roughly 1/d of the polynomials of
degree d are irreducible. Thus selecting degree d polynomials at random we should
expect to nd an irreducible one in O(d) selections. Verifying f is irreducible can
d
be done by checking, via the Euclidean algorithm, that xn x 0 mod (n, f (x))
d/q
and xn x is a unit in Z[x]/(n, f (x)) for all primes q dividing d. Once we have
found f we know that Z[x]/(n, f (x)) is a eld. The elements of Z[x]/(n, f (x)) can
be represented by the polynomials v(x)
mod n of degree < d. The proportion
of these that satisfy (b) and (c) is p|e (1 1/p) > 1/2 ln ln e, and so selecting
such v(x) at random we should expect to nd v(x) satisfying (b) and (c) in O(d)
selections.
d
d
The main part of the running time comes in verifying that (t 1)n tn 1
e
mod (n, f (x), t v(x)), which will take d log n steps, each of which will cost
O(de(log n)1+o(1) ) bit operations, giving a total time of O(d2 e(log n)2+o(1) ) bit operations. The conditions d 1, e > (2 log n)2 imply that the running time cannot
be better than O((log n)4+o(1) ), and we will indicate in the next section how to nd
d and e so that we obtain this running time.
6.3. More analytic number theory. To nd an almosteld when n is prime we
need to nd d and e for which e divides nd 1 and with d and e satisfying certain
conditions. Constructions typically give e as a product of primes p which do not
divide n and for which p 1 divides d, since then p divides nd 1 by Fermats Little
Theorem, and thus e divides nd 1.
However, to ensure that e is large, for instance e > (2d log n)2 as required in
the hypothesis of Bernsteins result, we need to use the ideas of analytic number
theory. Our general construction looks as follows: For given z < y, with z y for
xed > 0, let d be the least common multiple of the integers up to z and e be the
of p 1 are
product of all primes p y such that all prime power divisors q a
z.
Note that d = exp(z+o(z)) by the prime number theorem and e = py p
pP p,
where P is the set of primes p y for which p 1 has a prime power divisor q a
which is > z.
If p P write p 1 = kq a with q a > z, so that k < y/z 1/. Now the number
of q a (z, y) with a 2 is O( y), and so there are O( y/) values of p P for

a
which p 1 has a prime power divisor q with a 2.
In our rst construction we take y = 4z, so that if a = 1, then we have a prime
pair of the form q, kq + 1 with k < 4, and so k = 2. For this we use a well-known
bound on the number of prime pairs of the form q, 2q + 1:
Lemma 6.1. There exists a constant c > 0 such that there are 2cx/(log x)2
primes q x for which 2q + 1 is also prime, for all x 2.
Therefore |P | = O(y/(log y)2 ) and so e = exp(y + o(y)) by the prime number
theorem. If we take y = (4+) log log n, then we get e > (2d log n)2 as required, and
the values of d and e can, in practice, be found quickly. However, by the remarks
of the previous section the running time will be O((log n)8+O() ), so we need to
choose d and e slightly dierently.
31Here (m), the Mobius function, is dened to be (1)k if m is squarefree and has k distinct
prime factors, and 0 otherwise.
27
This time we take z = y with y = (2 + 3) log log n. We need the generalization
of Lemma 6.1 to prime pairs of the form q, kq + 1:
Lemma 6.2. There exists an absolute constant c > 0 such that there are
c(k/(k))(x/(log x)2 ) primes q x for which kq + 1 is also prime, for all even
integers k and all x 2.
In this case, corresponding to each prime p P with a = 1, we have a prime
pair q, kq + 1 with k 1/ and q y/k. For given k 1/ there are
cy/(log(y))2 such prime pairs, by Lemma 6.2 with x = y/k, since (k) 1. There
fore |P| = O(y/((log y)2 ) + y/) = o(y/ log y), so the product of the primes in P
is y |P| = exp(o(y)). Thus e = exp(y + o(y)) by the prime number theorem, and
so e > (2d log n)2 as required; but now the running time will be O((log n)4+O() ),
and letting 0 we get the desired result.
6.4. Bernsteins construction was originally analyzed using a beautiful result
of Prachar. If p 1 divides d for every prime p dividing e, as
above, then how
large can e be in terms of d? An obvious upper bound for e is m|d (m + 1) in the
case that one more than each divisor of d is a prime. So if (d) is the number of
divisors of d, then e < d (d) . Can we obtain e which is anywhere near this big?
Prachars idea is to look for primes of the form mk + 1 for some small integer k,
as m runs through the divisors of d. Under the assumption of the Generalized
Riemann Hypothesis, for each xed m|d, there are > d2 /2 log d primes of the form
mk + 1 with d2 < k < 3d2 , so their product is > exp(d2 ). Therefore there exists
some value of k < 3d2 such that the product e of the primes of the form km + 1
with m|d is > exp( (d)/2) (and we replace d above by D := kd < 3d3 ). If we took
the original d to be the product of the primes z, then, by the prime number
theorem, d = exp(z + o(z)) and log e > 2(z)1 so that D < (log e)c log log log e , for
some constant c > 0. This argument can be justied without assumption, since as
we discussed in section 5.2, we almost always get the expected number of primes
in an arithmetic progression in the relevant ranges (see [2] for the technical details).
6.5. Lenstras 1985 nite eld primality test uses many of the same ideas as
the AKS test and these variants. It is surprising how close researchers were twenty
years ago to obtaining a polynomial time primality test.
Lenstra. For a given almosteld Z[x]/(n, f (x)) with parameters (e, v(x)), if

ni
(d) g(T ) := d1
i=0 (T v(x) ) (Z[x]/(n, f (x)))[T ],
then p nj (mod e) for some j, 0 j d 1, for each prime p dividing n.
Proof. Let h(x) be an irreducible factor of f in (Z/pZ)[x], and dene F :=
Z[x]/(p, h(x)). Now g(v(x)) = 0 by (d), so that g(v(x)p ) = g(v(x))p = 0 in F,
i
by the Childs Binomial Theorem. Therefore v(x)p = v(x)n in F for some i. This
implies that p ni mod the order of v(x) in F. The result follows since the order
of v(x) in F is divisible by e, by (b) and by (c).

We use this as a primality test by selecting d and e as in section 6.4 so that n2 >
e > n and D < (log n)O(log log log n) and then nding an almosteld as described in
section 6.2. If (d) does not hold, then n is composite (since our choice of f (x) is
irreducible if n is prime). If d does hold, then the candidates to be prime factors
28
ANDREW GRANVILLE
of n by Lenstras theorem are the least residues of n, n2 , n3 , . . . , nd1 (mod e), and
these are easily checked by trial division. The running time of this random test
for the primality of n is thus O((log n)O(log log log n) ), just a smidgin slower than
polynomial time. (This is a simplication of Lenstras test, which was actually
a little more involved but compensated by being faster in terms of the constant
implicit in the O in the exponent.)
7. Stop the press: (log n)6 achieved
7.1. Lenstra and Pomerance obtain the desired running time. Lenstra
and Pomerance signicantly modied the AKS algorithm so that it will, in theory,
work as fast as can be hoped for, namely in O((log

n)6 ) bit operations. Moreover it
is possible to give explicit constants in this statement, in other words a computable
upper bound on the running time that holds for all n.
Their key idea is to replace the polynomial r (x) in AKS by a polynomial f (x)
with certain properties: For a given monic polynomial f (x) with integer coecients
of degree d, and positive integer n, we say that Z[x]/(n, f (x)) is a pseudoeld if
(a) f (xn ) 0 mod (n, f (x)),
d
(b) xn x 0 mod (n, f (x)), and
d/q
(c) xn x is a unit in Z[x]/(n, f (x)) for all primes q dividing d.
When n is prime and f (x) is irreducible mod n, then these criteria are all true and
Z[x]/(n, f (x)) is a eld.
Lenstra and Pomerance. For given integer n 2 let d be an integer in
((log n)2 , n) for which there exists a monic polynomial f (x) of degree d with integer coecients such that Z[x]/(n, f (x)) is a pseudoeld. Then n is prime if and
only if
n does not have any prime factor d,
(x + a)n xn + a mod (n, f (x)) for each integer a, 1 a A := d log n.

One can easily see that this theorem has its genesis in the work of Agrawal, Kayal
and Saxena, yet is more general. This generality is what allows them to prove that
it can be achieved in far fewer steps.
Evidently for a given f one can quickly determine whether one gets a pseudoeld, and if so check the criteria of the theorem. Thus if we can quickly nd
an f which gives a pseudoeld, this approach will lend itself to a quick primality
test. Lenstra and Pomerances construction of f comes back full circle to Gausss
Disquisitiones and his construction of regular n-gons, in particular what are now
known as Gaussian periods.
In the next subsection we will give Agrawals proof of Lenstra and Pomerances
theorem. In section 7.3 we will introduce Gaussian periods and discuss how to
construct f with the required properties. In sections 7.4 and 7.5 we sketch the
proof that this can be done in the required number of steps.
7.2. Proof that this is a characterization of the primes. Suppose that n
is composite and satises the three hypotheses. Let p be a prime dividing n and
let h(x) be an irreducible factor of f (x) (mod p), so that F Z[x]/(p, h(x)) is
isomorphic to a nite eld.
29
As in section 4, let H be the elements mod (p, f (x)) generated multiplicatively

by x, x + 1, x + 2, . . . , x + [A]; let G be the (cyclic) subgroup of F generated multiplicatively by x, x + 1, x + 2, . . . , x + [A]; and let S be the set of positive integers of
the form pi nj with i, j 0. Dene r to be the order of x mod (p, f (x)) so that d
is the order of n mod r by the denitions (b) and (c) of a pseudoeld, and further
0
1
d1
xn , xn , . . . , xn
are distinct mod (p, f (x)) and even mod (p, h(x)). There d1
i
fore the polynomial g(T ) := i=0 (T xn ) F[T ] has distinct roots; moreover
j
g(xp ) = g(x)p = 0 in F and therefore xp must equal xn in F for some j. This
j
implies that p n (mod r), and so if R is the subgroup of (Z/rZ) generated by
n and p, then R is in fact generated by n alone and so has d elements.
The same proof works for Lemma 4.1 except for the observation that xr 1
divides xkr 1 for any k S. We replace this by the observation that f (xk ) 0
(mod (p, f (x))) for all k S, which holds since f (xn ) 0 (mod (p, f (x))) by
(a), and f (xp ) f (x)p (mod p) 0 (mod (p, f (x))) by the Childs Remainder
Theorem (mod p).
Lemma 4.2 follows with much the same proof (with (p, xr 1) replaced by
(p, f (x))) since (p, f (x)) divides (p, xr 1). The remarks after (4.4) follow simply by replacing the rst sentence there by (b) and otherwise replacing (p, xr 1)
by (p, f (x)). Therefore (4.4) holds and Lemma 4.3 holds, sojust as in section 4.2,
we note that |G| 2B+1 1 where B := [A], so that |G| > n |R| 1, contradicting
(4.4).
7.3. The construction of f . Let r =e2i/r for prime r. If q divides r 1,
we dene the Gaussian period to be = jJ rj where J = Jr,q := {j (mod r) :
j (r1)/q 1 (mod r)} is the set of residue classes (mod r) which are coprime to
r and which are qth powers 32 (mod r). Now J is a subgroup of the cyclic group
(Z/rZ) , and so J = {g qi : 0 i < (r 1)/q} for a generator g of (Z/rZ) .
Moreover J has q cosets in (Z/rZ) , namely g i J for 0 i q 1. Thus = 0 has

gi j
conjugates33 i : =
for i = 0, 1, 2, . . . , q 1. The minimum polynomial
jJ r
q1
for over Q is f (x) = i=0 (x i ) which has degree q, is monic, and has integer
coecients.
Let p be prime, dierent from r. In the eld Q(r ), primes over Q (like p)
might factor further34 into prime ideals, so suppose P is a prime ideal factor of
p in Q(r ). Also f (x) might factor (mod p), so let g(x) be that factor of f (x)
(mod p) for which g() 0 (mod P). Now, the Childs Binomial Theorem tells us
that g(xp ) g(x)p (mod p), and hence mod P, and therefore g( p ) g()p 0
(mod P). Thus p k (mod p) where k g k J is a root of g(x) mod P, and by
similar arguments we nd that ik mod q are also, for i = 0, 1, . . . , q 1. These are
all distinct exactly when the order of p(r1)/q (mod r) is q. From this it is possible
to deduce that f (x) is irreducible (mod p) if and only if the order of p(r1)/q
(mod r) is q.
In this way we can construct an irreducible polynomial of degree q over Fp , the
minimum polynomial of the Gaussian period . Furthermore, if we have several
32That is, there exists i (mod r) for which iq j (mod r).
33The automorphisms of the eld Q( ) are each induced by a map k for some k, 1
r
r
r
k r 1.
34For example, 5 = (1 + 2 )(1 2 ) in Q( ).
4
4
4
30
ANDREW GRANVILLE
primes r1 , r2 , . . . , rk and pairwise coprime positive integers q1 , q2 , . . . , qk for which

qi divides ri 1 for each i, then f (x), the minimum polynomial of 1 2 k over
Q, has degree q1 q2 . . . qk and is irreducible (mod p) if and only if the order of
p(ri 1)/qi (mod r) is qi for each i. This leads to our construction of f : For given n
we look for qi and ri as above, though with p replaced by n. Note that if n is prime,
then Z[x]/(n, f (x)) is a pseudoeld. If n is composite, then either Z[x]/(n, f (x))
fails to be a pseudoeld, in which case we have a proof that n is composite, or it is
a pseudoeld, in which case we can apply the theorem of Lenstra and Pomerance,
provided q1 q2 . . . qk > (log n)2 , to test whether n is prime. In fact they prove that
such f exist with q1 q2 . . . qk just a little bigger than (log n)2 :
Proposition 7.1. There exists a computable constant n0 such that if n n0 , then
there exist primes r1 , r2 , . . . , rk < (log n)2 and pairwise coprime positive integers
q1 , q2 , . . . , qk with qi dividing ri 1 and for which the order of n(ri 1)/qi (mod r)
is qi for each i, such that (log n)2 < q1 q2 . . . qk < 4(log n)2 .
To determine whether the order of n(ri 1)/qi (mod r) is qi we need only check
that nri 1 1 (mod ri ) whereas n(ri 1)/qi 1 (mod ri ). Using Proposition 7.1
naively we can determine suitable values for ri and qi in O((log n)3 ) steps. Moreover
all of the results used in analyzing the steps of this algorithm can be written down
with the constants explicitly given. In other words the running time of the Lenstra
and Pomerance algorithm can be given totally explicitly.
7.4. The continuous postage stamp problem. Post oces issue stamps of
dierent value to allow the sender to ax a suitable amount for postage. If S is
the nite set of positive integers which gives the values of the dierent stamps (in
cents), then to be able to make up any price (in cents) using an assortment of
stamps, it is necessary that gcd{s S} = 1. On the other hand if gcd{s S} = 1,
then we know that every suciently large integer can be represented as a nonnegative integral linear combination of elements of S, but not necessarily every
positive integer. The Frobenius postage problem is to determine the largest integer
that cannot be so represented. If |S| = 2, then this is easy, but there is no known
formula for arbitrary S; in fact this problem is NP-hard.35
Perhaps the most surprising part of Lenstra and Pomerances proof of Proposition 7.1 is that, ultimately, their analytic argument depends on a continuous
analogy to the Frobenius postage problem: Let S be an open subset of the positive
reals that is closed under addition. One might expect that if S is suciently dense,
then that forces every
suciently large real number to be in S; here suciently
dense is in terms of 0tx, tS dt
t . They conjectured the following criterion, which
was proved by Daniel Bleichenbacher:
Lemma 7.2.
If S is an open subset of the positive reals that is closed under addition
for which 0tu, tS dt
t > u for some u (0, 1], then 1 S.
7.5. The analytic part of the argument, the proof of Proposition 7.1, has several
complicated details, so we will just sketch the main ideas. Let x = (log n)1+3
where = 1/ log log log n.
One can prove that for most
primes r, much of the size of r 1 is made up of
large primes, that is, that q|r1, q>x2 q > x1 for all but O(x/(log x)3 ) of the
35That is, it is at least as dicult as any other NP problem.
31
primes r x. If we have such an r, then it is very likely that n(r1)/q (mod r) has
order q for some such q dividing r 1, for if not, then n has order < x (mod r),
but there are few such r as can be shown by an argument like that in the proof of
Lemma 4.5.
2
Let Q be the set of primes q (x , x1/2 ] such that there exists a prime r x
which is 1 (mod q) and for which q is the order of n(r1)/q (mod r). Using
the above observations, together
with a sharpened form of the Brun-Titchmarsh
theorem, one can show that qQ 1/q > 3/11 o(1).
Let = 2 /2 log log n. For each q Q let q := log q/2 log log n and dene
S0 = qQ (q , q ) and then S to be the closure of S0 under addition. By the

prime number theorem, x <q<x , q prime 1/q ln(/) = dt
t , from which we

easily deduce that 0t1/4+, tS dt
>
{1
+
o(1)}
1/q
>
3/11
o(1) by the
qQ
t
above.
Therefore 1 S by Lemma 7.2 so that there exists V S0 such that

v
= 1. Let U be the set of q Q for which v (q , q ) as we run through
vV
v V (this simple argument
may be modied to show that dierent v give rise to

dierent q). Thus 0 < qU log q 2 log log n < 2 as required for Proposition 7.1.
8. Minor improvements and tempting ideas
In this section we begin with four areas in which there are elegant ideas for improving the proof of section 4, selected for the beauty of the mathematics involved.
Next we discuss Agrawal and his students quest for a polynomial time primality
test, and then develop their ideas to rewrite the AKS test without polynomials.
8.1. Varying the elements of G. In section 4 we took G to be the subgroup of
F generated by V := {x + 1, . . . , x + [A]} and veried that n S = SG (by checking
(4.1) for a = 1, 2, . . . , A). If we can deduce from this knowledge that (4.1) holds
for other values of a, then we can expand the size of V without changing the proof
except for showing that G is larger than before. So for any a, 1 a A verify
that an a (mod n) (else n is composite), and then select integer b so that ba 1
(mod n). Now, (4.1) can be rewritten as (x + a)n (xn + a) = nu(x) + (xr 1)v(x)
for certain u(x), v(x) Z[x], and one can show that deg u < n, deg v < n r using
the Euclidean algorithm in Z[x]. Replacing x by 1/x and multiplying through
by (bx)n we obtain (b + abx)n (bn + (abn )xn ) = nU (x) (xr 1)V (x) where
U (x) := bn xn u(1/x), V (x) := bn xnr v(1/x) Z[x]. Now (x + b)n (abx + b)n
(mod n) and xn (abn )xn (mod n), so we have proved that (4.1) also holds when
we replace a by b. Thus we may replace V by V {a1 (mod n) : 1 a A}, a
set that will be almost twice as large (and thus our lower bound on the size of G
will be squared).
Just as we replaced x by x1 in (4.1) to nd another case of the Childs Binomial
Theorem, so we may also replace x by xk for any integer k since xr 1 divides
xkr 1, and thus we may generate G using all polynomials of the form
{xk + a : k N, 1 a A} {xk + b : k N, 1 a A, b a1
(mod n)}.
From Lemma 4.3 we deduced, in section 4.2, that each polynomial generated as
a product of the polynomials {x, x + 1, . . . , x + [A]} which has degree < |R| belongs
to a dierent element of G, and thus we obtained a lower bound on the size of
G by counting the number of such polynomials without repeated roots. Voloch
suggested a cleverer counting argument. Suppose that G contains k distinct linear
32
ANDREW GRANVILLE
polynomials x + a1 , . . . , x + ak . We wish to nd the largest possible subset U of Zk

with the property that if u, v U , then

max{ui vi , 0} and
max{vj uj , 0} ,
(8.1)
1ik
1jk
for = |R| 1. Letting mi = minuU ui , the polynomials
{gu =
(x + ai )ui mi : u U }
1ik
are all distinct elements of G, for if gu = gv in F, then
(x + ai )max{ui vi ,0} =
(x + aj )max{vj uj ,0}
i
which are both of degree < |R|, and this contradicts Lemma 4.3. Thus |G| |U |.
Determining u(k, ), the size of the largest such set U , is an elegant and open36
combinatorial problem about
which we know the following:
s and
Dene Us,t := {u Zk :
1ik max{ui , 0}
1ik max{ui , 0}
k
t}. Evidently U,0 = {u Z : Each ui 0 and
1ik ui } satises (8.1).
satises
(8.1)
for
any
0
, for if u, v Us,s , then

Voloch
noted
that
U
s,s

max{u
v
,
0}
(max{u
,
0}
+
max{v
i
i
i
i , 0}) s + ( s) = .
1ik
1ik
On the other hand if we x v U , then U v + U, by (8.1). Thus we deduce that
|U, | u(k, ) max0s |Us,s |. We can determine the number of elements of
Us,t by counting those n Us,t with exactly j values of i for which ui 0 to obtain

k

k
s+j
t
|Us,t | =
.
j
j
kj
j=0
8.2. Upper bounds on |G|, revisited. In section 4.1 the pigeonhole principle
was used to nd two elements of R which are congruent mod r. Poonen and Lenstra
replaced this by
Minkowskis convex body theorem. Let be a lattice which is a sublattice of
Z2 . Let U be a convex 37 body in R2 which is symmetric about the origin. If the area
of U is at least 4 times the determinant 38 of , then U contains a lattice point of
other than the origin.
We also note the following lemma which is left as an elementary exercise for the
reader:
Lemma 8.1. For positive a, b, c dene T = {(x, y) R2 : x, y 0, and ax+by c}
and U = {t t : t, t T }. Then U = T T V V , where V = {(x, y)
R2 : 0 ax, by c}, so U is a symmetric convex body in R2 , whose area is 6 times
the area of T .

Take T = (x, y) R20 : px (n/p)y n |R|/3 in Lemma 8.1 with a = log p, b =

log(n/p) and c = |R|/3 log n. Then Area(U
) = 3c2 /ab 4|R|, "since
! ) = 6Area(T
2
1
2
i
log p log(n/p) 4 log n. The lattice = (i, j) Z : p (n/p)j 1 (mod r) has
determinant |R|, so by Minkowskis convex body theorem, T contains two distinct
36As far as I know.
37That is, U contains the line segment connecting any two points in U .
38That is, the area of the smallest parallelogram of .
33
points, call them (i, j) and (I, J), which are congruent
mod r. As in section 4.1, |G|
|R|/3
i
j
I
J
1, an improvement on (4.3).
divides p (n/p) p (n/p) , so that |G| n
8.3. Using polynomials of higher degree in G. Above we obtained lower
bounds on the size of G by determining a large set of polynomials in G of degree
< |R|, since Lemma 4.3 guarantees that they will be distinct. Voloch showed that
few pairs of polynomials in G of slightly larger degree can be equal, and so improved
the lower bounds on |G|. His method revolves around one of the most remarkable
results in the arithmetic of function elds:
abc-theorem for polynomials. If a, b, c C[t] and not all in C, with a + b = c,
where a and b have no common factors, then the degrees of a, b and c are each less
than the total number of distinct roots of abc.
For a proof, see [13]. By the example in Theorem 1, tp + 1 = (t + 1)p , this result
does not carry over to elds of characteristic p, for in this example the maximum
degree is p and yet there are just 2 distinct roots. However it is not dicult to
modify the proof in [13] to show that if a + b = c where a, b, c are genuine coprime
polynomials in a eld of characteristic p, then either a = Ap , b = B p and c = C p
where A + B = C or the above conclusion holds. From this we deduce
Lemma 8.2. Let K be a eld of characteristic p. Suppose a, b, c K[t] and have
no common factor, and that no two of u, v, w K[t] have a common factor, where
au + bv = cw and g = (au, bv, cw). Then g = (a, b)(a, c)(b, c) and either au, bv and
cw each equal g times the pth power of a polynomial in K, or
"
!
2 max {deg u, deg v, deg w} > max(deg a, deg b, deg c)# distinct roots of abc/g 2 .
Proof. Note that (a, b) divides au + bv = cw and has no common factor with
c, so divides w. Similarly (b, c) divides u, and (a, c) divides v. Therefore g =
(a, b)(a, c)(b, c) and g divides uvw.
Applying the above abc-theorem for polynomials in a eld of characteristic p
we nd that either au/g, bv/g and cw/g are each a pth power or the degrees of
au/g, bv/g and cw/g are all less than the ! number of distinct roots
" of
(au/g)(bv/g)(cw/g), which is deg(uvw/g) + # of distinct roots of abc/g 2 , and
the result follows.

Corollary 8.3. Let K be a eld of characteristic p. If a, b, c K[t] and have no
common factor, and m K[t] with a b c mod m, where a and m have no
common factor, then
max{deg a, deg b, deg c}

1
2
(deg m + p), 2 deg m #{distinct roots of abc/g }
min
2
where g = (a, b)(a, c)(b, c).
Proof. Let U = (b c)/m, V = (c a)/m and W = (b a)/m. Let h = (U, V, W )
and U = hu, V = hv, W = hw, so that au + bv = cw. Let G = (au, bv, cw). Note
that (a, b, c) = (u, v, w) = 1 by denition. Also (u, v, c) divides (b c, c a, c) =
(b, a, c) = 1, so (u, v, c) = 1 and similarly (u, w, b) = (v, w, a) = 1. Therefore the
hypothesis of Lemma 8.2 holds with g = G. Note that, by denition,
(8.2)
max{deg u, deg v, deg w} max{deg a, deg b, deg c} deg(hm).
34
ANDREW GRANVILLE
If au/g, bv/g and cw/g are not all pth powers (or all belong to K), then by substituting (8.2) into Lemma 8.2, we get the second lower bound here. If au/g, bv/g
and cw/g are all pth powers but are not all in K, then
p max{deg au, deg bv, deg cw} 2 max{deg a, deg b, deg c} deg(m)
by (8.2).
Suppose that p is prime and h(x) is irreducible mod p where deg h(x) p/3.
Suppose that V is a set of linear polynomials. Consider the set V of polynomials
of degree < 2 deg h |V |, which are products of elements of V . By Corollary 8.3
(with K = Fp , m = h and a, b, c V ) we deduce that there are at least 12 |V |
distinct polynomials in V mod h(x). This allows us to improve the lower bound
on G that we achieved in section 4 and may be combined with the ideas in section
8.1.
8.4. Fermats Last Theorem, again. Fouvrys result, mentioned in section
3b.4, was inspired by its application to the rst case of Fermats Last Theorem,
that is, integer solutions x, y, z to xp + y p = z p where p does not divide xyz. In
1910 Wieferich proved that if this has a solution, then 2p1 1 (mod p2 ), which
seems to be true very rarely. It turns out that knowledge about such congruences
can be used to reduce the value of r above.
Lemma 8.4. Let be a given prime and let L = if is odd, L = 4 if = 2. Let k
be the order of n mod L. Suppose that e is the highest power of dividing nk 1.
Then kj is the order of n mod e+j .
Proof. If nK 1 = E s where does not divide integer s, and E > 2, then, by the
binomial theorem, nK = (1 + E s) 1 + E+1 s (mod E+2 ). The result follows
from a suitable induction hypothesis.

Thus if n 3 (mod 8), then n has order 2i2 mod 2i , for all i 3, so n has
order > log2 n for some r < 8 log2 n, taking r to be an appropriate power of two.
More generally if n 1 (mod 2s ) for any s 2, then n mod r has order > log2 n
for some r < 2s log2 n. This argument may be extended to odd primes :
Corollary 8.5. If n1
1 (mod 2 ) for odd prime , then there exists r
2
2
log n for which n mod r has order > log2 n.
Proof. We may assume is odd and so, in Lemma 8.4, we have e = 1. Select
j as small as possible so that kj > log2 n, and thus j kj log2 n. Then

r := j+1 2 log2 n.
8.5. Undergraduate theses. Manindra Agrawal had worked for several years
with students trying to nd a criteria for primality of this sort. In April 2001,
Pashant Pandey and Rajat Bhattacharjees bachelors thesis Primality Testing
studied what would happen if
(8.3)
(x 1)n xn 1 mod (n, xr 1)
holds for odd composite n and odd prime r. Their hope was that if this happens,
then n 1, 0 or 1 (mod r). However such a neat conclusion seems unlikely,
even for r = 5: To prove that there are innitely many Carmichael numbers, one
constructs integers n with prime factors that have certain extraordinary divisibility
properties. In a similar vein Lenstra and Pomerance conjecture that there are
35
innitely many squarefree integers n with 4k + 1 prime factors (with k 1) such

that every prime p which divides n is 3 mod 80 and has p 1 divides n 1,
and p + 1 divides n + 1. Given such an n, note that p2 1 or 1 mod 5, and
2
2
so (x 1)p xp 1 x 1 or x1 (x 1) mod (p, x5 1), respectively;
2
either way (x 1)10(p 1) 1 mod (p, x5 1). But n p mod 10(p2 1) and so
(x 1)n (x 1)p xp 1 xn 1 mod (p, x5 1), and thus (8.3) holds with
r = 5 by the Chinese Remainder Theorem.
The following summer, Kayal and Saxena noted that (8.3) implies some of the
more classical criteria:
Proposition 8.6. Assume that (x 1)n xn 1 mod (m, xr 1) where m, n, r
are positive integers with n odd and (r, n) = 1. Then rn r (mod m).
r1
Proof. Let = e2i/r so that r = 1 but = 1. Now j=1 (t j ) = (tr 1)/(t 1)
r1
= 1 + t + t2 + + tr1 , so taking t = 1 gives j=1 (1 j ) = r. Taking x = j
in the hypothesis implies that (1 j )n = 1 jn (mod m). Multiplying these
equations together for j = 1, 2, . . . , r 1 we obtain the result by noting that
{jn (mod r) : 1 j r 1} = {j (mod r) : 1 j r 1} since (n, r) = 1.
They also show that if (r, m) = 1, then r(n1)/2 must be 1 or 1 (mod m), that
is a square root39 of 1 mod m. For r = 5 they determine the value of F(n1)/2
(mod n) where Fk is the kth Fibonacci number. Moreover that if (8.3) holds for
each r 2(ln n)2 , then n is squarefree. There may yet be much more of interest to
deduce from (8.3).
Taking x = 1 in (4.1) we obtain that (a+1)n a+1 mod n. By taking x = j in
(4.1) as in the proof of Proposition
8.6, we deduce ( j +a)n jn +a (mod n). Now

if d divides r, then d (x) = (x j ) where the product is over those j (mod r)
for which (j, r) = r/d; so multiplying the above congruence over such j we obtain
d (a)n d (a) (mod n) assuming that n is odd and (n, r) = 1. Thus we get
many congruences of the form (2.3) and even an improvement of Proposition 8.6,
since we get pn p (mod n) for each prime p dividing r, as p (1) = p.
8.6. An equivalent formulation: Recurrence sequences. It seems strange
that a primality test should be formulated in terms of polynomials! Inspired by
the results mentioned in section 8.5, we now show how to reformulate the AKS test
entirely in terms of integers.
Lemma 8.7. Suppose that (n, r) = 1. We have (x + a)n xn + a mod (n, xr 1)
if and only if ( + a)n n + a mod n for all satisfying r = 1.
Proof. If (x + a)n xn + a mod (n, xr 1), then we get ( + a)n n + a mod n
for each satisfying r = 1 simply by substituting x = . If ( + a)n n + a
mod n, then we may write (x + a)n xn + a mod (n, x ). The ideals (n, x )
are coprime (for the gcd of any two contains (n, ) which divides (n, r) = 1).
The result follows by the Chinese Remainder Theorem in Z[, x].

Thus the criteria is reformulated in terms of algebraic integers in a number eld,
but we desire criteria in terms of rational integers. To do this we introduce linear
39In fact, they determine which square root of 1 in terms of the Legendre-Jacobi symbol, one
of many relevant things that we have decided not to discuss in this article.
36
ANDREW GRANVILLE
recurrence sequences (the best known example being the Fibonacci numbers); we
give, without proof, the properties we need in the following paragraph.

i
Suppose that j=1 (x j ) = x 1
i=0 ui x Z[x] has no repeated roots.

Let c(t) Q(t). Dene fk = i=1 ci ki for each k 0, where cj = c(j ). An
alternate way to dene fk is by the values for k = 0, 1, 2, . . . , 1 and thereafter by
1
fm+ = i=0 ui fm+i for all m 0. The Fibonacci numbers make up the example
where our polynomial is x2 x 1 and c(t) = t/(t2 + 1), with F0 = 0, F1 = 1.
We claim that if we know 1 , 2 , , and f0 , . . . , f1 , then we can recover

j
c1 , , c . For this information gives us the linear equations
i=1 i ci = fj
for j = 0, 1, , 1 in the unknowns ci , and the matrix of
coecients is the
j
Vandermonde {i }1i, 0j1 , which has determinant 1i<j (j i ),
and this is non-zero since the i are distinct. In fact this argument works mod n
provided the determinant is coprime to n.
Take = r and i = i + a for each i where = e2i/r is a primitive rth root
of unity. The Vandermonde determinant above is then a power of r (by the proof
of Proposition 8.6) and so coprime to n. Choose c(t) so that (c(i ), n) = 1 for
all i, and dene C(t) = c(t)((t + a)n (tn + a)). If ( i + a)n in + a mod n
for all i, then Fj 0 mod n for all j, evidently. Moreover if Fj 0 mod n
for 0 j r 1, then we can deduce, as above, that C( i ) 0 mod n for all
i, and
therefore ( i + a)n in + a mod n. Now Fj = fn+j afj gj where
gj := ri=1 ci in ( i + a)j . We haveenormous freedom in choosing c(t); we select
c(t) = (t a)/r. Given that (1/r) 0ir1 ij = 0 unless r divides j, in which
case it
equals 1, we deduce that fj = 0 for 0 j r 2 and fr1 = 1; whereas
gj = kj ajk where k is the least non-negative residue of (n + 1) mod r, and we

dene kj = 0 if j < k. Thus we have proved the following:
Proposition 8.8. Suppose that (n, r) = 1. Dene the sequence {fm }m0 of integers by f0 = f1 = = fr2 = 0, fr1 = 1 and
r1

r
fm+r =
(1)j arj fm+j + (ar + 1)fm for all m 0.
j
j=1
Then (x + a)n xn + a mod (n, xr 1) if and only if

j jk
a
+ afj mod n for 0 j r 1.
fn+j
k
In the special case a = 1 we can work with a linear recurrence of order =
(r 1)/2, because (1 i )n 1 in mod n is trivial for i = 0 and since it
gives (1 i )n 1 in mod n by multiplying through by in . So take
i = (1 2i )(1 2i ) for 1 i , with c(i ) = i + i (which can be attained
by an appropriate polynomial c(t)). We dene C(t) so that C(i ) = c(i )(i )
where
(n+1)/2
(i ) = i
+ (1)(n1)/2 ( i i )( in in )
= (1)(n1)/2 ni ( i i )((1 2i )n 1 + 2in ).

Working through a similar argument to that above we then obtain the following.
Dene

(x (1 2i )(1 2i )) = x(r1)/2
gr,i xi .
gr (x) :=
1i(r1)/2
0i(r3)/2
37
Note that gr (2 y 2 y 2 ) = (1)(r1)/2 (y r y r )/(y y 1 ).

Proposition 8.9. Suppose that (n, r) = 1 with r an odd prime, and let k be the
least residue, in absolute value, of n/2 mod r. Dene the sequence {fm }m0 of
integers by f0 = 1, f1 = = f(r3)/2 = 0 and
(r3)/2
fm+(r1)/2 =
gr,j fm+j for all m 0.
j=0
Then (x 1)n xn 1 mod (n, xr 1) if and only if

n1
2j + 2
k
+k
2
f(n+1)/2+j (1)
r
mod n for 0 j (r 3)/2.
j+1 j+k+1
With dierent choices of c(t) we will obtain dierent recurrence sequences with
the same basic property.
Acknowledgements: I would like to thank Matt Baker, Dan Bernstein, Michael Guy,
Eric Pine, Francois Dorais, Felipe Voloch, the referee and particularly Manindra
Agrawal and Carl Pomerance for reading a draft version of this article and providing numerous suggestions. I am also indebted to Dan Bernsteins ne website
http://cr.yp.to/primetests.html, which contains a wealth of information and
ideas.
References
1. Leonard M. Adleman and Ming-Deh A. Huang, Primality testing and abelian varieties over
nite elds, Lecture Notes in Mathematics, vol. 1512, Springer-Verlag, Berlin. MR1176511
(93g:11128)
2. Leonard M. Adleman, Carl Pomerance and Robert S. Rumely, On distinguishing prime numbers from composite numbers, Annals of Mathematics 117 (1983), 173206. MR0683806
(84e:10008)
3. Manindra Agrawal, Neeraj Kayal and Nitin Saxena, PRIMES is in P (to appear).
4. W. R. Alford, Andrew Granville and Carl Pomerance, There are innitely many Carmichael
numbers, Annals of Mathematics 139 (1994), 703722. MR1283874 (95k:11114)
5. Roger C. Baker and Glynn Harman, The Brun-Titchmarsh Theorem on average, Progr. Math.
138 (1996), 39-103. MR1399332 (97h:11096)
6. D. J. Bernstein, Proving primality in essentially quartic random time (to appear).
7. Pedro Berrizbeitia, Sharpening PRIMES is in P for a large family of numbers (to appear).
8. Dan Boneh, Twenty years of attacks on the RSA Cryptosystem, Notices Amer. Math. Soc.
46 (1999), 203-213. MR1673760
9. Richard Crandall and Carl Pomerance, Prime numbers. A computational perspective,
Springer-Verlag, New York, 2001. MR1821158 (2002a:11007)
10. Whiteld Die and Martin E. Hellman, New directions in cryptography, IEEE Transactions
on Information Theory 22 (1976), 644654. MR0437208 (55:10141)
11. Etienne Fouvry, Th
eor`
eme de Brun-Titchmarsh; application au th
eor`
eme de Fermat, Invent.
Math 79 (1985), 383407. MR0778134 (86g:11052)
12. Dorian M. Goldfeld, On the number of primes p for which p + a has a large prime factor,
Mathematika 16 (1969), 23-27. MR0244176 (39:5493)
13. Andrew Granville and Thomas Tucker, Its as easy as abc, Notices Amer. Math. Soc. 49
(2002), 1224-1231. MR1930670 (2003f:11044)
14. Sha Goldwasser and Joe Kilian, Almost all primes can be quickly certied, Proceedings of the
18th annual ACM symposium on theory of computing, Association for Computing Machinery,
New York, 1986.
15. Donald E. Knuth, The art of computer programming, volume 2: Seminumerical algorithms,
Addison-Wesley, Reading, 1969. MR0286318 (44:3531)
38
ANDREW GRANVILLE
16. H.W. Lenstra, Jr., Galois theory and primality testing, Lecture Notes in Mathematics 1142
(1985), 169-189. MR0812498 (87g:11171)
17. H.W. Lenstra, Jr., and Carl Pomerance, Primality testing with Gaussian periods (to appear).
18. Yu. V. Matijasevich, Hilberts Tenth Problem, MIT Press, Cambridge, MA, 1993. MR1244324
(94m:03002b)
19. Preda Mihailescu and Roberto Avanzi, Ecient quasi - deterministic primality test improving AKS (to appear).
20. G
erald Tenenbaum and Michel Mend`es France, The Prime Numbers and Their Distribution,
Student Mathematical Library 6 (2000), Amer. Math. Soc. MR1756233 (2001j:11086)
21. Paulo Ribenboim, The new book of prime number records, Springer-Verlag, New York, 1995.
MR1377060 (96k:11112)
22. Jose Felipe Voloch, On some subgroups of the multiplicative group of nite rings (to appear).
matiques et Statistique, Universit
D
epartement de Mathe
e de Montr
eal, CP 6128
succ. Centre-Ville, Montr
eal, QC H3C 3J7, Canada
E-mail address: andrew@dms.umontreal.ca

It Is Easy To Determine Whether A Given Integer Is Prime

Uploaded by

Copyright:

Available Formats

It Is Easy To Determine Whether A Given Integer Is Prime

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

It Is Easy To Determine Whether A Given Integer Is Prime

Uploaded by

Copyright:

Available Formats

BULLETIN (New Series) OF THE

AMERICAN MATHEMATICAL SOCIETY

IT IS EASY TO DETERMINE WHETHER A GIVEN INTEGER

Abstract. The problem of distinguishing prime numbers from composite

(x + a)n xn + a mod (n, xr 1) for each integer a, 1 a r log n.

IT IS EASY TO DETERMINE WHETHER A GIVEN INTEGER IS PRIME

1.1. Our objective is to nd a quick foolproof algorithm to determine whether

IT IS EASY TO DETERMINE WHETHER A GIVEN INTEGER IS PRIME

for all integers a and primes n; and thus

for all integers x, y and primes n.

IT IS EASY TO DETERMINE WHETHER A GIVEN INTEGER IS PRIME

Fermats Little Theorem (1637). If n is prime, then n divides an a for all

but 341 = 11 31. Note though that by taking a = 3 above we get

(mod n), a(n1)/2

(mod n), . . . , a(n1)/2

(where n 1 = 2u v with v odd) is not equal to either 1, 1, . . . , 1 or 1, 1, . . . , 1, 1,

IT IS EASY TO DETERMINE WHETHER A GIVEN INTEGER IS PRIME

IT IS EASY TO DETERMINE WHETHER A GIVEN INTEGER IS PRIME

If n is composite let p be a prime dividing n. In the expansion

mod (n, xr 1)?

Then f (x) g(x)

gcd(b 1, n) gcd(b + 1, n) = gcd(b2 1, n) = n,

(as b2 1 (mod n)) whereas gcd(b 1, n), gcd(b + 1, n) < n (since b  1 or 1

IT IS EASY TO DETERMINE WHETHER A GIVEN INTEGER IS PRIME

In this context, read time as a synonym for bit operations.

take time O(r(

mod (n, xr 1).

Thus we have computed (x + a)n mod (n, xr 1) in a1 + 2 log n such steps,

in no more than O((log

Determine whether (x+a)n xn +a mod (n, xr 1) for a = 1, 2, . . . , [ r log n];

each such congruence takes O(r(log

run in fewer than O((log

IT IS EASY TO DETERMINE WHETHER A GIVEN INTEGER IS PRIME

leads to a running time of O((log

for each integer a, 1 a A where we take A = r log n. By Theorem 1 we know

mod (p, h(x))

for each integer a, 1 a A, since (p, h(x)) divides (p, xr 1).

Let H be the elements mod (p, xr 1) generated multiplicatively by x, x + 1,

((x + a)n )ea

|G| |ni pj nI pJ | (np) |R| 1 < n2 |R| 1.

Since n has order d mod r, nd 1 (mod r) and so xn x (mod xr 1).

IT IS EASY TO DETERMINE WHETHER A GIVEN INTEGER IS PRIME

|G| 2B+1 1 > n |R| 1,

for n 6, giving a contradiction.

and thus N < R12 (log n)/(log R) c()R/ log R as desired.

One easily deduces

IT IS EASY TO DETERMINE WHETHER A GIVEN INTEGER IS PRIME

mentioned in section 4.3. (Here and throughout section 5 we 

(ln p)/pa (2/x)

26We write u(x) v(x) if u(x)/v(x) 1 as x .

5.2. Primes in arithmetic progressions. We are interested in estimating

This is believed28 to be true when Q = x1 for any > 0.

in which the last estimate may easily be deduced

by the Brun-Titchmarsh Theorem for q a R, and then by extending the sum to

IT IS EASY TO DETERMINE WHETHER A GIVEN INTEGER IS PRIME

by (5.2). Finally we apply the Brun-Titchmarsh Theorem to note that

where c() := 1/4ln(1/(12)), which is > 0 provided <

IT IS EASY TO DETERMINE WHETHER A GIVEN INTEGER IS PRIME

(as b2 1 (mod n)) whereas gcd(b 1, n), gcd(b + 1, n) < n (since b 1 or 1

mentioned in section 4.3. (Here and throughout section 5 we

of q a (z, y) with a 2 is O( y), and so there are O( y/) values of p P for