Is N A Prime Number

Is n a Prime Number?
Manindra Agrawal
IIT Kanpur
March 27, 2006, Delft
Manindra Agrawal (IIT Kanpur) Is n a Prime Number? March 27, 2006, Delft 1 / 47
Overview
1 The Problem
2 Two Simple, and Slow, Methods
3 Modern Methods
4 Algorithms Based on Factorization of Group Size
5 Algorithms Based on Fermat’s Little Theorem
6 An Algorithm Outside the Two Themes
Outline
1 The Problem
3 Modern Methods
The Problem
Given a number n, decide if it is prime.
Easy: try dividing by all numbers less than n.
The Problem
Given a number n, decide if it is prime.
Easy: try dividing by all numbers less than n.
The Problem
Given a number n, decide if it is prime efficiently.
Not so easy: several non-obvious methods have been found.
The Problem
Given a number n, decide if it is prime efficiently.
Not so easy: several non-obvious methods have been found.
Efficiently Solving a Problem
There should exist an algorithm for solving the problem taking a

polynomial in input size number of steps.
For our problem, this means an algorithm taking logO(1) n steps.
Caveat: An algorithm taking log12 n steps would be slower than an

algorithm taking loglog log log n n steps for all practical values
of n!
Notation:
log is logarithm base 2.
O˜(logc n) stands for O(logc n log logO(1) n).


of n!
Notation:


of n!
Notation:


of n!
Notation:
Outline
1 The Problem
3 Modern Methods
The Sieve of Eratosthenes
Proposed by Eratosthenes (ca. 300 BCE).
1 List all numbers from 2 to n in a sequence.

2 Take the smallest uncrossed number from the sequence and cross out
all its multiples.
3 If n is uncrossed when the smallest uncrossed number is greater than
√
n then n is prime otherwise composite.
Time Complexity
√
If n is prime, algorithm crosses out all the first n numbers before
giving the answer.
√
So the number of steps needed is Ω( n).
Time Complexity
√
If n is prime, algorithm crosses out all the first n numbers before
giving the answer.
√
So the number of steps needed is Ω( n).
Wilson’s Theorem
Based on Wilson’s theorem (1770).
Theorem
n is prime iff (n − 1)! = −1 (mod n).
Computing (n − 1)! (mod n) naı̈vely requires Ω(n) steps.

No significantly better method is known!
Wilson’s Theorem
Based on Wilson’s theorem (1770).
Theorem
n is prime iff (n − 1)! = −1 (mod n).
Computing (n − 1)! (mod n) naı̈vely requires Ω(n) steps.

No significantly better method is known!
Outline
1 The Problem
3 Modern Methods
Fundamental Idea
Nearly all the efficient algorithms for the problem use the following idea.
Identify a finite group G related to number n.

Design an efficienty testable property P(·) of the elements G such
that P(e) has different values depending on whether n is prime.
The element e is either from a small set (in deterministic algorithms)
or a random element of G (in randomized algorithms).
Fundamental Idea

Fundamental Idea

Fundamental Idea

Groups and Properties
The group G is often:

I A subgroup of Zn∗ or Zn∗ [ζ] for an extension ring Zn [ζ].
I A subgroup of E (Zn ), the set of points on an elliptic curve modulo n.
The properties vary, but are from two broad themes.


Theme I: Factorization of Group Size
Compute a complete, or partial, factorization of the size of G

assuming that n is prime.
Use the knowledge of this factorization to design a suitable property.
Theme I: Factorization of Group Size
Compute a complete, or partial, factorization of the size of G

assuming that n is prime.
Use the knowledge of this factorization to design a suitable property.
Theme II: Fermat’s Little Theorem
Theorem (Fermat, 1660s)

If n is prime then for every e, e n = e (mod n).
Group G = Zn∗ and property P is P(e) ≡ e n = e in G .

This property of Zn is not a sufficient test for primality of n.
So try to extend this property to a neccessary and sufficient condition.
Theme II: Fermat’s Little Theorem
Theorem (Fermat, 1660s)

If n is prime then for every e, e n = e (mod n).
Group G = Zn∗ and property P is P(e) ≡ e n = e in G .

This property of Zn is not a sufficient test for primality of n.
So try to extend this property to a neccessary and sufficient condition.
Outline
1 The Problem
3 Modern Methods
Lucas Theorem
Theorem (E. Lucas, 1891)

Qt di
Let n − 1 = i=1 pi where pi ’s are distinct primes. n is prime iff there is
n−1
an e ∈ Zn such that e n−1 = 1 and gcd(e pi
− 1, n) = 1 for every
1 ≤ i ≤ t.
The theorem also holds for a random choice of e.

We can choose G = Zn∗ and P to be the property above.
The test will be efficient only for numbers n such that n − 1 is smooth.
Lucas Theorem

Qt di
n−1
1 ≤ i ≤ t.

Lucas Theorem

Qt di
n−1
1 ≤ i ≤ t.

Lucas-Lehmer Test
√
G is a subgroup of Zn∗ [ 3] containing elements of order n + 1.
n+1 √
The property P is: P(e) ≡ e 2 = −1 in Zn [ 3].
Works only for special Mersenne primes of the form n = 2p − 1, p
prime.
For such n’s, n + 1 = 2p .
√
The property needs to be tested only for e = 2 + 3.
Lucas-Lehmer Test
√
n+1 √
prime.
√
Lucas-Lehmer Test
√
n+1 √
prime.
√
Time Complexity
√ n+1
Raising 2 + 3 to 2 th power requires O(log n) multiplication
operations in Zn .
Overall time complexity is O˜(log2 n).
Time Complexity
√ n+1
Raising 2 + 3 to 2 th power requires O(log n) multiplication
operations in Zn .
Overall time complexity is O˜(log2 n).
Pocklington-Lehmer Test
Theorem (Pocklington, 1914)

n−1
If there exists a e such that e n−1 = 1 (mod n) and gcd(e pj − 1, n) = 1
Qp1 , p2 , . . . , pt dividing n − 1 then every prime factor of n
for distinct primes
has the form k · tj=1 pj + 1.
Similar to Lucas’s theorem.

Let G = Zn∗ and property P precisely as in the theorem.
The property is tested for a random e.
√
For the test to work, we need tj=1 ≥ n.
Q

n−1
for distinct primes

√
Q

n−1
for distinct primes

√
Q
Time Complexity
Depends on the difficulty of finding prime factorizations of n − 1

√
whose product is at least n.
Other operations can be carried out efficiently.
Time Complexity
Depends on the difficulty of finding prime factorizations of n − 1

√
whose product is at least n.
Other operations can be carried out efficiently.
Elliptic Curves Based Tests
Elliptic curves give rise to groups of different sizes associated with the
given number.
With good probability, some of these groups have sizes that can be
easily factored.
This motivated primality testing based on elliptic curves.
given number.
easily factored.
given number.
easily factored.
Goldwasser-Kilian Test
This is a randomized primality proving algorithm.

Under a reasonable hypothesis, it is polynomial time on all inputs.
Unconditionally, it is polynomial time on all but negligible fraction of
numbers.

numbers.

numbers.
Consider a random elliptic curve over Zn .

By a theorem of Lenstra (1987), the number of points of the curve is
√ √
nearly uniformly distributed in the interval [n + 1 − 2 n, n + 1 + 2 n]
for prime n.
Assuming a conjecture about the density of primes in small intervals,
it follows that there are curves with 2q points, for q prime, with
reasonable probability.

√ √
for prime n.

√ √
for prime n.
Theorem (Goldwasser-Kilian)
Suppose E (Zn ) is an elliptic curve with 2q points. If q is prime and there
exists A ∈ E (Zn ) 6= O such that q · A = O then either n is provably prime
or provably composite.
Proof.
√
Let p be a prime factor of n with p ≤ n.
We have q · A = O in E (Zp ) as well.
If A = O in E (Zp ) then n can be factored.
Otherwise, since q is prime, |E (Zp )| ≥ q.
√
If 2q < n + 1 − 2 n then n must be composite.
√ √
Otherwise, p + 1 + 2 p > n2 − n which is not possible.
Proof.
√
√
√ √
Proof.
√
√
√ √
Proof.
√
√
√ √
1 Find a random elliptic curve over Zn with 2q points.

2 Prove primality of q recursively.
3 Randomly select an A such that q · A = O.
4 Infer n to be prime or composite.



Analysis
The algorithm never incorrectly classifies a composite number.

With high probability it correctly classifies prime numbers.
The running time is O(log11 n).
Improvements by Atkin and others result in a conjectured running
time of O˜(log4 n).
Analysis

Analysis

Analysis

Adleman-Huang Test
The previous test is not unconditionally polynomial time on a small

fraction of numbers.
Adleman-Huang (1992) removed this drawback.
They first used hyperelliptic curves to reduce the problem of testing
for n to that of a nearly random integer of similar size.
Then the previous test works with high probability.
The time complexity becomes O(logc n) for c > 30!
Adleman-Huang Test

Adleman-Huang Test

Outline
1 The Problem
3 Modern Methods
Solovay-Strassen Test
A Restatement of FLT
n−1
If n is odd prime then for every e, 1 ≤ e < n, e 2 = ±1 (mod n).
n−1
When n is prime, e is a quadratic residue in Zn iff e 2 = 1 (mod n).
Therefore, if n is prime then
e n−1
=e 2 (mod n).
n
n−1
n−1
e n−1
=e 2 (mod n).
n
n−1
n−1
e n−1
=e 2 (mod n).
n
Proposed by Solovay and Strassen (1973).

A randomized algorithm based on above property.
Never incorrectly classifies primes and correctly classifies composites
with probability at least 12 .


1 If n is an exact power, it is composite.

2 For a random e in Zn , test if
e n−1
=e 2 (mod n).
n
3 If yes, classify n as prime otherwise it is proven composite.
The time complexity is O˜(log2 n).

e n−1
=e 2 (mod n).
n

e n−1
=e 2 (mod n).
n

e n−1
=e 2 (mod n).
n
Analysis
Consider the case when n is a product of two primes p and q.

Let a, b ∈ Zp , c ∈ Zq with a residue and b non-residue in Zp .
n−1 n−1
Clearly, < a, c > 2 =< b, c > 2 (mod q).
n−1 n−1
If < a, c > =6 < b, c >
2 (mod n) then one of them is not in
2
{1, −1} and so compositeness of n is proven.

Otherwise, either
< a, c > n−1
6=< a, c > 2 (mod n),
n
or
< b, c > n−1
6=< b, c > 2 (mod n).
n
Analysis

n−1 n−1
n−1 n−1
If < a, c > =6 < b, c >
2

Otherwise, either
< a, c > n−1
6=< a, c > 2 (mod n),
n
or
< b, c > n−1
6=< b, c > 2 (mod n).
n
Analysis

n−1 n−1
n−1 n−1
If < a, c > =6 < b, c >
2

Otherwise, either
< a, c > n−1
6=< a, c > 2 (mod n),
n
or
< b, c > n−1
6=< b, c > 2 (mod n).
n
Analysis

n−1 n−1
n−1 n−1
If < a, c > =6 < b, c >
2

Otherwise, either
< a, c > n−1
6=< a, c > 2 (mod n),
n
or
< b, c > n−1
6=< b, c > 2 (mod n).
n
Miller’s Test
Theorem (Another Restatement of FLT)

If n is odd prime and n = 1 + 2s · t, t odd, then for every e, 1 ≤ e < n,
s−1 s−2
the sequence e 2 ·t (mod n), e 2 ·t (mod n), . . ., e t (mod n) has either
all 1’s or a −1 somewhere.
Miller’s Test
This theorem is the basis for Miller’s test (1973).

It is a deterministic polynomial time test.
It is correct under Extended Riemann Hypothesis.
Miller’s Test

Miller’s Test

Miller’s Test

s−1
2 For each e, 1 < e ≤ 4 log2 n, check if the sequence e 2 ·t (mod n),
s−2
e 2 ·t (mod n), . . ., e t (mod n) has either all 1’s or a −1 somewhere.
3 If yes, classify n as prime otherwise composite.
The time complexity of the test is O˜(log4 n).
Miller’s Test

s−1
s−2
Miller’s Test

s−1
s−2
Miller’s Test

s−1
s−2
Rabin’s Test
A modification of Miller’s algorithm proposed soon after (1974).

Selects e randomly instead of trying all e in the range [2, 4 log2 n].
Randomized algorithm that never classfies primes incorrectly and
correctly classifies composites with probabilty at least 34 .
Time complexity is O˜(log2 n).
The most popular primality testing algorithm.
Rabin’s Test

Rabin’s Test

Rabin’s Test

AKS Test
Theorem (A Generalization of FLT)

If n is prime then for every e, 1 ≤ e < n, (ζ + e)n = ζ n + e in Zn [ζ],
ζ r = 1.
A test proposed in 2002 based on this generalization.

The only known deterministic, unconditionally correct, polynomial
time algorithm.
AKS Test

ζ r = 1.

time algorithm.
AKS Test

ζ r = 1.

time algorithm.
AKS Test
1 If n is an exact power or has a small divisor, it is composite.

2 Select a small number r carefully, let ζ r = 1 and consider Zn [ζ].
√
3 For each e, 1 ≤ e ≤ 2 r log n, check if (ζ + e)n = ζ n + e in Zn [ζ].
4 If yes, n is prime otherwise composite.
AKS Test

√
AKS Test

√
AKS Test

√
Analysis
Suppose n has at least two prime factors and let p be one of them.
Let S ⊆ Zp [ζ] such that for every element f (ζ) ∈ S, f ζ)n = f (ζ n ) in
Zp [ζ].
It follows that for every f (ζ) ∈ S, f (ζ)m = f (ζ m ) for any m of the
form ni · p j .
Since n is not a power of p, this places an upper bound on the size of
S.
√
If ζ + e ∈ S for every 1 ≤ e ≤ 2 r log n, then all their products are
also in S.
This makes the size of S bigger than the upper bound above.
Analysis
Zp [ζ].
form ni · p j .
S.
√
also in S.
Analysis
Zp [ζ].
form ni · p j .
S.
√
also in S.
Analysis
Zp [ζ].
form ni · p j .
S.
√
also in S.
Analysis
Zp [ζ].
form ni · p j .
S.
√
also in S.
Time Complexity
Number r is O(log5 n).

Time complexity of the algorithm is O˜(log12 n).
An improvement by Hendrik Lenstra (2002) reduces the time
complexity to O˜(log15/2 n).
Lenstra and Pomerance (2003) further reduce the time complexity to
O˜(log6 n).
Bernstein (2003) reduced the time complexity to O˜(log4 n) at the
cost of making it randomized.
Time Complexity

O˜(log6 n).
Time Complexity

O˜(log6 n).
Time Complexity

O˜(log6 n).
Time Complexity

O˜(log6 n).
Outline
1 The Problem
3 Modern Methods
Adleman-Pomerance-Rumeli Test
Proposed in 1980.
Is conceptually the most complex algorithm of them all.
Uses multiple groups, ideas derived from both themes, plus new ones!
It is a deterministic algorithm with time complexity logO(log log log n) n.
Was speeded up by Cohen and Lenstra (1981).
Proposed in 1980.
Proposed in 1980.
Proposed in 1980.
Overview of the Algorithm
Tries to compute a factor of n.

√
Let p be a factor of n, p ≤ n.
Find two sets of primes {q1 , q2 , . . . , qt } and {r1 , r2 , . . . , ru } satisfying:
Qt
I
i=1 qi = logO(log log log n) n.
I For
Qu each j √≤ u, rj − 1 is square-free and has only qi ’s as prime divisors.
I
j=1 rj > n.

√
Qt
I
I For
I
j=1 rj > n.

√
Qt
I
I For
I
j=1 rj > n.

√
Qt
I
I For
I
j=1 rj > n.

√
Qt
I
I For
I
j=1 rj > n.

√
Qt
I
I For
I
j=1 rj > n.
Let gj be a generator for the group Fr∗j .

γ
Let p = gj j (mod rj ) and γj = δi,j (mod qi ) for every qi | rj − 1.
Compute ‘associated’ primes rji ∈ {r1 , r2 , . . . , ru } for each qi .
Cycle through all tuples (α1 , α2 , . . . , αt ) with 0 ≤ αi < qi .
From a given tuple (α1 , α2 , . . . , αt ), derive numbers βi,j for
1 ≤ j ≤ u, 1 ≤ i ≤ t and qi | rj − 1 such that
I If p = gjαi i (mod rj ) for every j then δi,j = βi,j for every j and for every
i with qi | rj − 1.

γ

γ

γ

γ
From δi,j ’s, p can be constructed easily:

I Use Chinese remaindering to compute γj ’s fromQδi,j ’s.
u γ
I Use Chinese remaindering to compute p (mod j=1 rj ) from gj j ’s.
Qu √
I Since j=1 rj > n ≥ p, the residue equals p.
βi,j ’s are computed using higher reciprocity laws in extension rings
Zn [ζi ], ζiqi = 1.
For most of the composite numbers, the algorithm will fail during
computation of βi,j ’s.

u γ
Qu √

u γ
Qu √

u γ
Qu √

u γ
Qu √

u γ
Qu √
Outstanding Questions
Is there a ‘practical’, polynomial time deterministic primality test?

Is there a ‘practical’, provably polynomial time, primality proving test?
Are there radically different ways of testing primality efficiently?



Is N A Prime Number

Uploaded by

Document Informationclick to expand document information

Copyright:

Available Formats

Is N A Prime Number

Uploaded by

Document Information

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

Is N A Prime Number

Uploaded by

Copyright:

Available Formats

Is n a Prime Number?

March 27, 2006, Delft

2 Two Simple, and Slow, Methods

4 Algorithms Based on Factorization of Group Size

5 Algorithms Based on Fermat’s Little Theorem

6 An Algorithm Outside the Two Themes

2 Two Simple, and Slow, Methods

4 Algorithms Based on Factorization of Group Size

5 Algorithms Based on Fermat’s Little Theorem

6 An Algorithm Outside the Two Themes

Given a number n, decide if it is prime.

Easy: try dividing by all numbers less than n.

Given a number n, decide if it is prime.

Easy: try dividing by all numbers less than n.

Given a number n, decide if it is prime efficiently.

Not so easy: several non-obvious methods have been found.

Given a number n, decide if it is prime efficiently.

Not so easy: several non-obvious methods have been found.

There should exist an algorithm for solving the problem taking a

Caveat: An algorithm taking log12 n steps would be slower than an

There should exist an algorithm for solving the problem taking a

Caveat: An algorithm taking log12 n steps would be slower than an

There should exist an algorithm for solving the problem taking a

Caveat: An algorithm taking log12 n steps would be slower than an

There should exist an algorithm for solving the problem taking a

Caveat: An algorithm taking log12 n steps would be slower than an

2 Two Simple, and Slow, Methods

4 Algorithms Based on Factorization of Group Size

5 Algorithms Based on Fermat’s Little Theorem

6 An Algorithm Outside the Two Themes

Proposed by Eratosthenes (ca. 300 BCE).

1 List all numbers from 2 to n in a sequence.

Based on Wilson’s theorem (1770).

Computing (n − 1)! (mod n) naı̈vely requires Ω(n) steps.

Based on Wilson’s theorem (1770).

Computing (n − 1)! (mod n) naı̈vely requires Ω(n) steps.

2 Two Simple, and Slow, Methods

4 Algorithms Based on Factorization of Group Size

5 Algorithms Based on Fermat’s Little Theorem

6 An Algorithm Outside the Two Themes

Identify a finite group G related to number n.

Identify a finite group G related to number n.

Identify a finite group G related to number n.

Identify a finite group G related to number n.

The group G is often:

The group G is often:

The group G is often:

Compute a complete, or partial, factorization of the size of G

Compute a complete, or partial, factorization of the size of G

Theorem (Fermat, 1660s)

Group G = Zn∗ and property P is P(e) ≡ e n = e in G .

Theorem (Fermat, 1660s)

Group G = Zn∗ and property P is P(e) ≡ e n = e in G .

2 Two Simple, and Slow, Methods

4 Algorithms Based on Factorization of Group Size

5 Algorithms Based on Fermat’s Little Theorem

6 An Algorithm Outside the Two Themes

Theorem (E. Lucas, 1891)