Vous êtes sur la page 1sur 9

Math 105 notes, week 4

C. Pomerance

1 Constants
[This section is not dealt with in Apostol. The notes below are based on the treatment in texts
of Tenenbaum and Pollack.]
We have seen that there is a constant c such that
X1  
1
HP (x) := = log log x + c + O , (1)
p≤x
p log x

but “c” was poorly identified. We earlier had introduced two constants: Euler’s constant γ and
the constant α defined as
X 1 
1
 X X
1
α := − + log 1 − = j
.
p
p p p j≥2
jp

The second expression for α is obtained from the Taylor series for log(1 − θ) for −1 < θ < 1
applied with θ = 1/p.) In this unit we’ll show that the constant c in the prime harmonic sum
(1) is γ − α. P1 P 1
The idea is to change the divergent sum p
to the convergent sum p1+σ
, where σ > 0 is
+
small, learn some truths about this new sum, and then let σ → 0 .
We begin with the function
X 1 
1
 X X
1
f (σ) := − 1+σ
+ log 1 − −1−σ = j(1+σ)
,
p
p p p j≥2
jp

where we again used the Taylor series. The double sum has positive terms and it is easy to see
that the sum over j is smaller than
 
X 1 1 1
= 1+σ 1+σ =O
j≥2
2pj(1+σ) 2p (p − 1) p2(1+σ)

for σ ≥ − 14 , so the series for f (σ) converges uniformly and absolutely for σ ≥ − 41 . Thus, f (σ)
is a continuous function for σ ≥ − 41 (actually, for σ > − 12 ). Then
lim f (σ) = f (0) = α. (2)
σ→0+

Next we look at f (σ) for σ > 0. Part of the definition of f (σ) is a sum of logs, and we
recognize this as the log of an Euler product:
X Y
log 1 − p−1−σ = − log 1 − p−1−σ = log ζ(1 + σ), σ > 0.
 

p p

1
We
P show that this expression is almost equal to a certain infinite series. Using ζ(1 + σ) =
−1−σ
n n , we can compare ζ(1 + σ) with an integral:
Z ∞ Z ∞
−1−σ
t dt < ζ(1 + σ) < 1 + t−1−σ dt.
1 1

Further, the integral evaluates to 1/σ. Thus, we have the very tidy approximate formula:
1 1
ζ(1 + σ) = + O(1) = (1 + O(σ)), σ > 0.
σ σ
Taking the log of this equation give us
 
1 1
log ζ(1 + σ) = log (1 + O(σ)) = log + log(1 + O(σ)) = − log σ + O(σ).
σ σ

Now, via the Taylor series, we have σ = 1 − e−σ + O(σ 2 )) for 0 < σ < 1, so that

log 1 − e−σ = log σ + O(σ 2 ) = log (σ(1 + O(σ))) = log σ + O(σ).


 

Combining the last two displays, we get that


X1
log ζ(1 + σ) = − log 1 − e−σ + O(σ) = e−nσ + O(σ), 0 < σ < 1.

(3)
n
n

Recall that  
X1 1
H(x) := = log x + γ + O .
n≤x
n x

We apply partial summation to the sum in (3):


X1  Z x 
−nσ −xσ −tσ
e = lim e H(x) + σe H(t) dt
n
n x→∞ 1
Z ∞
= σe−tσ H(t) dt.
1

Thus, Z ∞
log ζ(1 + σ) = σe−tσ dt + O(σ), 0 < σ < 1.
1
−1−σ
P
The other part of f (σ) is − p , and we can sum this also by partial summation using the
prime harmonic sum defined in (1). We have
X X1 Z ∞ Z ∞
−1−σ −σ −1−σ
p = p = σHP (u)u du = σHP (et )e−tσ dt,
p p
p 2 log 2

2
where for the last step we made the change of variables u = et . This change of variables was not
completely arbitrary, since the harmonic sum is exponentially larger than the prime harmonic
sum, and we are trying to compare them. Putting these last two displays into what we did
above, we have
X Z ∞ Z ∞
−1−σ −tσ
f (σ) = log ζ(1 + σ) − p = σH(t)e dt − σHP (et )e−tσ dt + O(σ), 0 < σ < 1.
p 1 log 2

The part of the second integral from log 2 to 1 is O(σ), so we get that
Z ∞
H(t) − HP (et ) e−tσ dt + O(σ).

f (σ) = σ
1

We now substitute what we know about H(t) and HP (et ). The first is log t + γ + O(1/t)
and the second is log et + c + O(1/ log et ), which simplifies to log t + c + O(1/t). Hence,
Z ∞  
1
f (σ) = σ γ−c+O e−tσ dt + O(σ).
1 t
The first part is
Z ∞
σ (γ − c)e−tσ dt = (γ − c)e−σ = (γ − c)(1 + O(σ)) = γ − c + O(σ), 0 < σ < 1.
1

The second part is O of


Z ∞ Z ∞ Z ∞
1 −tσ 2 −tσ
u
σ e dt = σ log t e dt = σ log e−u du,
1 t 1 σ σ
where we integrated by parts and then substituted u = tσ. This last integral is
 Z ∞ Z ∞
1 −u
σ log e du + σ e−u log u du.
σ σ σ

The first of these terms is σ log(1/σ)e−σ < σ log(1/σ). The second integral is smaller than the
same integral from 1 to ∞ and so with the factor σ, the second term is O(σ). We conclude
that
f (σ) = γ − c + O(σ log(1/σ), 0 < σ < 1.
We now let σ → 0+ (which implies too that σ log(1/σ) → 0), so by (2), we have that γ − c = α;
that is, c = γ − α. This is what we wanted to prove.
Theorem 1 (Mertens). We have
X1  
1
= log log x + γ − α + O , x ≥ 2.
p≤x
p log x

3
Here are two interesting corollaries.
Corollary 1 (Mertens). For x ≥ 2,
Y −1
Y p 1
= 1− = eγ log x + O(1).
p≤x
p − 1 p≤x
p

Recall that P (n) denotes the largest prime factor of the integer n > 1 and P (1) = 1.
Corollary 2. For x ≥ 2,
X 1
= (eγ − 1) log x + O(1).
n>x
n
P (n)≤x

We leave the proofs for homework.

2 Dirichlet’s theorem
Two thousand years after Euclid proved there are infinitely many primes, Dirichlet proved the
following strengthening.
Theorem 2. If k is a positive integer and a is an integer relatively prime to k, then there are
infinitely many primes p ≡ a (mod k).
Our next goal will be to prove this theorem. Eventually we will prove it in a stronger form
where we somewhat quantify the distribution of primes in the residue class a (mod k). Our
plan is to first look at some elementary cases of Theorem 2, then discuss the algebraic structure
of the ring Z/kZ, introducing characters, and finally proving the following stronger form of the
theorem.
Theorem 3. If k is a positive integer and a is an integer relatively prime to k, then for all
x ≥ 2,
X log p 1
= log x + O(1),
p≤x
p ϕ(k)
p ≡ a (mod k)

where the O-symbol implies a constant that may depend on the value of k.

3 Some easy cases of Theorem 2


First we consider the cases of k = 1 and 2. Since primes are integers, for every choice of a
and for every prime p, we have p ≡ a (mod 1), so Theorem 2 for k = 1 follows from Euclid.
Similarly, gcd(a, 2) = 1 if and only if a is odd, and then every odd prime satisfies p ≡ a (mod 2),
so Euclid again suffices.

4
The first nontrivial cases are k = 3 and 4. We handle k = 4 first. There are essentially two
values of a to check out: a = 1, a = −1. Assume there are only finitely many primes p ≡ −1
(mod 4), multiply them all together, call the product M. Consider 4M − 1. It is odd, so each
prime factor of 4M − 1 is either ≡ 1 (mod 4) or ≡ −1 (mod 4). If all were in the first category,
the product would be ≡ 1 (mod 4), which is not the case for 4M − 1. Thus, there is at least
one prime p ≡ −1 (mod 4) with p | 4M − 1, and this prime p ∤ M. Thus, our assumption
about only finitely many is wrong, and we have proved Theorem 2 in the case a = −1, k = 4.
The proof for a = −1, k = 3 is almost exactly the same, with “3” replacing “4” at each
juncture; try it!
Now say a = 1, k = 4. Assume there are only finitely many primes p ≡ 1 (mod 4), mulitply
them all together, and call the product M. Let n = 4M 2 + 1 and let p be a prime factor of n.
Let u = 2M. In the multiplicative group mod p, we have 2M being a square root of −1, since
p | N = u2 + 1, so u2 ≡ −1 (mod p). This means that u has order 4 in this group, so 4 divides
the order of the group. The multiplicative group mod p has order p − 1, so 4 | p − 1, that is,
p ≡ 1 (mod 4). But p ∤ M, so there must be infinitely many primes p ≡ 1 (mod 4).
As you can see, this last proof brought in a little algebra. It is a direction in which we will
need to travel further!
The proof for a = 1, k = 3 is similar. If there are only finitely many primes p ≡ 1 (mod 3),
let M be their product, and let n = (3M)2 + 3M + 1. If p is a prime factor of M it can be
checked that u = 3M has order 3 in the multiplicative group mod p. Indeed, u 6≡ 1 (mod p),
since otherwise n = u2 + u + 1 ≡ 3 6≡ 0 (mod p), but u3 − 1 = (u − 1)(u2 + u + 1) ≡ 0 (mod 3).
Thus, 3 | p − 1, giving a contradiction to the assumed finitude of primes p ≡ 1 (mod 3).
The following result shows that the residue class 1 (mod q) for q prime can be handled in
a simple way. fairly easy.

Proposition 1. For each prime q, there are infinitely many primes p ≡ 1 (mod q).

Proof. Assume there are only finitely many primes p ≡ 1 (mod q), multiply them all together,
and call the product M. Let n = (qM)q − 1. Then u = qM has order q in the multiplicative
group mod n, since uq ≡ 1 (mod n), u 6≡ 1 (mod n), and q is prime. So, we have that q divides
the order of this group, which is ϕ(n). Now think of ϕ(n) as a multiplicative function, say n has
the prime factorization pa11 pa22 . . . pavv , where p1 , p2 , . . . , pv are distinct primes and a1 , a2 , . . . , av
are positive integers. We have
v
Y v
Y
ϕ(n) = ϕ (pai i ) = piai −1 (pi − 1).
i=1 i=1

Now q clearly does not divide n, but since it divides ϕ(n), it thus must divide one of the
factors pi − 1. That is, we have found a prime pi ≡ 1 (mod q), and this prime, since it divides
n = (qM)q − 1, it does not divide M. This contradiction proves the proposition.

5
4 The ring Z/nZ and the group (Z/nZ)∗
Let us begin our gentle foray into algebra. First note that for n a positive integer, nZ is an
ideal in the ring Z, so that Z/nZ is a ring. This is the fancy way to think about it, but we’re
just talking about addition and multiplication modulo n. The familiar rules hold (it’s tedious
to list all of them), and the upshot is that Z/nZ is a commutative ring with 1. Some members
of this ring are invertible and some are not. The criterion is that a (mod n) is invertible if
and only if gcd(a, n) = 1. Indeed, if ab ≡ 1 (mod n), then clearly a must be coprime with n.
Conversely, if gcd(a, n) = 1, there are integers u, v with au + vn = 1, so that au ≡ 1 (mod n).
The invertible elements of Z/nZ thus are ϕ(n) in number, and they comprise a group under
multiplication modulo n. This group is denoted by (Z/nZ)∗ or sometimes, (Z/nZ)× . (In
general, if R is a commutative ring with 1, we may look at the unit group R∗ of invertible
elements.)
If R is a commutative ring with multiplicative identity, let 1R denote this multiplicative
identity, and let 0R denote the additive identity in R.
Note that if R, S are commutative rings with multiplicative identities, then we can construct
a new ring R × S. As a set, it is the cartesian product, as the notation suggests. The addition
and multiplication in R × S is done coordinate-wise, the multiplicative identiy is the ordered
pair (1R , 1S ), and the additive identiy is (0R , 0S ). We see that (R × S)∗ = R∗ × S ∗ .
We would like to understand the ring Z/nZ and the group (Z/nZ)∗ a little better. To this
end, we prove the following fundamental result.
Proposition 2 (Chinese remainder theorem). If m, n are relatively prime positive integers,
then there is a canonical isomorphism from Z/(nm)Z to Z/nZ × Z/mZ which induces an
isomorphism from (Z/(nmZ)∗ onto (Z/nZ)∗ × (Z/mZ)∗ .
Proof. The canonical isomorphism F is easy to write down. For a (mod nm) in Z/(nm)Z,
let F (a) be the ordered pair (a (mod n), a (mod m)) in Z/nZ × Z/mZ. One checks that F
is well-defined; that is, the values of F do not depend on the choice of representative of the
residue class, just the residue class. So, if a ≡ b (mod nm), then a ≡ b (mod n) and a ≡ b
(mod m). Cool, that was easy.
We next check that F preserves addition and multiplication, try it. Thus, F is a ring
homomorphism.
We next check that F is one-to-one. Indeed, if F (a (mod nm)) = F (b (mod nm)), then

a ≡ b (mod n), a ≡ b (mod m).

So, both n and m divide a − b, so that since n, m are coprime, their product nm divides a − b.
Thus, a ≡ b (mod nm), which shows that F is one-to-one.
Finally, since the domain and range of F both have nm elements and F is one-to-one, it
must be onto. This completes the proof that F is a ring isomorphism and that Z/(nm)Z is
isomorophic to Z/nZ × Z/mZ. The final statement about the unit groups is now clear as
well.

6
By induction we can decompose the ring Z/nZ into the product ring
(Z/pa11 Z) × (Z/pa22 Z) × · · · × (Z/pakk Z),
where n has the usual prime factorization pa11 pa22 . . . pakk . And we have the unit group also
decomposed:
(Z/nZ)∗ ∼ a
= (Z/pa11 Z)∗ × (Z/pa22 Z)∗ × · · · × (Z/pkk Z)∗ ,
We will be particularly interested in this unit group decomposition, and it raises the question
about what we can say about (Z/pa Z)∗ for p prime.
Proposition 3. If p is prime, the group (Z/pZ)∗ is cyclic.
Proof. For d | p − 1, let f (d) be the number of elements of (Z/pZ)∗ of order d. Then
X
f (d) = p − 1.
d|p−1

We’ll
P show that each f (d) ≤ ϕ(d). If this is true, then comparing the above equation with
d|p−1 ϕ(d) = p − 1, we see that f (d) = ϕ(d) for each d | p − 1. In particular, f (p − 1) > 0, so
there are elements of order p − 1, which is the assertion of the proposition.
So it remains to prove that each f (d) ≤ ϕ(d). This is true if f (d) = 0, so assume that
f (d) > 0, that is, there is an element g of order d. Then g, g 2, . . . , g d are distinct residues
(mod p) and they are all roots of the polynomial xd − 1 in the finite field Z/pZ. (A field is
merely a commutative ring with multiplicative identity such that each nonzero element has a
multiplicative inverse.) Over a field, no polynomial has more roots than its degree, so in fact
every element of order d in the group is found in the list g, g 2, . . . , g d. If g j is in the list and
u = gcd(j, d) > 1, then (g j )d/u ≡ 1 (mod p), so that g j has order dividing d/u. In particular,
g j is definitely not counted among the f (d) elements of order d. That leaves just the ϕ(d)
possibilities g j for 1 ≤ j ≤ d and gcd(j, d) = 1. We have proved that f (d) ≤ ϕ(d), and we’ve
seen this is sufficient for the proposition.
Proposition 4. If p is an odd prime and a is a positive integer, then (Z/pa Z)∗ is cyclic.
Proof. By Proposition 3, this holds for a = 1. We now prove this for a = 2. Let g be an integer
such that g (mod p) is a cyclic generator of (Z/pZ)∗ . It is very possible that g (mod p2 ) is a
generator of (Z/p2 Z)∗ , in which case we’d be done. So assume that it is not such a generator.
Its order in (Z/p2 Z)∗ is a proper divisor of p(p − 1) but not a proper divisor of p − 1 (since we
are assuming that g has order p − 1 when working mod p). Thus, the order of g when working
mod p2 is exactly p − 1. Consider the element g + p. It also has order p − 1 when working
mod p. But
(g + p)p−1 ≡ g p−1 + (p − 1)g p−2p (mod p2 ),
by the binomial theorem. But we are assuming that g p−1 ≡ 1 (mod p2 ), and the second term
above is not 0 (mod p2 ), so g + p does not have order p − 1 when working mod p2 , which shows
that g + p must have order p(p − 1) when working mod p2 .

7
Now we assume that a ≥ 2 and there is a cyclic generator g for (Z/pa Z)∗ . We’ll show
that this same g also works for (Z/pa+1 Z)∗ . First note that g is a generator for (Z/pZ)∗ , for
if not, we’d have g (p−1)/h ≡ 1 (mod p) for some integer h | p − 1 with h > 1. Then, writing
g (p−1)/h = 1 + kp for some integer k, we would have
a−1 (p−1)/h a−1
gp = (1 + kp)p ≡ 1 (mod pa ),
using the binomial theorem (check it out!). This contradicts our inductive assumption about
g. So,
a−1 (p−1) a−2 (p−1)
gp ≡ 1 (mod pa ), g p 6≡ 1 (mod pa ),
a−1 (p−1)/h
gp 6≡ 1 (mod pa ) for h | p − 1, h > 1.
Since the order of (Z/pa−1 Z)∗ is pa−2 (p − 1), we have
a−2 (p−1)
gp ≡1 (mod pa−1 ).
Putting this together with the above, we have
a−2 (p−1)
gp = 1 + kpa−1 for some integer k 6≡ 0 (mod p).
Thus, raising this equation to the pth power, we get
a−1 (p−1)
gp = (1 + kpa−1 )p ≡ 1 + kpa (mod pa+1 ),
using that a ≥ 2, p ≥ 3. Thus, the order of g in (Z/pa+1 Z)∗ does not divide pa−1 (p − 1), and
as above, it does not divide pa (p − 1)/h for h > 1, but it does divide pa (p − 1). It thus must be
exactly pa (p − 1), which shows that g is indeed a cyclic generator for (Z/pa Z)∗ .
Proposition 5. If a ≥ 3, then 5 has order 2a−2 in (Z/2a Z)∗ , and the cyclic subgroup generated
by 5 does not contain −1.
Proof. It is true for a = 3. Assume that a ≥ 3 and it is true for a. It is clear that the
subgroup generated by 5 in (Z/2a+1 Z)∗ has only elements that are 1 (mod 4), so −1 is not in
the subgroup. It remains to consider the order of 5. We are assuming that
a−3
52 6≡ 1 (mod 2a ),
but since this power is congruent to 1 (mod 2a−1 ), we may write
a−3
52 = 1 + 2a−1 k, k odd.
Then squaring both sides, we get
a−2
52 ≡ 1 + 2a k (mod 2a+1 ).
Thus, the order of 5 in (Z/2a+1 Z)∗ does not divide 2a−2 . Therefore its order is either 2a−1 or
2a . In the latter possibility, this says that (Z/2a+1 Z)∗ is cyclic, but as ±1, 2a ± 1 comprise 4
elements of order dividing 2, we see the group cannot be cyclic. This leaves the only choice for
the order being 2a−1 , and the proposition is true by induction.

8
This last result is saying that when a ≥ 3, the group (Z/2a Z)∗ is isomorphic to the cartesian
product of a cyclic group of order 2a−2 and a cyclic group of order 2. In fact, we have (Z/2a Z)∗
isomorphic to the cyclic subgroup h5i of order 2a−2 times the cyclic subgroup h−1i of order 2.
Summary.
Every finite abelian group can be written as a product of cyclic groups. But for the group
(Z/nZ)∗ we have the factorization suggested by the prime factorization of n. It is as follows.
Suppose n is odd and has the prime factorization pa11 pa22 . . . pakk , where p1 , p2 , . . . , pk are distinct
primes and a1 , a2 , . . . , ak are postive integers. Each of the groups (Z/pai i Z)∗ is cyclic (of order
piai −1 (pi −1) and (Z/nZ)∗ is isomorphic to the product of these cyclic groups. For 2n with n odd,
we have (Z/2nZ)∗ isomorphic to (Z/nZ)∗ . For 4n with n odd, we have (Z/4nZ)∗ isomorphic to
a cyclic group of order 2 times (Z/nZ)∗ . And for 2a n with a ≥ 3 and n odd, we have (Z/2a nZ)∗
isomorphic to the product of a cyclic group of order 2, a cyclic group of order 2a−2 and (Z/nZ)∗ .
In group theory, there is a certain canonical factorization of a finite abelian group G into
cyclic groups, where the order of each factor divides the order of the next larger factor, if there
is one. The factorization of (Z/nZ)∗ given above is not necessarily this canonical factoriza-
tion. However, it is not so hard to find the canonical factorization from the number theoretic
factorization, if you wanted to know it.
For example, say n = 105, which has the prime factorization 3 · 5 · 7. Then (Z/nZ)∗ has
the number theoretic factorization C2 × C4 × C6 , where Cm denotes a cyclic group of order m.
This comes from 3 − 1 = 2, 5 − 1 = 4, 7 − 1 = 6. The canonical factorization is C2 × C2 × C12 .
Actually, this is the invariant factor decomposition. (Another canonical factorization is into
cyclic groups of prime or prime power order. The group of the example has the factorization
C2 × C2 × C3 × C4 .)
The invariant factor decomposition has the nice feature that the largest cyclic group gives
the exponent of the group, the least number h such that each element raised to the power h
is the identity. In number theory, for the group (Z/nZ)∗ , the exponent is denoted λ(n) and is
known as the Carmichael lambda-function. It has a simple definition. We have λ(n) equal to
the least common multiple of λ(pa ), where pa is a prime or prime power divisor of n. Further,
for p odd or p = 2, a ≤ 2, λ(pa ) = ϕ(pa ) (since the group (Z/pa Z)∗ is cyclic. And λ(2a ) = 2a−2
for a ≥ 3.
In the example above, we have λ(105) = lcm[2, 4, 6] = 12.

Vous aimerez peut-être aussi